{ "schema": "aner.metal-greedy-evidence/1", "date": "2026-11-01", "development_version": "0.0.6-dev.1", "scope": { "implementation": "Detached cached Metal greedy generation for bounded native/GPT-2 reference decoders; not a pretrained Gemma or PyTorch comparison.", "hardware": "Mac Metal measurements; separate V100 CUDA regressions.", "publication": "Local source and AnerWeb update only; deployment deferred.", "final_policy": "Prefill/selection boundary; decode1,2,4,... capped8 and clipped by remaining budget. Small one-token outputs (<65536 packed logits) retain the previous scan path.", "limitations": [ "Finished positive-budget lanes remain fixed across the entire request; masked GEMM work can continue.", "Short/EOS workloads can be slower; all measured negative profiles remain included.", "Prefill workspace and request states are retained; no general device-memory reduction claim.", "The backend mutex serializes a complete request; no fair interleaving or overlap claim.", "First device/model initialization is excluded from natural request timing; additional specialization cold-start cost is unmeasured.", "Four process pairs per profile; confidence intervals resample whole process clusters." ] }, "specialization_comparison": { "schema": "aner.metal-greedy-direct-specialization/1", "status": "PASS", "protocol": { "matched_process_pairs": 4, "timed_calls_per_process": 3, "excluded_warmups_per_process": 2, "process_order": "AB,BA,BA,AB; rotating three-profile order", "scope": "Warm complete public calls; identical producer object and all other libraries. Only the NN Metal implementation and Gen build-identity metadata differ. Cold startup and the extra pipeline-state construction are outside timing.", "confidence_interval": { "point_estimator": "median(all 12 pre-specialization calls) / median(all 12 specialized calls)", "resampling_unit": "whole matched pair of fresh process clusters", "inner_calls": "all three calls from each selected process retained together; no inner-call resampling", "method": "exhaustive enumeration of all ordered four-cluster draws with replacement", "draws": 256, "coverage": 0.95, "percentile_method": "linear interpolation at probability * (256 - 1) in the ordered bootstrap ratios" } }, "producer_sha256": "834ce647498123f3dd7f95cd56af3c4b7e0441f8cc14900821007be4b8e0ad96", "producer_object_sha256": "d43db9dd616f21aea3c9f2557c1bb0672adb19c430b6bc3fcdffb4cb808c0935", "metal_source_sha256": { "pre_specialization": "8326fce362ca42114f2be2ece059feeb9dfa6e783440aff9f3d53899750c535b", "specialized": "706c9b9bd30d8c3bafe0e77c7a236d4448a99eb7483669ac0d6920f5456f212c" }, "variant_library_sha256": { "pre_specialization": { "libaner_nn.a": "b847c38c8c18aa5dc4c4352265e27820a9c67aaa4aa92ba43d6a99196d398282", "libaner_generative.a": "788ba84d282f3dd19cb6d854f80417a3392b4b141fae5f05c18d2428a56fda6b" }, "specialized": { "libaner_nn.a": "b7a26984b11c2c5efb4b34ae6286ce9201681ffb75a283694d4bf6f76260c29d", "libaner_generative.a": "1ef2bce1c7aeccf6e1a577e41d9e23815c77b6927b570f1aa8dedc173d14d38b" } }, "identical_other_library_sha256": { "libaner_basic_types.a": "07e301399e324e5d539f415cb2375512eea3f4a3dc305e3cade281e58f995a6b", "libaner_string.a": "31c2a127d37eb7206dbfa582663fc65fc801074305bd3ab2829ce51e9812d583" }, "binary_sha256": { "pre_specialization": "7c1a43681d09e2a9f9705b333b22b77e11e84fee9ce77bfacdfb5e29da835d54", "specialized": "36b87a82f1e46bb87b2aacf5af964f5d04abb860e31eabe379b001feb25fc6d3" }, "parameter_value_sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107", "summary": [ { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "median_seconds": { "pre_specialization": 0.0163952295, "specialized": 0.0171809795 }, "pre_over_specialized_ratio_of_medians": 0.9542662861567353, "whole_process_pair_bootstrap_percentile95": [ 0.8145446333810126, 1.116951996428988 ], "exact_ids_and_parameter_values_equal": true }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "median_seconds": { "pre_specialization": 0.0058237915, "specialized": 0.0053369375 }, "pre_over_specialized_ratio_of_medians": 1.0912234778840861, "whole_process_pair_bootstrap_percentile95": [ 1.0256089464064375, 1.3572748924428963 ], "exact_ids_and_parameter_values_equal": true }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "median_seconds": { "pre_specialization": 0.0062986455, "specialized": 0.006070312 }, "pre_over_specialized_ratio_of_medians": 1.0376147881690432, "whole_process_pair_bootstrap_percentile95": [ 0.9568090653535901, 1.1754391334291938 ], "exact_ids_and_parameter_values_equal": true } ], "records": [ { "outer": 0, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "capture": "round0-pre_specialization-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "f04b22b7183bbfb44beea3c4cf5441cee2d7397e5f0d73f3e9b95699965d510a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.017676, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015747542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.014681084, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "specialized", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "capture": "round0-specialized-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "a5194491518198a19f99a398bbe6c55d2bf28d87d5a0f385c82224c293ffca2c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014413333, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.014250167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.01476475, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round0-pre_specialization-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "fa0454337bc07aff672c28d4fad62dfeab3a00c0b72a0f10fdcce02b0f942c68", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006098792, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005794041, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003699458, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "specialized", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round0-specialized-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "4cf0a8315774c1825b0f43b85639feca64930152c618091e6e64dab0a732eb3b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006537583, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005322542, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.0051135, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round0-pre_specialization-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "f7375478e3f7f47dca724b302b7d3d3f87a0dcd2a3f1c406eac8066b77eaaa3f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007000291, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006674917, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00605675, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "specialized", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round0-specialized-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "84a907558ba881c28c0966ae34846d7b114155f47bfae2f585501b163148554c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006546167, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005906958, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005880791, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "specialized", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round1-specialized-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "3a6c468cdcc72595df30a69871947d656915efdfa4337a4d239bf3aadea0c309", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005997, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005351333, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005024167, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round1-pre_specialization-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "b338ef389e3634b3c729fe7fa86a115537ecfd52323d069c46f86578ee9c4b05", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006190834, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005488375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005259584, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "specialized", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round1-specialized-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "35c912cf14f33524df48ec20e90e574f96c3ad200a82a503e2fd7a8032609190", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006761375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006462583, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005952583, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round1-pre_specialization-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "d5bfa02cb5fc4438929e994fed508328bd60fc8d69ad2a90604c6dcfa0bac47d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006664084, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006183458, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006035833, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "specialized", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "capture": "round1-specialized-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "7dc4ffe02ec2c3ca789473f791b064f27c8e5c469ed8c7a430e7d981924c3c2d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.017651292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.016710667, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.016318875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "capture": "round1-pre_specialization-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "0e42cc4d981905d051bfbf8a34e41ddcbf2bd2e0027732e2cd42274ff77cdc24", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.02089225, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.017042917, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015511083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "specialized", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round2-specialized-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "d587f0b522a097aa7148a020e40758c6ba9cb994c339a4a743c0fd6c0f0347fe", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005728209, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004730125, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005184875, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round2-pre_specialization-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "afc5f968b7e519b9b32587adeccf776d17c4e5dd9d379ee7faaf56cab3d3c7d6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00698875, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006413833, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006138459, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "specialized", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "capture": "round2-specialized-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "a85e44246ebedc5a73e5f877749bc81458f3dbb679623958ffab6558f7e23878", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.020914041, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.018129125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.016493792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "capture": "round2-pre_specialization-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "4f30da56b820ec831c7b866a6c1f71c63cbf915a712914a1f7a624477b4d0f63", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014880083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.014822958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.01513325, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "specialized", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round2-specialized-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "8570689c190ea4a9d996396b695e46ac7c354e46b32ce06d401f8afd318e3bb8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005384208, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003616917, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00364525, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round2-pre_specialization-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "796b068871fce29091f272f6cce85eaf52b25f0b13d35472baa85d7eff5acdbb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005910791, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005949791, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005392958, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "capture": "round3-pre_specialization-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "310dc12052cc33ee774f282c887f5ba6d5bc866ef802a4a77a29f48847824e13", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.026001333, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.022708834, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0192275, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "specialized", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "capture": "round3-specialized-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "590022431318b0d68ff4e6e12e2be276edfb3ea03c32af3a07392b74441b3b9d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.022120791, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.021269958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024724791, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round3-pre_specialization-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "aaac5efec06023d150f5b74e628d93d21ee1de1bc3d8c23c3261693453121736", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006433375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00538125, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005853542, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "specialized", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round3-specialized-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "c3258bedfeaa0f4b4619742ab862b525ce2d32632798b347983c0ae7b0afcd70", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005771416, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005676791, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005064542, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "pre_specialization", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round3-pre_specialization-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "945c6bbdf7734084920dc7e1cab3e94115e821173908f24cbb62f520850ec4d6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006508875, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006167916, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006162208, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "specialized", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "capture": "round3-specialized-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "4c8a42c0a7422028181c05e60bb0907bb94c6dbe880f84675383c890e5b4714e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006776292, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006372, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006188041, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] } ], "ci_estimand_correction": { "raw_timings_unchanged": true, "reason": "Previous intervals bootstrapped medians of process medians while the point statistic used medians of all calls; corrected draws preserve whole process pairs and recompute the same all-call statistic.", "original_summary_sha256": "4d4cfcba6ac0ddd5e10de4d4d47f72c75469cd8b75146d95bb85930d75cb28a6" } }, "ci_correction_receipt": { "schema": "aner.metal-greedy-ci-correction-audit/1", "status": "PASS", "raw_timings_unchanged": true, "confidence_interval": { "point_estimator": "median(all 12 baseline calls) / median(all 12 candidate calls)", "resampling_unit": "whole matched pair of fresh process clusters", "inner_calls": "all three calls from each selected process retained together; no inner-call resampling", "method": "exhaustive enumeration of all ordered four-cluster draws with replacement", "draws": 256, "coverage": 0.95, "percentile_method": "linear interpolation at probability * (256 - 1) in the ordered bootstrap ratios" }, "captures": [ { "capture": "local-capture-01", "original_summary_sha256": "01c741682dad6cdcac37f9dfaded132364fd745ea16511bd5ee28883d4de1cee", "corrected_summary_sha256": "3f1876ab2e9b09e149884878c199da11436977a8c8e237df079de55c190b3564", "raw_record_storage": "measurements.json", "raw_record_sha256": "892f492cd7b426c292b1084b7dff9e54d8e11ee3a6d0a369631fc11d99f6558f", "raw_records_unchanged": true, "point_estimates_unchanged": true, "all_intervals_reproduced_with_corrected_collector": true, "corrected_interval_count": 23, "audit_receipt_sha256": "907686dcb5b3a15c29223ab639a5c73b1ffd3ecf3f2dccc0390b8b98d6c8fd3d" }, { "capture": "local-capture-02", "original_summary_sha256": "be9c239b18fe5a8377d59e94827361628e8d5c269126742b9d035b2d6e6194d2", "corrected_summary_sha256": "90f2331c6ddec6f88f56f43808dd08d7550a96940d3bdd45696d37afcb6321f5", "raw_record_storage": "measurements.json", "raw_record_sha256": "b52e760f9bb2822370283fe038399d64a31d58195f8f1a623bed5cda3dcb54b0", "raw_records_unchanged": true, "point_estimates_unchanged": true, "all_intervals_reproduced_with_corrected_collector": true, "corrected_interval_count": 31, "audit_receipt_sha256": "67dc8de5e03cec2695d5e2a3975f271cce5f8352b02f213f8646b07b97c1c218" }, { "capture": "local-capture-03", "original_summary_sha256": "160e4db31ed236ccacc5be425cff616aec42d7de5f0a1dfddecd106d7ed80c98", "corrected_summary_sha256": "08b7ca43e0ff6c039d8f030ba05db88f04e629a4dfb250bb868f73ef38b2f715", "raw_record_storage": "measurements.json", "raw_record_sha256": "34a1e747d54390a023b91ab5d8ec52d03be13e6baa6122636b998332bb7f295d", "raw_records_unchanged": true, "point_estimates_unchanged": true, "all_intervals_reproduced_with_corrected_collector": true, "corrected_interval_count": 31, "audit_receipt_sha256": "9832e79ca6fbc7f544dff0fc4b2681a182c3a481751f36b7818d88e8502a4325" }, { "capture": "specialization-direct", "original_summary_sha256": "4d4cfcba6ac0ddd5e10de4d4d47f72c75469cd8b75146d95bb85930d75cb28a6", "corrected_summary_sha256": "255dfdc307dd465f1e730396f1b764a53879606168a39692fe555714169d4945", "raw_record_storage": "embedded records in archived and corrected summaries", "raw_record_sha256": "4d4cfcba6ac0ddd5e10de4d4d47f72c75469cd8b75146d95bb85930d75cb28a6", "raw_records_unchanged": true, "point_estimates_unchanged": true, "all_intervals_reproduced_with_corrected_collector": true, "corrected_interval_count": 3, "audit_receipt_sha256": "36bf041d4258624d0c8a96e5351643adfd11142c73ec44b55b7644116bea85e9" } ], "correction_script_sha256": "fb3d1bfa1b8ade25913fa02bc2a85c18e73832680a9f81d6a5d9614f955fe4ce", "corrected_collector_sha256": "a8f6375b967b8245c13bc7575edc711c9b8463bd3e8ab19451c28dcb0781aedc", "capture_audit_script_sha256": "e44b8bd650f29fd18951ef336b048c37d9ad47595630206b7b279e0e3527d67b", "direct_audit_script_sha256": "c9d93c7728607fac9de0e87ff1b4b48e38622b90a5d8d14fed4978aec1e222d9" }, "capture_audit_receipts": { "01": { "schema": "aner.metal-greedy-capture-audit/1", "status": "PASS", "profiles": 23, "fresh_timing_processes": 184, "retained_timed_calls": 552, "excluded_warmup_calls": 368, "resident_cap_and_kv_cases": 80, "all_natural_ids_and_counts_equal": true, "all_parameter_values_equal": true, "all_confidence_intervals_reproduced_exactly": true, "all_retained_numeric_bytes_zero": true, "all_resident_full_logit_readback_counters_zero": true, "final_id_bytes_equal_four_times_actual_emitted_tokens": true, "maximum_step_scaled_error": 3.230860285372708e-07, "maximum_resident_kv_probe_scaled_error": 2.3617566891512637e-07, "parameter_receipts": [ { "capture": "baseline-weights-cpu-tiny-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" }, { "capture": "baseline-weights-mps-eos1-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "e3e6ff6db789c848e78a759a6fc69b3e2c9329dff3ea951843eb0b348bf82f2a" }, { "capture": "baseline-weights-mps-eos10-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "ad522891246ca60c51e52ff565c3016a616a50b164ddd6127c9353a5e8ed1b56" }, { "capture": "baseline-weights-mps-eos2-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "ed9153be24577314bb4e31f0283b576ff51df1335d415bb9a6d1e361b1ad1d08" }, { "capture": "baseline-weights-mps-eos8-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "6b16fedfaa1cd79eccddbe930376a7c04833dd460280fad86699368b2331c5a9" }, { "capture": "baseline-weights-mps-eos9-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "bf13ad8b95d255bb2ae96a8f20a80f80154beebf9165d17e9d16642217f96818" }, { "capture": "baseline-weights-mps-mixed_eos-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "ae87e82e1f0040a28a821ff69662ed1773bc854740534880fa8f1670d11d1973" }, { "capture": "baseline-weights-mps-tiny-b1-w128-l3.stdout", "parameter_tensors": 28, "parameter_scalar_values": 637056, "sha256": "7432a93bd25dd41f127bc7621a6b14e5e15fba11c703df9d7c8ebfc28316c9d8" }, { "capture": "baseline-weights-mps-tiny-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" } ], "measurements_sha256": "892f492cd7b426c292b1084b7dff9e54d8e11ee3a6d0a369631fc11d99f6558f", "public_summary_sha256": "3f1876ab2e9b09e149884878c199da11436977a8c8e237df079de55c190b3564", "diagnostic_public_summary_sha256": "9334ec74069b18ffb566d5a7d8fb9577b08b59d49f85cad446c07eaadeb24709", "build_receipt_sha256": "923c521fd4db046a484039d052232ad59218b79e92e0e2968d6c11179e149974" }, "02": { "schema": "aner.metal-greedy-capture-audit/1", "status": "PASS", "profiles": 31, "fresh_timing_processes": 248, "retained_timed_calls": 744, "excluded_warmup_calls": 496, "resident_cap_and_kv_cases": 112, "all_natural_ids_and_counts_equal": true, "all_parameter_values_equal": true, "all_confidence_intervals_reproduced_exactly": true, "all_retained_numeric_bytes_zero": true, "all_resident_full_logit_readback_counters_zero": true, "final_id_bytes_equal_four_times_actual_emitted_tokens": true, "maximum_step_scaled_error": 3.230860285372708e-07, "maximum_resident_kv_probe_scaled_error": 2.3617566891512637e-07, "parameter_receipts": [ { "capture": "baseline-weights-cpu-tiny-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" }, { "capture": "baseline-weights-mps-eos1-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "e3e6ff6db789c848e78a759a6fc69b3e2c9329dff3ea951843eb0b348bf82f2a" }, { "capture": "baseline-weights-mps-eos10-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "ad522891246ca60c51e52ff565c3016a616a50b164ddd6127c9353a5e8ed1b56" }, { "capture": "baseline-weights-mps-eos2-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "ed9153be24577314bb4e31f0283b576ff51df1335d415bb9a6d1e361b1ad1d08" }, { "capture": "baseline-weights-mps-eos3-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "0fe8452022f38407bdad8c24d202a81f68d5b0d431f919295c67ad8d0449a505" }, { "capture": "baseline-weights-mps-eos4-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "7c24715177c00583dcbc5cd02c511fbe8439752dd4ddb1561eb79bdf060f04d6" }, { "capture": "baseline-weights-mps-eos5-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "3b57c548c9d63e1a6514ae59badabbb062de7b99c8321f84af893c5d563ba079" }, { "capture": "baseline-weights-mps-eos6-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "5c516784a0fe783125b7b5cb1e52e702d95c2adab654d2cbf376a0f79c02043d" }, { "capture": "baseline-weights-mps-eos8-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "6b16fedfaa1cd79eccddbe930376a7c04833dd460280fad86699368b2331c5a9" }, { "capture": "baseline-weights-mps-eos9-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "bf13ad8b95d255bb2ae96a8f20a80f80154beebf9165d17e9d16642217f96818" }, { "capture": "baseline-weights-mps-mixed_eos-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "ae87e82e1f0040a28a821ff69662ed1773bc854740534880fa8f1670d11d1973" }, { "capture": "baseline-weights-mps-tiny-b1-w128-l3.stdout", "parameter_tensors": 28, "parameter_scalar_values": 637056, "sha256": "7432a93bd25dd41f127bc7621a6b14e5e15fba11c703df9d7c8ebfc28316c9d8" }, { "capture": "baseline-weights-mps-tiny-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" } ], "measurements_sha256": "b52e760f9bb2822370283fe038399d64a31d58195f8f1a623bed5cda3dcb54b0", "public_summary_sha256": "90f2331c6ddec6f88f56f43808dd08d7550a96940d3bdd45696d37afcb6321f5", "diagnostic_public_summary_sha256": "feb34de7a2bcde9fba0e977a700f7519cb2ed5ab742d724b183fe27c5409e5ab", "build_receipt_sha256": "0afafc84727042949125ed668080a4b3369423eca5fbdc1e31050185af5027f7" }, "03": { "schema": "aner.metal-greedy-capture-audit/1", "status": "PASS", "profiles": 31, "fresh_timing_processes": 248, "retained_timed_calls": 744, "excluded_warmup_calls": 496, "resident_cap_and_kv_cases": 112, "all_natural_ids_and_counts_equal": true, "all_parameter_values_equal": true, "all_confidence_intervals_reproduced_exactly": true, "all_retained_numeric_bytes_zero": true, "all_resident_full_logit_readback_counters_zero": true, "final_id_bytes_equal_four_times_actual_emitted_tokens": true, "maximum_step_scaled_error": 3.230860285372708e-07, "maximum_resident_kv_probe_scaled_error": 2.3617566891512637e-07, "parameter_receipts": [ { "capture": "baseline-weights-cpu-tiny-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" }, { "capture": "baseline-weights-mps-eos1-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "e3e6ff6db789c848e78a759a6fc69b3e2c9329dff3ea951843eb0b348bf82f2a" }, { "capture": "baseline-weights-mps-eos10-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "ad522891246ca60c51e52ff565c3016a616a50b164ddd6127c9353a5e8ed1b56" }, { "capture": "baseline-weights-mps-eos2-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "ed9153be24577314bb4e31f0283b576ff51df1335d415bb9a6d1e361b1ad1d08" }, { "capture": "baseline-weights-mps-eos3-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "0fe8452022f38407bdad8c24d202a81f68d5b0d431f919295c67ad8d0449a505" }, { "capture": "baseline-weights-mps-eos4-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "7c24715177c00583dcbc5cd02c511fbe8439752dd4ddb1561eb79bdf060f04d6" }, { "capture": "baseline-weights-mps-eos5-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "3b57c548c9d63e1a6514ae59badabbb062de7b99c8321f84af893c5d563ba079" }, { "capture": "baseline-weights-mps-eos6-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "5c516784a0fe783125b7b5cb1e52e702d95c2adab654d2cbf376a0f79c02043d" }, { "capture": "baseline-weights-mps-eos8-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "6b16fedfaa1cd79eccddbe930376a7c04833dd460280fad86699368b2331c5a9" }, { "capture": "baseline-weights-mps-eos9-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "bf13ad8b95d255bb2ae96a8f20a80f80154beebf9165d17e9d16642217f96818" }, { "capture": "baseline-weights-mps-mixed_eos-b8-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "ae87e82e1f0040a28a821ff69662ed1773bc854740534880fa8f1670d11d1973" }, { "capture": "baseline-weights-mps-tiny-b1-w128-l3.stdout", "parameter_tensors": 28, "parameter_scalar_values": 637056, "sha256": "7432a93bd25dd41f127bc7621a6b14e5e15fba11c703df9d7c8ebfc28316c9d8" }, { "capture": "baseline-weights-mps-tiny-b1-w32-l2.stdout", "parameter_tensors": 20, "parameter_scalar_values": 36256, "sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" } ], "measurements_sha256": "34a1e747d54390a023b91ab5d8ec52d03be13e6baa6122636b998332bb7f295d", "public_summary_sha256": "08b7ca43e0ff6c039d8f030ba05db88f04e629a4dfb250bb868f73ef38b2f715", "diagnostic_public_summary_sha256": "36290938b046b892883afafb92f71e062e78af6491b8739b8e99ae1c52e2c326", "build_receipt_sha256": "3e45c0cf5bfb4077d9563ccab9e25f45933a179b88cccc0411c9d9320186aa71" } }, "statistical_correction": "All reported intervals resample matched process pairs and retain each selected process\u2019s three calls; the all-call median ratio is recomputed for each of256 exhaustive ordered draws. Original derived summaries are archived; raw measurements are unchanged.", "specialization_primary_source": "https://developer.apple.com/documentation/metal/mtlfunctionconstantvalues", "final_capture": "03", "source_sha256": { "include/aner/gpu.hpp": "54944d4a92c7944f8b2a3001d769da696c53e16661d1d08ff6c99c01fd01a75a", "include/aner/nn.hpp": "eab656a1461b0e8597428129c41ac201617bebdad4a5fa2e895efdf3cc19ca40", "include/aner/gen.hpp": "65f857502551ecb35d47d6437f321af0b9e9cac8bb0658b09f65ba0f3bffe82e", "src/gpu_mps.mm": "706c9b9bd30d8c3bafe0e77c7a236d4448a99eb7483669ac0d6920f5456f212c", "src/nn.cpp": "cf7677f37790a02dc93570e7e7e6257597535da6e79b2402c0a48ee23e7f3f01", "src/gen.cpp": "9c0102ab44db1a08af6d187ef5a7e8820b7eed2ab401c4666c06cf80743131da", "tests/gen_cache_tests.cpp": "19272c30b741a6e2e61b4208980dbfdfd9a683d423a486160055541807dcd018", "tests/gpu_tensor_tests.cpp": "79eed8ae2e78f49ec56b79b64857d3e9ab0ae6f3a0b98d7976bae7132c4eeca5" }, "mac_validation": { "groups": 283, "passed": 274, "skipped": 9, "failed": 0, "research_campaigns_excluded": 6, "wall_seconds": 286.87, "skips": [ "cnn_nccl", "cnn_cuda_training_strict", "cnn_cuda_training_fp16", "cnn_cuda_training_bf16", "cnn_cuda_training_tf32", "aner.db_mysql_live", "aner.db_mssql_live", "aner.db_mongodb_live", "aner.db_postgres_live" ], "resident_checks": 11990172, "resident_max_scaled_error": 5.78247e-08, "existing_mps_single_checks": 2392539, "gpu_provider_checks": 267, "sanitizers": { "address": true, "undefined_behavior": true, "leak_detector_enabled": false, "resident_mps": "PASS", "gpu_provider": "PASS", "cpu_float32": "PASS" }, "apple_metal_validation": { "api_enabled": true, "shader_enabled": true, "status": "PASS" }, "logs": { "mac-final-specialization-ctest.log": { "sha256": "423af74a4cee7eb3c1d98569553e77b7ffd7ce8e2736ca851693858771a8547b", "bytes": 37322 }, "specialization-resident.log": { "sha256": "074e0435d8d44f940a23e33f634d6ab3ccd8d891414b2f30f0d48e1bbd063eeb", "bytes": 112 }, "specialization-single.log": { "sha256": "0c2071b2bd0858bafc480bd9eac4b7ec95a3308f0d4d60579c6f8b3b704a7985", "bytes": 157 }, "specialization-focused-ctest.log": { "sha256": "81b1b41eec8de0cf16d684522f65b1ed3d14ca46a4039b7f0c5337f04b352f10", "bytes": 497 }, "sanitize-specialization-resident.log": { "sha256": "074e0435d8d44f940a23e33f634d6ab3ccd8d891414b2f30f0d48e1bbd063eeb", "bytes": 112 }, "sanitize-specialization-provider.log": { "sha256": "877066f6099b2b6753bbcaffdc265337ad12f200f1462998956e1728999163b4", "bytes": 108 }, "sanitize-final-cpu.log": { "sha256": "909fa79a6e5ca20a3584d3847262250debccfa46c736a59d8d6b1dd61435d2b6", "bytes": 157 }, "metal-specialization-validation.stdout.log": { "sha256": "074e0435d8d44f940a23e33f634d6ab3ccd8d891414b2f30f0d48e1bbd063eeb", "bytes": 112 }, "metal-specialization-validation.stderr.log": { "sha256": "19c8c436bd6bfa2b6f401ab3568d310eb44deb0640f8d74bc4aabeb317c1023d", "bytes": 178 } }, "policy": "gradual1,2,4,...cap8; small one-token packed-logit scan fallback below65536", "final_source": { "src/gpu_mps.mm": "706c9b9bd30d8c3bafe0e77c7a236d4448a99eb7483669ac0d6920f5456f212c", "src/gen.cpp": "9c0102ab44db1a08af6d187ef5a7e8820b7eed2ab401c4666c06cf80743131da", "src/nn.cpp": "cf7677f37790a02dc93570e7e7e6257597535da6e79b2402c0a48ee23e7f3f01", "include/aner/gpu.hpp": "54944d4a92c7944f8b2a3001d769da696c53e16661d1d08ff6c99c01fd01a75a", "include/aner/nn.hpp": "eab656a1461b0e8597428129c41ac201617bebdad4a5fa2e895efdf3cc19ca40", "include/aner/gen.hpp": "65f857502551ecb35d47d6437f321af0b9e9cac8bb0658b09f65ba0f3bffe82e", "tests/gen_cache_tests.cpp": "19272c30b741a6e2e61b4208980dbfdfd9a683d423a486160055541807dcd018", "tests/gpu_tensor_tests.cpp": "79eed8ae2e78f49ec56b79b64857d3e9ab0ae6f3a0b98d7976bae7132c4eeca5" }, "kernel_mask_variants": { "specialized": true, "additional_pipeline_states": 7, "cold_start_performance_measured": false } }, "v100_regression_validation": { "schema": "aner.metal-greedy-v100-regression/1", "status": "PASS", "scope": "CUDA Float64/Float32 shared-code regression only; Metal resident performance is measured separately on the Mac. The Metal source file is not compiled in this Linux CUDA build.", "host": { "gpu": "Tesla V100-PCIE-32GB", "memory_mib": 32768, "compute_capability": "7.0", "driver": "580.178.04", "system": "Linux x86_64" }, "build": { "configuration": "Release", "lto": false, "cxx_compiler": "GNU 13.4.0", "cuda_compiler": "12.9.86", "cublas": "12.9.1.4", "cuda_architectures": "70", "metal_enabled": false, "flags": { "generative": { "CXX_FLAGS": "-O3 -DNDEBUG -std=c++20 -Wall -Wextra -Wpedantic -fno-fast-math -ffp-contract=off" }, "neural": { "CXX_FLAGS": "-O3 -DNDEBUG -std=c++20 -Wall -Wextra -Wpedantic -fno-fast-math -ffp-contract=off", "CUDA_FLAGS": "-O3 -DNDEBUG -std=c++20 \"--generate-code=arch=compute_70,code=[compute_70,sm_70]\" --fmad=false" } }, "toolchain_receipt_sha256": "541b0d27630bfcff19e83b8638b797a41539f086665016a41c2c98bb96b421b6" }, "source_files_verified": 6530, "source_manifest_sha256": "8f79ad62bc5ea8b364e23346c87251ad312896dcde4bee8c72a181fc80edbfc4", "compiled_shared_source_sha256": { "CMakeLists.txt": "2aa45ff88b08801a2afdb8ceb32978972db18b379587c05048c3889f9dd501db", "src/gen.cpp": "9c0102ab44db1a08af6d187ef5a7e8820b7eed2ab401c4666c06cf80743131da", "src/nn.cpp": "cf7677f37790a02dc93570e7e7e6257597535da6e79b2402c0a48ee23e7f3f01", "src/gpu.cpp": "c69be6b251398e1c664df1bb3ae75f9e4a6c597678d800752671a8027be68f1b", "src/gpu_cuda.cu": "96a814e38c9b1a92634683498a60d754bc7c7c53a8b0f686fe162d070ade4f44", "include/aner/gen.hpp": "65f857502551ecb35d47d6437f321af0b9e9cac8bb0658b09f65ba0f3bffe82e", "include/aner/nn.hpp": "eab656a1461b0e8597428129c41ac201617bebdad4a5fa2e895efdf3cc19ca40", "include/aner/gpu.hpp": "54944d4a92c7944f8b2a3001d769da696c53e16661d1d08ff6c99c01fd01a75a", "tests/gen_cache_tests.cpp": "19272c30b741a6e2e61b4208980dbfdfd9a683d423a486160055541807dcd018", "tests/gpu_tensor_tests.cpp": "79eed8ae2e78f49ec56b79b64857d3e9ab0ae6f3a0b98d7976bae7132c4eeca5" }, "native_groups": [ { "group": "nn", "exit": 0, "reported_checks": 4483, "stdout_sha256": "487d688b402eaee17a3f22295085c571db8e67472b23faa8c5d65b7c7b0122ea" }, { "group": "memory", "exit": 0, "reported_checks": 33, "stdout_sha256": "b37a3c76e6faa61fe0112d73cdc2d126b18cab7c137105f032554641e5bac32c" }, { "group": "sequence", "exit": 0, "reported_checks": 912, "stdout_sha256": "8534ae856e4cc429199b5925233a3fe68effe7a2ab1f6c11ec4d8bc32fa69124" }, { "group": "gen", "exit": 0, "reported_checks": 3150, "stdout_sha256": "59f3fb7c873c4c64dd6ba6365ee0e2d99f7ce5ce3259dced5f6d503efdb7740f" }, { "group": "tensor", "exit": 0, "reported_checks": 321, "stdout_sha256": "a23bda36558b93df92e4d036b54db93e47f82cf378cd7050189dbcdf708376e2" }, { "group": "sequence-device-Float64", "exit": 0, "reported_checks": 141277, "stdout_sha256": "b992fa1ffbe5aac714fd3f45f27f099f81040327ed34b6bb8beefa4d540681aa" }, { "group": "gen-device-Float64", "exit": 0, "reported_checks": 19250, "stdout_sha256": "0f8f5a8127ee688497ba5267cd74b7f24840daab9448f95d525ede0527f3778d" }, { "group": "cache-Float64", "exit": 0, "reported_checks": 2397275, "stdout_sha256": "af65442f5beb7652ddbc51563b0b07dd49d163990e069b30fddde56f34e68a16" }, { "group": "batch-Float64", "exit": 0, "reported_checks": 20341919, "stdout_sha256": "dc08b61ab7d1da241d9471fedb13837ecb04e07ed79f3482fe23da0c763a429d" }, { "group": "sequence-device-Float32", "exit": 0, "reported_checks": 141275, "stdout_sha256": "fb0b1fed7733aaba8c643c0fdfbe480ceec672371429d9b62dbcd412ff02bc36" }, { "group": "gen-device-Float32", "exit": 0, "reported_checks": 19250, "stdout_sha256": "f122df5b9f33860a856bae80934bc78411496f282b3e4057ed73c0a95f35f506" }, { "group": "cache-Float32", "exit": 0, "reported_checks": 2397275, "stdout_sha256": "330908d71fa6ad7d73e4592312a47efd86c6cb9ad7356db99969dac231922a10" }, { "group": "batch-Float32", "exit": 0, "reported_checks": 20341919, "stdout_sha256": "6d949e3b06dfdd318c1c6867fbb26b2e6d806f052932da768097a773aa3717d0" } ], "native_groups_passed": 13, "compute_sanitizer": [ { "group": "sanitizer-tensor", "exit": 0, "memcheck_errors": 0, "stdout_sha256": "b3240ea4ebe84210dd3705236adfcd710583c211f88c4540d09683733bc37460" }, { "group": "sanitizer-batch-Float64", "exit": 0, "memcheck_errors": 0, "stdout_sha256": "44212fdc9b09c789adae06198da101ab27d37f1a888efdf1ec31d77d146464f7" }, { "group": "sanitizer-batch-Float32", "exit": 0, "memcheck_errors": 0, "stdout_sha256": "cb4432b570430227fc0ef7d0fe59732db85d1a99944015e1040ca88be8a8fc7a" } ], "producer_smokes": [ { "group": "cache-producer-Float64", "exit": 0, "stdout_sha256": "9e38c5f075a52c63ccc0bb4551f04c5e417b8c0067c246b157a282dcb6133e01" }, { "group": "batch-producer-Float64", "exit": 0, "stdout_sha256": "021b8cf18374d4094d91489adf3522206428da9a99c7ddab6a675c9d54de9557" }, { "group": "cache-producer-Float32", "exit": 0, "stdout_sha256": "7689450cbe7d532b0cf2f8742f9f6d17866f0d07b83040717253e662d2339c49" }, { "group": "batch-producer-Float32", "exit": 0, "stdout_sha256": "7c716cd62d7684d68e2ea974943ffa0c42e7e873d1e707f3d804c1fb6f3d32ec" } ], "final_gpu_memory_used_mib": 0, "final_gpu_utilization_percent": 0, "task_pipeline_finished": true, "execution_binaries_sha256": { "aner_gen_cache_tests": "dd25d72d954da054c48f17599e9763ceb3498cd270ed68623df954d4cf50acb9", "aner_gen_device_tests": "28015499ffe77fa4661505cd4095448dff8c6d750bddc842486328aeab0cb8df", "aner_gpu_tensor_tests": "ed38d0bc3af6ce42d31a882be684b7e8f82fd25ce80cf581edfb92a50ad25507", "aner_nn_tests": "fbd5c3fc52d310946830cf05591cdc478af4632ae58043c9e68ca2dc49faa081", "aner_nn_memory_tests": "6c3865d263a7581963da3b1803cd95d8eca86fd9015494c8f1430ba4954194bb", "aner_sequence_device_tests": "8859cb24b74a9cfadf584eb52e323ffcf87dd9c8ecfb041773e837999c2b6229", "aner_sequence_tests": "3e63a9150c28f8d5eabb608b31cd8f119468572d790e69b815bbffa297b31eb8", "aner_gen_tests": "297b563b2ab910db2a89caf143abf5bc1acfa499362e2d913f025f21c38864aa", "aner_gen_cache_benchmark": "5d4ae8a40bfff6023e2db43335ce5a2399992e8e8225d17da19e3e4d8ad5643f", "aner_gen_batch_benchmark": "639b05038e061b669a3d245cad1d87d50fa3d63d3a9275808df058927db12a76" }, "final_metal_only_followup": { "tested_snapshot_metal_source_sha256": "8326fce362ca42114f2be2ece059feeb9dfa6e783440aff9f3d53899750c535b", "accepted_final_metal_source_sha256": "706c9b9bd30d8c3bafe0e77c7a236d4448a99eb7483669ac0d6920f5456f212c", "all_recorded_compiled_shared_cuda_source_hashes_match_final": true, "metal_compiled_in_cuda_build": false, "scope": "The accepted follow-up changes only the uncompiled Metal backend. Every recorded compiled shared/CUDA source hash above matches the final accepted source; these CUDA results qualify shared-code regressions, not the Metal specialization." } }, "experiments": [ { "capture": "01", "label": "rejected fixed eight-step startup", "is_final": false, "summary": { "schema": "aner.metal-greedy-local-summary/1", "protocol": { "matched_process_pairs": 4, "timed_calls_per_process": 3, "excluded_warmups_per_process": 2, "process_order": "AB,BA,BA,AB; profile order rotates by outer round", "primary_phase": "natural", "natural_policy": "greedy complete requests; actual IDs, EOS and counts retained; baseline/candidate ID equality is measured, not assumed", "controls": "CPU Float32 tiny request and unchanged nonzero-temperature MPS sampling; all-zero and one-token budgets retained as functional/latency controls", "memory_scope": "tracked owned numeric capacities; readback counter includes metadata and Float32 widening; excludes stack vectors, allocator/driver/command-buffer/BLAS overhead and RSS", "thread_environment_request": { "VECLIB_MAXIMUM_THREADS": "1", "OPENBLAS_NUM_THREADS": "1", "OMP_NUM_THREADS": "1" }, "confidence_interval": { "point_estimator": "median(all 12 baseline calls) / median(all 12 candidate calls)", "resampling_unit": "whole matched pair of fresh process clusters", "inner_calls": "all three calls from each selected process retained together; no inner-call resampling", "method": "exhaustive enumeration of all ordered four-cluster draws with replacement", "draws": 256, "coverage": 0.95, "percentile_method": "linear interpolation at probability * (256 - 1) in the ordered bootstrap ratios" } }, "producer_sha256": "f2b16c3927a79162e1d4165cc907c0b4f6b8453b23d47aaae696aad16dfb0838", "binary_sha256": { "baseline": "d626cfa793ded8eec98c20d664130c6db37143c47b71580af0ca83dd60e4ff3c", "candidate": "6e6b6c14db78bea4491b9f3aa476484ab64d86320949072f08d93b0fb2c48f9a" }, "build": { "configuration": "Release", "lto": false, "devices": [ "mps", "cpu" ], "dtype": "Float32", "compiler": { "id": "AppleClang", "version": "21.0.0.21000334" }, "compiler_flags": [ "-O3", "-DNDEBUG", "-std=c++20", "-mcpu=apple-m1", "-fno-fast-math", "-ffp-contract=off" ] }, "measurement_sha256": "892f492cd7b426c292b1084b7dff9e54d8e11ee3a6d0a369631fc11d99f6558f", "oracle": { "schema": "aner.metal-greedy-local-oracle/1", "profiles": [ { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 1.3992326293094704e-07, "candidate_max_scaled_error_vs_cpu_float64": 1.3992326293094704e-07, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 32, "baseline_max_scaled_error_vs_cpu_float64": 9.61704129314485e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.61704129314485e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 3.0323179185145576e-07, "candidate_max_scaled_error_vs_cpu_float64": 3.0323179185145576e-07, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 3.230860285372708e-07, "candidate_max_scaled_error_vs_cpu_float64": 3.230860285372708e-07, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 9.487577588236166e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.487577588236166e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 0, "baseline_max_scaled_error_vs_cpu_float64": 0, "candidate_max_scaled_error_vs_cpu_float64": 0, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 0, "baseline_max_scaled_error_vs_cpu_float64": 0, "candidate_max_scaled_error_vs_cpu_float64": 0, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 1, "baseline_max_scaled_error_vs_cpu_float64": 6.854355312918667e-08, "candidate_max_scaled_error_vs_cpu_float64": 6.854355312918667e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 1, "baseline_max_scaled_error_vs_cpu_float64": 6.68301477395295e-08, "candidate_max_scaled_error_vs_cpu_float64": 6.68301477395295e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 2.907853657679229e-08, "candidate_max_scaled_error_vs_cpu_float64": 2.907853657679229e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 } ], "exact_parameter_values_equal": true, "parameter_values": [ { "device": "cpu", "width": 32, "layers": 2, "model_fixture": "random", "parameter_serialization_sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos1", "parameter_serialization_sha256": "e3e6ff6db789c848e78a759a6fc69b3e2c9329dff3ea951843eb0b348bf82f2a" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos10", "parameter_serialization_sha256": "ad522891246ca60c51e52ff565c3016a616a50b164ddd6127c9353a5e8ed1b56" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos2", "parameter_serialization_sha256": "ed9153be24577314bb4e31f0283b576ff51df1335d415bb9a6d1e361b1ad1d08" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos8", "parameter_serialization_sha256": "6b16fedfaa1cd79eccddbe930376a7c04833dd460280fad86699368b2331c5a9" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos9", "parameter_serialization_sha256": "bf13ad8b95d255bb2ae96a8f20a80f80154beebf9165d17e9d16642217f96818" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "mixed_eos", "parameter_serialization_sha256": "ae87e82e1f0040a28a821ff69662ed1773bc854740534880fa8f1670d11d1973" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "random", "parameter_serialization_sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" }, { "device": "mps", "width": 128, "layers": 3, "model_fixture": "random", "parameter_serialization_sha256": "7432a93bd25dd41f127bc7621a6b14e5e15fba11c703df9d7c8ebfc28316c9d8" } ] }, "summary": [ { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.005225520500000001, "candidate": 0.0033351665000000003 }, "baseline_over_candidate_ratio_of_medians": 1.5667944913694714, "whole_process_pair_bootstrap_percentile95": [ 1.052321214279843, 2.0790521785425446 ], "min_max_seconds": { "baseline": [ 0.003226541, 0.006930292 ], "candidate": [ 0.00248975, 0.005100166 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0057469375, "candidate": 0.0046075414999999995 }, "baseline_over_candidate_ratio_of_medians": 1.2472893624506693, "whole_process_pair_bootstrap_percentile95": [ 1.1378676292432073, 1.6410822058691021 ], "min_max_seconds": { "baseline": [ 0.0054005, 0.006847541 ], "candidate": [ 0.002873542, 0.005473708 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0291624165, "candidate": 0.017615729 }, "baseline_over_candidate_ratio_of_medians": 1.6554759953448421, "whole_process_pair_bootstrap_percentile95": [ 1.4032191912125807, 2.062249056624835 ], "min_max_seconds": { "baseline": [ 0.022761458, 0.042271292 ], "candidate": [ 0.013728792, 0.0204765 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 64 ], "candidate": [ 64 ] } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.025086146, "candidate": 0.016284333499999998 }, "baseline_over_candidate_ratio_of_medians": 1.5405079980706613, "whole_process_pair_bootstrap_percentile95": [ 1.2720896375964463, 1.6978757471430954 ], "min_max_seconds": { "baseline": [ 0.023542334, 0.0287385 ], "candidate": [ 0.01468425, 0.020731791 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "candidate": [ 64, 64, 64, 64, 64, 64, 64, 64 ] } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0136113125, "candidate": 0.00989425 }, "baseline_over_candidate_ratio_of_medians": 1.3756790560173837, "whole_process_pair_bootstrap_percentile95": [ 1.1891641526975996, 1.9563509209181764 ], "min_max_seconds": { "baseline": [ 0.012212834, 0.026738334 ], "candidate": [ 0.007757291, 0.012938541 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "candidate": [ 32, 1, 16, 2, 24, 3, 8, 4 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0059986664999999995, "candidate": 0.0051807915 }, "baseline_over_candidate_ratio_of_medians": 1.1578668047150709, "whole_process_pair_bootstrap_percentile95": [ 1.0832637288135594, 2.221133387824667 ], "min_max_seconds": { "baseline": [ 0.00476275, 0.0163265 ], "candidate": [ 0.003492916, 0.006495042 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0053203750000000005, "candidate": 0.0045909160000000004 }, "baseline_over_candidate_ratio_of_medians": 1.158891820281617, "whole_process_pair_bootstrap_percentile95": [ 1.0977288355017951, 1.3284487691467077 ], "min_max_seconds": { "baseline": [ 0.004942292, 0.006873292 ], "candidate": [ 0.003781583, 0.005862792 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.001012313, "candidate": 0.0010996455 }, "baseline_over_candidate_ratio_of_medians": 0.9205812236761756, "whole_process_pair_bootstrap_percentile95": [ 0.839985321526175, 0.9450892531876138 ], "min_max_seconds": { "baseline": [ 0.000660875, 0.001175917 ], "candidate": [ 0.00105725, 0.001619625 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1 ], "candidate": [ 1 ] } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0010231455, "candidate": 0.001054375 }, "baseline_over_candidate_ratio_of_medians": 0.9703810314167161, "whole_process_pair_bootstrap_percentile95": [ 0.9223755490882137, 1.096972614953509 ], "min_max_seconds": { "baseline": [ 0.000958416, 0.001150292 ], "candidate": [ 0.000657333, 0.001246 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "candidate": [ 1, 1, 1, 1, 1, 1, 1, 1 ] } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0019259165, "candidate": 0.0045396245000000005 }, "baseline_over_candidate_ratio_of_medians": 0.4242457718694574, "whole_process_pair_bootstrap_percentile95": [ 0.3671691164610896, 0.4376863678012799 ], "min_max_seconds": { "baseline": [ 0.000939375, 0.002387125 ], "candidate": [ 0.003485459, 0.006074709 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 2 ], "candidate": [ 2 ] } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0020029375, "candidate": 0.004233583500000001 }, "baseline_over_candidate_ratio_of_medians": 0.47310688451048616, "whole_process_pair_bootstrap_percentile95": [ 0.4217573708624892, 0.5916043354285235 ], "min_max_seconds": { "baseline": [ 0.001252458, 0.002036041 ], "candidate": [ 0.002524209, 0.004886167 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "candidate": [ 2, 2, 2, 2, 2, 2, 2, 2 ] } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0062818335, "candidate": 0.0044305005 }, "baseline_over_candidate_ratio_of_medians": 1.4178609166165312, "whole_process_pair_bootstrap_percentile95": [ 1.1189860910435223, 1.8342887605734275 ], "min_max_seconds": { "baseline": [ 0.005442875, 0.007438708 ], "candidate": [ 0.002762167, 0.005509625 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.006332208000000001, "candidate": 0.0055088545 }, "baseline_over_candidate_ratio_of_medians": 1.1494600193198061, "whole_process_pair_bootstrap_percentile95": [ 0.9336407616072373, 1.2075281822707096 ], "min_max_seconds": { "baseline": [ 0.004763125, 0.007168875 ], "candidate": [ 0.00511275, 0.00730425 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "candidate": [ 9, 9, 9, 9, 9, 9, 9, 9 ] } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.006812271, "candidate": 0.0054214375 }, "baseline_over_candidate_ratio_of_medians": 1.2565433060880993, "whole_process_pair_bootstrap_percentile95": [ 1.2228527870864208, 1.3575386881640268 ], "min_max_seconds": { "baseline": [ 0.005604833, 0.007936375 ], "candidate": [ 0.004891375, 0.006026709 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "candidate": [ 10, 10, 10, 10, 10, 10, 10, 10 ] } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.026035625, "candidate": 0.017423167 }, "baseline_over_candidate_ratio_of_medians": 1.4943107071177129, "whole_process_pair_bootstrap_percentile95": [ 1.3505529691667033, 1.8591231052046249 ], "min_max_seconds": { "baseline": [ 0.023342125, 0.040042541 ], "candidate": [ 0.014258167, 0.020778125 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "candidate": [ 64, 9, 8, 7, 6, 5, 3, 1 ] } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.023505645999999998, "candidate": 0.015795625 }, "baseline_over_candidate_ratio_of_medians": 1.488111170023345, "whole_process_pair_bootstrap_percentile95": [ 1.2662124163878183, 1.5790778594854191 ], "min_max_seconds": { "baseline": [ 0.023141166, 0.02872675 ], "candidate": [ 0.015353625, 0.020203584 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "candidate": [ 0, 1, 0, 8, 64, 1, 0, 16 ] } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 1.85415e-05, "candidate": 2.25835e-05 }, "baseline_over_candidate_ratio_of_medians": 0.8210197710718001, "whole_process_pair_bootstrap_percentile95": [ 0.7523127485190294, 1.1003466306327478 ], "min_max_seconds": { "baseline": [ 1.4208e-05, 3.6834e-05 ], "candidate": [ 1.45e-05, 2.7709e-05 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 0 ], "candidate": [ 0 ] } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 2.2666499999999997e-05, "candidate": 2.0458e-05 }, "baseline_over_candidate_ratio_of_medians": 1.1079528790693127, "whole_process_pair_bootstrap_percentile95": [ 0.7591384310134309, 1.5590950797109762 ], "min_max_seconds": { "baseline": [ 1.8208e-05, 5.7e-05 ], "candidate": [ 1.4875e-05, 4.1417e-05 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "candidate": [ 0, 0, 0, 0, 0, 0, 0, 0 ] } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0009224375000000001, "candidate": 0.0010165414999999999 }, "baseline_over_candidate_ratio_of_medians": 0.907427291458342, "whole_process_pair_bootstrap_percentile95": [ 0.8470436579826006, 0.9408940710714343 ], "min_max_seconds": { "baseline": [ 0.000532667, 0.0009715 ], "candidate": [ 0.000982166, 0.001210292 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1 ], "candidate": [ 1 ] } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.000992396, "candidate": 0.001039771 }, "baseline_over_candidate_ratio_of_medians": 0.9544370827807276, "whole_process_pair_bootstrap_percentile95": [ 