{"schema":"aner.metal-decoder/1","study_date_local":"2026-11-01","development_version":"0.0.6-dev.1","platform_scope":"Mac-only qualification; no new remote CUDA execution in this study.","measurements":{"schema":"aner.metal-decoder-local-summary/1","protocol":{"matched_process_pairs":4,"timed_calls_per_process":3,"excluded_warmups_per_process":2,"process_order":"AB,BA,BA,AB; phase/profile order rotates by outer round","primary_equal_work_phase":"forced","natural_policy":"greedy complete requests; actual IDs, EOS and counts retained; baseline/candidate ID equality is measured, not assumed","forced_policy":"one prefill plus per-row budget minus one fixed token appends; no EOS shortcuts or timed result readback; setup, synchronization and release included","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":"115d34cf800e0aace5e3c4c4093fbc53432e199d81ae4a811c5902899ec25519","binary_sha256":{"baseline":"0015908c2bb3528db20ec60032e390701e8b6e4774bf024512e22bc3185c6b84","candidate":"1349d4adca7288d738cc0681eb287835a6bb3be0f965ab82f2678d543bd8df38"},"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":"4fc1aa6d45dd54ab5a325a11aaa28300241e7e19af6e6576055a52158af01ae2","oracle":{"schema":"aner.metal-decoder-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":9.903150023335168e-08,"candidate_max_scaled_error_vs_cpu_float64":9.903150023335168e-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":9.903150023335168e-08,"candidate_max_scaled_error_vs_cpu_float64":9.903150023335168e-08,"candidate_vs_baseline_max_scaled_error":8.529741228051439e-08,"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":1.058903150680582e-07,"candidate_max_scaled_error_vs_cpu_float64":1.058903150680582e-07,"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":1.1464932408034221e-07,"candidate_max_scaled_error_vs_cpu_float64":1.427624163084606e-07,"candidate_vs_baseline_max_scaled_error":1.106141578033328e-07,"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":9.341542941083764e-08,"candidate_max_scaled_error_vs_cpu_float64":9.341542941083764e-08,"candidate_vs_baseline_max_scaled_error":8.14813470841168e-08,"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":2.87037996357919e-07,"candidate_max_scaled_error_vs_cpu_float64":2.87037996357919e-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":2.9591418225332806e-07,"candidate_max_scaled_error_vs_cpu_float64":3.268849948883582e-07,"candidate_vs_baseline_max_scaled_error":2.6260975558954124e-07,"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":3.968364477611226e-08,"candidate_max_scaled_error_vs_cpu_float64":3.968364477611226e-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,"parameter_serialization_sha256":"1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107"},{"device":"mps","width":32,"layers":2,"parameter_serialization_sha256":"1fdc2462e56376b1e6f1f254950366e2ef9bc64c7e5a85bf1689a4bbc14c3107"},{"device":"mps","width":128,"layers":3,"parameter_serialization_sha256":"7432a93bd25dd41f127bc7621a6b14e5e15fba11c703df9d7c8ebfc28316c9d8"}]},"summary":[{"profile":{"device":"mps","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","calls_per_implementation":12,"median_seconds":{"baseline":0.0327604375,"candidate":0.005348104},"baseline_over_candidate_ratio_of_medians":6.125617134595737,"whole_process_pair_bootstrap_percentile95":[6.1035734124946295,8.96896046213039],"min_max_seconds":{"baseline":[0.030540125,0.035197334],"candidate":[0.003616416,0.007122625]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"candidate":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0}},"statistics":{"baseline":[{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}],"candidate":[{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":72,"compute_encoders":136,"packed_output_rows":8},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":72,"compute_encoders":136,"packed_output_rows":8},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":72,"compute_encoders":136,"packed_output_rows":8},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":72,"compute_encoders":136,"packed_output_rows":8}]}},{"profile":{"device":"mps","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"natural","calls_per_implementation":12,"median_seconds":{"baseline":0.0309747085,"candidate":0.0049596875000000006},"baseline_over_candidate_ratio_of_medians":6.24529438598702,"whole_process_pair_bootstrap_percentile95":[5.973038256503749,21.938277122706364],"min_max_seconds":{"baseline":[0.028955,0.162709375],"candidate":[0.003460375,0.006503084]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"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},"phase":"forced","calls_per_implementation":12,"median_seconds":{"baseline":0.2403085835,"candidate":0.004964312},"baseline_over_candidate_ratio_of_medians":48.407228131511474,"whole_process_pair_bootstrap_percentile95":[48.20507613750588,71.64142061723265],"min_max_seconds":{"baseline":[0.231402625,0.262323459],"candidate":[0.003048584,0.006049209]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"candidate":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0}},"statistics":{"baseline":[{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}],"candidate":[{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