0.929868785160194, 1.0837442497261776 ], "min_max_seconds": { "baseline": [ 0.000967875, 0.001023 ], "candidate": [ 0.000871084, 0.001098375 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "candidate": [ 1, 1, 1, 1, 1, 1, 1, 1 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0051710415, "candidate": 0.0055467295 }, "baseline_over_candidate_ratio_of_medians": 0.9322685557318056, "whole_process_pair_bootstrap_percentile95": [ 0.7621407990621402, 1.5140357936048772 ], "min_max_seconds": { "baseline": [ 0.004312667, 0.007508833 ], "candidate": [ 0.00343425, 0.007678875 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0057861875, "candidate": 0.006339146 }, "baseline_over_candidate_ratio_of_medians": 0.9127708211800137, "whole_process_pair_bootstrap_percentile95": [ 0.7624053108026554, 1.009014700087362 ], "min_max_seconds": { "baseline": [ 0.004079, 0.006530792 ], "candidate": [ 0.005226416, 0.007204333 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0007288539999999999, "candidate": 0.0007145 }, "baseline_over_candidate_ratio_of_medians": 1.0200895731280615, "whole_process_pair_bootstrap_percentile95": [ 0.9962426536839265, 1.095985779773599 ], "min_max_seconds": { "baseline": [ 0.000689167, 0.00079325 ], "candidate": [ 0.00065325, 0.000822958 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } } ], "scope": "One Mac, bounded Float32 reference decoders. Matched whole public requests include all emitted IDs, actual counts and EOS behavior. CPU and nonzero-temperature sampling are controls; numerical oracles and exact parameter proofs are outside timing. This fixed-cap experiment exposed early-EOS regressions and is preserved alongside the subsequent ramp-policy study.", "policy": { "status": "superseded fixed-cap experiment; all results retained", "initial_decode_chunk": "fixed requested cap, without a warmup ramp", "default_chunk_cap": 8, "public_one_token_policy": "resident" }, "numerical_reference_weights": "Exact placed Float32 parameter values widened to CPU Float64 before reference arithmetic", "ci_estimand_correction": { "raw_timings_unchanged": true, "reason": "Previous intervals bootstrapped medians of process medians while the point statistic used medians of all calls; corrected draws preserve whole process pairs and recompute the same all-call statistic.", "original_summary_sha256": "01c741682dad6cdcac37f9dfaded132364fd745ea16511bd5ee28883d4de1cee" } }, "natural_observations": { "schema": "aner.metal-greedy-local-measurements/1", "protocol": { "matched_process_pairs": 4, "timed_calls_per_process": 3, "excluded_warmups_per_process": 2, "process_order": "AB,BA,BA,AB; profile order rotates by outer round", "primary_phase": "natural", "natural_policy": "greedy complete requests; actual IDs, EOS and counts retained; baseline/candidate ID equality is measured, not assumed", "controls": "CPU Float32 tiny request and unchanged nonzero-temperature MPS sampling; all-zero and one-token budgets retained as functional/latency controls", "memory_scope": "tracked owned numeric capacities; readback counter includes metadata and Float32 widening; excludes stack vectors, allocator/driver/command-buffer/BLAS overhead and RSS", "thread_environment_request": { "VECLIB_MAXIMUM_THREADS": "1", "OPENBLAS_NUM_THREADS": "1", "OMP_NUM_THREADS": "1" } }, "producer_sha256": "f2b16c3927a79162e1d4165cc907c0b4f6b8453b23d47aaae696aad16dfb0838", "binary_sha256": { "baseline": "d626cfa793ded8eec98c20d664130c6db37143c47b71580af0ca83dd60e4ff3c", "candidate": "6e6b6c14db78bea4491b9f3aa476484ab64d86320949072f08d93b0fb2c48f9a" }, "records": [ { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "8b8f346279f9473a1d842c8e5d551de9fafb23372e41cd129e34c7bf8dd519c7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00545875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004801084, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003304375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "8ada8fbe0d680a645317c300e24a8fc41fcbb8267970af851595046a231e5ffd", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005100166, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003537125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004644458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "5ae75fe973a36f6718f8f907c7bb3829dfd074d0150986a251b0c8c790f74ef4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006291792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006492041, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005769375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "4b180686fb7e1e84139b1f65ec1012d409e0c0ebbff23aaa9f4bc710e336cfa7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005095875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004601458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004477916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "b0bf52559eb064e4b819a9864a4c46ffee371890a32d399f618d06b797e48b53", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.031440958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.025677084, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.022761458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "1891ac7bf8977c40df24315e40d2c21976f38748ee89dcfb086f6269be6cd859", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.01408775, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.013728792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.013769042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "e24b7bbd271ac4629a74c4c8b84be3bf703d4736b0046c7a526ac58f3187dbb0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.026751417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.025333125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0243495, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "7576fa05922fdf8e52a2fc40ff4ab17a94a78cf833b201465bc9d26d5d7603f5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018372125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015758625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015054291, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "7d5845305d9bc36b3fcf242a268abbb7385c8c20d821070df0384dfd79e0b7e1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.026738334, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.020982042, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.018265292, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "9e39df2661b7eeb1da026f996e8a3369ec908e1938c62023ee4a7b381af8154f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.010242, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00975775, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.01003075, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "1c8bc614f402fc329a2703a266f701cbeefd5c25c4097c9b951a9d589f21c559", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005885416, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.006111917, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005577, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "ff8f55c0b2e5cf6042b3b1ae4a46b48f31135826f7fe34f65ed3a0af70c279f9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005503334, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005320916, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004678083, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "607106abecc9e2c4127d845304bdf2470ee448e3007a537095526874e1c7988a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005797459, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004959458, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004942292, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "143007010dca2cd076fc782497a37f23c263100995c89595446165bcc3fa0147", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0053985, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003873167, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004327291, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "cc18f157ee8f90774e4488f5967c8b44db396fe6f9c193c6236dafe6aed14596", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0009285, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000973292, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000940334, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "132c11555743197777ae36682cc642a75788276f11a9c30f155e23d66270eb56", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001139083, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001101291, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001144041, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "93105dc053509684cf654ccda9997cc616502e36190e7ec43723b84c82b11f4c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001072, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001150292, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001081083, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "d500b3a5545bd704219e4f707037757dd79210846e89a5e6e2ab091d47e357d9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001157791, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000657333, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000981375, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "ef7d19b8aa31cea6eac302b16394d509898513de10bf9bbe168fde000b8583fc", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002004375, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001946958, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001918125, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "4d557f75ad93a7a2548c6f6d6df707d06f10237f0050ebff4f70480edb19276d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004724792, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004804042, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004734583, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "63152780b70c443127695729ff4c7ced9dd1d0e40b284f5e03d6a8c44525f960", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002011541, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002027167, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002011708, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "19cfd804e4253366108f0c739931c5a7f09dc47e9ca08926a7cee36289ac9502", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004674667, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004886167, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004864583, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "39b9cb8b7c8a49ca2e0b87008c2d05d758492ff2bc24992ed4640e9ba67bbc23", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00615125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006305542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00575825, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "72de91fba38b4f49918ed0d1c2020d3a2db58971ef7574925132ea6e7dae356b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003649667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003123917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002762167, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "65c3c31283dc1dc3927b5197c630b1a6e498c018903a52cd5371613a816063ed", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006389541, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006274875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007168875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "a17f9dd4ba869738aa59553264f8b4a49039438039cc82e408fbc949eb737121", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00561025, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005335375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0051855, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "1e0b6d9812f7377e0d14a58a94d818358d539744f3729dccb7601aecf54047d3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00782525, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0067505, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006874042, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "aef51a35ec76f8b962304db66e14648ff8c40bc0820ad381af1f2f0065d1a581", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005665417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004939917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004896, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "02146b11fe7a5b65bb445aa758c440db46c8a80c4639a82c008cee4a0577d851", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.040042541, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.030544292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.02669075, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "1f618c9c83fadf087e29122c94f85ef9a9f9357c3c5a60d0871de087dd071ef0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0175425, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.017410417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015399708, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "a68e917486cdc9dbdd965ca9b0e80c9755a08b4b9701d0df4fb6499bd6392453", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.026024083, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023834667, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023283667, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "374f9ae44b06cf7deac7c2cd818b30732999915ad23a54f6f2a4c50bc63b0227", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.020065583, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.019446667, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.01733275, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "3c7199f822ed47927b2a73e642ef49758808c7294cc49880c17bce413024c36f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.275e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.7041e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 3.6834e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "5cbdb287e94a15152e576576f9e40deec64782703bd44bda481e48d41e3d7186", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.3875e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.7125e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.4667e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "c997e39251cb1083e1ff5119db6f7109565e40f158ba5bce79be415d9fe58fcb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 5.7e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.825e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 1.8208e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "fe97a408054d408965f35ae2897efc3c8f2155b7992ee65c9286a3f6cc2bac28", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.6291e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.4875e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 1.6292e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "dbcc5128fb2635b71fc4d5abc64fb4391e4434c987f9e9cb4745c6e0c2db8cdd", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00091325, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000532667, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000829583, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "eec69a89cc36f43629e523cc46c7a7532347c1b8d0de8901a291ac345a3e122a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001018667, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001073792, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001026708, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "173a949345510ba3fbf84868b49474f3ca331a7425ec7a30fa6e0c6d517f2ad0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000987875, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000992833, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000993833, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "ddf735b4515a0eff7ed6d7de145fbc6ff315bcd12fc0fc27e2d7975e6793edeb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001036375, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001046084, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001026708, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "1e7e297e9a8930e4210cf8a4d6e2762be009eee383007b8e74bb10314be6746f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006495833, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005877708, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004411791, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "4a299fcaa938ab331e89a49d8e917517052b95f4f20cad97baf6a18d8443ec73", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00343425, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003880667, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003883625, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "3c2b899e5cc09399807459191caa726055658e8d37f5efdf03848a28e7db5d5e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0056705, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00544625, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005627, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "8b69d42a3e997b007618607cfa4441557dc7b992f596bad05663aa5bb4abeb04", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006911792, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006542041, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006174375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "5ae00f17e6fe487219136acb2e7d8ee4a6ec05673827f1f10efc09dfc9ee666c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000710209, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000726083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000731625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "1504d0b50a3bafafbd70ab19ad7d49e972289ce8021e32bfae0ca4893eaecfe5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000724709, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0007275, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000753875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "d6065979a185aa2e9672d8ca9ac6873b7738bbb373ea55daecc1cb4c22b4c394", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005473708, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004792417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005055875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "89b03b6cbaace1d8b803535662dedd22d74e3bf42617d127c6172b205a30fc9e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006549292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005736458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0054005, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "4c7c6a7dfef0d12565cd22a34697ebf3d60028db8e45eb3a94e0a4500bedfb07", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.017980833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.017610583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.017226084, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "d2a16083c827a1df7e0fd90e2ab06901380a1547a1661981f834641bb1f5a4cf", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.025701416, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.023736042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.022819375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "2ed574de1a98d08b437cb89d9d70c0d93f485bfac95e9da5b4b9f69a0175cc3a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018595333, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0168015, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015073917, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "36422a5635c9bb5939e7654b047293d1d84681d21b272abb76477d7893268ab3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.025966417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0242495, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023542334, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "9a5cca6c79c1519ca059956bf43ee86853752574856b0d2b14bd4d5e93219a62", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.012938541, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.012174459, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.010919583, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "a70e4de8da2c7434c2cbc99c82d81d71f14b68c88c75726e69c1df1e224edcd8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014758708, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.013759666, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.012417667, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "ed3bf34fb0f591128df43330b1917c5c096d30c655bfc525600a7093cfb86d9f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005430833, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005768125, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005086125, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "9af84325a2e8c35cf760d552fe8541405d9a4ee55e0bb71fdd3c1a0fc82a8430", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.012274792, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.011850417, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.0163265, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "2669525202eb6f2394d2313a17314c70d6edb7707efd59da677438a31e4922a1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004187208, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003794125, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003781583, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "91c8862cb44d757855f789262127918a4b57484b1f9d70293e0ccfba5070599c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006604791, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005169958, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005195792, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "10f3ac2e0fc2998c24d2b174d024fdbdddaa747afe244f44bbda31db980bacec", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001084459, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001075333, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00105725, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "8be00ea4c4bde2bd4d993552ac41f7320553114fb055b00d1792023169328096", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001175917, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000660875, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000915042, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "3f5ca404038848da315218685fdb89f62673d480f708ebc6c8010e1ae416bfdd", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001246, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000861208, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001100916, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "5bb1c698a563ef94f618db93feec7b4764f472ae74c0fc721de58619e7ed2339", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001008583, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001022333, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000991834, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "6af19253970c82f1fe8b12b1bd5af0a35fc2c108febf9b2a41bf52a2aee13350", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004471333, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003936333, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005036084, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "405572fcef4724cdaf72bd937141fb824005a2f04b83b80eddaef24edce19606", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001967125, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001641917, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001988542, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "464af2109e8b2513f11f6d5660d48d25fa16654dca6a7dc4d8d4c4e1c76353ea", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003372417, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003255542, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002971125, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "c67dd0e56431c4f2c1a8c977a45714a028efc9af24df5010089d9c96fadb6c66", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001900834, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001936416, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001252458, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "7c59c2642aacf3238571a17ac9656bd49e8c66b1e09955331b970f92a7aa4f79", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004952083, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005193458, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004577334, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "4f37e4e82cb3b1c2a4a793ac9c68c398e2e22591981354e7964263f71ceee2f9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006579625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006441208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006273459, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "32d5240bb39e8901c8813a8633ffccf695e6dd736d60865de53ea3c12c305112", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005583708, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005318167, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005436666, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "5797e5bc7369c6583db481265c8662907b3115f1647ec5012498a6a8be8aad3c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006591333, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006074917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00609425, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "5e9e2ce49748d1f96eb8b5106106582e1c778a5e74d078e15fb2620979f2f261", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005992125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004891375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005096292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "9766fc6013f1a772339b7d612f8dce1e212a53be3dde5f544a50aaa99c89a6f4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007936375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00674175, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006611875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "4c263b3e3de3ec1f03466c52782c3c258b10524dcbf88e9619ff608569d00d72", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.020778125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.018903375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.016955875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "2095d09a95b6db08434fc489308e75eeafe52a2d7a5fae71bd02fc72c184f5f6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.028506, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024306916, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023654625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "3402bedb9e54b0d4d7e728d3527f6fec047d896753ae52a655c81ab3119f5c9f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.020203584, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.017498458, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015726334, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "545874fb6b342331fab8b36285017c1aa7213ae009c1ee0945c5f5523d04b59f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.023446084, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023260416, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023334292, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "f41fc4adc3293c3631e343f6e8d8d397f4337793b147f2266fb2c05a0f78cb59", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 1.6542e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.6167e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 1.75e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "a3029fcbccdf5b60d762b2962a50e89667ef7156fda7f6f4fa62966a2264a63a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 1.6334e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.9e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 1.8458e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "d223eb58b310039a25d9ade1971b086eff9b0501c5c96b57c2c30bf38ad01854", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 1.6208e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.3125e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 4.1417e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "d5be9fcb88cddd7b8e25ee525b8ff1f6ba1f09ca43f70544b3f4f649bcfaec70", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 1.9916e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.8417e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 1.8625e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "cd58875bd8cdf3d0ef864325454331c0a93af71ce0227fc31b1c008cc2f89621", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001210292, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000995375, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001014416, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "7b3c6d29e1cf9336034618cf75d8aa9deb3b7cf63aa15938e9ae7e6249c5ee2d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000954458, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000947959, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.0009715, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "e468479ec8cacd575165ae6df7133288da90cb8714089928f5872b1751aa2103", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000920708, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000905292, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000884667, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "1cf822d4cb051e029c30042eaac5e96353aed14ac19273f5824f134b95f68768", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000976375, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000991959, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000967875, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "2c25bf4f39986c3c622c75d8f25e61f3e2e7ea4a97f9b2387942970d37bb5a6b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.007678875, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00540625, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005676042, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "f5a1068f6e89e6ba8b6224cd876e2e23973d3fee9223a588262a2cf8d53b3569", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.007508833, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005612292, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005249875, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "be0d541ebf176b73d1578c705e6b3e67f72259cba4a376c691dc995ef07e419a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006312541, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007109292, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006324375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "e4f73d84aa042a5ff1a2d8423081411d59b387106f12489f77273f218c64f09a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006530792, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006043334, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006431542, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "fde1d6b305a2df86dfd7fea1cee7888d5a63eeae07e8dc0b6951174e7cc3b3cd", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000822958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000719916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000712, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "ee1eda70af667ea9be64d2170b7b2981678cefeef9e4adca8a42007cb95e3b15", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000739208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000765958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000740625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "700a876d0bcea76f2f7ccdd3bf9ace40518acb70280d348bf84cb18267183b7a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002775542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00308275, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00248975, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "8e982f4dc1f535c5c8e280d7a2df17302332ca67ef6f1f93db262c53d0ce958e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003226541, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003938916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003246167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "e543151a416417b469f1a16b3d032a86cedca871794e468c2be3e552714d7b2c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0204765, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.01869825, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.016012417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "82c20b148ff80c7450789cb363898b61ba680160ccc32fc07c386844b07a0bd0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.033604666, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.037636416, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.042271292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "501a94ebb70a83dc20b5db362b7802c4635e5552fe613218e2d09e1750c10a7f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.020731791, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.020680417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015767167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "39f0fc0e91e5f3a64270afc644ebf482eaac9d3826fc3c74665b300ae94838ca", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0287385, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024839167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024816083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "31c5b2d2a6c0cdf03d2cbcb7457d760da57ba676da53355a98aef60972294bb8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008887417, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008078583, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007757291, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "8f6d8a1155b013152c25bd830fff80a7feca13dd0bd8adf79be4bb8ac0e45272", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014376209, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.013462959, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.01242275, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "3327828f18354059f2dddd82b89b982c04c49c39149f85bd7886081c36d680ed", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006495042, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004047583, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003492916, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "d41d2aaf82166c3af1de827104029b6f3aac0520b8b5ff688f475507f19b421e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0067215, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005296916, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00476275, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "6be8ee78406792d69ecae14d6472de1eea17ce6d2b00ed4d5101049bc5495d5e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005862792, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005320625, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004600541, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "f9d48ecd97e0b3b0bc436129c8216a27dbe59d8034d0197280fd3ea0d40f2088", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006873292, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00548375, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005407, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "26ae30db7e30f9ca78e1b147764215fe09537f4dd5de02b8d6bd234e6de77f67", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001098, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001619625, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001137417, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "5502d451fe751156e84fd827a0e75a33c1dc151e79886a70a5bc78d14b464708", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001043292, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001051416, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001037708, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "cb4d0a916d88c4eb7ac34a12c5cb0ae60c200478531eb65af152a3e3ca56ae0d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001034625, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0009525, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00093825, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "23d651cb22b2641f1a7e4c0c9c9427eaba811c6e74f65aa90b2103cc24445e0f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001040083, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000958416, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000978625, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "81ac2e79d8cb25b2cc01e8bade8a4d0df38e40bcd2b456e3769c3e5dd9576375", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006074709, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004607916, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003485459, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "e026cd3670b01b0bfbe7d10d6b7f5e21c0d8391e0aa2aea5bb032f115040c819", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001933708, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001623541, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001661792, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "9db0441b4517e3558eda0ddfef3d6582b8abe49d0a991ae10b9329476ecc676a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003065125, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00386825, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002524209, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "aa2859e2a71186b6979b46c7df46f8164154dc6a33feb2e4d3ab885c7288e5a0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0018385, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002003708, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001767958, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "32018105cd2f0e31dcd44e3ce877c158f07a988c7c245c740d557016e2fab2f9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004283667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003832542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003843792, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "58028e4c19ae626cc0db9266404aa9b7ca3f9324fb7db247d7c1e9ec8c5b0e4e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006290208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007438708, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006894125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "dc7334e11eaf3037363edccc1b57e4dd2075874ac47afa98c2bef758c3169432", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006876209, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005451, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006442209, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "1a9b7a4b13009f57b04c6398da1a16fab736f0573fe8707cf18a20967ea099bd", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006027, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005225667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004763125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "576e264114872f9d67bde514dcc123bc2b3692a9a5b34a71c90ac1c59dbe4f2a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005546125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005659291, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0055925, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "32a4b3da014760bae92fafe008eac5595df60755c3416ab3b98f51d7085e4488", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007706542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00671775, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007054875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "51c9427219feb1377d1e4e98e47c9f968369d03225ef93409568ebb5ddd42940", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018440709, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.01778075, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.017435917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "d98de5f47d5e3c8ade5bb8dd7371a1b4d11926c8de2d8cbb390306a575b121d5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.027872875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024612083, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023342125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "39b9f46cb4e090a82b6ca94a73a0a98f0ca77139ad36da03fb4559d3cd3a6225", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0158015, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015353625, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015549709, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "80b576abafd4a3dbb68e21e6cd0f3476b2c0a41018cfebc533503320bdcc0f47", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.02872675, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023261584, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024753, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "dc1601da010fc3fe3841a9010d6c1ec1e03718c6844273447a2f24f7f26db0c7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 1.9792e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.45e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.25e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "cd1e684723f694b160cfa5d252678ddad332d4ef3ffe58cffea005f22d4101f3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 1.8541e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.4875e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 1.825e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "023df1b38405e5d7d41d058db4e3cf7e6ad284b15d4c8c6d0027abd8650ac232", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 1.8542e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.9541e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 1.6334e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "1a424477683695151dbe1dd63c7db65ef7de475ae43d9032f7472ce6e9135a32", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.1958e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.6125e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 2.8791e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "2ca50f34cea2b5ede3ec7ce765596e5845252fcab19858dc5591947a21e7123b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00104425, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000986291, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001002791, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "1e738120fe57d3f30d3afcb00daeec594bda3b22ffacb01d4a6e39c1b86cb2d2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000815917, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0009465, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000920083, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "52103facce444630ea8aecd6ca6cfe381d571a6035cab9302a933c362ca34911", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001098375, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001057416, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001080833, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "6a9a073bc596116442ea1798b3ad0ec5de8a758475522541387a1d3b223e736a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001023, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00098275, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000994458, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "36f5683199d28cd48c51071f21220198ad322ac0d0dbdacc87a0b86960528b6f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00690125, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005417417, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005878917, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "80e19687a8a9d669090c6ebaeace872b195b66d33ebb99a3d363e553e834d496", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004382375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004312667, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00457875, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "dee94164e417626cc892a74a8b9b5246c2e6d98d7a55921728f5dcba07656cee", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007204333, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00621375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005226416, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "e7953ee696cab45cd7e78ff6bb08032759b5a43f8900664673a27d581d7de198", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005261959, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004212833, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004079, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "2e47efeda44f878b9ef44f99e7cae35196a3a382ec797f8bfb0d449d3279754c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000676334, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00067525, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000680125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "17bac796e6f2eea952b01ff6b6d45fdff36162d8e62aa34fbe335cdcc648dede", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000689167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00068975, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00070425, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "a9fbcbc4796636d13701e43148409d3ebdb4f82efad700ba6d3030e634d6aa05", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002772208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003133208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002669209, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "723ab93087f3e5410db821f5df9cb1fce66a9c551dd9a27e8b71ad4e91d6fea7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006138834, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.006930292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005558834, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "e158f8e5d18004a1bae30e8faf5918d4893e06edfc02af964a4af09473e0c36b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003810292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002948459, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002873542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "8352e98d126cc057a9a011c494b170759300e31d7cb0f922b58662f7585c8014", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005545833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0054895, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005757417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "63921e0c6eb744d420b3474598d672828d7732adbd2dc8222b0f7b1487899804", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.026071916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.026923375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023831167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "fa70714ea64ab471497df96bfc4d9b90c59bbe5ef26071c5c18d5aef91b4ef83", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.017482958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015324541, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.01468425, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "923b82750fdfec7912441a451ea3725aef482776cd983655aa4ff598bc294c6b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.01258, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.012237042, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.012212834, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "7b1a587801f60231575b3774e1252dea6e313d715a6f1578405ddfe235d8343b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.010101625, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.009415, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.008251459, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "1c4c4690f86cd1f4b022dfe47e4f701e22365e924f3ba4a3bc6c8f1a4dbe41f6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.008144625, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004784292, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004808791, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "838323945f054dd7bfe9e18c9589ee5898b531037f9ed1793c7c215233d07776", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00479375, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005275458, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004381708, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "cd6e073a1652d0cce48497e4f36f3112266b401e943c2248e2fa8cbe0df633c2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004965875, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005526167, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00523375, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "d2834532bc41460ff5e0bc29eca2b9b7cbf1ade339b76ab54d8a71dcc3b6ba22", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004581291, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004658084, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004626125, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "b229335c3b63363e4636bafc4a5825b916f5c5cdeab5bfdd0477bc009d221523", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001029833, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001017459, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001007167, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "a16fa46bbf0d6d5ba31e888876ec016b2a4a3b68c5867ce9cddb26be62dd2cc7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001123542, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001063083, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001087458, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "f8446f22f852dbb7bd52771cf54da93d26f39e42960484bc95b379bddf14cd4d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001043875, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001023958, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001009333, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "320c48f6efe34bbacb9eeb770089fe3a51708aa44cc81866e90e9fd190c197e2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001112916, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001074125, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001082875, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "d8391bdd27dcb7ceb1e4dd6e3186cbe0e7754e777ccff1e544a52cc5b6febf31", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002387125, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001738292, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000939375, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "cf53623ae2c965dae2d6b0dc1342359556b49f722182ea3d523d83a518e77e2e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004444, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004425416, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004435167, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "a3eb5c30f4da653b4db5d413cf4f978d4b0f32a4b72851db9becb3260a998077", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002036041, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0020025, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002003375, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "1a2f7cd1dc2364944d425bf397d1f4b258b98a70ef3713e5245bbb3babcce1a0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004881042, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004621667, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004598917, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "4a0fbeeafadd6a60d0d9788adf972c54fe2536dbb32736173fd6dd00a917e6d5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006007083, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00594975, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005442875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "fa0c4ec9cfcfa24febd630833058860c9e594fab0feac4fa39417d97eea50046", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005509625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005337792, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005347625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "3c1c9222bc612b84e83546cc3073598ae47b854d358c000076598b2880591582", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006948041, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006477583, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006721958, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "9857e0a7eeabeb40e565fd47e0e78f99e040d2c9ff636b842a690da130577f1f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005566709, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00730425, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00511275, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "9a0a6fc0efbae91c570b6a906febdd5e79a3b4277d599cea4f98e6c06cbea947", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005604833, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006342416, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007214292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "13f6fab647e26ba4e82de024f07e1114427e5b640141f663dc23799e9ec7e2c5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006026709, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00529675, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005274417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "d901e7e900ab3964f303436885aeb38a0a0012c791e0f02e6db2035e047161e4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.031132417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024639625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0253805, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "d91b28278d624f13d412b3418574f25ff2dd5816302ae85ebd3a6a38c66af8d5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014258167, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015386334, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.01445825, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "e67f455ccb4bb824a0672e9cd2f566748cf0058bb006dc3039a85c8c9dc7022a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.025326875, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023565208, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023141166, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "f362d13d32513751f126c9c829e11335bdec42f2e7bfbf3b33a8f506650620e3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.015646084, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.01578975, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015365917, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "c84243bbc8ae187a1e1a8c7e3c624bb167762648bd105207b04da09eead22648", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 