":72,"compute_encoders":136,"packed_output_rows":64},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":72,"compute_encoders":136,"packed_output_rows":64},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":72,"compute_encoders":136,"packed_output_rows":64},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":72,"compute_encoders":136,"packed_output_rows":64}]}},{"profile":{"device":"mps","case":"tiny","batch_size":8,"width":32,"layers":2},"phase":"natural","calls_per_implementation":12,"median_seconds":{"baseline":0.24364583350000002,"candidate":0.005298417},"baseline_over_candidate_ratio_of_medians":45.98464664068533,"whole_process_pair_bootstrap_percentile95":[39.31800350629788,65.6579523324506],"min_max_seconds":{"baseline":[0.232350791,0.259229209],"candidate":[0.002989333,0.006514958]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"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":"mps","case":"long","batch_size":1,"width":32,"layers":2},"phase":"forced","calls_per_implementation":12,"median_seconds":{"baseline":0.2376521875,"candidate":0.022231354000000002},"baseline_over_candidate_ratio_of_medians":10.689955614039523,"whole_process_pair_bootstrap_percentile95":[9.390148690512534,14.651386232623452],"min_max_seconds":{"baseline":[0.232304959,0.363072083],"candidate":[0.020939833,0.026451708]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"candidate":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0}},"statistics":{"baseline":[{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}],"candidate":[{"available":true,"forward_calls":64,"command_buffers":64,"status_reads":64,"gemm_calls":576,"compute_encoders":1088,"packed_output_rows":64},{"available":true,"forward_calls":64,"command_buffers":64,"status_reads":64,"gemm_calls":576,"compute_encoders":1088,"packed_output_rows":64},{"available":true,"forward_calls":64,"command_buffers":64,"status_reads":64,"gemm_calls":576,"compute_encoders":1088,"packed_output_rows":64},{"available":true,"forward_calls":64,"command_buffers":64,"status_reads":64,"gemm_calls":576,"compute_encoders":1088,"packed_output_rows":64}]}},{"profile":{"device":"mps","case":"long","batch_size":1,"width":32,"layers":2},"phase":"natural","calls_per_implementation":12,"median_seconds":{"baseline":0.23925958349999998,"candidate":0.021750375000000002},"baseline_over_candidate_ratio_of_medians":11.000250961190323,"whole_process_pair_bootstrap_percentile95":[10.395167611238394,11.262929164127327],"min_max_seconds":{"baseline":[0.228350042,0.282106792],"candidate":[0.021268417,0.026850334]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"candidate":{"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":{"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":"forced","calls_per_implementation":12,"median_seconds":{"baseline":1.9289212290000002,"candidate":0.0229731255},"baseline_over_candidate_ratio_of_medians":83.96424896560114,"whole_process_pair_bootstrap_percentile95":[71.54479775556575,85.47254135238633],"min_max_seconds":{"baseline":[1.861422042,2.448900833],"candidate":[0.021940583,0.032314083]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"candidate":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0}},"statistics":{"baseline":[{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}],"candidate":[{"available":true,"forward_calls":64,"command_buffers":64,"status_reads":64,"gemm_calls":576,"compute_encoders":1088,"packed_output_rows":512},{"available":true,"forward_calls":64,"command_buffers":64,"status_reads":64,"gemm_calls":576,"compute_encoders":1088,"packed_output_rows":512},{"available":true,"forward_calls":64,"command_buffers":64,"status_reads":64,"gemm_calls":576,"compute_encoders":1088,"packed_output_rows":512},{"available":true,"forward_calls":64,"command_buffers":64,"status_reads":64,"gemm_calls":576,"compute_encoders":1088,"packed_output_rows":512}]}},{"profile":{"device":"mps","case":"long","batch_size":8,"width":32,"layers":2},"phase":"natural","calls_per_implementation":12,"median_seconds":{"baseline":1.9189476875,"candidate":0.022764604},"baseline_over_candidate_ratio_of_medians":84.29523691692594,"whole_process_pair_bootstrap_percentile95":[79.86059240682833,85.56132037821796],"min_max_seconds":{"baseline":[1.884974958,1.965077042],"candidate":[0.022145,0.026189166]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"candidate":{"retained_increment_bytes":0,"all_peak_increment_bytes":365636,"device_peak_increment_bytes":349060,"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":[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":"forced","calls_per_implementation":12,"median_seconds":{"baseline":0.3429870835,"candidate":0.0127916875},"baseline_over_candidate_ratio_of_medians":26.813278818764143,"whole_process_pair_bootstrap_percentile95":[20.048292665187827,28.797012971547655],"min_max_seconds":{"baseline":[0.327299541,0.354910875],"candidate":[0.011270666,0.020506541]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"candidate":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0}},"statistics":{"baseline":[{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}],"candidate":[{"available":true,"forward_calls":32,"command_buffers":32,"status_reads":32,"gemm_calls":288,"compute_encoders":544,"packed_output_rows":90},{"available":true,"forward_calls":32,"command_buffers":32,"status_reads":32,"gemm_calls":288,"compute_encoders":544,"packed_output_rows":90},{"available":true,"forward_calls":32,"command_buffers":32,"status_reads":32,"gemm_calls":288,"compute_encoders":544,"packed_output_rows":90},{"available":true,"forward_calls":32,"command_buffers":32,"status_reads":32,"gemm_calls":288,"compute_encoders":544,"packed_output_rows":90}]}},{"profile":{"device":"mps","case":"ragged","batch_size":8,"width":32,"layers":2},"phase":"natural","calls_per_implementation":12,"median_seconds":{"baseline":0.334721979,"candidate":0.0128918965},"baseline_over_candidate_ratio_of_medians":25.963750096814692,"whole_process_pair_bootstrap_percentile95":[23.094896161125813,28.91405136123],"min_max_seconds":{"baseline":[0.323350041,0.360259],"candidate":[0.011264959,0.016217583]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"candidate":{"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":{"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":"forced","calls_per_implementation":12,"median_seconds":{"baseline":0.0419528125,"candidate":0.0058100625},"baseline_over_candidate_ratio_of_medians":7.220716214326438,"whole_process_pair_bootstrap_percentile95":[6.969942276358212,10.969066976284159],"min_max_seconds":{"baseline":[0.04082325,0.068284167],"candidate":[0.004746166,0.00694125]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"candidate":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0}},"statistics":{"baseline":[{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}],"candidate":[{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":104,"compute_encoders":192,"packed_output_rows":8},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":104,"compute_encoders":192,"packed_output_rows":8},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":104,"compute_encoders":192,"packed_output_rows":8},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":104,"compute_encoders":192,"packed_output_rows":8}]}},{"profile":{"device":"mps","case":"tiny","batch_size":1,"width":128,"layers":3},"phase":"natural","calls_per_implementation":12,"median_seconds":{"baseline":0.044777478999999995,"candidate":0.0055975834999999995},"baseline_over_candidate_ratio_of_medians":7.999430289874193,"whole_process_pair_bootstrap_percentile95":[7.116235371496262,8.02807947987489],"min_max_seconds":{"baseline":[0.040461833,0.05850975],"candidate":[0.00470675,0.008265084]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"candidate":{"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":{"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":"forced","calls_per_implementation":12,"median_seconds":{"baseline":0.3385915625,"candidate":0.006923833},"baseline_over_candidate_ratio_of_medians":48.90232946115252,"whole_process_pair_bootstrap_percentile95":[43.6671542136566,61.67830683050235],"min_max_seconds":{"baseline":[0.332666042,0.356680584],"candidate":[0.005004541,0.012429709]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"candidate":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0}},"statistics":{"baseline":[{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}],"candidate":[{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":104,"compute_encoders":192,"packed_output_rows":64},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":104,"compute_encoders":192,"packed_output_rows":64},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":104,"compute_encoders":192,"packed_output_rows":64},{"available":true,"forward_calls":8,"command_buffers":8,"status_reads":8,"gemm_calls":104,"compute_encoders":192,"packed_output_rows":64}]}},{"profile":{"device":"mps","case":"tiny","batch_size":8,"width":128,"layers":3},"phase":"natural","calls_per_implementation":12,"median_seconds":{"baseline":0.345526354,"candidate":0.0062356454999999995},"baseline_over_candidate_ratio_of_medians":55.411481297325196,"whole_process_pair_bootstrap_percentile95":[51.21315402919383,64.61456655135771],"min_max_seconds":{"baseline":[0.330891,0.402531125],"candidate":[0.005067834,0.006964875]},"memory":{"baseline":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"candidate":{"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":{"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":"forced","calls_per_implementation":12,"median_seconds":{"baseline":0.0007337495,"candidate":0.000715542},"baseline_over_candidate_ratio_of_medians":1.0254457460218966,"whole_process_pair_bootstrap_percentile95":[1.0052343816520266,1.1325404986109608],"min_max_seconds":{"baseline":[0.0006915,0.00085625],"candidate":[0.000663792,0.000730459]},"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}],"candidate":[{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0},{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}]}},{"profile":{"device":"cpu","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"natural","calls_per_implementation":12,"median_seconds":{"baseline":0.0007204585,"candidate":0.000685479},"baseline_over_candidate_ratio_of_medians":1.0510292802551209,"whole_process_pair_bootstrap_percentile95":[0.955622362955844,1.0757932077040038],"min_max_seconds":{"baseline":[0.000673292,0.00076775],"candidate":[0.000667666,0.000764292]},"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, Metal