1.4208e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.8542e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.15e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "d867d6bf0b045e13de2b50fd03e380ea2af903aaf30f1c591657ea9b801ecc51", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.7709e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.6625e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.2667e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "0325f3d07ca77cfe11af9f879e8b3ea7aab94c48f47f9c65f9e54d1b25850b01", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 1.9875e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.3458e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 2.3375e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "487675207c5bb2df0e9c049edf9a1a00f4920c754fdbd245bffd7974c04add41", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.9291e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.1375e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 3.4583e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "2df031226c6e4fbe6ccedf64793a4db7293f98da4153b664be12aa419a8f3ab8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000962333, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000924792, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00090975, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "2b46c795586dd5930f9d8adaa73c64556ba744e6de29596da7ebb2e699a2d665", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001053917, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001003375, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000982166, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "c80c3d1cf6ab93aa2e91aef127cd428c6bdc53446eb3b8ee0c3160eff6aa5538", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000986958, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001004083, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000996458, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "c4eb60e5d61c5b71b58d3bdb9fbe4acf84e76207de3eb640a0b35095ce4d12b6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000871084, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001051875, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001043167, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "9144b95d90e16a112f399a4de6f7b9a2583bb833c353f80b2e9459e5d9ccf8e6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006310208, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005092208, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004978041, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "9c14544e161931a4b8684dbab00c80470144395a4081c8474abbc30b14668416", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005898417, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00653325, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005259583, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "acec48f4479f109364471f5b6ec0ee386579e8660fbb8177c7771c9bceb63b5f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00644675, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006361041, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005901875, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "5a056d852fc9591a517f2a5ac108956cbe316d3cfbab8a1bae79d29f438d5e68", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006353917, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006364417, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006159583, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "84ea05971fe5d480baaf23c8419e2283cb7186d563ff5d9a352ae81c0b681d4f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000708041, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000756333, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00079325, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "1853c4c50110cb7696ff95fc99ad3dd59171830c2f9c9dc83c77088add7c7129", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000660875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00065325, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000717, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "50dcb1c167f47414fd70b292c9c8b0c9bd916f532569d0fe57f6f6302b004c00", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005987792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.006414833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004992291, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "5b29bebc4a556dd8f710fedac3653d8fe69b35a5d881df6fde79657adfe05706", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004971083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004480292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00467975, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "0b68326a43a2071efdc67c97de14198d62888308386aa173262a5405360de165", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006847541, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0055395, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0056175, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "c3dfcc158b2a80eddebab1fa64cd0c6d66e0d3c880ef24712664ddc1d866772e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004613625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005100542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004572584, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "e5cc05f273a9a7704758ef002a569e849f23c5c886554f741e04e5338fa7f8e3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.031905125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.030633083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.02769175, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "471c5df86ad8efd0178634076e6053083922b227059f4de4c0f588a16d74ce89", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.018265917, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.018485042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.017620875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] } ] }, "diagnostics": { "schema": "aner.metal-greedy-diagnostics/1", "timing_quotable": false, "records": [ { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006671542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.007827959, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005837125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.006197208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004568875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003942209, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005795167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.006381667, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007298583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006858625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006248542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005913834, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006188791, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005094375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0051575, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005488917, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.02936975, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.02655875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.018048125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.017337709, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.019239958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.018306417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.017051084, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.01416325, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.027423459, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.026025417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.022468042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.021739791, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018461375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.014837125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.020437375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.01790225, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.022672666, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 32, "status_reads": 32, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 32, "greedy_active_reads": 32, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.018349166, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 32, "status_reads": 32, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 32, "greedy_active_reads": 32, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 32, "status_reads": 32, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 32, "greedy_active_reads": 32, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.013441375, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 12, "status_reads": 12, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 12, "greedy_active_reads": 12, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.009917709, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 12, "status_reads": 12, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 12, "greedy_active_reads": 12, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 12, "status_reads": 12, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 12, "greedy_active_reads": 12, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.010414792, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 5, "status_reads": 5, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.010108167, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 5, "status_reads": 5, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 5, "status_reads": 5, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008474875, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 2, "status_reads": 2, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007144083, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 2, "status_reads": 2, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 2, "status_reads": 2, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.009214083, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005939542, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.010119084, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.009335375, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00540175, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00632875, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00898725, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.008024042, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.011652291, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007086667, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008227666, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006042541, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006001417, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004729333, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.010755709, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005382958, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 2, "status_reads": 2, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001270375, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002186416, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000607667, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001899167, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001226083, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001288708, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001155208, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001116625, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001266875, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000971625, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001161542, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001143792, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00118225, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001155583, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001212167, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001312292, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002118834, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002120792, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003186, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 2, "status_reads": 2, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002720292, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 2, "status_reads": 2, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 2, "status_reads": 2, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004017083, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 9, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005493375, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 9, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 9, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.009542584, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 33, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.009829, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 33, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 33, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002164917, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0021575, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002932583, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 2, "status_reads": 2, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002709958, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 2, "status_reads": 2, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 2, "status_reads": 2, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003224459, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003315625, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008753291, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008762625, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008696, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006548833, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004991667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 4, "status_reads": 4, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004846791, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 4, "status_reads": 4, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 4, "status_reads": 4, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005738125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006013292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008773958, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008456792, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008666834, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 9, "status_reads": 9, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007304583, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 9, "status_reads": 9, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 9, "status_reads": 9, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007056125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 4, "status_reads": 4, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006992625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 4, "status_reads": 4, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 4, "status_reads": 4, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005622625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005575709, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 2, "status_reads": 2, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008135875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008172916, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.009657542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 10, "status_reads": 10, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 10, "greedy_active_reads": 10, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008565125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 10, "status_reads": 10, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 10, "greedy_active_reads": 10, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 10, "status_reads": 10, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 10, "greedy_active_reads": 10, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007267958, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 4, "status_reads": 4, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007169875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 4, "status_reads": 4, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 4, "status_reads": 4, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0094535, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 17, "command_buffers": 3, "status_reads": 3, "gemm_calls": 153, "compute_encoders": 322, "packed_output_rows": 136, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.009764583, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 17, "command_buffers": 3, "status_reads": 3, "gemm_calls": 153, "compute_encoders": 322, "packed_output_rows": 136, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 17, "command_buffers": 3, "status_reads": 3, "gemm_calls": 153, "compute_encoders": 322, "packed_output_rows": 136, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.011115291, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.011356459, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 33, "command_buffers": 2, "status_reads": 2, "gemm_calls": 297, "compute_encoders": 626, "packed_output_rows": 264, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 1, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.031397041, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 1, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.027504833, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 1, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 3, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018936458, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 3, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.017506375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 3, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.016196917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015008041, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 32, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.017412958, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 32, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.016232916, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 32, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.032116125, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.026183166, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.017745375, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.018484125, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 22, "status_reads": 22, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 22, "greedy_active_reads": 22, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.02152675, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.016813833, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 9, "status_reads": 9, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018870541, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.018071667, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 3, "status_reads": 3, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 3.4792e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.9333e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 3.5333e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.7292e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 3.4708e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 3.7042e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 3.4167e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 3.3083e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 3.8334e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.6708e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 3.5666e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 3.4542e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 4.0333e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 3.4e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 3.4834e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 3.1834e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001044917, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000742583, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000957583, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0008975, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001062459, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001022584, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001688416, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001051625, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001376333, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000895875, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001142709, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001207292, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001401958, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000899125, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001239542, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001043167, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } } ] }, "full_diagnostic_json_sha256": "433111948678a0362b4836310ab6c771ecfff1deffa168322f82d6231e8b80cd", "numerical_vector_policy": "Full vectors retained in the research capture; portable Float64 little-endian digests replace repeated vectors here. All output IDs, timings and counters are retained.", "original_producer_cpp": "#include \"aner/gen.hpp\"\n#include \"aner/gpu.hpp\"\n#include \"aner/nn.hpp\"\n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n\nnamespace gen = aner::gen;\nnamespace nn = aner::nn;\nnamespace token = aner::token;\nusing IDs = std::vector>;\nusing Clock = std::chrono::steady_clock;\n\nnamespace {\nvoid require(bool value, const std::string &message) {\n if (!value) throw std::runtime_error(message);\n}\ntemplate void array(const std::vector &values) {\n std::cout << '[';\n for (std::size_t i = 0; i < values.size(); ++i) std::cout << (i ? \",\" : \"\") << values[i];\n std::cout << ']';\n}\nvoid nested(const IDs &rows) {\n std::cout << '[';\n for (std::size_t i = 0; i < rows.size(); ++i) { if (i) std::cout << ','; array(rows[i]); }\n std::cout << ']';\n}\nstruct Options {\n std::string device = \"mps\", dtype = \"Float32\", phase = \"natural\", name = \"tiny\";\n std::int64_t batch = 8, width = 32, layers = 2;\n int repetitions = 3, warmups = 2;\n std::size_t chunk_steps = 8;\n double temperature = 0, top_p = 1;\n std::int64_t top_k = 0;\n bool native_plan = true;\n};\nstruct Work {\n IDs prompts;\n std::vector lengths, budgets, capacities, seeds;\n std::int64_t eos_position = -1;\n};\nWork work(const Options &options) {\n Work result;\n if (options.name == \"ragged\") {\n require(options.batch == 8, \"ragged workload requires eight logical rows\");\n result.lengths = {1, 3, 7, 31, 2, 5, 13, 47};\n result.budgets = {32, 1, 16, 2, 24, 3, 8, 4};\n } else if (options.name == \"mixed_eos\") {\n require(options.batch == 8, \"mixed EOS workload requires eight rows\");\n result.lengths = {1, 57, 58, 59, 60, 61, 63, 65};\n result.budgets = {64, 31, 31, 31, 31, 31, 31, 31};\n result.eos_position = 64;\n } else if (options.name == \"zero_mixed\") {\n require(options.batch == 8, \"mixed zero budgets require eight rows\");\n result.lengths = {0, 1, 8, 3, 8, 0, 31, 5};\n result.budgets = {0, 1, 0, 8, 64, 1, 0, 16};\n } else if (options.name == \"all_zero\" || options.name == \"one\") {\n result.lengths.assign(static_cast(options.batch), 0);\n result.budgets.assign(static_cast(options.batch), options.name == \"one\" ? 1 : 0);\n } else if (options.name.starts_with(\"eos\")) {\n const auto count = std::stoll(options.name.substr(3));\n require(count == 1 || count == 2 || count == 8 || count == 9 || count == 10, \"unsupported EOS boundary\");\n result.lengths.assign(static_cast(options.batch), 8);\n result.budgets.assign(static_cast(options.batch), 64);\n result.eos_position = 8 + count - 2;\n } else {\n require(options.name == \"tiny\" || options.name == \"long\", \"unknown workload\");\n result.lengths.assign(static_cast(options.batch), 8);\n result.budgets.assign(static_cast(options.batch), options.name == \"tiny\" ? 8 : 64);\n }\n for (std::size_t row = 0; row < result.lengths.size(); ++row) {\n std::vector prompt;\n if (result.lengths[row] > 0) prompt.push_back(token::bos_id);\n for (std::int64_t step = 1; step < result.lengths[row]; ++step)\n prompt.push_back(97 + (37 * step + 11 * static_cast(row) + 5) % 26);\n result.prompts.push_back(std::move(prompt));\n result.capacities.push_back(std::max(std::int64_t(1), std::max(std::int64_t(1), result.lengths[row]) + result.budgets[row] - 1));\n result.seeds.push_back(7301 + 101 * static_cast(row));\n }\n return result;\n}\nIDs normalized_prompts(const Work &workload) {\n auto result = workload.prompts;\n for (auto &row : result) if (row.empty()) row = {token::bos_id};\n return result;\n}\ngen::Model fixture_model(const Options &options, const Work &workload) {\n auto result = gen::decoder(token::byte_tokenizer(), {96, options.width, 4, options.layers, 20261015, 0});\n if (workload.eos_position < 0) return result;\n // With zero transformer projections, the residual is token + position.\n // Ordinary token 65 has direction u; EOS has -u. Positions switch from\n // +4u to -4u at a declared absolute position, yielding a robust EOS margin.\n for (const auto &name : gen::parameter_names(result)) {\n const auto parameter = gen::parameter(result, name);\n std::vector values(static_cast(parameter.rows() * parameter.cols()), 0);\n if (name.ends_with(\".gain\")) std::fill(values.begin(), values.end(), 1);\n if (name == \"embedding.token\") {\n values[65 * static_cast(options.width)] = 1;\n values[65 * static_cast(options.width) + 1] = -1;\n values[static_cast(token::eos_id * options.width)] = -1;\n values[static_cast(token::eos_id * options.width) + 1] = 1;\n }\n if (name == \"embedding.position\")\n for (std::int64_t position = 0; position < 96; ++position) {\n const auto sign = position < workload.eos_position ? 4. : -4.;\n values[static_cast(position * options.width)] = sign;\n values[static_cast(position * options.width + 1)] = -sign;\n }\n result = gen::with_parameter(result, name, nn::tensor(parameter.rows(), parameter.cols(), std::move(values)));\n }\n return result;\n}\ngen::Model cpu_reference_model(const Options &options, const Work &workload) {\n // Match the actual placed dtype first, then widen those exact parameter\n // values for an independent Float64 mathematical reference.\n return gen::to(gen::to(fixture_model(options, workload), \"cpu\", options.dtype), \"cpu\", \"Float64\");\n}\nvoid verify_analytic(const IDs &output, const Work &workload) {\n if (workload.eos_position < 0) return;\n for (std::size_t row = 0; row < output.size(); ++row) {\n const auto eos_at = std::max(std::int64_t(1), workload.eos_position - std::max(std::int64_t(1), workload.lengths[row]) + 2);\n const auto count = std::min(workload.budgets[row], eos_at);\n std::vector expected(static_cast(count), 65);\n if (count > 0 && count == eos_at) expected.back() = token::eos_id;\n require(output[row] == expected, \"position-driven EOS fixture differs from its analytic sequence\");\n }\n}\nIDs forced_inputs(const Work &workload, std::int64_t step) {\n IDs result(workload.prompts.size());\n for (std::size_t row = 0; row < result.size(); ++row)\n if (step < workload.budgets[row] - 1)\n result[row] = {97 + (7 * step + 13 * static_cast(row) + 3) % 26};\n return result;\n}\ntemplate\nCache single_cache(const gen::Model &model, std::int64_t capacity, bool native) {\n if constexpr (std::is_constructible_v)\n return Cache(model, true, capacity, native);\n else {\n require(native, \"baseline does not expose the primitive-plan testing selector\");\n return Cache(model, true, capacity);\n }\n}\ntemplate\nCache batch_cache(const gen::Model &model, const Work &workload, bool native) {\n if constexpr (std::is_constructible_v &, bool, bool, bool>)\n return Cache(model, workload.capacities, true, false, native);\n else {\n require(native, \"baseline does not expose the primitive-plan testing selector\");\n return Cache(model, workload.capacities, true, false);\n }\n}\nstruct Statistics {\n bool available = false;\n std::uint64_t forwards = 0, buffers = 0, reads = 0, gemms = 0, encoders = 0, packed_rows = 0;\n bool greedy_available = false;\n std::uint64_t chunks = 0, active_reads = 0, result_reads = 0, id_bytes = 0, emitted = 0;\n};\ntemplate Statistics statistics(const Cache &cache) {\n if constexpr (requires { cache.statistics(); }) {\n const auto source = cache.statistics();\n Statistics result{source.available, source.forward_calls, source.command_buffers, source.status_reads,\n source.gemm_calls, source.compute_encoders, source.packed_output_rows};\n if constexpr (requires { source.greedy_chunks; }) {\n result.greedy_available = true; result.chunks = source.greedy_chunks;\n result.active_reads = source.greedy_active_reads; result.result_reads = source.greedy_result_readbacks;\n result.id_bytes = source.greedy_id_bytes; result.emitted = source.greedy_emitted_tokens;\n }\n return result;\n } else return {};\n}\nvoid statistics_json(const Statistics &stats) {\n std::cout << \"{\\\"available\\\":\" << (stats.available ? \"true\" : \"false\")\n << \",\\\"forward_calls\\\":\" << stats.forwards << \",\\\"command_buffers\\\":\" << stats.buffers\n << \",\\\"status_reads\\\":\" << stats.reads << \",\\\"gemm_calls\\\":\" << stats.gemms\n << \",\\\"compute_encoders\\\":\" << stats.encoders << \",\\\"packed_output_rows\\\":\" << stats.packed_rows\n << \",\\\"greedy_available\\\":\" << (stats.greedy_available ? \"true\" : \"false\")\n << \",\\\"greedy_chunks\\\":\" << stats.chunks << \",\\\"greedy_active_reads\\\":\" << stats.active_reads\n << \",\\\"greedy_result_readbacks\\\":\" << stats.result_reads << \",\\\"greedy_id_bytes\\\":\" << stats.id_bytes\n << \",\\\"greedy_emitted_tokens\\\":\" << stats.emitted << '}';\n}\nvoid prefix(const Options &options, const Work &workload) {\n std::cout << \"{\\\"schema\\\":\\\"aner.metal-greedy-benchmark/1\\\",\\\"phase\\\":\\\"\" << options.phase\n << \"\\\",\\\"device\\\":\\\"\" << options.device << \"\\\",\\\"dtype\\\":\\\"\" << options.dtype\n << \"\\\",\\\"case\\\":\\\"\" << options.name << \"\\\",\\\"batch_size\\\":\" << options.batch\n << \",\\\"width\\\":\" << options.width << \",\\\"layers\\\":\" << options.layers\n << \",\\\"heads\\\":4,\\\"context\\\":96,\\\"vocabulary\\\":259,\\\"model_seed\\\":20261015,\\\"dropout\\\":0\"\n << \",\\\"temperature\\\":\" << options.temperature << \",\\\"top_k\\\":\" << options.top_k << \",\\\"top_p\\\":\" << options.top_p\n << \",\\\"eos_position\\\":\" << workload.eos_position << \",\\\"chunk_steps\\\":\" << options.chunk_steps\n << \",\\\"prompt_lengths\\\":\";\n array(workload.lengths); std::cout << \",\\\"budgets\\\":\"; array(workload.budgets);\n std::cout << \",\\\"row_seeds\\\":\"; array(workload.seeds);\n}\nvoid memory_json(const nn::MemoryStats &before, const nn::MemoryStats &after) {\n require(after.all_buffers.current_bytes == before.all_buffers.current_bytes,\n \"generation retained tracked numeric buffers after release\");\n std::cout << \",\\\"memory\\\":{\\\"retained_increment_bytes\\\":0\"\n << \",\\\"all_peak_increment_bytes\\\":\" << after.all_buffers.peak_bytes - before.all_buffers.current_bytes\n << \",\\\"device_peak_increment_bytes\\\":\" << after.device_buffers.peak_bytes - before.device_buffers.current_bytes\n << \",\\\"host_peak_increment_bytes\\\":\" << after.host_buffers.peak_bytes - before.host_buffers.current_bytes\n << \",\\\"readback_peak_increment_bytes\\\":\" << after.host_readback_cache.peak_bytes - before.host_readback_cache.current_bytes\n << '}';\n}\nIDs public_request(const gen::Model &model, const Work &workload, const Options &options) {\n if (options.batch == 1)\n return {gen::generate(model, workload.prompts[0], workload.budgets[0], options.temperature, options.top_k, options.top_p, workload.seeds[0], true)};\n return gen::generate_batch(model, workload.prompts, workload.budgets, options.temperature, options.top_k, options.top_p, workload.seeds, true, false);\n}\nvoid reference_request(const gen::Model &model, const Work &workload, const Options &options) {\n const auto output = public_request(model, workload, options);\n if (options.temperature == 0) verify_analytic(output, workload);\n prefix(options, workload); std::cout << \",\\\"output_ids\\\":\"; nested(output); std::cout << \"}\\n\";\n}\nvoid natural(const gen::Model &model, const Work &workload, const Options &options) {\n IDs expected;\n for (int iteration = -options.warmups; iteration < options.repetitions; ++iteration) {\n nn::synchronize(options.device);\n const auto before = nn::memory_stats(); nn::reset_memory_peaks();\n const auto started = Clock::now();\n const auto output = public_request(model, workload, options);\n nn::synchronize(options.device);\n const auto seconds = std::chrono::duration(Clock::now() - started).count();\n const auto after = nn::memory_stats();\n require(output.size() == workload.prompts.size(), \"natural generation changed row inventory\");\n for (std::size_t row = 0; row < output.size(); ++row)\n require((workload.budgets[row] == 0 ? output[row].empty() : !output[row].empty()) && output[row].size() <= static_cast(workload.budgets[row]),\n \"natural generation violated an output budget\");\n if (options.temperature == 0) verify_analytic(output, workload);\n if (expected.empty()) expected = output;\n else require(output == expected, \"one runtime produced inconsistent repeated greedy IDs\");\n if (iteration < 0) continue;\n prefix(options, workload);\n std::cout << \",\\\"iteration\\\":\" << iteration << \",\\\"seconds\\\":\" << seconds << \",\\\"output_ids\\\":\";\n nested(output);\n std::vector counts, ended;\n for (const auto &row : output) { counts.push_back(static_cast(row.size())); ended.push_back(!row.empty() && row.back() == token::eos_id); }\n std::cout << \",\\\"actual_new_tokens_per_row\\\":\"; array(counts);\n std::cout << \",\\\"ended_eos_per_row\\\":\"; array(ended);\n std::cout << \",\\\"actual_total_new_tokens\\\":\" << std::accumulate(counts.begin(), counts.end(), std::int64_t(0));\n memory_json(before, after);\n std::cout << \",\\\"statistics\\\":{\\\"available\\\":false}}\\n\";\n }\n}\ntemplate bool resident_available(const Cache &cache) {\n if constexpr (requires { cache.supports_resident_greedy(); }) return cache.supports_resident_greedy();\n else return false;\n}\ntemplate IDs resident_call(Cache &cache, const Work &workload, std::size_t chunk) {\n if constexpr (requires { cache.generate_greedy(workload.prompts, workload.budgets, chunk); })\n return cache.generate_greedy(workload.prompts, workload.budgets, chunk);\n else throw std::runtime_error(\"resident greedy testing seam unavailable in this runtime\");\n}\nstd::vector expected_positions(const IDs &output, const Work &workload) {\n std::vector result;\n for (std::size_t row = 0; row < output.size(); ++row)\n result.push_back(output[row].empty() ? 0 : std::max(std::int64_t(1), workload.lengths[row]) + static_cast(output[row].size()) - 1);\n return result;\n}\nvoid resident(const gen::Model &model, const Work &workload, const Options &options) {\n for (int iteration = -options.warmups; iteration < options.repetitions; ++iteration) {\n nn::synchronize(options.device); const auto before = nn::memory_stats(); nn::reset_memory_peaks();\n const auto started = Clock::now(); IDs output; Statistics stats; std::vector positions;\n {\n auto cache = batch_cache(model, workload, true);\n require(resident_available(cache), \"explicit resident request unsupported; no fallback\");\n output = resident_call(cache, workload, options.chunk_steps);\n stats = statistics(cache); positions = cache.positions();\n require(positions == expected_positions(output, workload), \"resident committed EOS/final token or masked tail to KV\");\n }\n nn::synchronize(options.device);\n const auto seconds = std::chrono::duration(Clock::now() - started).count();\n const auto after = nn::memory_stats();\n verify_analytic(output, workload);\n if (iteration < 0) continue;\n prefix(options, workload); std::cout << \",\\\"iteration\\\":\" << iteration << \",\\\"seconds\\\":\" << seconds << \",\\\"output_ids\\\":\";\n nested(output); std::cout << \",\\\"positions\\\":\"; array(positions);\n memory_json(before, after); std::cout << \",\\\"statistics\\\":\"; statistics_json(stats); std::cout << \"}\\n\";\n }\n}\nvoid kv_oracle(const gen::Model &model, Work workload, const Options &options) {\n for (std::size_t row = 0; row < workload.capacities.size(); ++row)\n workload.capacities[row] = std::max(std::int64_t(1), workload.lengths[row]) + workload.budgets[row];\n auto cache = batch_cache(model, workload, true);\n require(resident_available(cache), \"resident KV oracle unsupported; no fallback\");\n const auto output = resident_call(cache, workload, options.chunk_steps);\n verify_analytic(output, workload);\n const auto positions = cache.positions();\n require(positions == expected_positions(output, workload), \"resident cache position differs before probe\");\n const auto stats = statistics(cache);\n const auto reference_model = cpu_reference_model(options, workload);\n std::vector expected;\n const auto normalized = normalized_prompts(workload);\n for (std::size_t row = 0; row < output.size(); ++row) {\n std::vector history;\n if (!output[row].empty()) {\n history = normalized[row];\n history.insert(history.end(), output[row].begin(), output[row].end() - 1);\n }\n history.push_back(97);\n gen::detail::GenerationCache reference(reference_model, true, workload.capacities[row]);\n const auto logits = reference.append(history);\n const auto &values = logits.values(); expected.insert(expected.end(), values.begin(), values.end());\n }\n const auto actual = cache.append(IDs(output.size(), {97})).values();\n require(actual.size() == expected.size(), \"resident KV probe shape differs\");\n double maximum = 0;\n for (std::size_t i = 0; i < actual.size(); ++i) {\n require(std::isfinite(actual[i]) && std::isfinite(expected[i]), \"nonfinite resident KV probe\");\n maximum = std::max(maximum, std::abs(actual[i] - expected[i]) / (1 + std::abs(expected[i])));\n }\n require(maximum <= 5e-4, \"resident KV state differs from independently reconstructed CPU Float64 history\");\n prefix(options, workload); std::cout << \",\\\"output_ids\\\":\"; nested(output);\n std::cout << \",\\\"positions_before_probe\\\":\"; array(positions);\n std::cout << \",\\\"max_scaled_logit_error\\\":\" << maximum << \",\\\"scaled_tolerance\\\":0.0005,\\\"actual_logits\\\":\"; array(actual);\n std::cout << \",\\\"reference_logits\\\":\"; array(expected); std::cout << \",\\\"statistics\\\":\"; statistics_json(stats); std::cout << \"}\\n\";\n}\nvoid forced(const gen::Model &model, const Work &workload, const Options &options) {\n const auto steps = *std::max_element(workload.budgets.begin(), workload.budgets.end()) - 1;\n std::vector inputs;\n for (std::int64_t step = 0; step < steps; ++step) inputs.push_back(forced_inputs(workload, step));\n for (int iteration = -options.warmups; iteration < options.repetitions; ++iteration) {\n nn::synchronize(options.device);\n const auto before = nn::memory_stats(); nn::reset_memory_peaks();\n const auto started = Clock::now();\n Statistics stats;\n std::uint64_t bytes = 0, elements = 0, allocations = 0;\n if (options.batch == 1) {\n auto cache = single_cache(model, workload.capacities[0], options.native_plan);\n auto output = cache.append(workload.prompts[0]);\n for (const auto &input : inputs) { output = nn::Tensor(); output = cache.append(input[0]); }\n require(cache.position() == workload.capacities[0], \"forced single final position differs\");\n stats = statistics(cache); bytes = cache.key_value_bytes(); elements = cache.key_value_elements(); allocations = cache.key_value_allocations();\n } else {\n auto cache = batch_cache(model, workload, options.native_plan);\n auto output = cache.append(workload.prompts);\n if (!inputs.empty()) cache.prepare_decode(static_cast(std::count_if(inputs[0].begin(), inputs[0].end(), [](const auto &x) { return !x.empty(); })));\n for (const auto &input : inputs) { output = nn::Tensor(); output = cache.append(input); }\n require(cache.positions() == workload.capacities, \"forced batch final positions differ\");\n stats = statistics(cache); bytes = cache.key_value_bytes(); elements = cache.key_value_elements(); allocations = cache.key_value_allocations();\n }\n nn::synchronize(options.device);\n const auto seconds = std::chrono::duration(Clock::now() - started).count();\n const auto after = nn::memory_stats();\n if (iteration < 0) continue;\n prefix(options, workload);\n std::cout << \",\\\"iteration\\\":\" << iteration << \",\\\"seconds\\\":\" << seconds\n << \",\\\"timed_logit_readback\\\":false,\\\"forced_next_logit_rows\\\":\"\n << std::accumulate(workload.budgets.begin(), workload.budgets.end(), std::int64_t(0))\n << \",\\\"kv_bytes\\\":\" << bytes << \",\\\"kv_elements\\\":\" << elements\n << \",\\\"kv_allocations\\\":\" << allocations;\n memory_json(before, after);\n std::cout << \",\\\"statistics\\\":\"; statistics_json(stats); std::cout << \"}\\n\";\n }\n}\nvoid oracle(const gen::Model &model, const Work &workload, const Options &options) {\n const auto reference_model = cpu_reference_model(options, workload);\n std::vector> references;\n for (auto capacity : workload.capacities)\n references.push_back(std::make_unique(reference_model, true, capacity));\n std::unique_ptr single;\n std::unique_ptr packed;\n if (options.batch == 1)\n single = std::make_unique(single_cache(model, workload.capacities[0], options.native_plan));\n else\n packed = std::make_unique(batch_cache(model, workload, options.native_plan));\n const auto steps = *std::max_element(workload.budgets.begin(), workload.budgets.end()) - 1;\n const double tolerance = options.dtype == \"Float64\" ? 