Float32 with a CPU Float32 control, bounded new-weight reference decoders. Full natural and equal-work forced timings are separate; the numerical oracle and exact parameter parity are outside timing. Four matched process pairs support each interval.","host":{"machdep.cpu.brand_string":"Apple M2 Max","hw.model":"Mac14,6","hw.memsize":"68719476736","hw.logicalcpu":"12","macos_product_version":"26.7.2"},"runtime_source_sha256":{"baseline":{"CMakeLists.txt":"2aa45ff88b08801a2afdb8ceb32978972db18b379587c05048c3889f9dd501db","include/aner/gen.hpp":"caa1686e00da92c335ed9b2285395a2880b156e470d4eb3c0f765fdd14b03f24","include/aner/gpu.hpp":"6e365ce74baf6e9299a3bfc43fe19f0dbf13252436ffb0caa7c3dc6d8a16f372","include/aner/nn.hpp":"1331f2f929a2f8b9f04041302290f9eafae17a08919377131d780ab6be728dbe","release/version.json":"68a916a6ca3cba82ba4581c18034185d9098d485b89c89ad8ccea4807488ca4b","src/gen.cpp":"ae389fac8b0e237569a79114f7d0951d5d8b9bb405de2f2d3f4309a223c18360","src/gpu.cpp":"c69be6b251398e1c664df1bb3ae75f9e4a6c597678d800752671a8027be68f1b","src/gpu_cuda.cu":"96a814e38c9b1a92634683498a60d754bc7c7c53a8b0f686fe162d070ade4f44","src/gpu_mps.mm":"1693f359a6309a2c3e79763e2648561a3a1180df8388a4653049c9bb29601bb7","src/nn.cpp":"6c493614977ebf4c348e741e80ce0afd44be80b62356e9e609cabe428f9fd303"},"candidate":{"CMakeLists.txt":"2aa45ff88b08801a2afdb8ceb32978972db18b379587c05048c3889f9dd501db","include/aner/gen.hpp":"afa6007d2f8ae0aedd469c8f1070220683fc587e0e076b04e53d80fd41e0acc8","include/aner/gpu.hpp":"bc69e447fc24bdc525e26046b6d252033d747fa7b0319580229621aa8ee265f3","include/aner/nn.hpp":"040b2f42b3011950711b3595334f7e5b8f9dd3ecb11c7bee425049e8a2e634df","release/version.json":"68a916a6ca3cba82ba4581c18034185d9098d485b89c89ad8ccea4807488ca4b","src/gen.cpp":"444656bdb5b27a2b6d2ca0f58cd5a8631d790aa4eefd36fa9662bf5445e2af2d","src/gpu.cpp":"c69be6b251398e1c664df1bb3ae75f9e4a6c597678d800752671a8027be68f1b","src/gpu_cuda.cu":"96a814e38c9b1a92634683498a60d754bc7c7c53a8b0f686fe162d070ade4f44","src/gpu_mps.mm":"ca0eaf88587e18aa33dd77a650281285631e8790b5b84342617d03f267af3d71","src/nn.cpp":"62a67a9e0d1d496b63dd8e47e97cb9ea7651b651ef551f6a76e2dd3e89773bc0"}},"source_and_library_provenance_checked_before_and_after_capture":true,"record_counts":{"fresh_timing_processes":128,"retained_timed_calls":384,"excluded_warmup_calls":256,"natural_calls":192,"forced_calls":192,"numerical_oracle_processes":16,"weight_receipt_processes":6},"numerical_summary":{"baseline_max_scaled_error_vs_cpu_float64":2.9591418225332806e-07,"candidate_max_scaled_error_vs_cpu_float64":3.268849948883582e-07,"candidate_vs_baseline_max_scaled_error":2.6260975558954124e-07,"all_single_request_logits_bitwise_equal":true,"all_natural_ids_and_counts_exactly_equal":true}},"raw_process_captures":[{"outer":0,"implementation":"baseline","profile":{"device":"mps","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round0-baseline-forced-mps-tiny-b1-w32-l2.stdout","capture_sha256":"91b881c708b756b1e40a3fca675f38e5cadaf93c06b80987ea666d8f682356c4","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.033879917,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.03369825,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.032818,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":0,"implementation":"candidate","profile":{"device":"mps","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round0-candidate-forced-mps-tiny-b1-w32-l2.stdout","capture_sha256":"37f74e6ec4df1e20d2c165c955e927351594eef13552e0f5b32344d1eb5a4f22","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.003757208,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.006824375,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.003616416,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}}]},{"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":"3a08136cc304672033d1ee3c7e0fb76e77ef98b5c0b02c6f9f511b22112f0e5a","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.142004625,"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":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.162709375,"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":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.086943917,"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":24064,"device_peak_increment_bytes":24064,"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":"fe980f58fed28f153551342e25767f6b1b42280cbe831f14039aa305b71be82c","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.005533416,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.006503084,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.006472916,"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":"baseline","profile":{"device":"mps","case":"tiny","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round0-baseline-forced-mps-tiny-b8-w32-l2.stdout","capture_sha256":"8e0210d686719c50cc2c30c4be421e0de46016520ac685c4a7f4cb6adeb7c544","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.246224084,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.262323459,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.237752208,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":