1e-10 : 5e-4;\n for (std::int64_t step = -1; step < steps; ++step) {\n auto input = step < 0 ? normalized_prompts(workload) : forced_inputs(workload, step);\n if (step < 0) for (std::size_t row = 0; row < input.size(); ++row) if (workload.budgets[row] == 0) input[row].clear();\n const auto output = options.batch == 1 ? single->append(input[0]) : packed->append(input);\n const auto actual = output.values();\n std::vector expected; std::vector active;\n double maximum = 0;\n for (std::size_t row = 0; row < input.size(); ++row) if (!input[row].empty()) {\n const auto next = references[row]->append(input[row]);\n const auto &values = next.values();\n expected.insert(expected.end(), values.begin(), values.end()); active.push_back(static_cast(row));\n }\n require(actual.size() == expected.size(), \"oracle shape differs\");\n for (std::size_t i = 0; i < actual.size(); ++i) {\n require(std::isfinite(actual[i]) && std::isfinite(expected[i]), \"oracle observed nonfinite logit\");\n maximum = std::max(maximum, std::abs(actual[i] - expected[i]) / (1 + std::abs(expected[i])));\n }\n require(maximum <= tolerance, \"forced logits exceed the existing CPU Float64 reference tolerance\");\n prefix(options, workload); std::cout << \",\\\"step\\\":\" << step << \",\\\"active_rows\\\":\"; array(active);\n std::cout << \",\\\"max_scaled_logit_error\\\":\" << maximum << \",\\\"scaled_tolerance\\\":\" << tolerance << \",\\\"actual_logits\\\":\"; array(actual);\n std::cout << \",\\\"reference_logits\\\":\"; array(expected); std::cout << \"}\\n\";\n }\n}\nvoid weights(const gen::Model &model) {\n auto names = gen::parameter_names(model); std::sort(names.begin(), names.end());\n auto write64 = [](std::uint64_t value) { std::cout.write(reinterpret_cast(&value), sizeof(value)); };\n std::cout << \"aner.parameters.widened-f64/1\\n\" << gen::dtype(model) << '\\n'; write64(names.size());\n for (const auto &name : names) {\n const auto parameter = gen::parameter(model, name); const auto &values = parameter.values();\n write64(name.size()); std::cout.write(name.data(), static_cast(name.size()));\n write64(parameter.rows()); write64(parameter.cols()); write64(values.size());\n std::cout.write(reinterpret_cast(values.data()), static_cast(values.size() * sizeof(double)));\n }\n}\n} // namespace\n\nint main(int argc, char **argv) {\n try {\n Options options;\n for (int i = 1; i < argc; i += 2) {\n require(i + 1 < argc, \"option needs value\");\n const std::string flag = argv[i], value = argv[i + 1];\n if (flag == \"--device\") options.device = value;\n else if (flag == \"--dtype\") options.dtype = value;\n else if (flag == \"--phase\") options.phase = value;\n else if (flag == \"--case\") options.name = value;\n else if (flag == \"--batch-size\") options.batch = std::stoll(value);\n else if (flag == \"--width\") options.width = std::stoll(value);\n else if (flag == \"--layers\") options.layers = std::stoll(value);\n else if (flag == \"--repetitions\") options.repetitions = std::stoi(value);\n else if (flag == \"--warmups\") options.warmups = std::stoi(value);\n else if (flag == \"--chunk-steps\") options.chunk_steps = static_cast(std::stoull(value));\n else if (flag == \"--temperature\") options.temperature = std::stod(value);\n else if (flag == \"--top-k\") options.top_k = std::stoll(value);\n else if (flag == \"--top-p\") options.top_p = std::stod(value);\n else if (flag == \"--native-plan\") { require(value == \"true\" || value == \"false\", \"native-plan requires true/false\"); options.native_plan = value == \"true\"; }\n else throw std::runtime_error(\"unknown option\");\n }\n require(options.batch == 1 || options.batch == 8, \"batch must be one or eight\");\n require((options.width == 32 && options.layers == 2) || (options.width == 128 && options.layers == 3), \"unsupported bounded model\");\n require(options.repetitions >= 2 && options.repetitions <= 100 && options.warmups >= 1 && options.warmups <= 4, \"invalid repeat protocol\");\n require(gen::device_supported(options.device, options.dtype), \"explicit device/precision unavailable; no fallback\");\n const auto workload = work(options);\n const auto model = gen::to(fixture_model(options, workload), options.device, options.dtype);\n std::cout << std::setprecision(17);\n if (options.phase == \"natural\") natural(model, workload, options);\n else if (options.phase == \"reference\") reference_request(model, workload, options);\n else if (options.phase == \"forced\") forced(model, workload, options);\n else if (options.phase == \"oracle\") oracle(model, workload, options);\n else if (options.phase == \"resident\") resident(model, workload, options);\n else if (options.phase == \"kv_oracle\") kv_oracle(model, workload, options);\n else if (options.phase == \"weights\") weights(model);\n else throw std::runtime_error(\"unknown phase\");\n return 0;\n } catch (const std::exception &error) {\n std::cerr << error.what() << '\\n'; return 1;\n }\n}\n" }, { "capture": "02", "label": "gradual startup with dynamic mask", "is_final": false, "summary": { "schema": "aner.metal-greedy-local-summary/1", "protocol": { "matched_process_pairs": 4, "timed_calls_per_process": 3, "excluded_warmups_per_process": 2, "process_order": "AB,BA,BA,AB; profile order rotates by outer round", "primary_phase": "natural", "natural_policy": "greedy complete requests; actual IDs, EOS and counts retained; baseline/candidate ID equality is measured, not assumed", "controls": "CPU Float32 tiny request and unchanged nonzero-temperature MPS sampling; all-zero and one-token budgets retained as functional/latency controls", "memory_scope": "tracked owned numeric capacities; readback counter includes metadata and Float32 widening; excludes stack vectors, allocator/driver/command-buffer/BLAS overhead and RSS", "thread_environment_request": { "VECLIB_MAXIMUM_THREADS": "1", "OPENBLAS_NUM_THREADS": "1", "OMP_NUM_THREADS": "1" }, "confidence_interval": { "point_estimator": "median(all 12 baseline calls) / median(all 12 candidate calls)", "resampling_unit": "whole matched pair of fresh process clusters", "inner_calls": "all three calls from each selected process retained together; no inner-call resampling", "method": "exhaustive enumeration of all ordered four-cluster draws with replacement", "draws": 256, "coverage": 0.95, "percentile_method": "linear interpolation at probability * (256 - 1) in the ordered bootstrap ratios" } }, "producer_sha256": "834ce647498123f3dd7f95cd56af3c4b7e0441f8cc14900821007be4b8e0ad96", "binary_sha256": { "baseline": "d0522c5e8c5c52f3ebc0f6920ec48aa2d405bf4d8a82eb3fd1fc1dab8162e610", "candidate": "7c1a43681d09e2a9f9705b333b22b77e11e84fee9ce77bfacdfb5e29da835d54" }, "build": { "configuration": "Release", "lto": false, "devices": [ "mps", "cpu" ], "dtype": "Float32", "compiler": { "id": "AppleClang", "version": "21.0.0.21000334" }, "compiler_flags": [ "-O3", "-DNDEBUG", "-std=c++20", "-mcpu=apple-m1", "-fno-fast-math", "-ffp-contract=off" ] }, "measurement_sha256": "b52e760f9bb2822370283fe038399d64a31d58195f8f1a623bed5cda3dcb54b0", "oracle": { "schema": "aner.metal-greedy-local-oracle/1", "profiles": [ { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 1.3992326293094704e-07, "candidate_max_scaled_error_vs_cpu_float64": 1.3992326293094704e-07, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 32, "baseline_max_scaled_error_vs_cpu_float64": 9.61704129314485e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.61704129314485e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 3.0323179185145576e-07, "candidate_max_scaled_error_vs_cpu_float64": 3.0323179185145576e-07, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 3.230860285372708e-07, "candidate_max_scaled_error_vs_cpu_float64": 3.230860285372708e-07, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 9.487577588236166e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.487577588236166e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 0, "baseline_max_scaled_error_vs_cpu_float64": 0, "candidate_max_scaled_error_vs_cpu_float64": 0, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 0, "baseline_max_scaled_error_vs_cpu_float64": 0, "candidate_max_scaled_error_vs_cpu_float64": 0, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 1, "baseline_max_scaled_error_vs_cpu_float64": 6.854355312918667e-08, "candidate_max_scaled_error_vs_cpu_float64": 6.854355312918667e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 1, "baseline_max_scaled_error_vs_cpu_float64": 6.68301477395295e-08, "candidate_max_scaled_error_vs_cpu_float64": 6.68301477395295e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 2.907853657679229e-08, "candidate_max_scaled_error_vs_cpu_float64": 2.907853657679229e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 } ], "exact_parameter_values_equal": true, "parameter_values": [ { "device": "cpu", "width": 32, "layers": 2, "model_fixture": "random", "parameter_serialization_sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos1", "parameter_serialization_sha256": "e3e6ff6db789c848e78a759a6fc69b3e2c9329dff3ea951843eb0b348bf82f2a" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos10", "parameter_serialization_sha256": "ad522891246ca60c51e52ff565c3016a616a50b164ddd6127c9353a5e8ed1b56" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos2", "parameter_serialization_sha256": "ed9153be24577314bb4e31f0283b576ff51df1335d415bb9a6d1e361b1ad1d08" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos3", "parameter_serialization_sha256": "0fe8452022f38407bdad8c24d202a81f68d5b0d431f919295c67ad8d0449a505" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos4", "parameter_serialization_sha256": "7c24715177c00583dcbc5cd02c511fbe8439752dd4ddb1561eb79bdf060f04d6" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos5", "parameter_serialization_sha256": "3b57c548c9d63e1a6514ae59badabbb062de7b99c8321f84af893c5d563ba079" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos6", "parameter_serialization_sha256": "5c516784a0fe783125b7b5cb1e52e702d95c2adab654d2cbf376a0f79c02043d" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos8", "parameter_serialization_sha256": "6b16fedfaa1cd79eccddbe930376a7c04833dd460280fad86699368b2331c5a9" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos9", "parameter_serialization_sha256": "bf13ad8b95d255bb2ae96a8f20a80f80154beebf9165d17e9d16642217f96818" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "mixed_eos", "parameter_serialization_sha256": "ae87e82e1f0040a28a821ff69662ed1773bc854740534880fa8f1670d11d1973" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "random", "parameter_serialization_sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" }, { "device": "mps", "width": 128, "layers": 3, "model_fixture": "random", "parameter_serialization_sha256": "7432a93bd25dd41f127bc7621a6b14e5e15fba11c703df9d7c8ebfc28316c9d8" } ] }, "summary": [ { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.004763375, "candidate": 0.0048635205 }, "baseline_over_candidate_ratio_of_medians": 0.9794088459172733, "whole_process_pair_bootstrap_percentile95": [ 0.7309878752100268, 1.733447119328781 ], "min_max_seconds": { "baseline": [ 0.003574875, 0.006578666 ], "candidate": [ 0.002824666, 0.005773875 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0059503335, "candidate": 0.005903083 }, "baseline_over_candidate_ratio_of_medians": 1.0080043766960418, "whole_process_pair_bootstrap_percentile95": [ 0.9012929647164261, 1.0853575415225178 ], "min_max_seconds": { "baseline": [ 0.005293542, 0.006671458 ], "candidate": [ 0.005010208, 0.00667525 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.02371825, "candidate": 0.017612125 }, "baseline_over_candidate_ratio_of_medians": 1.3467000716835702, "whole_process_pair_bootstrap_percentile95": [ 1.2325345297348298, 1.6652528442593901 ], "min_max_seconds": { "baseline": [ 0.022562625, 0.040938125 ], "candidate": [ 0.014332583, 0.023700292 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 64 ], "candidate": [ 64 ] } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.024840229499999998, "candidate": 0.014913 }, "baseline_over_candidate_ratio_of_medians": 1.665676222088111, "whole_process_pair_bootstrap_percentile95": [ 1.626397836275976, 1.6928333765310815 ], "min_max_seconds": { "baseline": [ 0.023676625, 0.027904 ], "candidate": [ 0.014669, 0.016787666 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "candidate": [ 64, 64, 64, 64, 64, 64, 64, 64 ] } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0135544165, "candidate": 0.011151728999999999 }, "baseline_over_candidate_ratio_of_medians": 1.215454258258966, "whole_process_pair_bootstrap_percentile95": [ 1.1541437421524328, 1.472802001181782 ], "min_max_seconds": { "baseline": [ 0.011611458, 0.019992667 ], "candidate": [ 0.007760125, 0.012897083 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "candidate": [ 32, 1, 16, 2, 24, 3, 8, 4 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.005988875, "candidate": 0.004765729 }, "baseline_over_candidate_ratio_of_medians": 1.2566545433028191, "whole_process_pair_bootstrap_percentile95": [ 1.0385572548494943, 1.4737196566682822 ], "min_max_seconds": { "baseline": [ 0.00493475, 0.00889025 ], "candidate": [ 0.004055583, 0.006995125 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0062056455, "candidate": 0.0047730210000000006 }, "baseline_over_candidate_ratio_of_medians": 1.3001504707396006, "whole_process_pair_bootstrap_percentile95": [ 1.06928586178456, 1.5095332209498342 ], "min_max_seconds": { "baseline": [ 0.0050845, 0.007453667 ], "candidate": [ 0.00415975, 0.008266583 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.001006521, "candidate": 0.000914021 }, "baseline_over_candidate_ratio_of_medians": 1.1012011759029607, "whole_process_pair_bootstrap_percentile95": [ 0.6415272545233164, 1.1810695715323167 ], "min_max_seconds": { "baseline": [ 0.000541125, 0.001077334 ], "candidate": [ 0.000582666, 0.001814958 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1 ], "candidate": [ 1 ] } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0009886255, "candidate": 0.0010925625 }, "baseline_over_candidate_ratio_of_medians": 0.904868600194497, "whole_process_pair_bootstrap_percentile95": [ 0.7299926203792299, 0.9384061869347221 ], "min_max_seconds": { "baseline": [ 0.0007965, 0.001417333 ], "candidate": [ 0.00085325, 0.001165541 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "candidate": [ 1, 1, 1, 1, 1, 1, 1, 1 ] } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0017254165000000002, "candidate": 0.0020558955 }, "baseline_over_candidate_ratio_of_medians": 0.8392530165078916, "whole_process_pair_bootstrap_percentile95": [ 0.7387240753313031, 0.9713705784077661 ], "min_max_seconds": { "baseline": [ 0.000923875, 0.002113083 ], "candidate": [ 0.00157425, 0.002182417 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 2 ], "candidate": [ 2 ] } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0019849165, "candidate": 0.0021045205 }, "baseline_over_candidate_ratio_of_medians": 0.9431680518198801, "whole_process_pair_bootstrap_percentile95": [ 0.8890508039824188, 1.148252556600444 ], "min_max_seconds": { "baseline": [ 0.001441708, 0.002089875 ], "candidate": [ 0.001540958, 0.002696541 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "candidate": [ 2, 2, 2, 2, 2, 2, 2, 2 ] } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0026697705, "candidate": 0.0030994789999999996 }, "baseline_over_candidate_ratio_of_medians": 0.8613610545514263, "whole_process_pair_bootstrap_percentile95": [ 0.7914944996234695, 0.9155123117213796 ], "min_max_seconds": { "baseline": [ 0.001961667, 0.003094625 ], "candidate": [ 0.002776042, 0.003415792 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 3 ], "candidate": [ 3 ] } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0027489995, "candidate": 0.0032778125 }, "baseline_over_candidate_ratio_of_medians": 0.8386689293545619, "whole_process_pair_bootstrap_percentile95": [ 0.7795698997906831, 1.0245879163554288 ], "min_max_seconds": { "baseline": [ 0.002172708, 0.003215875 ], "candidate": [ 0.002207625, 0.003640667 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "candidate": [ 3, 3, 3, 3, 3, 3, 3, 3 ] } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0035154995, "candidate": 0.0032185415 }, "baseline_over_candidate_ratio_of_medians": 1.0922647727239185, "whole_process_pair_bootstrap_percentile95": [ 0.9841721530040626, 1.1366546906187625 ], "min_max_seconds": { "baseline": [ 0.002604334, 0.003873125 ], "candidate": [ 0.002907791, 0.003497458 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 4 ], "candidate": [ 4 ] } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.003582604, "candidate": 0.003114917 }, "baseline_over_candidate_ratio_of_medians": 1.1501442895589193, "whole_process_pair_bootstrap_percentile95": [ 0.9748704172227055, 1.3051239456095876 ], "min_max_seconds": { "baseline": [ 0.0027655, 0.004007417 ], "candidate": [ 0.002007584, 0.003823584 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "candidate": [ 4, 4, 4, 4, 4, 4, 4, 4 ] } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.003449229, "candidate": 0.0051999585 }, "baseline_over_candidate_ratio_of_medians": 0.6633185630231472, "whole_process_pair_bootstrap_percentile95": [ 0.5767914301623753, 0.7771981742167198 ], "min_max_seconds": { "baseline": [ 0.002088167, 0.004607834 ], "candidate": [ 0.003053167, 0.005948542 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 5 ], "candidate": [ 5 ] } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0042621455, "candidate": 0.005171042000000001 }, "baseline_over_candidate_ratio_of_medians": 0.8242333943526274, "whole_process_pair_bootstrap_percentile95": [ 0.7316289441987827, 1.3076121433239813 ], "min_max_seconds": { "baseline": [ 0.003225084, 0.005426083 ], "candidate": [ 0.003269833, 0.006238333 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "candidate": [ 5, 5, 5, 5, 5, 5, 5, 5 ] } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0044446875, "candidate": 0.005157353999999999 }, "baseline_over_candidate_ratio_of_medians": 0.8618154774715873, "whole_process_pair_bootstrap_percentile95": [ 0.8372654405380184, 1.1214520767125846 ], "min_max_seconds": { "baseline": [ 0.003495916, 0.0054595 ], "candidate": [ 0.002652084, 0.00566375 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 6 ], "candidate": [ 6 ] } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.00480575, "candidate": 0.0058478539999999996 }, "baseline_over_candidate_ratio_of_medians": 0.8217971926111699, "whole_process_pair_bootstrap_percentile95": [ 0.6844699185871524, 1.0024027965705522 ], "min_max_seconds": { "baseline": [ 0.003021084, 0.005755875 ], "candidate": [ 0.004941125, 0.006142125 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "candidate": [ 6, 6, 6, 6, 6, 6, 6, 6 ] } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0058008125, "candidate": 0.005470042 }, "baseline_over_candidate_ratio_of_medians": 1.0604694625745106, "whole_process_pair_bootstrap_percentile95": [ 0.9574444895257211, 1.2923766919148585 ], "min_max_seconds": { "baseline": [ 0.005384, 0.006986417 ], "candidate": [ 0.003002291, 0.006637584 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0064755, "candidate": 0.0085199795 }, "baseline_over_candidate_ratio_of_medians": 0.7600370399952253, "whole_process_pair_bootstrap_percentile95": [ 0.5641808877711585, 0.9971192024401435 ], "min_max_seconds": { "baseline": [ 0.003928333, 0.007607208 ], "candidate": [ 0.004784375, 0.010346958 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "candidate": [ 9, 9, 9, 9, 9, 9, 9, 9 ] } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0072424585, "candidate": 0.0080662705 }, "baseline_over_candidate_ratio_of_medians": 0.8978695296667772, "whole_process_pair_bootstrap_percentile95": [ 0.7530038508686853, 1.3830722696974564 ], "min_max_seconds": { "baseline": [ 0.005973333, 0.007835875 ], "candidate": [ 0.004807917, 0.010687667 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "candidate": [ 10, 10, 10, 10, 10, 10, 10, 10 ] } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0243319375, "candidate": 0.0169313545 }, "baseline_over_candidate_ratio_of_medians": 1.4370933819854756, "whole_process_pair_bootstrap_percentile95": [ 1.3742902754454454, 1.6969738824947853 ], "min_max_seconds": { "baseline": [ 0.023179792, 0.02741275 ], "candidate": [ 0.014772125, 0.020097417 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "candidate": [ 64, 9, 8, 7, 6, 5, 3, 1 ] } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.024384708499999998, "candidate": 0.016653542 }, "baseline_over_candidate_ratio_of_medians": 1.4642355662236897, "whole_process_pair_bootstrap_percentile95": [ 1.3373703469800156, 1.5117138731968174 ], "min_max_seconds": { "baseline": [ 0.022731333, 0.032793958 ], "candidate": [ 0.015701917, 0.023265667 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "candidate": [ 0, 1, 0, 8, 64, 1, 0, 16 ] } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 1.9312500000000002e-05, "candidate": 2.4146000000000002e-05 }, "baseline_over_candidate_ratio_of_medians": 0.7998219166735692, "whole_process_pair_bootstrap_percentile95": [ 0.6808465485974456, 1.0224358974358974 ], "min_max_seconds": { "baseline": [ 1.5917e-05, 2.2125e-05 ], "candidate": [ 1.0417e-05, 2.9708e-05 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 0 ], "candidate": [ 0 ] } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 2.1853999999999998e-05, "candidate": 2.0311999999999998e-05 }, "baseline_over_candidate_ratio_of_medians": 1.0759157148483656, "whole_process_pair_bootstrap_percentile95": [ 0.6910164436562954, 1.8275134507859478 ], "min_max_seconds": { "baseline": [ 1.5166e-05, 4.5417e-05 ], "candidate": [ 1.7333e-05, 3.125e-05 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "candidate": [ 0, 0, 0, 0, 0, 0, 0, 0 ] } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0009407295, "candidate": 0.0009595624999999999 }, "baseline_over_candidate_ratio_of_medians": 0.9803733472285547, "whole_process_pair_bootstrap_percentile95": [ 0.9338419290770023, 1.0631615842972308 ], "min_max_seconds": { "baseline": [ 0.000659792, 0.00105325 ], "candidate": [ 0.00078, 0.001747459 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1 ], "candidate": [ 1 ] } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0009810415, "candidate": 0.0009969789999999998 }, "baseline_over_candidate_ratio_of_medians": 0.9840142069191027, "whole_process_pair_bootstrap_percentile95": [ 0.8246513624244001, 1.053759771086688 ], "min_max_seconds": { "baseline": [ 0.000712083, 0.001021167 ], "candidate": [ 0.000493959, 0.001062375 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "candidate": [ 1, 1, 1, 1, 1, 1, 1, 1 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0054511460000000005, "candidate": 0.005598583 }, "baseline_over_candidate_ratio_of_medians": 0.973665300666258, "whole_process_pair_bootstrap_percentile95": [ 0.6755659803515387, 1.0341121137259195 ], "min_max_seconds": { "baseline": [ 0.003488375, 0.006433375 ], "candidate": [ 0.005224083, 0.006718125 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0058805625, "candidate": 0.006588583 }, "baseline_over_candidate_ratio_of_medians": 0.8925382741630484, "whole_process_pair_bootstrap_percentile95": [ 0.7360079845217193, 0.9622441046659334 ], "min_max_seconds": { "baseline": [ 0.004452584, 0.007308791 ], "candidate": [ 0.00609225, 0.007461833 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0007320415000000001, "candidate": 0.0007257920000000001 }, "baseline_over_candidate_ratio_of_medians": 1.0086105936687095, "whole_process_pair_bootstrap_percentile95": [ 0.9874009971809367, 1.0861886259272342 ], "min_max_seconds": { "baseline": [ 0.000719542, 0.000831625 ], "candidate": [ 0.000701125, 0.000839 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } } ], "scope": "One Mac, bounded Float32 reference decoders. Matched whole public requests retain every emitted ID, actual count and EOS result, including latency regressions. CPU and nonzero-temperature sampling are controls; placed-Float32-to-CPU-Float64 numerical oracles and exact parameter proofs are outside timing. Four matched process pairs support each interval.", "ci_estimand_correction": { "raw_timings_unchanged": true, "reason": "Previous intervals bootstrapped medians of process medians while the point statistic used medians of all calls; corrected draws preserve whole process pairs and recompute the same all-call statistic.", "original_summary_sha256": "be9c239b18fe5a8377d59e94827361628e8d5c269126742b9d035b2d6e6194d2" } }, "natural_observations": { "schema": "aner.metal-greedy-local-measurements/1", "protocol": { "matched_process_pairs": 4, "timed_calls_per_process": 3, "excluded_warmups_per_process": 2, "process_order": "AB,BA,BA,AB; profile order rotates by outer round", "primary_phase": "natural", "natural_policy": "greedy complete requests; actual IDs, EOS and counts retained; baseline/candidate ID equality is measured, not assumed", "controls": "CPU Float32 tiny request and unchanged nonzero-temperature MPS sampling; all-zero and one-token budgets retained as functional/latency controls", "memory_scope": "tracked owned numeric capacities; readback counter includes metadata and Float32 widening; excludes stack vectors, allocator/driver/command-buffer/BLAS overhead and RSS", "thread_environment_request": { "VECLIB_MAXIMUM_THREADS": "1", "OPENBLAS_NUM_THREADS": "1", "OMP_NUM_THREADS": "1" } }, "producer_sha256": "834ce647498123f3dd7f95cd56af3c4b7e0441f8cc14900821007be4b8e0ad96", "binary_sha256": { "baseline": "d0522c5e8c5c52f3ebc0f6920ec48aa2d405bf4d8a82eb3fd1fc1dab8162e610", "candidate": "7c1a43681d09e2a9f9705b333b22b77e11e84fee9ce77bfacdfb5e29da835d54" }, "records": [ { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "e47830a0ed22798869e3c4091b8c711fd222069f9fac68a1e71b34ff6bebb844", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006139666, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0059765, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004209166, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "c53f7790e29b53b0f59b556643239f5976bf7462a35d6897fc38c426b00bb3f3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003544833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002824666, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00287775, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "f3b4bbbd39e314de5bb141fd81c56f87213e231cd40bbee521df4d28141686ad", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006208916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006671458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005811125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "11482fe6b066111afa1fd1a7481020b84d600b004b943b9fc802e82c36a4761f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005719625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005676375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005044709, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "594bd948ef09b13cc7aeccb8053804355b2930a1581bea6279b08116432977c4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.040938125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.031341625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.02636, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "9c5f55a7f221bbdd2e86893a7dec9cf5397e9adc688ab65341aa9faf0972a6fa", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.018045125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.018358625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.0182905, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "a7edc41ed51dd7d091a64e9dbf226f34353d56a16630c064bd2aa97eaa2eedff", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.027904, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024429125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024854875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "c1b8c806936a371ccdcf36362cd50e9947501f4a374c6c6c6e6f2acb202787b1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.016787666, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015324125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.014816542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "05f1bd83cfba0df0dab1c3dd51f3f1cdccc787a504dd303107a2c1f6da1b2bf4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.019992667, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.019460917, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.017286166, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "53b74a3d2159485bfdc3a91872444dfa8e8b10794b498b6539d5ec7b17910827", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.01273225, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.012397291, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.012553166, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "69c3abcdf3da2aa859840d62df6d59b99dd783fd3082a387bb88cbe5c1a6fac2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00889025, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00669575, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.006006291, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "7eac01c3715f248139a56bbdc471985ddb8b39006e299616370d33e09555c1e6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004496292, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004485459, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00408125, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "aef6a89434be9d91e7ba5f4e304a8fe57b9885c99b227927207f8b050c84b1c3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006742167, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00571725, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005239208, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "f716ba72389359ccd5efc8587f128825db82314edffe126f3dc9bafa73b0426d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004959667, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004836667, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00415975, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "3c1a6e7d1745999b29fa18cddb473f985d2751926425aea49793e8b3a1b33fdd", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000788667, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001016458, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001077334, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "8472aa529338d4e1702f134c9dca055ed4f347d957468e2a2971c6b40f596a61", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001135292, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000860625, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000842792, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "2668bab1301e889ac960026648aec2c1d8bef5ca1c242eaa141d5c5b75578ba4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000820542, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0007965, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001151042, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "541d45455fe838bff4fd80c6ef28ab1064e61ca34cdbc586bcfef8a213653b35", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001165541, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001133792, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001109042, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "c1974dd9e77d6395f56ccbc9f4a402a37a0166a6e32f3122d00a49377f251913", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001729875, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00125625, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000923875, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "b2cb62bebc2eb434c38ae6a56703ea835aa85b268f7a689d801ac60f26d66f1f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001866292, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002027291, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00157425, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "4b7359e8d05c40618033afa2c1fe56bd3bc52529a9efb8fe6f397e8eacfedd8e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001997791, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001985583, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00196175, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "b7dc829fd5b3b1b8cc90cc362a33aa77cdc22162d791ae4d955ba04e77c21165", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002142875, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002092833, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002154417, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "f508c9eeafbb8ef518ae0e13e1bc9273f583b59accc6eb619848679c7785d349", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002716541, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002751584, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003094625, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "8147d114bef2c7dfdfb622784413fd7550be03daa1d02d67e4a9640fffe80050", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003406208, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003095625, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003206333, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "4bc65bfa0a96da1b3505753d4e71c1e5678ddcfa1b429cad5d3ee6dd18a7fbf4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002343708, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002433833, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002918083, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "4fe717a919383efc30192df5b0a2545bdc798b168977b6e91476982451f365e3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003292167, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002242916, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002207625, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "99cd09b50e680217be9463f0c6d2e4188ffd37d47c7bf2083319c8a769c38292", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003873125, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003341833, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002604334, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "9c8a4c1732f28aff57b176cb8a7c9481053182437cc757a7a313720ce862e661", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003497458, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002967958, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003134083, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "8718c8494f8b350a789beecc1c1c1480cfcb46369e9027590a556d88ae2d269f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003539333, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003056167, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003328708, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "3552cf82b2e30b7620cb890467fce68722a19e438df3f4c9a759e810c8460f95", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002711875, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002696792, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002007584, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "1a4eb9092f6d0e0dde34a395b50038c151fc0667a9330aca8a38d2ae0b5d385c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003731166, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004037041, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00316325, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "9336fcbd33911dbe786b278112e62418e0f9cb98533f489ee9917e24557401f5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005527125, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005176459, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004940375, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "5a60a9396ddcac7bcdc6e08cd78a47afe506c413d20b89dc35e486f29f3db010", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0050645, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004066417, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004220166, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "929babf0a724a68fa91fef4945ac88c62f99a20aa0c20773b80542ff435d158d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006238333, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00560975, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004924958, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "7664bf93ae000221cab56fc21ddb8d75d9eec3beabc982f641825aae596788d7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005068916, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004119375, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005281459, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "57fb39ec1bffd6920c34cc389c5d1cb7bed5544b5de8c72f249e470ba7eb03cb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005586833, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004399333, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005282, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "e68f399281a440961181aa9cfb280f64ce55de25780255d28a8ed24192f14f18", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004495, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003021084, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003631083, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "ce9df0e9246eca940fadd6a0ceb68ae1aa9b562b0129e293c7a582dde5799cb5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005936041, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006142125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004941125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "e7cc1e4a7530a4392f6d48083b54d8a9a893f2beabffff0a16970feb401e1bff", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006727791, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0058065, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005447042, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "f1df8b411aba5c6065eb5ed339b029e943372da7618dcf1eb3e0174886984759", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006119209, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00541625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00506525, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "1b9e69a3e6c01f69b6fedbaf2b8fce7d98601476daeaf928f5a6f4662e5500f4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006365791, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006498875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007229709, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "840b4ff391fc23d7c16e9cb37b5c5ea9e9a5d0416f99a09fab367b0d715a4362", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.009251833, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007058709, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004784375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "5245c5e7b6dcb71e30a10747b4033502a70ee395dcb9dd37510693271d038b30", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007485166, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007334125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006833958, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "47241e325b44f03b6417a2fac260879eb4a2d6581050286996653639236c4f31", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00897725, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.009485959, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.008198791, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "473dfec1af4a46c827ab5b9c8f0e8d5bb789145a788ec2b5454ad8d6e29e7297", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.02741275, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.025210916, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024271208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "9e5d6a0badda9e5da45539be0042fb7e45bfbe3b596c01b96744c8e69fb90cd9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014772125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.014807375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.014905417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "8c09ab12fb1daa8ff0e84bd57c9ebd6d1e9ff2517501345acb7c328fdf72c193", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.025541333, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024064875, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024429125, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "b5dac4006a108245e8c34eccfd149dbcabeeedee0487ec3c5faad45ff845e9e3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.01625825, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015701917, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.016539125, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "2594d9853f5d59c126f07430216c2d3e0e13311de978dc26f38a78dc91b844bf", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.1792e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.775e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 1.7708e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "4fcb33e8de875e91fa858395b991d7d9097f87bdbd258b470910c4290784eafc", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.625e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.2042e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.4209e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "0ac046efc8e190c1fd9620196ac58dce3b5764ab222ac176864ece5eea10ef89", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.4917e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.6e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 1.6958e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "59de4f69795f5f0157b09d84f49aa901bd0f596bb469af32a9b27a98fe16de20", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.3791e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.7417e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 2.1666e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "0ea63693e5014492bc0a6381bca4337e542d69456a8428b2ebc068e46096fce8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00105325, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000659792, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00095375, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "16f4b56802ddabc466a470e56dbf07f9689f6c839239100220c313566fb2c2e8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000836292, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00078, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000963625, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "485b4c8521f94df777721b3977a539acba59a0dcefe7e2db756ad19175afa07e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000857458, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001018833, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001021167, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "8fcf43efdc650bcc90ac8b6a43ba6a266082f3c383fe6f056c573dd9c49e7196", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000493959, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001062375, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000927417, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "f0191696e3a44e714353a48b826f0fe80069b626d55ba90df9a46e63029bb0d8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006433375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005749583, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005188084, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "c7cbc35a2e5e135e98658c4aa644fd78b94721e54970b64f75ecffc6af3a5ab8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005601625, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005595541, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00523, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "58f6fc9b7084cefb478d25830b3c0d6b1b860c607cb39ec20d68e1b69d1998e3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006939583, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005952333, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006247292, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "a30a6c575c408a94a2b2901e825dafa687f202a7e94998a5d2498695bf0aef0f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00708225, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006519958, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00623975, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "636db918933fb99aa5ef4460055a6a1c8b270e5d0ccb6306b27ab8a6d471e4df", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000734708, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000726458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000726625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "9684eeba99a50655c9ff044ec9468eca7434e43d8d2a5f8665cbabaaac2288d9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000742167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000718625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000740833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "6f8620e62a7e2f17d1913251789d65902b929cd8e53898419da8b00f15d5e5a7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00667525, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006364166, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006467584, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "39f822c8a553346c65c02f3e87b71878e3e5846b780c5f3690f750c4fa90afcc", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005782583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005565125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005894167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "625cced7c8ca5b6494f731b8d4ff4570e2646dc91faed4537c12daecc9ec8dda", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.021047084, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.017179125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.015526792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "8a01ca2594c21e37f33631b334eb4a818f29da2969a8283b4d1b8d1d9d774189", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.022698334, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.022562625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.024091875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "47fa76d55faa410ae5d016efda3f2bf7cc409657a093d0ecb9320eb1444eb4cb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014809292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.01494425, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.014669, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "6485ed928f2ea414fe9dadb37d106bbc7218b80f92dc2ccc93b96192e0d2e42c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.027189042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024492709, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023964792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "e58423e2296c5476d180650f066edca58a8fbee0de47056205da0fc4e433034a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008448291, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007987083, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007760125, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "e2afc51eb795f3cf23413589b2caea530c985bf1ea32182389d9238fe8b93187", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.012158833, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.012023084, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.011611458, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "11813b4073c6ef2862d29835694042a18a6ceb2da474ff28a1a8c5a46a82bccc", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006995125, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005471292, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004535583, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "0c57d1d085f51fc281d143856bcff933b4011cd6af4c17b17e43fceecdcdffcc", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006540833, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00568225, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005058792, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "54306fde6928f99cb591fc441b4bc367b1cf86a9f9762ff0460270ffa44e8080", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008266583, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004709375, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00420025, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "f6a87465a730153c48b5e9acb40001c86c01499707c1eacd397361c00197c187", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007453667, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006264791, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005997791, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "59e78150ebae20996798ee9f70ab1e64b0449fdaaec1c2cd66e24cb6ff103e36", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000854959, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000853208, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001814958, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "816369e4f20228f42a5344438fc481d7de47df095ea009516f14e6e7541b1be2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000552375, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000544584, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000541125, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "9a579b10dac09dee630575021c53dd31fd40b7f3e833d0c780f763d6989899b5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00112975, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00085325, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000880875, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "86a0e439f60516781b97ab62da290ee646e7819332f230caa14f689600be5fb5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001417333, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000825958, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000817208, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "906b88247457834756c6263e82cb3b7134e39d4c81dbc95aeab5f607ecbede2e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0021425, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002182417, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001947209, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "711d2e048d475ae712929f00e6372cb5d65b9e5e8838b87f53482ffb7fd0545c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001560875, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0017015, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001936833, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "0ee5bc3ffb4eca63b9206edd7f1de8766433d88ddd0ab178075ea19c846dcb2b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002116208, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001564916, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001540958, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "1d43fd651fc714ef87910470f716a6122de959847d0257268f892cbb848fbfb5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002010542, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00198425, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002013667, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "acdfca9a13502e36009e5431d05acdf4d890a14dcb053e48bd945f3e344fecde", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003415792, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003103333, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002776042, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "d74ea89999c337a3a16eb44b5005afbcd8030de33b7e34e860a4173eb519bc62", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002332292, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002801917, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001961667, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "ad861a30d4bced48ca0e9fb385dd1ca270ad4fcd0263a8bf38d46914001def12", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002823709, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002961292, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003352542, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "126c877ec959dc1aea779e2d4c3d7c18205354d8ee38bdb73d6c14ac26ad8e04", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003215875, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002996708, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0030715, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "9d7bffa10504c2d80b5017aa124377aae5708ab28cf68fd1477af05c799248c3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003486375, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003243208, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00312425, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "b4711229dd9f619944b84f2f1f2f3e41956863909db846fc75c2bbae8db736aa", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003162625, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003572875, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003191875, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "32da64a8966567f40bd9d6c2e6434f1c13aac920d6f7558193af27b526e9eed0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003823584, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003705041, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003382208, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "3a3acbddad749be38497172d4506121eb86860bd6ada48d2d6b23cd2a7bb4a0d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004007417, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003957084, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003941833, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "e49dd5e3903777e7d009aba546b77795d3287060b8abc003b8b9843a4e1cc48d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005501792, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004930125, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005180792, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "80422cec75d540a5f789d06680a46e91088cac2a618a3340ada5a4233b2d8605", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003951, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003689708, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003021792, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "0cd8ff03ae236380468a177be42410cd1df9695f849728000ac41bbcab0c5632", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003269833, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003309667, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003335334, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "b345e1608f6b6fa4a85ca118bd34dffcd17678dbd5573103ece5b3c121250c2b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005019125, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003365375, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004344542, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "1cea6496071f145b1e22a6b7d4d7e53c5d1c56ba5424ef4d3ded151aa8a6cea4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003830875, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002652084, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003280292, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "e4b399f2afd6953b589b115a131a5c51b414883c2813be55ba97518dd81e8ae8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003495916, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004451833, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003796375, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "2645d35a3b9e10288e8ff453b1bf5cd8ebae364599397d7cfdc836f5b9bd809c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005923583, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005966458, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005174167, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "aecb306a3443a44c42d2b8e907f67ca407ce2f382c2e09681f017015a3d8f754", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005628709, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004936, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0046755, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "fbb674ef27b09713c70ca21c18c15ee729128da7a2c75e0ee78cd6db45f32f0b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005853625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006621292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0052535, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "f708c1ae20d06e5f795630e988ea9495cf69f5a347a486759945823293ba9091", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006986417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005516709, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005384, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "69509cac572f54f6cb3d5a711dacac9a709f7e19aa146faf19d50d847a5e5c84", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.010346958, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.009098542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.008441042, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "709dcc10d62954fa8e39f5f1e3f56ec59fac0b72c0c5cc95d6c3d0bd316421b7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007143542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006740166, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006082333, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "e71bb5964274b958e0c26d1c0d732969c66a2e00e7b19ae89fbfc9dd9b765fd0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005245166, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005302834, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005386833, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "0bb20846324ba14c9509023d3e0bd3a007f2469e952dfa25149fbc0e0ab055f9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007392291, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007022084, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007835875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "661104bb94d5ecee6a3962f22f6e307dcbe66aa2e213e077e3b3ac9d450f4320", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018306417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0170805, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.016374958, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "57898531728aa0937d9a99c937134aa1baf0b380c1e5bdd45e0b692edf9a4f80", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.025562834, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024277208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024495875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "73082a835a1acb53252ec2ee149fd5ae4e8be0a91edb8b6f0a000afc92a28ab1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.019866791, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.017368666, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.016767959, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "7d5b29b9e32823bc9abeed3331bc2be2dded861dd8f87422f2c8918ba0305ceb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.022731333, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023693292, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023121042, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "39c7d737c97a75fe6c89f79eb1c79c14adaf8fa25e40432736132bcf37805158", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.0084e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.0417e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.9708e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "6f2d6d4a9ae99df2f9e53db38557ebb044c473a0d5fed18cd9977e242d1b4b05", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.0542e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.1667e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 1.9333e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "eea966d59010c9bd966dc1f1c7b77da634b015a0df2e3f5870436b641434c0fe", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 1.8666e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.8375e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 1.7333e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "8a23a4fd67fd6d64529882afd6e427402b5ed49782f85b2e609d78a56f06f75c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.5292e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.2375e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 2.4959e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "75b09a6844b1568fb0e9848f6a6f350f52a232ec1273f222b160253b67524a69", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00095175, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001747459, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00094325, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "da1d5e5716dcda5f3349b7e96525d2692aef3ee30e178f357c7b53d14a5ca479", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000937375, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000948, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000932917, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "4b163b8afe923c76753973df68e8c6d0f734b33938663afedd7dfd8add666b91", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001062292, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001006666, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000976541, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "c5548541eee83c71919792b91258b9133baa5a53d6ccaee7c05ef0f462724855", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000995417, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000990042, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000985208, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "abe29d167ad7da99cc7f3092e9d4ac68426b2ebe465dd4620c47f6b7cd5aedd6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005707375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005650417, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005260458, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "6249be38e9923d585275267391a55a9bf4020ade143752e917dc8b0acaaf412f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003684625, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003949834, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003488375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "1ead4c7db6e27e3c63bf68689c546d8575129694d55adef9518011af77fcc053", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006696334, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006438958, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00609225, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "7301d4c22f7c05dd873c1f2e55f2e6e137dd7845ed8942750ec19e41907cc5ab", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004844458, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004452584, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004633791, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "a4ad0c54c3c53eceb5412a5bca6777c5eaea0d10ecc5829236eeeee8877f1ab8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000732959, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000733416, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000734125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "6805171e0984db7ac37c4c017c99583b6d958d2e105f060eb52f4adb64d684db", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00072975, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000734333, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.0007395, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "b2c9a869cada1cc3ed07b81cd2e977db253d1f78de886330ab33693baeec844f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005697083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005713708, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004808584, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "25a44b3fbbb0b35693b1a6591bc163698d0311e377b0656d4a97bf11e41875da", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006578666, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005916167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005662917, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "d55bc7ed22a478fb527c47b390d4e49ad91c184662897d705e90e051daf0a06e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.023700292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.019365166, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.016495083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "655a9fa701954a16fc2a32493571cda40d18676cd7fcfbfa9094558d109c8f5f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.023344625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.023127125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.02315125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "a1d157c4f8d92807b067972b269daf8783dec16d7b6873145ada51d47ee7e9f3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014968792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.014950625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.014762708, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "7b1b3e1a4760241cc415cb3c471e78c5933c57981db59b5ca341f04c4f5ad565", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.025757041, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024825584, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.025308917, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "63f83919ec2f31b8d218aed075b42592ef0056dfc90e87c806fc02d3ba7f40d4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.012025041, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.010002667, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.009283666, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "41f53fefb49fc8d73711ddea57da88bae30bf004ace8e68cd11949fb7cc28ff3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.01670475, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.013863333, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0132455, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "a8fc4f7960143c76a0a42da56d8131ad6f15d270d8b55cc4177f803012d94a9f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00507825, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005482083, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004454083, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "0787dc816a3cf4f6b50298ecd5d2a5aa6689382be781209db4a39c39aab9a040", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006582917, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005758083, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00493475, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "58afafc330fbbc51881f351acfc82619f9dbaee686a7932a6d04e60ac22e6cd0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005279042, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00430075, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004217292, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "42773fab77967d1f12c2c04517a7155d212ca17fb4b31a71a8754b967ad7839a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007103167, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006492125, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005627459, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "cb230328cc43711fa400564b3e4604b0b5703828128e79682098a5f3b2a7364e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000913792, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000933542, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000582666, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "fc2954dbfb77beaa7a67934cb6c484f1df6d656810337eab0354afb8200802b7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001018625, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001026583, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001019666, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "90218742d5a83c8967a33da943bfd5c8f949a1daec8b87192822d75ff01c91e0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001076083, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001055625, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001130917, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "21111ef84fe90159ef94a3648a8f8b61db797137a0a00381e6f52908618c407e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000996208, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000986459, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000990792, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "cd40b0a4f19b6a35f6a73b03e7d9c2e9833648530ebd2e0c6b074e2b5aaa900f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002105792, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002140667, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.0020845, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "1bbf84d0dac02d173424f2cf8f68ef10aac2c5c5d1d6d018e4645abfbaf750c9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001941791, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001914709, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001590875, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "4611b94b27a65eee3da58aec3e7ad4ba081bebd00c87d6b42a2c0f0c0c4c0940", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001975792, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001937, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002160333, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "f02be6849920a4c1390e5bfbb35d26c58887bc81ebb88b821ee71ae84f3079d7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002033625, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001979375, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001982, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "8c2f92fcfdbab847512669ec7920c6633e815693772b0f1164dee633b992a246", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002776667, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003346334, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002805, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "752705bd91400ce39dbd0521fe9a632632a3296116d9a3f9ea7f55c6509bf85a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002583042, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002623, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00251925, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "5a50238b2d8c9e3c2a7a54576e10724243c7767fa3cb420419d51678d386d235", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003602542, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0029575, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003263458, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "e363769e9e61c6971f6598a32662b8333a3d3dbed948213902f72d86756b705d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002711625, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002792208, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002436625, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "a62f1dee12ebe6ea8b5ba56c3f8acd651a94cf470d41ff7c09086e74c077f175", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003396917, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00296125, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00348575, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "c576a666a74f45d4e4dbc5d9b5f9e40cc46e49d40b12e53a43331b1e0f654bd0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00379775, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003535791, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003030875, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "c1e6132a2da56c4d2b12522c1a541d14af8f1ea863b1a80e7cc4e1430b2d0283", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003074834, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003656291, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003293875, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "8ab166362d838d9bfe38de4b4fc8c95d7da2358fbfe2346718dff7e4376907ea", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003963292, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003031667, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003104333, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "8dc8a265fecb12bc59b407b6f0b311a405238b904894a79b1204ad30f159e5ec", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005948542, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004225708, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003053167, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "d0d8ffe49dd43d2910d69beaa41e7603a5560ca449d834016b34e330764cade9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002088167, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002720209, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002976791, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "5f64bc5455041fab1ec92279e7d1e3b60f5bdcf0adc1b9e69876714d3a7907eb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005696, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005068875, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005164959, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "5d77eb9058754946a23304b222f0163c62d154f728847b043a0aaad955e72ee0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003225084, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004146292, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00349125, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "acbe90718d27eac0e06abdf070d82f1ec6d606309e5fc3f138357d882951df58", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005300042, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00566375, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005064083, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "97da889d5235cd48bc11b7ad2814a04acefcb73a37cd7017167bf75816803808", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0054595, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004437542, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004383792, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "c5aa58575bf0cdc878d1b0b56acf8ca8a0130dc318ed41d44fd0b8c58d2fb3da", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005176791, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005772125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00508325, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "9e3695c4252dad085712c160de3a6409b5c57118aa581099896b372457e110b5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005679667, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005197667, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00441075, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "4884e840b2f1d34bed4cf37e8da84486e1aef91f51939f4d48fe08b70b8344df", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006637584, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005855833, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005249584, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "4947adebf43a63a6ef62e2eb957d534a5ec66260780bfe32da1255904543bf05", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0060425, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006062292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005692333, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "bded6d814e63ba5b3f9c253423f4ae44aa768771d652d06686c2272defc4cac2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.009986542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.009474209, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.008598917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "4d3bac63470de6cb3f49224800e2a1fb96cc113ed9e94af32e89f475b26cf7a1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007607208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004199084, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003928333, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "e0894762e464a657ef8eb0db64999d3be1366637d44f8040bb04786994d96014", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.010204459, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008828917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00793375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "0009e9f025ccc4555bfdc9c5507c0c62af3555fbf0f29d63f3400cea44935406", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007408209, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005973333, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006338917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "f46be3b6e55b76790faba6820923e2e6c95fb00e45227bcfadde1c7162e279bb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.020003875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.017194875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.016782209, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "e9ee0462876a5749bd87df32e38afe6df713edc7e1c8a843a260b683b8c80121", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.023920416, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023341083, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023179792, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "b26c1cc3ad27a9e70de02623b335017458ef892f96327b47b2ec9f6788f90b77", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.023265667, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.019317791, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.017323958, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "1138440d9549900d70291ccd8e97395d5344cc5a28090a5cdbd4e805017bb85b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.032793958, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.027455, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024890125, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "38ea61c69eb0c8dde2d766563896c0ffc291b9cc3871c3cdb155b788856b4e79", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.7416e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.5917e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.4083e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "9e783b7498607943f702f045e6a4589d65fab37e80661d2cf1b63aca72df5a4c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 1.7583e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.6458e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.1667e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "eae363204d0a758f61d6e544e10903bc8e225bc39f82531b279d2f7ef9efc55a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 1.8958e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.8917e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 2.5583e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "28b80f8700a535f1bb7609040b39c2b60cc8dd4fc2aec2ed414214a5ebfe1b5c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 4.4375e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.1333e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 4.5417e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "4581fb2fa2f009b41ae6ed0438b1a3fbe3ee7f0056f548de7f99c2571e84c387", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000999708, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000946833, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.0009555, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "1ce63e42ae832de8189dad2ab9fa21e0de91c5eb73981cb99a89211a7fb04b4e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000920584, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000942, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000939459, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "51987dc512488a432d09f0e766d02d62205c9f8797e6514ff828133c8c539753", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001021667, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00073975, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000739208, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "9f564b68cc58dbe544750a4127cba707b51b43df9e8b5a12ad5b3ce40579839e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000720416, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000727625, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000712083, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "e87d70de4b58440c51b6e6a83d64d01555f211a389b90ad7d6c390b2853b4906", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005914375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005587, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00570025, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "d6fed0cfa06d81097b08d06e52b4c3a3993d63d87d925d55a729ea067dd8b7a7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005829542, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005961958, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005714208, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "9ada82bf4a6d7be4004d68430edac0897b3af766a570f43959552c5d3060c831", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006657208, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006362208, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006371875, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "c2756a3a3e32f1c226a755c7b5c34e48a81dc5e3cf37e5f05c885a42625b3e0d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006449958, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00579025, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005808792, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "50e9452783167a7cbe2082ef23cd0074c889cb7f8c8edb4e4a3a6a6b5cb507a7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000704959, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000702, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000839, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "bcdcc7e68693cbf37c642d8de285ce734d33b414ef7454c54822d4ff43f04f1b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000719542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000724458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000728834, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "441de2e7df4a2c61328823d9d3d0df7b7b49ef80cc131c75dd6d9ff4aa29ada4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005773875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004592208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00550525, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "3947a3b0dd0829ae10e047334ac9962863349e113094c16a75f84f4e99009764", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0036785, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004370042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003574875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "bd414a4c0a7eb03e9bcc99d9a441b762c7bbcd009a1ed8de96e326b9350a7531", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006194958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005935333, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005870833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "a62978ebe2af948996eaae47114947f40614d04847d0cee48584be4508edf483", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00650275, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006370875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005470208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "0d927404335837ac1af86a9f13558152ad434b2469261db689cd814946b060aa", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.026510667, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023978167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023676625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "fbbeee730c95dfa97fdf3c63b53d42035c062d2f81551850213e8148d47b37c9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014985167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.01488175, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.014825208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "67e5ac951897431fdbd80766019aa95160f01bc574b08f0eed6e4a87d976d037", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014195167, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.01324, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.012533334, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "20f32625443a36dfdca38eb31c6513e6086f9472195824be95b50bbaa55a82fe", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.012897083, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.011471708, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.01083175, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "502b7510ada4503f2074e3d9bf6b717548f0e740b536ce3e0e6843a23d946bf6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006436042, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005971459, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005183917, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "84d260f9e5a13741cbf57f6260d60ee732b60e13bb16d9669e104fb59afc5946", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006682417, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004995875, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004055583, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "a312eeadbfa89ad34438e0ede21592074f50c3d504a3c16e36c1e7baae53510b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006927208, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0061465, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0050845, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "50065f1a53f9182c63b0fbcfa26dbdcb9ca06276841d5179a28931bf4290dfa8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006244333, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006217416, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004471959, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "e27626f199515457d7ce690a538020b464047ea3b91994d8cd3117c9e62798f4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001027792, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00099575, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000996584, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "219e1811aae25c640c1c78edef4d443d46636b6a6d89b3e80d5cf080f889f26b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00091425, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000927917, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001002541, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "c1dedf231e940d5fb33fb69ba3ae826ccd43f5373274a42cf93a782eaded2f88", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000798042, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001012167, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001009791, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "6d2d2303def98558ca15631bc76ccdf3a8c381bdd59c9dbfe1f5cad710c67104", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001068833, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001114291, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000992125, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "f9cdd0256828bb9093ca7e77a12bf66250d006290747fb964483d52376ec57d8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002113083, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001720958, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001939625, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "432f071c9bf47a1f204a30b4ff498e02fb9b0b59e16fb8b38928d100a55b45e5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001660458, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001909084, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002085375, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "e5b9748f1a002d3e08f731ab0714a1a6149f8b2a75f0d486335e4efd22f4b56f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001441708, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001786125, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002089875, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "5deb564e6bec5d1f6f7e0217c5e9d3cc793bcb62f966a7e11bb09f5c1165b678", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001913292, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002117708, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002696541, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "72c13a95c4ba7907a1f1fef705d0d295b1aba88c8816eeb31a975a1279c64c46", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002738167, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00281575, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00258025, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "c93665640d6aa1c6d1af588e3b0d11404dea7e36b4f97bf1279c374e57a08deb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002994958, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002837834, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003153166, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "844875795049580d206906132e0e6313da215a35cfbdc841db111776ca5457fc", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002721041, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002172708, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002776958, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "34375c5da5380b84ce3a3a416a651dcecc217819b3ea6f2eda4e13b6088cd346", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003582459, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003640667, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003320292, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "e346698e3fde9492c48e4b55c56872e482c4395440b735e28dfc3a2efa474825", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003630333, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003850292, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003495208, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "9528d1f0140ead4ea7064368a472fb99596bde593b65c8d92fb91e13fdbed07c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003456458, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003193875, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002907791, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "e6b10364ffc38733364ee6a60cbc9cdfdfcae47487f504a6ba5d8ae5908cd127", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003951458, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003625875, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0027655, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "2412d22d21aef482729acfe7b26fbe7beaea916ea709677b9c64e2a1baddb950", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003155, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003013625, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002061083, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "e75ec5ee1ef5a4da2207fdf4ad5ee2c1cc2404016bc8a4dac812ba2fe29fd82d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004607834, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004532959, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00320875, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "6f7b5e8fab229f17c8fbd3c619f76a0bff8422f342fc0f09e152f03755648e1f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005458041, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005219125, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005740375, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "76bd799d1253bdebe928a8e0c48e61e41a48705d32e99c2fff4d8fb16f6ad06b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005426083, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004446, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004304125, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "962dd363ed611bf0df8f78f376f5efde00187294e25ad65bba6f3e7f3e937094", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005972166, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005627458, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005177125, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "143475d3ff9075fb8a3f77fa27855731d6147126b490195ad474b1587d1df7a2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004978708, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004178458, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004705166, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "125e966aab5e3b7cbce46b4f775879f1561e4b70fa1a8d5f70d3be473952f2d2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005250625, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005317625, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004469375, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "c30fee47962c7bf315976507818a0802fd09465d2028964d7961b2d2aaba63cf", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005755875, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00418125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005360042, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "205bbbf9509fa9ad33dfb0d0c00b4cd59fdfbac0dfbf22b2d21aaafac11c6be9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005926834, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006036791, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005193625, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "6262e5ba8034b2d99ee7314f557214e0b3c40fe2bfdba0cef2b886630f69726a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0066385, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005795125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005629416, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "75cc2a9dc610046857495974afd6eb1626bef0584f092e0791c777b5899533f1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005523834, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003002291, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003718583, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "32adb22dcc35f61de57826b76144519dbc4ac2fe5efd97bf3d457af1db596b1b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007577542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006452125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006416292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "aefc0889b6798ccac5f3e4589d84fe925a1a94df65821af8040c0a450da4b6ce", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00741275, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005279791, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005929708, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "5e653643fcd7d5401a55c1d94bb867b9213398498ac241766cc6ea2a73d6aec4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007150792, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0069575, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007529333, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "7e79aee66aad8298e65f7db0364a32394a74ec3ed8830a18dbdd82e2a9fbf4ba", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.010687667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006808042, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004807917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "d4481a110d14e1434d43a3ae1f1b7c53dc7a1f6bd135fcd36f5678b9ad0b0579", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.025899667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024212208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024386667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "1cc04bbcce63a42496e38754a99dbd288f5e38243ecebc5addb7d0b362589ba3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.020097417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.017347042, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015674375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "8c6754319e089c0c14e93318ed6f3b8037963a702b8499f1c17a96a0f35d603e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.024536125, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023607792, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024340292, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "bd32c6bb6db2f146bee460020ee964ccc7786361ddd029c891493c1aa1a5238e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.016056958, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.016274833, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015883875, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "14ae4d746426883f3f18cf6e039d6fae4ff319f5a74ea58ebc4e6df06dfffdc3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.2125e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.5917e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 1.9292e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "4f4d4ca9ae2c4ed71dffe055dd22b9cba708697ded126d6f5bf15378bf88fe1c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 1.8916e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.7459e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.6375e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "0f1dd5d57165aa672ac439328e36dc13544b3fd9b189e415949abdc932095a93", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 1.5166e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.7333e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 1.9375e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "b47305348b7a00049896de27a59f22e90991aa70d33c947ce0e431ab1245d6dc", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 3.125e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.5917e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 2.5833e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "b482ca581063607432a0748259ca27726a413b86369d58a801a76697faa86bb7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000942959, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000947792, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000916041, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "0d79619bdb42bb3c790c4e4c7124ce0f5a4232204f158131528b5b6057fc1a37", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001100167, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001018791, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.0009845, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "dd2af1cfb212607b776ae876159d46d141b4e4d33cc95332611a2bd7e06c665e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001006625, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000975333, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000976875, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "b292df325430df29053d8d599376806f43752c4f85fb3c804c9d9efc1ddbd432", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001044666, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00099925, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000994708, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "e274af8dc2d6971b843c27f72dd30aa465c9388a1bacd54520206ab020833321", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006358209, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005143125, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005158959, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "01bf336dd56cd2cb0f39688aaadca7c2e2bf7ae655c926abc3dc4223ee8fb05f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006718125, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005313541, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005224083, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "bf5997db1d483bc0fb3995ea761d9f1d3135daa872650cc606173224978b7e72", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007308791, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006834583, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0055405, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "1415587140c57c0f8b8d38f49e62c5a7ff6330d30fea40c992d3d089bae6c17b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007461833, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00684625, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006959541, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "477f014aa500a4df4774956688ddb3127a82fb4427b7b4579ef175ff762d7bf3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000831625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00076925, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00076825, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "b1c7d33144d3d6a8fcabd3fa18888965f62349b259a1c8e651c0d3db1e3460e0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000701125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000706167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000709333, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "0d47f2c6bc3da299ba84641cf8b9bcef4ad9da7c17cc1f8fd6b3153bd3ebdb33", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003736708, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005156708, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003634833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "3d0a8943759218cf4f29fd01c7011c94094876d7572c6ea07cb6f4743f865b1f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005088833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004902166, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004824875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "e4be936e6efb2a5d66a37f69ebe319fc6ea1a847af40158904dfbc5174ae5b0d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0060065, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006085667, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005293542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "6ef88ea4d0522e2b3da9b44df99055dee9a4f1bef6d5bdb83ea0f79949a9bedf", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005697541, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005993792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005010208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "e815daa71579f9df716ef8061041cd9b6fe37dd3e362afaef94add9aebc9a952", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.030659792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.025433042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.022918916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "b1acc9db82ee20252fc9440228d8c22f955f6c176471820d52d9bfc50c06cace", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.016277375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.014824834, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.014332583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] } ] }, "diagnostics": { "schema": "aner.metal-greedy-diagnostics/1", "timing_quotable": false, "records": [ { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.007625958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.006773, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0066165, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.006391959, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005818, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004932833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005590667, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005836583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008491292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007066458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005857042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005535625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005898875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0057465, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005714625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0065705, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.042401, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.03171975, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.022218417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.018317708, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.025752291, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.019951625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.022273583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.017617042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.03734075, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.028698125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.017686917, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.017077792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.015008208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.01489275, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014500541, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.014635583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.01532675, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 32, "status_reads": 32, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 32, "greedy_active_reads": 32, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.01488075, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 32, "status_reads": 32, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 32, "greedy_active_reads": 32, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 32, "status_reads": 32, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 32, "greedy_active_reads": 32, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.022549625, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 13, "status_reads": 13, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 13, "greedy_active_reads": 13, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.016115833, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 13, "status_reads": 13, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 13, "greedy_active_reads": 13, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 13, "status_reads": 13, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 13, "greedy_active_reads": 13, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.010162541, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 7, "status_reads": 7, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.01058425, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 7, "status_reads": 7, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 7, "status_reads": 7, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.017422958, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 6, "status_reads": 6, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.013791084, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 6, "status_reads": 6, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 6, "status_reads": 6, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.011165584, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.006158042, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0114035, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.008526791, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.010501166, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.010664458, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.008391916, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0063015, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.009738083, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007159958, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.011385334, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004843167, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007169625, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006058834, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.011626209, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008377125, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0011165, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001103667, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001198917, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001115042, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00104925, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000967834, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00111825, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001110667, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001128, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000857625, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001000834, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000936291, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001190708, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001109667, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001137333, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001091209, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002153917, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002075167, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001794292, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001771167, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0021385, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002098291, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002074458, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002849875, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001930792, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002179334, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002141, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001762625, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002294917, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002096625, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001995417, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002052542, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003174125, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 3, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003153958, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 3, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 257 ] ], "positions_before_probe": [ 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 3, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003421458, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002914541, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 257 ] ], "positions_before_probe": [ 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003424041, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003103917, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 257 ] ], "positions_before_probe": [ 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003785375, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003439375, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 257 ] ], "positions_before_probe": [ 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002988625, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 24, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003138333, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 24, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions_before_probe": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 24, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003298042, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003858625, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions_before_probe": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003596084, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003599208, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions_before_probe": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002935417, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003538083, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions_before_probe": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004108292, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003770916, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003594459, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003463292, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003453208, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003477375, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004217084, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003552959, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004224625, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003316584, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00241775, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002951667, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003681875, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003663417, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003700917, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003961667, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005033667, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 5, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004973375, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 5, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 5, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005446583, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005238458, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004939125, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005012084, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005512041, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004956167, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005180875, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 40, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005055417, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 40, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 40, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004925291, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004828667, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005936583, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005826167, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005615375, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00522, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005065916, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 6, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004515208, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 6, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 6, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005284, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004983208, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004835083, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005091125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004549667, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002755208, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005364125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 48, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004620417, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 48, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 48, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005213, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005769167, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005912, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005888791, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005686417, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005986917, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008379708, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007723375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006000833, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006688, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005828208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005843916, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006613709, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006166625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00753875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 9, "status_reads": 9, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006284208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 9, "status_reads": 9, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 9, "status_reads": 9, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007783834, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006820708, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00933575, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.009495875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.010780542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.009411291, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00890825, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 10, "status_reads": 10, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 10, "greedy_active_reads": 10, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008755, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 10, "status_reads": 10, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 10, "greedy_active_reads": 10, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 10, "status_reads": 10, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 10, "greedy_active_reads": 10, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007479959, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006351917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.009410375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.010243417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.009315083, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.009839042, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 1, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.030804042, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 1, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.025411542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 1, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 3, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.017425042, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 3, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.01701375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 3, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.015353292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.014969417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 32, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.015397, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 32, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015777167, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 32, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.036526583, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.029143291, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.01872525, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.018161292, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.02149975, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.020589166, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.015689417, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0168005, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 4.1583e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.875e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000152709, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 3.6375e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 3.65e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.925e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 3.5958e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.6292e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 3.575e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 3.1959e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 3.475e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 3.8666e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00012675, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 3.05e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 4.2083e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.8458e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00110525, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000797584, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001100167, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001041083, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001045875, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001014, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001070458, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001036, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00111625, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001035333, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001079208, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001036083, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001021041, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000949791, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001089916, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001066208, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } } ] }, "full_diagnostic_json_sha256": "4bd9d5d69f5699c245340a71ffc947c3c5ae86ac8fc33b79b9e08680776d9433", "numerical_vector_policy": "Full vectors retained in the research capture; portable Float64 little-endian digests replace repeated vectors here. All output IDs, timings and counters are retained." }, { "capture": "03", "label": "gradual startup with specialized mask", "is_final": true, "summary": { "schema": "aner.metal-greedy-local-summary/1", "protocol": { "matched_process_pairs": 4, "timed_calls_per_process": 3, "excluded_warmups_per_process": 2, "process_order": "AB,BA,BA,AB; profile order rotates by outer round", "primary_phase": "natural", "natural_policy": "greedy complete requests; actual IDs, EOS and counts retained; baseline/candidate ID equality is measured, not assumed", "controls": "CPU Float32 tiny request and unchanged nonzero-temperature MPS sampling; all-zero and one-token budgets retained as functional/latency controls", "memory_scope": "tracked owned numeric capacities; readback counter includes metadata and Float32 widening; excludes stack vectors, allocator/driver/command-buffer/BLAS overhead and RSS", "thread_environment_request": { "VECLIB_MAXIMUM_THREADS": "1", "OPENBLAS_NUM_THREADS": "1", "OMP_NUM_THREADS": "1" }, "confidence_interval": { "point_estimator": "median(all 12 baseline calls) / median(all 12 candidate calls)", "resampling_unit": "whole matched pair of fresh process clusters", "inner_calls": "all three calls from each selected process retained together; no inner-call resampling", "method": "exhaustive enumeration of all ordered four-cluster draws with replacement", "draws": 256, "coverage": 0.95, "percentile_method": "linear interpolation at probability * (256 - 1) in the ordered bootstrap ratios" } }, "producer_sha256": "834ce647498123f3dd7f95cd56af3c4b7e0441f8cc14900821007be4b8e0ad96", "binary_sha256": { "baseline": "d0522c5e8c5c52f3ebc0f6920ec48aa2d405bf4d8a82eb3fd1fc1dab8162e610", "candidate": "36b87a82f1e46bb87b2aacf5af964f5d04abb860e31eabe379b001feb25fc6d3" }, "build": { "configuration": "Release", "lto": false, "devices": [ "mps", "cpu" ], "dtype": "Float32", "compiler": { "id": "AppleClang", "version": "21.0.0.21000334" }, "compiler_flags": [ "-O3", "-DNDEBUG", "-std=c++20", "-mcpu=apple-m1", "-fno-fast-math", "-ffp-contract=off" ] }, "measurement_sha256": "34a1e747d54390a023b91ab5d8ec52d03be13e6baa6122636b998332bb7f295d", "oracle": { "schema": "aner.metal-greedy-local-oracle/1", "profiles": [ { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.202861309985695e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 1.3992326293094704e-07, "candidate_max_scaled_error_vs_cpu_float64": 1.3992326293094704e-07, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 32, "baseline_max_scaled_error_vs_cpu_float64": 9.61704129314485e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.61704129314485e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 3.0323179185145576e-07, "candidate_max_scaled_error_vs_cpu_float64": 3.0323179185145576e-07, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 3.230860285372708e-07, "candidate_max_scaled_error_vs_cpu_float64": 3.230860285372708e-07, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_max_scaled_error_vs_cpu_float64": 6.06905886771679e-09, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 64, "baseline_max_scaled_error_vs_cpu_float64": 9.487577588236166e-08, "candidate_max_scaled_error_vs_cpu_float64": 9.487577588236166e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 0, "baseline_max_scaled_error_vs_cpu_float64": 0, "candidate_max_scaled_error_vs_cpu_float64": 0, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 0, "baseline_max_scaled_error_vs_cpu_float64": 0, "candidate_max_scaled_error_vs_cpu_float64": 0, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 1, "baseline_max_scaled_error_vs_cpu_float64": 6.854355312918667e-08, "candidate_max_scaled_error_vs_cpu_float64": 6.854355312918667e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "steps": 1, "baseline_max_scaled_error_vs_cpu_float64": 6.68301477395295e-08, "candidate_max_scaled_error_vs_cpu_float64": 6.68301477395295e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 }, { "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "steps": 8, "baseline_max_scaled_error_vs_cpu_float64": 2.907853657679229e-08, "candidate_max_scaled_error_vs_cpu_float64": 2.907853657679229e-08, "candidate_vs_baseline_max_scaled_error": 0, "declared_cpu_reference_tolerance": 0.0005 } ], "exact_parameter_values_equal": true, "parameter_values": [ { "device": "cpu", "width": 32, "layers": 2, "model_fixture": "random", "parameter_serialization_sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos1", "parameter_serialization_sha256": "e3e6ff6db789c848e78a759a6fc69b3e2c9329dff3ea951843eb0b348bf82f2a" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos10", "parameter_serialization_sha256": "ad522891246ca60c51e52ff565c3016a616a50b164ddd6127c9353a5e8ed1b56" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos2", "parameter_serialization_sha256": "ed9153be24577314bb4e31f0283b576ff51df1335d415bb9a6d1e361b1ad1d08" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos3", "parameter_serialization_sha256": "0fe8452022f38407bdad8c24d202a81f68d5b0d431f919295c67ad8d0449a505" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos4", "parameter_serialization_sha256": "7c24715177c00583dcbc5cd02c511fbe8439752dd4ddb1561eb79bdf060f04d6" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos5", "parameter_serialization_sha256": "3b57c548c9d63e1a6514ae59badabbb062de7b99c8321f84af893c5d563ba079" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos6", "parameter_serialization_sha256": "5c516784a0fe783125b7b5cb1e52e702d95c2adab654d2cbf376a0f79c02043d" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos8", "parameter_serialization_sha256": "6b16fedfaa1cd79eccddbe930376a7c04833dd460280fad86699368b2331c5a9" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "eos9", "parameter_serialization_sha256": "bf13ad8b95d255bb2ae96a8f20a80f80154beebf9165d17e9d16642217f96818" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "mixed_eos", "parameter_serialization_sha256": "ae87e82e1f0040a28a821ff69662ed1773bc854740534880fa8f1670d11d1973" }, { "device": "mps", "width": 32, "layers": 2, "model_fixture": "random", "parameter_serialization_sha256": "1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107" }, { "device": "mps", "width": 128, "layers": 3, "model_fixture": "random", "parameter_serialization_sha256": "7432a93bd25dd41f127bc7621a6b14e5e15fba11c703df9d7c8ebfc28316c9d8" } ] }, "summary": [ { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.00556825, "candidate": 0.005148812 }, "baseline_over_candidate_ratio_of_medians": 1.0814630637125613, "whole_process_pair_bootstrap_percentile95": [ 0.9368580484905105, 1.1227882767377542 ], "min_max_seconds": { "baseline": [ 0.002870416, 0.0072655 ], "candidate": [ 0.004395542, 0.00594875 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.005730292, "candidate": 0.0054573125 }, "baseline_over_candidate_ratio_of_medians": 1.0500208664979327, "whole_process_pair_bootstrap_percentile95": [ 0.887879489683141, 1.1744996113069126 ], "min_max_seconds": { "baseline": [ 0.003464083, 0.007043916 ], "candidate": [ 0.004351917, 0.006302291 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.025156479000000002, "candidate": 0.0155632915 }, "baseline_over_candidate_ratio_of_medians": 1.616398369201014, "whole_process_pair_bootstrap_percentile95": [ 1.454319849709478, 1.7325735446540274 ], "min_max_seconds": { "baseline": [ 0.023221958, 0.029424583 ], "candidate": [ 0.013493916, 0.022352042 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 64 ], "candidate": [ 64 ] } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.024750604000000002, "candidate": 0.015017312500000001 }, "baseline_over_candidate_ratio_of_medians": 1.6481380406780508, "whole_process_pair_bootstrap_percentile95": [ 1.4025215727346236, 1.7337060142421676 ], "min_max_seconds": { "baseline": [ 0.023045667, 0.030632792 ], "candidate": [ 0.014393792, 0.022200916 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "candidate": [ 64, 64, 64, 64, 64, 64, 64, 64 ] } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.013126937, "candidate": 0.0091288125 }, "baseline_over_candidate_ratio_of_medians": 1.437967643655733, "whole_process_pair_bootstrap_percentile95": [ 1.0850397559922795, 1.6626344741673222 ], "min_max_seconds": { "baseline": [ 0.012244208, 0.018193125 ], "candidate": [ 0.007518833, 0.014329416 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "candidate": [ 32, 1, 16, 2, 24, 3, 8, 4 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.006026229, "candidate": 0.005181229 }, "baseline_over_candidate_ratio_of_medians": 1.1630887189120573, "whole_process_pair_bootstrap_percentile95": [ 1.0369554241327565, 1.3423912916186327 ], "min_max_seconds": { "baseline": [ 0.004773291, 0.009345333 ], "candidate": [ 0.003969083, 0.0085245 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0061602085, "candidate": 0.0050415205000000005 }, "baseline_over_candidate_ratio_of_medians": 1.2218949620456765, "whole_process_pair_bootstrap_percentile95": [ 1.206048422669938, 1.283567307930377 ], "min_max_seconds": { "baseline": [ 0.004986125, 0.008742584 ], "candidate": [ 0.004151833, 0.008376291 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0009944795, "candidate": 0.0010741875 }, "baseline_over_candidate_ratio_of_medians": 0.9257969395473323, "whole_process_pair_bootstrap_percentile95": [ 0.8839781189899942, 0.9827695960124236 ], "min_max_seconds": { "baseline": [ 0.000762292, 0.001735 ], "candidate": [ 0.000823833, 0.001176917 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1 ], "candidate": [ 1 ] } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0009631665, "candidate": 0.0011024795000000001 }, "baseline_over_candidate_ratio_of_medians": 0.8736366526543122, "whole_process_pair_bootstrap_percentile95": [ 0.7639024099265881, 0.9333490744455271 ], "min_max_seconds": { "baseline": [ 0.000766084, 0.001055084 ], "candidate": [ 0.001027, 0.001164166 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "candidate": [ 1, 1, 1, 1, 1, 1, 1, 1 ] } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.001937771, "candidate": 0.0019236255000000002 }, "baseline_over_candidate_ratio_of_medians": 1.0073535623228118, "whole_process_pair_bootstrap_percentile95": [ 0.8477093183296313, 1.0756325465853434 ], "min_max_seconds": { "baseline": [ 0.001351667, 0.002761 ], "candidate": [ 0.001444875, 0.002132708 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 2 ], "candidate": [ 2 ] } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0019936249999999997, "candidate": 0.002095458 }, "baseline_over_candidate_ratio_of_medians": 0.9514029868410627, "whole_process_pair_bootstrap_percentile95": [ 0.932677747163215, 1.0510218875029418 ], "min_max_seconds": { "baseline": [ 0.001434958, 0.002499667 ], "candidate": [ 0.001712875, 0.002220792 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "candidate": [ 2, 2, 2, 2, 2, 2, 2, 2 ] } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0025359375, "candidate": 0.0029990629999999997 }, "baseline_over_candidate_ratio_of_medians": 0.845576601758616, "whole_process_pair_bootstrap_percentile95": [ 0.6676345525138853, 0.9763674373910181 ], "min_max_seconds": { "baseline": [ 0.001953625, 0.003003625 ], "candidate": [ 0.002383625, 0.003499 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 3 ], "candidate": [ 3 ] } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.002959666, "candidate": 0.003000063 }, "baseline_over_candidate_ratio_of_medians": 0.9865346161063951, "whole_process_pair_bootstrap_percentile95": [ 0.8317624157133503, 1.0429620222073788 ], "min_max_seconds": { "baseline": [ 0.002454, 0.003308167 ], "candidate": [ 0.002306459, 0.004066417 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "candidate": [ 3, 3, 3, 3, 3, 3, 3, 3 ] } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0031859375, "candidate": 0.0031178335 }, "baseline_over_candidate_ratio_of_medians": 1.0218433729703653, "whole_process_pair_bootstrap_percentile95": [ 0.9520041139965344, 1.1092703297812523 ], "min_max_seconds": { "baseline": [ 0.002567167, 0.003799167 ], "candidate": [ 0.002465708, 0.003378958 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 4 ], "candidate": [ 4 ] } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0029805415000000004, "candidate": 0.0029985625 }, "baseline_over_candidate_ratio_of_medians": 0.9939901202659609, "whole_process_pair_bootstrap_percentile95": [ 0.79424280676404, 1.1900640617569345 ], "min_max_seconds": { "baseline": [ 0.001853875, 0.004266083 ], "candidate": [ 0.002709166, 0.003571292 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "candidate": [ 4, 4, 4, 4, 4, 4, 4, 4 ] } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0038451875, "candidate": 0.0045536455 }, "baseline_over_candidate_ratio_of_medians": 0.844419597441215, "whole_process_pair_bootstrap_percentile95": [ 0.7861984483462638, 1.0174185543929475 ], "min_max_seconds": { "baseline": [ 0.003202541, 0.005084541 ], "candidate": [ 0.003056666, 0.005197833 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 5 ], "candidate": [ 5 ] } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0041242295, "candidate": 0.0049892705 }, "baseline_over_candidate_ratio_of_medians": 0.8266197433071628, "whole_process_pair_bootstrap_percentile95": [ 0.7526489372631766, 0.8488143809639063 ], "min_max_seconds": { "baseline": [ 0.003431334, 0.005091875 ], "candidate": [ 0.004388666, 0.005403583 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "candidate": [ 5, 5, 5, 5, 5, 5, 5, 5 ] } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.004273438, "candidate": 0.004850396 }, "baseline_over_candidate_ratio_of_medians": 0.881049299892215, "whole_process_pair_bootstrap_percentile95": [ 0.8134642252419255, 0.9753517261152308 ], "min_max_seconds": { "baseline": [ 0.00386425, 0.005542834 ], "candidate": [ 0.004118708, 0.005307167 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 6 ], "candidate": [ 6 ] } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0040995630000000005, "candidate": 0.0052018125 }, "baseline_over_candidate_ratio_of_medians": 0.7881028007064846, "whole_process_pair_bootstrap_percentile95": [ 0.7628281911534233, 0.9971385063527963 ], "min_max_seconds": { "baseline": [ 0.003168375, 0.006377333 ], "candidate": [ 0.004567125, 0.005867875 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "candidate": [ 6, 6, 6, 6, 6, 6, 6, 6 ] } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.005204792, "candidate": 0.005301375 }, "baseline_over_candidate_ratio_of_medians": 0.9817815189455565, "whole_process_pair_bootstrap_percentile95": [ 0.8097256814542915, 1.0555264302517449 ], "min_max_seconds": { "baseline": [ 0.003762875, 0.006501 ], "candidate": [ 0.005058542, 0.006154042 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.006567229, "candidate": 0.0077387285 }, "baseline_over_candidate_ratio_of_medians": 0.8486186070489486, "whole_process_pair_bootstrap_percentile95": [ 0.8092131213993997, 0.9568589853680264 ], "min_max_seconds": { "baseline": [ 0.005620875, 0.007571709 ], "candidate": [ 0.005404834, 0.009734417 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "candidate": [ 9, 9, 9, 9, 9, 9, 9, 9 ] } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0064650625, "candidate": 0.00719 }, "baseline_over_candidate_ratio_of_medians": 0.8991742002781641, "whole_process_pair_bootstrap_percentile95": [ 0.811384236218904, 1.0425263830987048 ], "min_max_seconds": { "baseline": [ 0.004737166, 0.008048625 ], "candidate": [ 0.005160209, 0.00955325 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "candidate": [ 10, 10, 10, 10, 10, 10, 10, 10 ] } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.02421375, "candidate": 0.015028167 }, "baseline_over_candidate_ratio_of_medians": 1.6112244427414202, "whole_process_pair_bootstrap_percentile95": [ 1.4894429703197403, 1.7242539557785603 ], "min_max_seconds": { "baseline": [ 0.02329275, 0.029619666 ], "candidate": [ 0.013959208, 0.018649333 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "candidate": [ 64, 9, 8, 7, 6, 5, 3, 1 ] } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.023810688, "candidate": 0.0166898335 }, "baseline_over_candidate_ratio_of_medians": 1.4266582108203776, "whole_process_pair_bootstrap_percentile95": [ 1.190861225718049, 1.5454536377666284 ], "min_max_seconds": { "baseline": [ 0.0226045, 0.026623917 ], "candidate": [ 0.015123333, 0.020994083 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "candidate": [ 0, 1, 0, 8, 64, 1, 0, 16 ] } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 2.2499999999999998e-05, "candidate": 2.3791499999999997e-05 }, "baseline_over_candidate_ratio_of_medians": 0.9457159069415548, "whole_process_pair_bootstrap_percentile95": [ 0.7254833113020489, 1.2710252659048673 ], "min_max_seconds": { "baseline": [ 1.6416e-05, 4.1375e-05 ], "candidate": [ 2.1041e-05, 7.4167e-05 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 0 ], "candidate": [ 0 ] } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 2.0854e-05, "candidate": 2.0708e-05 }, "baseline_over_candidate_ratio_of_medians": 1.0070504152984354, "whole_process_pair_bootstrap_percentile95": [ 0.7438475739814665, 1.16334110294306 ], "min_max_seconds": { "baseline": [ 1.8958e-05, 2.7041e-05 ], "candidate": [ 1.5708e-05, 2.9625e-05 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "candidate": [ 0, 0, 0, 0, 0, 0, 0, 0 ] } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0009335834999999999, "candidate": 0.0009555835 }, "baseline_over_candidate_ratio_of_medians": 0.9769774174627334, "whole_process_pair_bootstrap_percentile95": [ 0.8210297687224611, 1.1026514917554595 ], "min_max_seconds": { "baseline": [ 0.00065875, 0.001529209 ], "candidate": [ 0.000665208, 0.001398417 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1 ], "candidate": [ 1 ] } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.000925396, "candidate": 0.0009929375 }, "baseline_over_candidate_ratio_of_medians": 0.9319780952980425, "whole_process_pair_bootstrap_percentile95": [ 0.7537749987913559, 1.7748042048833603 ], "min_max_seconds": { "baseline": [ 0.000748666, 0.001137375 ], "candidate": [ 0.000460167, 0.001811209 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "candidate": [ 1, 1, 1, 1, 1, 1, 1, 1 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.005659792, "candidate": 0.005459708 }, "baseline_over_candidate_ratio_of_medians": 1.0366473811419952, "whole_process_pair_bootstrap_percentile95": [ 0.8370530299354714, 1.0532618387851693 ], "min_max_seconds": { "baseline": [ 0.003234708, 0.006401417 ], "candidate": [ 0.004724333, 0.006261 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0064911875, "candidate": 0.006519499999999999 }, "baseline_over_candidate_ratio_of_medians": 0.9956572589922542, "whole_process_pair_bootstrap_percentile95": [ 0.9489065128574862, 1.0549862017626135 ], "min_max_seconds": { "baseline": [ 0.006006, 0.007263125 ], "candidate": [ 0.005537334, 0.00761275 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "candidate": [ 8, 8, 8, 8, 8, 8, 8, 8 ] } }, { "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "calls_per_implementation": 12, "median_seconds": { "baseline": 0.0006942915, "candidate": 0.0006762915 }, "baseline_over_candidate_ratio_of_medians": 1.0266157418805353, "whole_process_pair_bootstrap_percentile95": [ 0.9951840589058709, 1.0923821940771092 ], "min_max_seconds": { "baseline": [ 0.0006725, 0.000774417 ], "candidate": [ 0.000670875, 0.000845958 ] }, "memory": { "baseline": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "candidate": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 } }, "statistics": { "baseline": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ], "candidate": [ { "available": false }, { "available": false }, { "available": false }, { "available": false } ] }, "within_runtime_ids_repeated": { "baseline": true, "candidate": true }, "baseline_candidate_exact_ids_equal": true, "baseline_candidate_actual_counts_equal": true, "actual_new_tokens_per_row": { "baseline": [ 8 ], "candidate": [ 8 ] } } ], "scope": "One Mac, bounded Float32 reference decoders. Matched whole public requests retain every emitted ID, actual count and EOS result, including latency regressions. CPU and nonzero-temperature sampling are controls; placed-Float32-to-CPU-Float64 numerical oracles and exact parameter proofs are outside timing. Four matched process pairs support each interval.", "ci_estimand_correction": { "raw_timings_unchanged": true, "reason": "Previous intervals bootstrapped medians of process medians while the point statistic used medians of all calls; corrected draws preserve whole process pairs and recompute the same all-call statistic.", "original_summary_sha256": "160e4db31ed236ccacc5be425cff616aec42d7de5f0a1dfddecd106d7ed80c98" } }, "natural_observations": { "schema": "aner.metal-greedy-local-measurements/1", "protocol": { "matched_process_pairs": 4, "timed_calls_per_process": 3, "excluded_warmups_per_process": 2, "process_order": "AB,BA,BA,AB; profile order rotates by outer round", "primary_phase": "natural", "natural_policy": "greedy complete requests; actual IDs, EOS and counts retained; baseline/candidate ID equality is measured, not assumed", "controls": "CPU Float32 tiny request and unchanged nonzero-temperature MPS sampling; all-zero and one-token budgets retained as functional/latency controls", "memory_scope": "tracked owned numeric capacities; readback counter includes metadata and Float32 widening; excludes stack vectors, allocator/driver/command-buffer/BLAS overhead and RSS", "thread_environment_request": { "VECLIB_MAXIMUM_THREADS": "1", "OPENBLAS_NUM_THREADS": "1", "OMP_NUM_THREADS": "1" } }, "producer_sha256": "834ce647498123f3dd7f95cd56af3c4b7e0441f8cc14900821007be4b8e0ad96", "binary_sha256": { "baseline": "d0522c5e8c5c52f3ebc0f6920ec48aa2d405bf4d8a82eb3fd1fc1dab8162e610", "candidate": "36b87a82f1e46bb87b2aacf5af964f5d04abb860e31eabe379b001feb25fc6d3" }, "records": [ { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "271145868abb764e41843f840d6753c804b48409f67b54e8c5ab22360cd478f8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005092084, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00423625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002870416, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "2c9decb2302b88af067bb2271f9ec6756cf7c77e8b98513ff49ec0f2e079671e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004949875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005290458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004395542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "13f1891b3d62d14be1b44d45a4f79e10bd3eed6d304239d86d285dc29d96806b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007043916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00579825, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005479291, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "805b44c82cd6399db3a3ea6b55cb13c40dad7ce8b7702399641a4f8937424019", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006302291, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004956416, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004832375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "2bf55ad8ca815125cf77b79520077e4f85628e2701fafcb4c878d676da4a0732", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.029385959, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.023221958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.024469209, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "67792973449b08d538883a1ad10aedf881071951947c21dc6fe3b127f0fb2949", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.013789958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.014723792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.0135305, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "b991da2320a41cfcc8acef4c99a20aebf2992c53b8c33aa0af2bb4764d285062", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.024649417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023765208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023045667, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "9e4f32324eacd36cd8bcb75e9aa1bc575f79a7897b097e308f5a97359d4c9ece", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.016312625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015901041, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.014668916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "68c7b5129fda1450aed033d047458905a8b6aac1dbdc7bd3ac4048b5c4537cc1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.013422292, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.014060375, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.012690959, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "d95a89eb10886cc0d5fed501b6719568552e40be73ab467b34f7ef909bb46192", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014329416, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.012675417, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.011804792, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "bd38b8668c629b469ccf3b77d9a8cb8b0d65ec220c69d12134c52d1357fe8cfa", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.007755958, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005873834, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005031708, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "918c1ffd69da97c14df76dbc0bf9ed81d81d3dc581c4d02c4bcc7f99e5b89365", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005893209, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0056645, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004414667, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "f2f182b10b31ff5e9e566ed11383c6b5612169bfdbe8ecaa29d9ec665ac05799", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007240166, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006083459, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005139666, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "cbc207ad7594077975f2cf15c9b3de1503ef5c99918bcf09fb73a521f58bfb45", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006293875, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005044125, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004513667, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "2307e22c7b8b557c0eeadafa59457b0cbd38ecde488a7a561d7b10ce8be49016", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000996042, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0010095, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000977834, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "c01af7aa838e05ce80643656e871595e15d1b3d3c4b65bdb65d87e189466eb6f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00115625, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000825917, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001074375, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "3e3fb1ff89ae35bf670dbfe8a6b59033c8f263128cf12330ebb221e0236bdd19", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000850541, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000766084, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001033833, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "1a135ad6e33d8921a551e4616671458cc96cd97de1d0bd8ee70496225771138c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001067375, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00105225, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00104575, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "ffc6aedfd8a3acca932bd408cd5517d7cd5314e28523e6cefaea68864bcdde83", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0019345, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001351667, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001667417, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "f40c2a08beaade87df0cd382145fd1cc648be34528684f97e83b28b309f676c6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001544209, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0020815, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001821333, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "68cdcbcdd60c6c739f2b4626b2478462f1846d89860d91a697cd01c09faeded7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002007792, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00196375, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001971334, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "ddeecb3f8ed6be925e82ab26a21e0aed9bb98c1f92ea0025e89825acd09acf18", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002220792, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002117167, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002101958, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "31762c6a642b6443cd2661b533519f53220742c39286c7d39c6a0f3187d2bcd9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002895375, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001955584, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001953625, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "ecfa4075fb3e2448c165779da2f0079b81d80520c55e99d0f4d40f391c8b2fc1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00321175, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003218416, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003499, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "116290c65fd00cdc07fff78d3ed6388bfd0ffcfd51a6321b0a7c6c4a296c791d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003308167, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003024167, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002924041, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "5cd0bfa44e23c0c1950904be5dd62ab7dc2ddfab8b0a8c4f0b8a053a0d2ef6b5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003101584, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002548375, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003497125, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "1725a294c7d26db630b9583f315f4c065eb6f7c5028996f852a2303e3b8a32c4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003799167, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003428708, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002567167, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "478c1c9c13ec483a8b8d23dbc69815f2ec5488be97877f9cfbe8e37199509ebb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003287292, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003090958, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002465708, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "dc031f231e71f67789482a2690f331302b12d621954446c67d567a4a5d1393ec", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001853875, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0027535, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001942791, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "f07922d3f09af7a9d61b546c9d40388ca9c2ff2723e3aa62c53738f68b25fe95", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002709166, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003004583, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002956458, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "474531648e124ae0d39a3c810af20195e3e7d938670d406853e5635c8db06118", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005084541, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003512709, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003999417, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "f977424f09b77b8fe99eec389f1d6f1f8ef03e533cb908aab4b58d338f0c59dc", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004700791, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004876208, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004567583, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "b17edc4c655c148777b38d3fd7a5b1ab707e66b9ecd3d6d8389827bdb51dfdb8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003468375, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004564291, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003431334, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "bfff748e86f9445f38492685ef8bf8348603f7119cda2663cd1da98b9c5f304c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004756833, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005221791, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004388666, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "5e5520f1894cba971dffd69f387db412cfc545049997ed00bc189cb887e1ada1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004438625, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004175834, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004087625, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "9d94368b4eae34466bf9167465598b1cadf1ba43e4801e573ff3d991047673b6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005307167, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005177709, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004205208, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "ba35fa6e0a92072df7d508a901f90473573a84474a67296975c306708ee5d90a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005242125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004102917, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00405, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "db8e1eb38cb254e590b9f6f8e5391a2ad2d7d6c712de10293ed66478888ec7c4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004878292, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005114958, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005867875, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "719a78adc1f6c26d5112d3983aa3dab2d71444533230e2f8f49e5154bf887795", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005833958, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005643291, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005358917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "791bbc0205e16a4020c549091becded2e83a35b65bf98d8de3dd458c3c6e80d3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005089917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005698167, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005267208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "d397ffe23d5258a81ac223eb77083177971dc1cddf8c9d35d7f9df7ca5d07cf0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007184834, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006304208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006037333, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "f5dd6cd3896f6b1450788db089d26b1531b88fcf012417d961c6128d2aa8ac68", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007790541, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007686916, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.008098875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "874b1cb9e8603c3cff0ee37ca8a390b6d068599ac0a1c3b4217e2ea49f800f63", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007593334, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006473875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006356459, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "4c77da58c123b680cec269cbeefc0a7c97a4e3874745d907c4c10c58bec995ad", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007465834, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006206625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005160209, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "64f97c416a0b2f2771a1fb24e3c14fac3adb421a3b41716ddbea65d30ad3a8fe", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.026503333, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023827125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023866042, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "3475731481d9e1da37a0b1d5d3e8b5b3dcee0936a6e51829e57038b63b7314de", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014001459, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.013959208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.014710292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "9f3a3e94df9b90bc29c08050a9b67a6649b47330110dc373fea14aa405d00273", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.023290125, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023949459, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023336167, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "17fb8865b5389323c0835845e93e326e70abbb339d29470a4ddcbcf20b74d4c1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.019378042, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.016859459, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015681833, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "925c115e486cf210fe5801080b8e8094650b2e406efe43d13623b258017c06f2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 3.7666e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.3875e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 4.1375e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "102776b8beaec69d0bd741f218bf91c990a2610ededd081c68a1f5937b7d477a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.2875e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.5291e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 5.4834e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "594ca3a070d8dc856a1ed836c53653bbf5947481ebba0dc300dd25b6cb16924e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.075e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.0958e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 1.9459e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "20bbbef2ba0f81125a1040d74989e306104ead6d7ca7bafb3fc7fade5e22a94e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.8416e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.9291e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 2.9625e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "5216dbc4deac2cb8079cbddb4fa3b199b60eae5132b665086e7a3afe7564fdc9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000679042, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001319916, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000754333, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "6e5ef1f4bed4844799fc707ff4081dc76f98943afdde5a7fafa3f11ddb66413e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000949167, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001008375, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000962, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "9f6a84c43f2c7800eb74e2de1de7dfff4e33ae8a683ffa404462e99f6025a8a1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001022125, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000814625, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000787917, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "4538cfcef3364bf8e478a16624fe12c23839d2b18443d77270f21d27e620311f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00099425, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000988375, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001010625, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "917a8494ce4b418194e99d24ed104e3dffd5b8efd859c2b9b315343d2f302d96", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005673917, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0056515, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005999791, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "96183c62fe685b4e22cc4c351916a57300cdfaa7317664175a1ef8d39ee550ef", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005268791, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.006261, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005389916, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "285936dfb7b4cdf3432f1cf8b14667ee2ca0cdbbab7667cfe858d595f7d3ce71", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006985792, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006134708, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006585292, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round0-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "16c160bcad7b5db3af65f440313cd6b38b6f82ed8f491e7788a1cabf8fcc90e7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006817125, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006662292, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006918, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "02576b3cb638c005e683bdf47150cddcaba2f2112d8bcb0cd3d38f7e0c0685f0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000683291, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000677458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000677042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 0, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round0-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "a8aa532225c0cfcda7a8cb2be61a59c598cc3d9fd2f57fe0ecc01c709f190c61", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000679, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000676416, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000845958, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "87bfa750b5f0a6865612b3cbe30805396b9c878bf85408a4863fc857186bd671", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00550425, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006062125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0059345, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "8cda7b37f6a1abfa661c5be71a09223bb9bbde7928bb7e12f11d8f131fbfc166", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005662334, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006978125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005949583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "c5a13f2e0ff9fecf1d40cffe4a70291bf499330a9931ab64c3c9540cbcaea0da", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.018644833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.017380209, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.016884, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "9b151d57b78f50ab0b9d8425e9191d2b8155a0b5bcbb788771920a761629be36", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.029424583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.026276208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.024597291, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "85967ef9bc5068add96223e63066edda92a17b9f145ef47f74e56ee79d2f495d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.022200916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.018452333, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.017824709, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "866126b19983e89deae69fbe158cc389592da34d05ba1a675e5cd64f6c043250", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.030632792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.025439667, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024851791, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "d82f436f31aea982448e5fc180b6fbda6463c8e08f4f02374021f4077a8afef8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.011360625, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.010223042, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.009157834, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "775e5b1bfa0ef1f14f8a4a9ee6ebf9dd45ff4a950b6aca664bc2a0d0558eaf19", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018193125, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015713917, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.013263375, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "33640d16ab5c4ad84c97d0cc21bd41024dcc4c497f24ef54e959dd622ec1bdf0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00679525, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00510125, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004541, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "c3f849efcaf74b78d966e811e2029b39f18cc05310e5077c9a267ac92e1825ca", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.007279042, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005950708, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00610175, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "3918ba60b6fe50ba7e4fd45e5cea4791fdbe091e36f5031ebb83c11f049dad75", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005038916, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0045755, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004264625, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "ad55306bd5009a68d54f0826fae877dd5e308495c214828c8c29171f8dbcf791", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006236958, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005842459, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005734708, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "a7232062d8882ccafef8351ec9efe45b19117d3813d6d7affc8406c27d722d23", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001107208, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001091791, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001074, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "a32f02d6fe4e0047433c039fc7b89378a4459a2d1f72c90a59e01a2a44dc0fb7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001735, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00106925, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001076708, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "6d33d2157dc8dd035396b2c8dc05cffb663898b137a9b6c6a709a80f014744f4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001150083, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001078458, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001102667, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "ff5acbde8f87a9cffec33af6b3febed903f863c0323e1ca368e00af80b00341e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001055084, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000788666, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000783709, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "63998c27b9f625d7a8709456613e85797736ad735c68d9cb2fa6774a3c7201bd", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002097958, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002132708, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001702625, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "8f9487016951f8bad5a45578dc2d68b7c4358ead8f3ca32e47d1d878604fbfe6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002761, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001755041, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001721917, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "e8ecaac7f6f984b95f1b16e4a98825be94d869723b6e991953965c9d1fee7e41", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0021245, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002126834, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002088958, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "5240081232417c062e9cd7b2a0f65c30bb4a01e5d5893139f51e2780afa02991", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002094542, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002335125, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002499667, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "4b08cb0b7cb0b2314fd002b3b9fcd3a965de0cb81fab44eecb0139ecc1895c0b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002472208, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00277825, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003083834, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "c3f137f65197afe7d7151ec424536425ce6e3df3b7d2f18017716ea30229acd4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00289075, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002337417, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002524916, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "dc4334c7913dae2d735f18962c5cbb0899ba2350bd08633ccbe4aebb97e735d7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004066417, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002898542, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002306459, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "414573fa005153a8fde54d2e6108e245c52fd30abc1a721db3e8cc8571ef6609", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002454, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003037458, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002995291, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "d4e3cea2e3eae1e083e90075c1d29644a7904f3bee2291f6866162d95521ad18", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003152166, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003326667, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002527458, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "7dfef442b32cffddae4f0b8f06f4699a2d0673103351a096450662eb98079d89", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002830125, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003000875, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003265875, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "a8313baf7a9d28473eb41853f144b48b299d3df252a860dfb17e4e76203f9422", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002992542, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002960417, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003356375, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "0f099dd9093f8c6cddf9d46ddf0ee55150e81c2b210559e5ef998c546399aa31", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002821583, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003133583, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00365425, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "e970ab6b5424b9f17bfc58851ca4c971ae07a062b9e2102e1c70596ae417343f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003807167, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003056666, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004232708, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "4810733db1c950ff1ab52378093c1ae89ba15c3f279e5d483d4f1b0331b47f36", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004290083, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004089959, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003311709, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "42400340f3a4197f623fa1e29f5c73007a1dc1d6e174872d4f4cb3238746787b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005309791, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004660833, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005403583, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "4f3c2061d4989e9d4944b121bf8f676c4a09122878444c2832ec6bb6977df647", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004167916, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003963292, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003555958, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "2e1d73048aac8c6091b54993d62daab5695827f669b7b4a01c4ed70279028810", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004878125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005089083, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004118708, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "dd84a19dc6a270251685e84831d686cff39572272ffdfd13bc53a3577f22550a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004099708, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004871375, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00392825, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "3d5602b98786d0e7a0b2207c8254e89e04ab5237e90ccb36183a9e422444a44c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005191125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004567125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004881083, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "53760f547079be6901b10b429386634aad74dfdbf58d3723c60421857faef5aa", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004024791, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003938458, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003168375, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "7fe3844ba82d5aa333028d72dcb0685b14256cde46201836f0d3357862bdec4a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005234875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005285292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005317458, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "c9a4e016f1af8d8041489f00db0a8b72c8f64913326a149fffd103781842c15f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003762875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004817, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004503792, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "10089d51448c3d21a71d278f0905275cea5c0edc1b9a445b093e28e4acf6d729", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00812925, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008100667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007000917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "e37385341b235f3d5acdaef5e763de8667f2d5ed8e57ad948a2956047b6b272a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006731583, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006342917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006913667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "dc30e246e083052e7bfbd856e67acc2d0ef2591dd1decf0e4344664e8cec598e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00955325, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007540958, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007328875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "299b5582cb3505f4aad3f7e8fc1e2890d201033992438051eefd61d7b52f7992", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00729125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006814459, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007460375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "74434f854226e86dfc755068eaac2cb747bf702399c1955bede51656ff1adb61", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018649333, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.016837541, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015676292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "2a64e385e10eacf964d73997ef70e24cbcc50dc309ed03a89f15089d0840969e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.029619666, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024506292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023921208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "05bf0e1051a349217e4726cc70dcde6d8af8ff16abfc8e8ec09222d0b9227f7d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018509458, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.016520208, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015131667, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "9c2ba19ad94244198477217e283001aa13f4a24f037dc76a351465fbdebb8e31", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.026623917, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024225833, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023671917, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "0285096cc119c590d845a3e7d83ae993574247bcb1672eb62cd74d6db69f1c58", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.575e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.1167e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.3375e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "7ffdc3e2655f41682117931ca6bc3d0444dfd610d3465d9d3f11d0cdcc621aa2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.0334e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.6416e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.1125e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "51b65de50d7fc43719c14a7cc961dec81e5f88b6d5e24e5880d20ea2e94e7c33", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 1.9291e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.725e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 2.1083e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "4459aecc512b0712bb137d3e2118389da96393c2633302d90d3de945c99eb0db", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.1833e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.2709e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 2.7041e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "b8d7b74c86717e11e42c397d392db9b03961fabf3150bceb5f482fd659f63bbb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001006458, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001398417, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001078458, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "4de83e48bc9c5741f1694eeca0d488b4b9ef4d30aecce06a2347d03d48c1219f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000955875, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000927042, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000918833, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "0f0085a118d8333ae90c9f04112b2dd13f1c4c7ec3d3c680188fb6d08d65a60b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001027833, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001019916, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0011035, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "ee95ffde3ec6c5c82b8d3fcb68ed8699ee084e45076d78d76c2362e6f8d8bded", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001002584, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000748666, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000755625, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "13621363c2314d1d28f59d2f3642fec6d26628025931e286ff25a2347c0ef056", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006100625, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0055295, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005975167, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "ce12a2e8dfdc53dae7eb953088adfde53f0a39e3e3a6eaba9b62bb856af94a65", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006401417, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005905666, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005116333, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "cfc4a7b7e3a61edd8884cfb4f9feaf25c621d4c4bb2ff277d47d12aa5fe00801", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006200792, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007103084, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006171792, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "362499822c8d1fe8287d24e406ff92d73a65c887263bc8ef7654dcd9e162b92e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007263125, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006440625, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00654175, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "8f41f6482bc53f936a5b90d42a66dc23864abccefffaf10596ff9c6abf2d7d54", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000672125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0006745, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000761833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "6605387327369233ae3b917159c8875b3fc4be8633870e473faf3e980d5352a2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0006725, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000673292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000774417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "f59bcd321ed61649346ca6b5328ed1ba823b24185cec27540587cf3a2f5b4190", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005007166, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005504667, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005893708, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 1, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round1-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "6cceecb57bcfef3d2da283035f5cebcfdbb9f729f8f26c1eb948ec14a94fb88b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006203458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005468583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005842334, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "ab53a7b31d82adc44736232ff167a04eb27e65534dc2fcdf6a9ff4c315406bfc", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.022352042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.01836975, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.016402791, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "a07c759ae4ef53bc559ddd636f621b0b488a8612bdf9f350a79e6608fb2eb239", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.029155417, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.025715667, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.024419334, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "3452153ea0102a641d9c700b1626ae002f227b4c745058d0a3374e3fdaf121f3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014393792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.014754125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.014957083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "c5484b0472064d130c92dfe75a15f00d391db35ccf945876a58acc2623e7a78f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.026776375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0246195, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.02575525, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "bf047e1a97c2505ec3652b68de8b8bc421588f1763446c723e39b6faf23b5260", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00760225, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00764975, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007518833, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "efd7f25a85faf1e3f92e8b7c42014191d9eb49ee9a7dd0a752d1e80922f6e1bb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.012667542, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.012396084, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.012757375, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "f97ecd47935c9962329b4c2f508063adf4f9475dfd10afdb0c4ba1d46ddaab7a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.007992042, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003998084, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003969083, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "94aef8d8fa46f2523fe29202f70327b8eda59828dd0221d491eeb53bff7d165f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.008047084, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005512084, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004787292, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "4da3a85b5c00ffc5b95bc7f6428cba9c2eef7e63674ce85767d1e5ccf9400048", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008376291, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005144167, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004214084, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "1b1209bc7f13ecbfb1494c73d0ee362e6b73bc8b50a736537246909006e58d91", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00690425, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006321375, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.00543175, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "20717514a421bc478ce5592b36e7b0cca8af60c0bb7d66c156e542f721d3a1f7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000823833, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001062291, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00104525, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "6ab8316f3c0b38ee57a498367751e1c99f3db4b02792bf24797e182d1dc79b16", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000928791, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000939042, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000908625, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "500d7b1c119c942cd95be3ca622752b421a1e2f9b43420b548c739806ce27912", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001102292, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001164166, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0011425, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "0dfa71a5d3fe8615e7e58721c9a866fcd0006892a74c7ab58bed341dc5ce9427", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000953667, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000972666, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000941166, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "00f233d84a9720a0b094c8e77b4338177cd570181f5f7b3bde38979ec4d4370f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001499375, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001831209, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002091041, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "8f28d02ba115818f57a7ed14a576b7734a177852ef0640e5efba8d51ed462c59", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00240225, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001969708, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001941042, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "98162dfe638764cad9c86fb817fb5defe83b56eb2d62574a985ff44eba063104", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002065417, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002074125, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001712875, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "28670ad8ca5f20b0878bddbe9cadf5854b797b1da59f95142d983f36920299bd", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001979458, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002018958, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001434958, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "3255e0b441e198be4609dc15194aec52d2da18032969d5f383bb0ba7c8911085", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003220416, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002383625, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002963792, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "374116f94dee45cd0a5d7b6e202e76f215df113c30b1956de7f237eb3f3031d6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001988667, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003003625, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002939292, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "ccc7a800f2d2f47cff686c7c0e64fd235b58561a203918babf3a16a492ce4668", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003492208, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003178291, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0026165, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "c68aa4c369cbfd2e2832c6ece31d4fc5830ad61bed5f7f10ee491519d2933322", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0030105, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002643583, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002536125, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "bbfa99916a55b4478f1e0a7fecec53de2d6c6350dfc446ca7ee8f3973311de71", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002529375, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003111667, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003378958, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "ba91a05b7b925e3e52972942d6b346c6490f1520e83e92c23101195c279a5b23", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003350041, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003689625, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00262025, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "f2e0a28b02ada1d33f2f6b5a57b9e924689bcc553c3a20e5d8e666859fd0b7bb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002929417, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003571292, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003366625, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "d22fa9643556424fb56f0a79decc576ce376e98bb2c14b55ac66439a5708d322", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003653542, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002819375, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003964709, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "2af866fa9ce4087327dbe7e8a373fa89b29a2e7ae93f28653a19d491ae9dc373", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004539708, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004591875, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005131375, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "ae93f20d06d2f5f51ddb7db4f6debbeae6edae537bd4ef3f05e350a9291b8e16", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003828333, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003610125, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003202541, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "91172687fa437c41e0ec50c40f594f68f1834231b54c38edf87f16904c6ab7eb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005266125, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005132042, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004677292, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "928dfbd89c6221f3324e1c06c5a5a2f8bf8694c5f838911174914e609bada3c0", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00415175, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00426275, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005091875, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "e91b2b32be477c7c7712120ec66469bcd20afd0764015cbca32c9d293997a524", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005069208, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004412875, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005306, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "af24bc1c2b1c41e0a396b884312c4c1d105c9b84d7de567abcedf66386ff850f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004984166, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00386425, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005503709, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "ef15f98c032a2ef97f72f3a1c80466034b9ae41fb7cfe22b2af72526e32ab9b5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0052125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00474425, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005707125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "bbd3412f1d196ecda1bdee05f4503d006d4791ba613dbe9f86b7f9950bf039b2", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006377333, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004096209, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005263084, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "8a18084cd0cff16abf04d9e5d0c91d41825c12dd0ef81555b15409be22bb920d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005058542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005486792, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005393625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "8ef19fd57df4b7c4435d0f3d233b1dfd6a336282172ad76dbf732a8fdc689567", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006501, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0056095, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005406583, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "cc9c142bab9eed356fc8e1175669e41e52e5a00185dd7e96fb03077f840bd5a4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.009734417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006465875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005404834, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "a57f61a55b6ba12a904f5b46edf0458be569edd66b994874f345da501abeee00", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007571709, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00601, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006402875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "a8f3839103b606fc1cbb236e9c8093d4d0a685454924901c591589637f97e59f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.009444917, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00775625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007051125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "fdd44b0b4d084d9a112448b9eb5f4e231f715afa011987842b517eabd19756bd", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008048625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0059945, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004981667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "34f3d951a52c9ea11988b045c8d4783345b545512c8671f9b15b140a01169b7e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.017800708, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015389, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.015346042, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "5c731c58c51002a8366c7932ef8103b3681cd0e8322bf3b3761ed04331e9e33b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.023891083, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024607208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.02369575, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "63e5979afc5a488429c12edede3fbbe4aa43b483eb33f6c559799f1e53257fe7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.020994083, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0208325, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.019623166, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "3e626e8a60ca77920ec669db37fcb7bfcce5a0322454f27a96499e37eef079ee", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.024088542, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024945083, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023508583, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "c52eb3686d67f7b3cbd09aada2d5f471d9a5d4bab424fdcf42c2356c264c253d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.2959e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.1041e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 7.4167e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "9b1bfc70c74b1c6c8fbb4611396f72d9f6abf07d4fe72d01ff4d85898e8dbebb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.6625e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.6e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 1.7875e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "d2c2041b490d2040d036cf097057d976ee4dc2f9be58a2fa14735aa2a44faf16", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.0333e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.7375e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 1.975e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "37dab4963324e91b73990fccf21c917b3fbcbe51a0fdcae17f0a22c09f41ad0d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 2.5917e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "0809bf4dd8ca8307dfaa993139ff22cdf9dbc6ce26ba2f2ac6a53b26bd2ca48e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0009385, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000929208, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000665208, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "1937ae93e0abbf129755f3a163f6e522f365e3605a48d3f12ed83e4606dedfd4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000940125, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001529209, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000814792, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "c09a62693a025cc0185e763164683b6dc6551974a9b62ebf8511a31d4c5716cb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000508292, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000460167, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000494542, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "47be3180007a1efdcb6ec67be9c5d7bf5fd7b4493c3c35f70a86283145c2fa85", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.000988458, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000889917, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000759834, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "d73ec86b61dfd2de3c37f3a935229b9cbc2308cb0dc3c0e112662ebb18778bdb", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005238416, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005937833, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005362792, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "812690b5e65e5834c56e01e43c04950dc55b6a090965ef3c22aea25075f1aa5b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005386875, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005758958, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005110708, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "2aaf5882dbaa7bccd9d5307e02cfb42fa2648c74f1978e012d2c779d7df9bb45", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00761275, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006340833, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005950791, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "2370aa28dc9477b07b808067685fd79affe08b96d18c9f52a98499817c933968", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006169375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006006, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006893834, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "bbbdfd8c6bb42c54c4c6c09f8ae2b7f58b8187a246fc0e671e8416b51c663638", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000671334, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000670916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000670875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "9fad3484bf9d85e760b31bd6c4f0e148b48982fa11efcb5afc744a35f784a4d9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000733333, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000733125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000725167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "e8fa6446d985aadef986fb51c4f7431662618fa9a100854746cfec231c91412c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00594875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004687833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004831083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "5bb445a3e49accf02ac576cdef80a842e98b04f82b936a79cdbd1fc088007c4c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005952959, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.0052745, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004422583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "be66a24c51e4eed91f72587987cfbbb64417e71e0c6cf25371626a0ecf14fa4a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005410375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005615833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004597917, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 2, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round2-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "c48ebe51458605a7e250c505af2cc8ac184be3db631ac98745a987a948f5d1e8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006792709, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006377542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005300584, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "a4de32979d92e4925ddf0d74488f08bf154b5b6996478da8b7383c73f7f009a4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.026705625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.023925666, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.024433333, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-long-b8-w32-l2.stdout", "capture_sha256": "54f97cff16e155d062c0aec36bc54c5391313bf800ca3a6599e6d0cee5038794", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014990375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.01504425, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.014451583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "actual_new_tokens_per_row": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 512, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "8e372b5dd4f105f9df0a72776b1fc85d3b653036d9605053a168d98dcbb3026a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.013114166, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.013139708, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.012244208, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 207712, "device_peak_increment_bytes": 191136, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-ragged-b8-w32-l2.stdout", "capture_sha256": "b41f66fa0111f65587e353639a87d7963f79839fae7135df5c52ba6569cf0fa6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008939667, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.009099791, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.007907417, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "actual_new_tokens_per_row": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "c0ae6d0121bcac37fc57fd7803a7a97641f18fc349554ec331d6e3f1d7ff1b26", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.009345333, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.006301334, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004773291, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 73868, "device_peak_increment_bytes": 71796, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b1-w128-l3.stdout", "capture_sha256": "1b4d89d06bb416892b376b921c8c1b0f507cce2ce946941219648dd21f0ee244", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0085245, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004095041, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005261208, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "63a97ea0c3afcf31680c3bdb79c6f21ef91caf280e8c61d543842890f9f10401", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008742584, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00648025, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004986125, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 615624, "device_peak_increment_bytes": 599048, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b8-w128-l3.stdout", "capture_sha256": "95e105aafb72838855639cd9a2fba27c2b7cd063e9c2a72005ee353e9c3c7891", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006697166, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005048625, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004151833, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "96913c3293974a6a3d30fb2aa14c6e7ba7f8bc73917c7c05418af19a97a5f033", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001001125, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000992917, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000762292, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos1-b1-w32-l2.stdout", "capture_sha256": "ff1f43aa353abe061978008632013c593b2c30664ae8ffdef8d41938302e26df", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001094917, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001176917, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001052542, "output_ids": [ [ 257 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "7df04a47990f7becba299bf9cc5b507c4e0ff892aec42aa3bb7ddb16375d5b7d", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001018334, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001017417, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001041833, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos1-b8-w32-l2.stdout", "capture_sha256": "0c4d3978ed2a78c20ba790f683593ef7d7c56b52daf640fb1cb17eead9433843", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001148958, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001120292, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001027, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "2855524b358fce79db034a00e5f8f764f0aed3b557e997a586cb4792ca01a4a6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0026005, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001951208, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00161825, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos2-b1-w32-l2.stdout", "capture_sha256": "f2478549a6add539da03316febe3f6662ec33703bd957d384fcddc16b1e4d0c9", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002042709, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002016042, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001444875, "output_ids": [ [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 2, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "bc35e6b9f7d3ac1cd426e8a4a794d3a2efe0c9dfa7cd71f6bbb2495252e86350", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002130667, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001778416, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.0017075, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos2-b8-w32-l2.stdout", "capture_sha256": "c2551415b59d732f5d89ff03e4e3f5efe88a1799f9da52a3f6e902be66dee027", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002201292, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001753292, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001766875, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "actual_new_tokens_per_row": [ 2, 2, 2, 2, 2, 2, 2, 2 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 16, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "938476851c3925d609478d97327a76b49bcbd26cdd6e3fe44cd53654a2c9e99a", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002513208, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002546959, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.002848208, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos3-b1-w32-l2.stdout", "capture_sha256": "80a18def298cd87d54c841e34b0bec56523579a4a8c6b07dbdd476588997cb96", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002773041, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003034334, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.0027, "output_ids": [ [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 3, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "e94af6e39076134fbb9957ea59dda7faa8c3ba1312b79fb3c338f06f265ada90", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002523917, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002869583, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003056125, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos3-b8-w32-l2.stdout", "capture_sha256": "f8e129a973806005bb349e90855cf59fac0d16c8ff291c536bb91620dcb913f4", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003428042, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.002776959, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.002599708, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 3, 3, 3, 3, 3, 3, 3, 3 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 24, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "55eedfb9f149710415990304a4062a4fd02474ff2a70ab5ed3495d3b5d4044b3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003777709, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003106, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003046666, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos4-b1-w32-l2.stdout", "capture_sha256": "5ce1f2b5cb3a9df0abfecba0eb2a32ed83e9788ac85a54a17ef0c06a17ae415e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003124, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002661291, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003141209, "output_ids": [ [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 4, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "709d7229b3d4ebe21844acd2d1ba1d32cf5e341aae24014de304e49fcd144f63", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004266083, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0028275, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003636042, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos4-b8-w32-l2.stdout", "capture_sha256": "2c7708e73dd5851705ba6138f3de321bd788d9297271fc774fba567991fa8a07", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00270925, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003055333, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.003322583, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 4, 4, 4, 4, 4, 4, 4, 4 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 32, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "d95c87506a853e0f47a76ab051852075776f1a8edb33fc9f8af1a7d854bc8ff7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003862042, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004395584, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003217417, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos5-b1-w32-l2.stdout", "capture_sha256": "1d4aca61c0cc4fba7edc3d207cbbb87546c32510c4218b0d9ceda7ddd2e09638", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003989917, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005197833, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003980375, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 5, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "6e8937d65643ceca3cb3394974f248c3607e91997e88abed778caf66bc4eefb7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004472959, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003804459, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004096709, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos5-b8-w32-l2.stdout", "capture_sha256": "48bcbbb4d39721a09e1188fdd984d839d8e9eeb022e06f922705cf13e47ae572", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00501775, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004960791, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004573416, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 5, 5, 5, 5, 5, 5, 5, 5 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 40, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "3400f575c53b3c8a1cc9c6a11100d041dff1e2070de3d885e6b749a8d25116e5", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004371042, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005542834, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003866209, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos6-b1-w32-l2.stdout", "capture_sha256": "6f42a6348af558671bb420fd310f348b07a243a6ca0fc78f1499359ac23be640", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004768958, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00465575, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004822667, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6 ], "ended_eos_per_row": [ 1 ], "actual_total_new_tokens": 6, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "2d21e3cee622ef9154c72e4e472b4f587d5e9259d5b590320f8d96df1439f361", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005457958, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005060583, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004000792, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos6-b8-w32-l2.stdout", "capture_sha256": "bb4efabd0357c1c3c76526b7978fa3b6825320bc5b7f0c020dc5191a11839498", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005409083, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005343875, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005357334, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 6, 6, 6, 6, 6, 6, 6, 6 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 48, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "8fc9ae4ffa18bb92faadd60ff3cf287c460c59a9a44765670476817a91ab8d02", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005050667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00446, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004104083, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos8-b8-w32-l2.stdout", "capture_sha256": "85b6bb89c87bde93b8e407010e759199f79d62e0d8d93b2f217a78e3c209aa2c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005209416, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006154042, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006047375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "db74c9fb263de9b216caa77b2b38785dbf7d4dc6bb31a12ced49c5113819265e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006779333, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007173416, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005620875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos9-b8-w32-l2.stdout", "capture_sha256": "f93d7e1b396626086a4a57a83d957567119525fe4dddd3aa800a7ffbc60017ce", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007864875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007308958, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006467584, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 72, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "a20b7efbe97b7dab4c7344faacf51f5a98cda71fb32ef58eba6327b287a97c4e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00645625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005314417, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004737166, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 365636, "device_peak_increment_bytes": 349060, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-eos10-b8-w32-l2.stdout", "capture_sha256": "2c1046fc9858446f53af6d411c71b6c3cd5dbd00331516f5477dc1470f4ebda1", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006345334, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005195458, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005795542, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "actual_new_tokens_per_row": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "ended_eos_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 80, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "4cf8ad31d5bea8b7db8c00a649326049cfc3580814f270dfd3c5b4275bd5b92e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.027538291, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.024952792, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.02329275, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 717980, "device_peak_increment_bytes": 701404, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-mixed_eos-b8-w32-l2.stdout", "capture_sha256": "d7a0b83b60f14129adf10710cb9849a7a26680c840a5aca49141891c4f37df72", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014340083, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.014603208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.013973292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "actual_new_tokens_per_row": [ 64, 9, 8, 7, 6, 5, 3, 1 ], "ended_eos_per_row": [ 0, 1, 1, 1, 1, 1, 1, 1 ], "actual_total_new_tokens": 103, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "1455c5611ffaca86df1fc9fc239384cc0bd3817a9d17a721f79047133a691b6f", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.025157458, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0226045, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.023227708, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 82340, "device_peak_increment_bytes": 71980, "host_peak_increment_bytes": 10360, "readback_peak_increment_bytes": 10360 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-zero_mixed-b8-w32-l2.stdout", "capture_sha256": "66c100ba43d7ac2552cd33deff89c84ea7325d1a830b3884402d10db977fd0c8", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.015173417, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015123333, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.01566275, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "actual_new_tokens_per_row": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 90, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 72564, "device_peak_increment_bytes": 72564, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "8c6b50b2018e72ac1cc599b70d3d99dbdd54ff9640ecc11bc7aaedf7c4be1e37", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 1.7041e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 1.725e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.7291e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-all_zero-b1-w32-l2.stdout", "capture_sha256": "3d80c9c012d509848a278626d81c2fa7a789f4b0a152402a5801b62127e0f8b7", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.4208e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 2.2792e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 2.6333e-05, "output_ids": [ [] ], "actual_new_tokens_per_row": [ 0 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "7b178696c001a32cdff4d345d1c402189188d8ce02c4dd20b371f529ce6e7c90", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.5e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.8958e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 1.925e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-all_zero-b8-w32-l2.stdout", "capture_sha256": "4df454175810a86803b1d1edc35edacdadc677eadb6d870e28648cb70e829fbe", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 2.7875e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 1.5708e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 2.1292e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "actual_new_tokens_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 0, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "811d99f69adbf327ef959a79aef874d3ee0afc3139da31a719a262be977b144b", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000954792, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000952208, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.00065875, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-one-b1-w32-l2.stdout", "capture_sha256": "43598d029f179c0f6a37e5a87d1113f12b33577b197103b01fce247233b78c79", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000786958, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000684, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.001036458, "output_ids": [ [ 232 ] ], "actual_new_tokens_per_row": [ 1 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 1, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 4408, "device_peak_increment_bytes": 2336, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "bfd4a9e9d835d3ce67420840cf07aa314e6f1f77aa2cf179485b3c526f45fb36", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001001875, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000960875, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.001137375, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-one-b8-w32-l2.stdout", "capture_sha256": "a9ff01b0cd9d6cea8c2ee0c155d52e01359da491faaa165c420331d069113ac6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001811209, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000739958, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.000991625, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "actual_new_tokens_per_row": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 35236, "device_peak_increment_bytes": 18660, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "cda6a8b478c227d0daa55cee8a8c629e59a501075737f8dc073eb006ed09f842", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005668084, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003234708, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.003515542, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b1-w32-l2-sample.stdout", "capture_sha256": "0c9f9a072d625c90e9b09fb4d5f703bb13584b0c59c5954f8cc37fc861004c5c", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00595625, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005067084, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.004724333, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "c7700815e20043da7c2ae64b0ce70275571e1ed595af1485115b0fa365fd28b3", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006305, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006763375, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.006152458, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "role": "sampling_control" }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b8-w32-l2-sample.stdout", "capture_sha256": "23c85f464732d882de316c53947dfd2cd0ce7fa4a6f2cf1f0a9cb86024f17cb6", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006601083, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006437917, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0.7, "top_k": 8, "top_p": 0.95, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.005537334, "output_ids": [ [ 42, 42, 85, 211, 92, 101, 211, 232 ], [ 61, 165, 243, 251, 42, 165, 42, 157 ], [ 250, 165, 211, 232, 157, 211, 101, 165 ], [ 218, 165, 42, 232, 232, 157, 232, 165 ], [ 165, 42, 157, 42, 232, 232, 114, 211 ], [ 114, 232, 114, 211, 232, 211, 211, 211 ], [ 250, 103, 35, 42, 82, 211, 211, 218 ], [ 232, 42, 35, 232, 232, 211, 232, 42 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "965c3d8bc72da74481a4dee7d4fee9fe788d46fa623bca06e3a3f13be82c6133", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000705292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000673833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.0007715, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "cpu", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-cpu-tiny-b1-w32-l2.stdout", "capture_sha256": "821fedc94240cc4e3de04273afd24cdaca9f663c482eaa808b28bfcec8a45831", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000688334, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000676167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "cpu", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.000743291, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 169088, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 169088, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "1c98cd17e2d1397468fb843074a091562e8132073da96136f231d3db4de350fc", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0072655, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.006128833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005667917, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 17036, "device_peak_increment_bytes": 14964, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b1-w32-l2.stdout", "capture_sha256": "ec3b19159d5b92536e9cb1c4daf9579cc553477e71d25ec0cd413458d6631abf", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005458583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004892375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.005820833, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "actual_new_tokens_per_row": [ 8 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 8, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "eeab9d12072f61c654eed4b2a8ea30c7ff0ceb5cde3406efb9c0eb70007b195e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003464083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005097125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004057875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 136260, "device_peak_increment_bytes": 119684, "host_peak_increment_bytes": 16576, "readback_peak_increment_bytes": 16576 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-tiny-b8-w32-l2.stdout", "capture_sha256": "9dcc151e0ca15ed81d6f79c5dd9f3aaa8a9b8f4a2a1035cb161834be266cfc9e", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005155542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005763167, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 2, "seconds": 0.004351917, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "actual_new_tokens_per_row": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "ended_eos_per_row": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "baseline", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-baseline-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "8e227e0f39d6342b632fc3e7b1dd4cf297f59f7e01e5ac35a01231beeddfad75", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.026317583, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.023764916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.023720625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 45708, "device_peak_increment_bytes": 43636, "host_peak_increment_bytes": 2072, "readback_peak_increment_bytes": 2072 }, "statistics": { "available": false } } ] }, { "outer": 3, "implementation": "candidate", "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "phase": "natural", "capture": "round3-candidate-natural-mps-long-b1-w32-l2.stdout", "capture_sha256": "444aeae1ce2b7ddedb4b9c7ff37f766d816e47dbf749a41f6f33b289593e7561", "records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.014453125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.014049625, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "natural", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 2, "seconds": 0.013493916, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "actual_new_tokens_per_row": [ 64 ], "ended_eos_per_row": [ 0 ], "actual_total_new_tokens": 64, "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": false } } ] } ] }, "diagnostics": { "schema": "aner.metal-greedy-diagnostics/1", "timing_quotable": false, "records": [ { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006400542, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005412834, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004777541, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003798375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00564, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005158959, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005502375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00636125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 15044, "device_peak_increment_bytes": 15044, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 5.223554712113906e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a867f4048d846995e5a150bb387cb5b9b3de476e533b5c342e943221dc02c3e8" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "352093e50054a2f247ad56bd2e4d5c1dade894be3ab824fd0c0a9991d0fd07ba" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007876292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005892875, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00586025, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006010917, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005252, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00511575, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004614458, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00322925, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 120296, "device_peak_increment_bytes": 120296, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 211, 92 ], [ 232, 211, 232, 211, 211, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 218, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 42, 211, 218 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.776524883542186e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "48eeff223a6da4669c31a906a6552dc97e58f3a6eba9cbe5fe5d4891075934ab" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "48a37c497405b6775633be24182cf52022a24b94f0ee6591d3e85add0484db12" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.031969292, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.03239075, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0170365, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.015673333, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.019227792, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.017385083, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.017985708, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.015404042, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions": [ 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ] ], "positions_before_probe": [ 71 ], "max_scaled_logit_error": 5.853771208176165e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "32dff5d35f39e7e46963f7ec1ecdc65842ffd24b4094beeea7871086e343a81e" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "cd02a323c59bf4a95c094d9094139be7b9a6a91d89c4c4488ba96c98334a18a8" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.023231791, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.02222825, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.03918375, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.029766125, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014328708, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.014183416, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "long", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.017579166, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.015607208, "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "long", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 232, 232, 152, 92, 152, 218, 232, 232, 232, 232, 92, 211, 218, 232, 232, 152, 232, 232, 218, 19, 218, 211, 211, 211, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 24, 24, 103, 24, 103, 24, 24, 24, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 211, 92, 92, 211, 211, 67, 211, 211, 211, 211, 24, 103, 24, 24, 103, 24, 103, 211, 24, 103, 103, 103, 211, 24, 103, 24, 24, 103, 24, 103, 24, 103, 103, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 103, 24, 103, 24, 24, 24, 24, 24, 24, 103, 24, 24, 24, 38, 24 ], [ 232, 211, 232, 211, 211, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 218, 218, 218, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 38, 38, 38, 38, 38, 61, 24, 24, 103, 24, 24, 24, 103, 24, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 218, 232, 218, 92, 218, 103, 211, 211, 211, 218, 19, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 61, 103, 103, 103, 24, 103, 24, 132, 24, 132, 24, 38, 38, 24, 24, 132, 24, 24, 24, 38, 24, 132, 24, 132, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 92, 211, 67, 211, 246, 178, 246, 10, 10, 246, 178, 246, 178, 67, 178, 67, 178, 178, 178, 178, 67, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 103, 103, 103, 24, 103, 24 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 157, 157, 157, 178, 67, 211, 211, 218, 67, 246, 246, 246, 178, 67, 178, 246, 178, 246, 67, 178, 246, 246, 10, 246, 246, 246, 246, 250, 246, 250, 250, 246, 250, 250, 250, 250, 246, 246, 178, 246, 250, 250, 246, 250, 250, 246, 178, 246, 250, 250, 246, 178, 246 ], [ 232, 232, 232, 232, 232, 42, 211, 218, 218, 246, 211, 211, 211, 211, 211, 211, 211, 211, 218, 103, 103, 61, 103, 211, 211, 211, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 103, 103, 24, 24, 103, 103, 24, 103, 103, 24, 103, 24, 24, 24, 103, 24, 24, 38, 38, 38 ] ], "positions_before_probe": [ 71, 71, 71, 71, 71, 71, 71, 71 ], "max_scaled_logit_error": 6.877026435809321e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 2048, "greedy_emitted_tokens": 512 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "5e7667ba3b4618f666696f0586e7027cca786d7a8d8e7a878a3c96058035aa27" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "0f036ae70db15961934702197a318cc4988bedc7626d204902646ff0edb9df50" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.022621083, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 32, "status_reads": 32, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 32, "greedy_active_reads": 32, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.018747458, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 32, "status_reads": 32, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 32, "greedy_active_reads": 32, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 32, "status_reads": 32, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 32, "greedy_active_reads": 32, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014107292, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 13, "status_reads": 13, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 13, "greedy_active_reads": 13, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.011235833, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 13, "status_reads": 13, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 13, "greedy_active_reads": 13, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 13, "status_reads": 13, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 13, "greedy_active_reads": 13, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008376208, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 7, "status_reads": 7, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008554333, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 7, "status_reads": 7, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 7, "status_reads": 7, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.014503291, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 6, "status_reads": 6, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.012226584, "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 191852, "device_peak_increment_bytes": 191852, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 6, "status_reads": 6, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "ragged", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 1, 3, 7, 31, 2, 5, 13, 47 ], "budgets": [ 32, 1, 16, 2, 24, 3, 8, 4 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232, 218, 232, 103, 212, 103, 212, 103, 103, 103, 24, 103, 24, 38, 24, 103, 103, 24, 103, 103, 24, 103, 24, 103, 24, 24, 103, 24, 38, 24, 38, 24 ], [ 232 ], [ 232, 232, 232, 232, 232, 92, 218, 218, 232, 232, 92, 211, 218, 211, 19, 19 ], [ 211, 211 ], [ 189, 189, 212, 232, 232, 232, 103, 103, 103, 103, 103, 24, 24, 38, 24, 103, 24, 24, 24, 24, 24, 103, 24, 24 ], [ 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232 ], [ 211, 211, 211, 211 ] ], "positions_before_probe": [ 32, 3, 22, 32, 25, 7, 20, 50 ], "max_scaled_logit_error": 6.861486816488932e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 32, "command_buffers": 6, "status_reads": 6, "gemm_calls": 288, "compute_encoders": 607, "packed_output_rows": 256, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "8924da54cc6e32e3b0a6846fd8c01405e2c56f78906d2f457a17cf127353f790" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "d87ea65d78ce96f82863a7006e592916a6013c965896d970effa725f38ed9c79" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.008905042, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005775417, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.011383334, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00523575, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.009132416, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00527525, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.011098583, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.009424417, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions": [ 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 71876, "device_peak_increment_bytes": 71876, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 1, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 8 ], "row_seeds": [ 7301 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ] ], "positions_before_probe": [ 15 ], "max_scaled_logit_error": 2.033690377542414e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "ae850dbd3a73724fae5aa402988ee15b01f3fde2e3cc20d22ca28559945c2d83" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "1feb2dacb8d0c0de039e862951615ebe3cb69c5aab46b861b92aceecc71abd4f" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.011828833, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005150958, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004725834, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005094708, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 5, "status_reads": 5, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.009814833, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006310959, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.010176458, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.004195416, "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 574952, "device_peak_increment_bytes": 574952, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "tiny", "batch_size": 8, "width": 128, "layers": 3, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 134, 134, 134, 23, 23, 23, 23, 23 ], [ 134, 134, 23, 23, 23, 23, 23, 23 ], [ 134, 23, 23, 23, 23, 23, 23, 23 ], [ 129, 23, 23, 23, 23, 23, 23, 23 ], [ 205, 205, 205, 205, 205, 205, 23, 23 ], [ 23, 23, 23, 23, 23, 23, 23, 23 ], [ 175, 175, 205, 205, 23, 23, 23, 23 ], [ 205, 205, 205, 23, 23, 211, 23, 211 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 2.3617566891512637e-07, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 104, "compute_encoders": 207, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "822537d6942d5d8921f8d1cf1ca12fc607596cd9f99524cbfe954e5e98f5061f" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "e6960d6519c79666660e8a4e5354526942713bbe3eb08e4561346eadbb974e0e" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000879875, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001472375, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001149625, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001086875, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001162417, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000852334, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00111175, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00080075, "output_ids": [ [ 257 ] ], "positions": [ 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 257 ] ], "positions_before_probe": [ 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001139458, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001039, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001181667, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001309083, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001196958, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.000907166, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001165041, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001224792, "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos1", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 7, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ], [ 257 ] ], "positions_before_probe": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.00213025, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002272083, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002109916, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002033, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001839583, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001058, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001597334, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00179875, "output_ids": [ [ 65, 257 ] ], "positions": [ 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 257 ] ], "positions_before_probe": [ 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 2, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 8, "greedy_emitted_tokens": 2 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002234666, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001540333, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002110917, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001746458, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002058417, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001538375, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002126, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00206875, "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos2", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 8, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ], [ 65, 257 ] ], "positions_before_probe": [ 9, 9, 9, 9, 9, 9, 9, 9 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 2, "command_buffers": 2, "status_reads": 2, "gemm_calls": 18, "compute_encoders": 37, "packed_output_rows": 16, "greedy_available": true, "greedy_chunks": 2, "greedy_active_reads": 2, "greedy_result_readbacks": 1, "greedy_id_bytes": 64, "greedy_emitted_tokens": 16 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003002833, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 3, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.002736334, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 3, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 257 ] ], "positions_before_probe": [ 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 3, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003549917, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003224333, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 257 ] ], "positions_before_probe": [ 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002840792, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003568875, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 257 ] ], "positions_before_probe": [ 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002820542, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00297025, "output_ids": [ [ 65, 65, 257 ] ], "positions": [ 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 257 ] ], "positions_before_probe": [ 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 12, "greedy_emitted_tokens": 3 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002886334, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 24, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0031185, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 24, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions_before_probe": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 3, "command_buffers": 3, "status_reads": 3, "gemm_calls": 27, "compute_encoders": 56, "packed_output_rows": 24, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003522791, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0034255, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions_before_probe": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003490292, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003601709, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions_before_probe": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002711542, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003427083, "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos3", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 9, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ], [ 65, 65, 257 ] ], "positions_before_probe": [ 10, 10, 10, 10, 10, 10, 10, 10 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 96, "greedy_emitted_tokens": 24 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.002802625, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00382375, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003360916, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003440209, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.003359417, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003511916, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004039209, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.003445291, "output_ids": [ [ 65, 65, 65, 257 ] ], "positions": [ 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 4, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 16, "greedy_emitted_tokens": 4 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004008125, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003787542, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 4, "status_reads": 4, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.002311084, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001699292, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003257708, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.003341375, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.003600667, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00293725, "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos4", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 10, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ], [ 65, 65, 65, 257 ] ], "positions_before_probe": [ 11, 11, 11, 11, 11, 11, 11, 11 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 4, "command_buffers": 3, "status_reads": 3, "gemm_calls": 36, "compute_encoders": 75, "packed_output_rows": 32, "greedy_available": true, "greedy_chunks": 3, "greedy_active_reads": 3, "greedy_result_readbacks": 1, "greedy_id_bytes": 128, "greedy_emitted_tokens": 32 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005028709, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 5, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004690708, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 5, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 5, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004631333, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004583625, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005377167, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004866583, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.004572292, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004284834, "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions": [ 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 20, "greedy_emitted_tokens": 5 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004981541, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 40, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0046075, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 40, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 5, "command_buffers": 5, "status_reads": 5, "gemm_calls": 45, "compute_encoders": 94, "packed_output_rows": 40, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004705291, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005125833, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004909333, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005212333, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005498334, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005205417, "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos5", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 11, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 12, 12, 12, 12, 12, 12, 12, 12 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 160, "greedy_emitted_tokens": 40 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005935125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 6, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004691709, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 6, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 6, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005471459, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004872958, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 7, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.006054959, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.004409083, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.005425, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.005416084, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 43940, "device_peak_increment_bytes": 43940, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8 ], "budgets": [ 64 ], "row_seeds": [ 7301 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 24, "greedy_emitted_tokens": 6 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "a533a25c75029f2344cbffaaae7531bbda0057b4c3d9f7ff0ffa481389371733" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "112f13ff1a5ee370919a3327f573b80ed2d8505ccc153aeaa25ebce3cea44977" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005998667, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 48, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006754416, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 48, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 6, "command_buffers": 6, "status_reads": 6, "gemm_calls": 54, "compute_encoders": 113, "packed_output_rows": 48, "greedy_available": true, "greedy_chunks": 6, "greedy_active_reads": 6, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.004846958, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005157125, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 7, "command_buffers": 4, "status_reads": 4, "gemm_calls": 63, "compute_encoders": 132, "packed_output_rows": 56, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005198333, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005407209, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006243167, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005846708, "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos6", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 12, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 13, 13, 13, 13, 13, 13, 13, 13 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 192, "greedy_emitted_tokens": 48 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006695959, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005920875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 8, "status_reads": 8, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 8, "greedy_active_reads": 8, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0071525, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006406292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.00609875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006034167, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.005781125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.005381875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos8", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 14, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 15, 15, 15, 15, 15, 15, 15, 15 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 8, "command_buffers": 4, "status_reads": 4, "gemm_calls": 72, "compute_encoders": 151, "packed_output_rows": 64, "greedy_available": true, "greedy_chunks": 4, "greedy_active_reads": 4, "greedy_result_readbacks": 1, "greedy_id_bytes": 256, "greedy_emitted_tokens": 64 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.009329458, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 9, "status_reads": 9, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006845208, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 9, "status_reads": 9, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 9, "command_buffers": 9, "status_reads": 9, "gemm_calls": 81, "compute_encoders": 170, "packed_output_rows": 72, "greedy_available": true, "greedy_chunks": 9, "greedy_active_reads": 9, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.007245375, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.006739416, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008957666, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.009083541, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008249292, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008862291, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos9", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 15, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 16, 16, 16, 16, 16, 16, 16, 16 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 288, "greedy_emitted_tokens": 72 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008591834, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 10, "status_reads": 10, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 10, "greedy_active_reads": 10, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.007720334, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 10, "status_reads": 10, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 10, "greedy_active_reads": 10, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 1, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 10, "status_reads": 10, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 10, "greedy_active_reads": 10, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.006381333, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00689625, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 3, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 10, "command_buffers": 5, "status_reads": 5, "gemm_calls": 90, "compute_encoders": 189, "packed_output_rows": 80, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.0093955, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.00908475, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 8, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.008697584, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.008820583, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 351464, "device_peak_increment_bytes": 351464, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "eos10", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 16, "chunk_steps": 32, "prompt_lengths": [ 8, 8, 8, 8, 8, 8, 8, 8 ], "budgets": [ 64, 64, 64, 64, 64, 64, 64, 64 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 257 ] ], "positions_before_probe": [ 17, 17, 17, 17, 17, 17, 17, 17 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 16, "command_buffers": 5, "status_reads": 5, "gemm_calls": 144, "compute_encoders": 303, "packed_output_rows": 128, "greedy_available": true, "greedy_chunks": 5, "greedy_active_reads": 5, "greedy_result_readbacks": 1, "greedy_id_bytes": 320, "greedy_emitted_tokens": 80 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 1, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.025868667, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 1, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.025382125, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 1, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 3, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.016981541, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 3, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.016343291, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 3, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018422666, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.017895833, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 8, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 32, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.019795458, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 32, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.016214875, "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 697388, "device_peak_increment_bytes": 697388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "mixed_eos", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": 64, "chunk_steps": 32, "prompt_lengths": [ 1, 57, 58, 59, 60, 61, 63, 65 ], "budgets": [ 64, 31, 31, 31, 31, 31, 31, 31 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65 ], [ 65, 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 65, 257 ], [ 65, 65, 65, 65, 257 ], [ 65, 65, 257 ], [ 257 ] ], "positions_before_probe": [ 64, 65, 65, 65, 65, 65, 65, 65 ], "max_scaled_logit_error": 6.06905886771679e-09, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 512, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 412, "greedy_emitted_tokens": 103 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "164a7da394fe794c9cae820690729c82f69caf1c58242b0059850d5ce6b64bec" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "fce5252db83f699c18da4abcb756b198e85255ef997b359f8ceab2092dcdf836" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.029360625, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.025469292, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 64, "status_reads": 64, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 64, "greedy_active_reads": 64, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.018385167, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.0184025, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 23, "status_reads": 23, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 23, "greedy_active_reads": 23, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.020406375, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.019204333, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 11, "status_reads": 11, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 11, "greedy_active_reads": 11, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.02248225, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.019564125, "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 92032, "device_peak_increment_bytes": 92032, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "zero_mixed", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 1, 8, 3, 8, 0, 31, 5 ], "budgets": [ 0, 1, 0, 8, 64, 1, 0, 16 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [ 232 ], [], [ 211, 212, 211, 212, 232, 232, 232, 232 ], [ 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 232, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 103, 24, 103, 103, 24, 24, 103, 24, 103, 103, 24, 24, 38, 38, 38, 38, 38, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 38, 38, 38 ], [ 232 ], [], [ 232, 232, 232, 218, 178, 178, 103, 211, 103, 103, 103, 103, 103, 103, 103, 103 ] ], "positions_before_probe": [ 0, 1, 0, 10, 71, 1, 0, 20 ], "max_scaled_logit_error": 7.185593771612324e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 64, "command_buffers": 7, "status_reads": 7, "gemm_calls": 576, "compute_encoders": 1215, "packed_output_rows": 320, "greedy_available": true, "greedy_chunks": 7, "greedy_active_reads": 7, "greedy_result_readbacks": 1, "greedy_id_bytes": 360, "greedy_emitted_tokens": 90 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "83b26fff91ca65c2199c3ca8b70409f67a0f7aa728123275893b873852672859" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "3edb83dac499e2193c8fd22f90ad653960e208c290d917e9f4b59fe00ea564ca" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 3.5042e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 4.6375e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 3.4042e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 3.225e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 2.9417e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 3.4375e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 5.4666e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 3.5459e-05, "output_ids": [ [] ], "positions": [ 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 0 ], "row_seeds": [ 7301 ], "output_ids": [ [] ], "positions_before_probe": [ 0 ], "max_scaled_logit_error": 5.6858956700939205e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "6b8780d5d1bc5a3883bcc406079d67f182a51dea84980d869c90d3d2ded0d24b" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "b947668c92d0acfc17c17e9c1e42a84c1a15b4a928c2e2c5787fd7f9a9ddd9fd" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 3.4333e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 3.55e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 4.6958e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 4.2417e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 3.8583e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.55e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 3.2708e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 2.6834e-05, "output_ids": [ [], [], [], [], [], [], [], [] ], "positions": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 0, "device_peak_increment_bytes": 0, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "all_zero", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [], [], [], [], [], [], [], [] ], "positions_before_probe": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "max_scaled_logit_error": 5.9048992593384384e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 0, "command_buffers": 0, "status_reads": 0, "gemm_calls": 0, "compute_encoders": 0, "packed_output_rows": 0, "greedy_available": true, "greedy_chunks": 0, "greedy_active_reads": 0, "greedy_result_readbacks": 0, "greedy_id_bytes": 0, "greedy_emitted_tokens": 0 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "9ff5779700eafe21a506231219df7137da4ed60bfff489e5e59735b50d089655" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "633881b838de0f8c030b04208e1f8b9abcf2d2aee4cd089f7ae2ca5d70cca34e" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.0007885, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.001454916, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001015208, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000986834, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.000777125, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.000717667, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 1, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 0, "seconds": 0.001045625, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "iteration": 1, "seconds": 0.00098, "output_ids": [ [ 232 ] ], "positions": [ 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 2388, "device_peak_increment_bytes": 2388, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 1, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0 ], "budgets": [ 1 ], "row_seeds": [ 7301 ], "output_ids": [ [ 232 ] ], "positions_before_probe": [ 1 ], "max_scaled_logit_error": 5.172642832127705e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 1, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 4, "greedy_emitted_tokens": 1 }, "actual_logits_float64_le_digest": { "count": 259, "sha256": "fda8694e4f1fbaeb130318a171bee5289177cfda1a6eca573a50b5528614c4a7" }, "reference_logits_float64_le_digest": { "count": 259, "sha256": "8087682e846da97a35a44bfbcdac1bd23f3dc08431c9aae23016e659e8af22e8" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 1, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001061791, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001017167, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 1, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 3, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001834167, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001086916, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 3, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 8, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001062625, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001024791, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 8, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } }, { "profile": { "device": "mps", "case": "one", "batch_size": 8, "width": 32, "layers": 2 }, "chunk_steps": 32, "resident_records": [ { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 0, "seconds": 0.001093459, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } }, { "schema": "aner.metal-greedy-benchmark/1", "phase": "resident", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "iteration": 1, "seconds": 0.001082292, "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "memory": { "retained_increment_bytes": 0, "all_peak_increment_bytes": 19048, "device_peak_increment_bytes": 19048, "host_peak_increment_bytes": 0, "readback_peak_increment_bytes": 0 }, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 } } ], "kv_oracle": { "schema": "aner.metal-greedy-benchmark/1", "phase": "kv_oracle", "device": "mps", "dtype": "Float32", "case": "one", "batch_size": 8, "width": 32, "layers": 2, "heads": 4, "context": 96, "vocabulary": 259, "model_seed": 20261015, "dropout": 0, "temperature": 0, "top_k": 0, "top_p": 1, "eos_position": -1, "chunk_steps": 32, "prompt_lengths": [ 0, 0, 0, 0, 0, 0, 0, 0 ], "budgets": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "row_seeds": [ 7301, 7402, 7503, 7604, 7705, 7806, 7907, 8008 ], "output_ids": [ [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ], [ 232 ] ], "positions_before_probe": [ 1, 1, 1, 1, 1, 1, 1, 1 ], "max_scaled_logit_error": 5.7325762537475555e-08, "scaled_tolerance": 0.0005, "statistics": { "available": true, "forward_calls": 1, "command_buffers": 1, "status_reads": 1, "gemm_calls": 9, "compute_encoders": 18, "packed_output_rows": 8, "greedy_available": true, "greedy_chunks": 1, "greedy_active_reads": 1, "greedy_result_readbacks": 1, "greedy_id_bytes": 32, "greedy_emitted_tokens": 8 }, "actual_logits_float64_le_digest": { "count": 2072, "sha256": "94dc8e3dcf8e29d2c6ef957c9991915dd4cb3fb033f90ac790f2b2d888064e52" }, "reference_logits_float64_le_digest": { "count": 2072, "sha256": "1d93b36aa089e2ad14a7d7e3f0adb2666bcd5a9999a1d8727967efe5f7302cb3" } } } ] }, "full_diagnostic_json_sha256": "67a7ace41e3db6850156ec49427ea614a95a5bc4045136080bb141b1b702319d", "numerical_vector_policy": "Full vectors retained in the research capture; portable Float64 little-endian digests replace repeated vectors here. All output IDs, timings and counters are retained." } ], "reproduction": { "producer": "metal_greedy_benchmark_20261101.cpp", "producer_sha256": "834ce647498123f3dd7f95cd56af3c4b7e0441f8cc14900821007be4b8e0ad96", "raw_capture_directories": [ "research/metal-greedy-20261101/validation/local-capture-01", "research/metal-greedy-20261101/validation/local-capture-02", "research/metal-greedy-20261101/validation/local-capture-03" ], "protocol": "Identical producer source against sealed baseline/candidate Release libraries; strict floating-point flags, LTO off;4matched fresh-process pairs \u00d73timed calls,2excluded warmups; AB,BA,BA,AB order and rotating profiles.", "reference": "Exact placed Float32 weights widened to CPU Float64, full score oracles, strict independent selected-ID checks and post-resident KV continuation probes." } }