0,"implementation":"candidate","profile":{"device":"mps","case":"tiny","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round0-candidate-forced-mps-tiny-b8-w32-l2.stdout","capture_sha256":"da25fb3caa312951c58d1a10d41c85a362de63acf07d1d1ab9b602c83ba6ee91","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.005100583,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.003999166,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.0041135,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}}]},{"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":"e590da5bb00a77149abcbcaf6553c524a0047cd490ffbd628afab2a216a21328","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.258122375,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.2431915,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.232350791,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"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":"37304586254a592bdbac263a5905cf88b04b8245d455ac44a2151b0fe32c4048","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.00582075,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.003764459,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.002989333,"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":"baseline","profile":{"device":"mps","case":"long","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round0-baseline-forced-mps-long-b1-w32-l2.stdout","capture_sha256":"268568337f01ac44432f795cd048bbd1d083dc907f2b372a918f204c36a4d022","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.255503958,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.363072083,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.326250959,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":0,"implementation":"candidate","profile":{"device":"mps","case":"long","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round0-candidate-forced-mps-long-b1-w32-l2.stdout","capture_sha256":"216f00fa08df1f523ef40d3dc61fa322ef984e6ee2e7d08fd4585b44bb2dfcc7","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.0253345,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.022267583,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.021417834,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}}]},{"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":"77597c2281095f04486b29c1fee2293fea55d092fd6c4baddac9770a99c1ce31","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.254194,"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":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.258621375,"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":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.282106792,"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":52736,"device_peak_increment_bytes":52736,"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":"733cc0b2a6640ba579643fdea4c530d0d4306e22ef1688385ffa6d421862743d","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.026850334,"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-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.024879,"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-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.021763125,"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":"baseline","profile":{"device":"mps","case":"long","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round0-baseline-forced-mps-long-b8-w32-l2.stdout","capture_sha256":"f21622a7e32fab83951c3e9dd70184ee1bbbcfcec42c926bd1d0b34c02dba7d0","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":1.915210791,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":1.937332958,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":2.448900833,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":0,"implementation":"candidate","profile":{"device":"mps","case":"long","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round0-candidate-forced-mps-long-b8-w32-l2.stdout","capture_sha256":"7985d4eb7e92df50b2ce36d1c520e987fd0075929e9bd48f09c7db316b5b1a3d","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.023302416,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.022557333,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.022698042,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}}]},{"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":"ca2130f39e283533a76970350acebfa818e0f1a45b3d65e24c9336b771fa6a99","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":1.890802375,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":1.953981416,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":1.8871505,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"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":"409826edffa5a0ddd3c9835089c98d9d0fdcc60592d6afdaf9c7f8b72ad30674","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.022145,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.022430791,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.022502916,"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":"baseline","profile":{"device":"mps","case":"ragged","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round0-baseline-forced-mps-ragged-b8-w32-l2.stdout","capture_sha256":"8ba52ecf56fa47474e265861f37faab35ec590495fc6196345b2633bbc7b899e","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.341641417,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.350431208,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.34968925,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":0,"implementation":"candidate","profile":{"device":"mps","case":"ragged","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round0-candidate-forced-mps-ragged-b8-w32-l2.stdout","capture_sha256":"0851574a3db1582e0735988eee9e07a2d548eb3d923dd6da55ce0fdd444793f2","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.01666975,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.015133125,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.013020791,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}}]},{"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":"082f8488bed8616d7e9c2501113fc5f66ffc1687a1a0ead192491314b3cf9217","records":[{"schema":"aner.metal-decoder-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,"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":0.342402625,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"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":0.334540958,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"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":0.359684959,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"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":"4bfc1db1bd63cc28eea9b6a7940fc9e153a0bbffcc82b457129496dfcfcc5203","records":[{"schema":"aner.metal-decoder-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,"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":0.011892208,"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-decoder-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,"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":0.011842084,"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-decoder-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,"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":0.011264959,"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":"baseline","profile":{"device":"mps","case":"tiny","batch_size":1,"width":128,"layers":3},"phase":"forced","capture":"round0-baseline-forced-mps-tiny-b1-w128-l3.stdout","capture_sha256":"477dfe4edd46a8d9741a0e94e882330b0b34b829628cd5ea0834988628872811","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.041749833,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.041186959,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.041386667,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":0,"implementation":"candidate","profile":{"device":"mps","case":"tiny","batch_size":1,"width":128,"layers":3},"phase":"forced","capture":"round0-candidate-forced-mps-tiny-b1-w128-l3.stdout","capture_sha256":"36c4769bc522d791e8d1a47b90f85feb7ae3b9fd3b3e78b5708af1cc49f4128d","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.006433167,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.0058065,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.004746166,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}}]},{"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":"aecdcca3582ae8f2ed209e3f2109034dc915eaf8ab8939d06a316b4a7b86bfef","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.040461833,"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":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.044368291,"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":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.043516542,"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":111616,"device_peak_increment_bytes":111616,"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":"dbf955a94679773999723cbd83bcbabc1d9caadaf26510fa4cb77a1caa0e347e","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.00541025,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.00584075,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.005420542,"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":"baseline","profile":{"device":"mps","case":"tiny","batch_size":8,"width":128,"layers":3},"phase":"forced","capture":"round0-baseline-forced-mps-tiny-b8-w128-l3.stdout","capture_sha256":"0cc4d6983f2004a1226760a2075750c44c98a0ba92dc1eb2ebc0c8b8a9efe798","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.334459917,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.338263542,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.335898667,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":0,"implementation":"candidate","profile":{"device":"mps","case":"tiny","batch_size":8,"width":128,"layers":3},"phase":"forced","capture":"round0-candidate-forced-mps-tiny-b8-w128-l3.stdout","capture_sha256":"78a177228a1c5eedbf2fd54adc169492709219fe167d3c04e6f2ce353f680a66","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.012429709,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.00769225,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.006054542,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}}]},{"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":"ce38adcda6f2bd5e8ab68bbb16881d3b5650132da821452c3712b9c2fdc0b80c","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.330891,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.349589542,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.353205916,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"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":"486eb2770366098866b50da70383b5788c09a72dd3c2cfffce6464d58ff033a3","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.006826167,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.006964875,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.005274583,"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":"baseline","profile":{"device":"cpu","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round0-baseline-forced-cpu-tiny-b1-w32-l2.stdout","capture_sha256":"bc3701e45cb6ba3b79810a532fc15fbce2515cc2d759a825cac7ea762454b81a","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000720167,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.00085625,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.0006915,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":0,"implementation":"candidate","profile":{"device":"cpu","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round0-candidate-forced-cpu-tiny-b1-w32-l2.stdout","capture_sha256":"7b16849f3d5d79ed2f484abf6e2100ebcc07dbe12cc965e7f060458e596f2bd6","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000714667,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000722917,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000716417,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"f7074fd824c9017d3f4cf893047266fd6736046a624ca19d9a98e591e9dfb699","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000737667,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000734875,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000735125,"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":"fb84813fad16c35aa46def988b1f416782993fb90864571de0ea982145db8318","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000680292,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000683333,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000717417,"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":"1b0afb393cff2970c7a898e22fa288e5d4f50469357a9e075a41b2ad2b0e862a","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.003460375,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.004572583,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.003654708,"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":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":"33f88c770fe8b6e8b6e03c8a27ac336b451a199ce87d5e58543ade581474bd14","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.031576833,"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":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.029784125,"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":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.02922125,"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":24064,"device_peak_increment_bytes":24064,"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},"phase":"forced","capture":"round1-candidate-forced-mps-tiny-b8-w32-l2.stdout","capture_sha256":"987a7da82a4f4d6a5e6abaa47e402c45875bdc1d20b5ccf4a21a41b14b8f6443","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.005178042,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.005155917,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.005437209,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}}]},{"outer":1,"implementation":"baseline","profile":{"device":"mps","case":"tiny","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round1-baseline-forced-mps-tiny-b8-w32-l2.stdout","capture_sha256":"3395dba09942c91089e4fbda21fa966d40f51139ed9f14c2759a06890e09bcf1","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.251720875,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.242411167,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.250388917,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"b1821329ba34547375a9c913654b22c619ae3df241c7cca0514eba54f2ce7724","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.006030292,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.006514958,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.005329375,"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":"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":"fee7417a744ad327ff0eb1d8f6d38a5589b6b0e6983943e9a54d051afb138b6f","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.236129167,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.237099042,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.259229209,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"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":"forced","capture":"round1-candidate-forced-mps-long-b1-w32-l2.stdout","capture_sha256":"ded4e84ef31ae131da9ee8253f6c4710bb17e735939fa9cbec01cbcd6a0d8641","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.026451708,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.024976375,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.023218541,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}}]},{"outer":1,"implementation":"baseline","profile":{"device":"mps","case":"long","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round1-baseline-forced-mps-long-b1-w32-l2.stdout","capture_sha256":"9f21240ffeac38825f64f6f3bb1d4adfd5c8a5ebe31062e72d37b5c94cc212d5","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.244049458,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.234531875,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.233589125,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"0e8f3d1c4d4b76f27749e9375d7bc40ec6bc78ce4c22f40d1ce06dbf63cab657","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.026033083,"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-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.021857542,"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-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.021592,"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":"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":"0221b1499ea7a63122421065fbb9de805d7915e5b9f2473d140ceff205973d90","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.235844458,"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":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.240203125,"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":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.236034208,"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":52736,"device_peak_increment_bytes":52736,"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":"forced","capture":"round1-candidate-forced-mps-long-b8-w32-l2.stdout","capture_sha256":"8989759482fa2ce03a5e778cf6ab1538c959425e42b1f1387da36d0ceb5f31f1","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.023248209,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.022687083,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.021940583,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}}]},{"outer":1,"implementation":"baseline","profile":{"device":"mps","case":"long","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round1-baseline-forced-mps-long-b8-w32-l2.stdout","capture_sha256":"aed5e4c798dd79023896a2eb6591c6373b131ce1ffb84793f31e0982d5496606","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":1.95258975,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":1.861422042,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":1.924357416,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"e15c6571d965000f3046aa6d7c693a99d9d166a526e869d3918fd772b3929d83","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.023342792,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.022996916,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.023853125,"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":"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":"540f1c7a713bbc37a3ce3bf7ced59197c0e683afababf3bfcb581e3137def132","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":1.919254625,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":1.938617667,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":1.965077042,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"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":"forced","capture":"round1-candidate-forced-mps-ragged-b8-w32-l2.stdout","capture_sha256":"1c9b5249754de7590fa1edc1f99d8a0c5efd99950bffde20a6ec265084bce178","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.012562584,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.0117565,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.011270666,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}}]},{"outer":1,"implementation":"baseline","profile":{"device":"mps","case":"ragged","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round1-baseline-forced-mps-ragged-b8-w32-l2.stdout","capture_sha256":"a63366433175c32d3f4d7ab2fe878b548553b5b018f29f042034a8d29b295f18","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.354910875,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.338552083,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.328007875,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"311b949fdd9975cafea6bee74ac4dae4de8bf59c02887cc71e2464a3dad8bb82","records":[{"schema":"aner.metal-decoder-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,"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":0.013012333,"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-decoder-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,"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":0.012112709,"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-decoder-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,"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":0.011495708,"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":"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":"4f827b76ebff4e6e561960928f43646535211dd384585843a4ecebb567976c28","records":[{"schema":"aner.metal-decoder-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,"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":0.330245709,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"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":0.329166,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"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":0.331443708,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"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":"forced","capture":"round1-candidate-forced-mps-tiny-b1-w128-l3.stdout","capture_sha256":"7dfaf6a667f240be2dbd7dc7766f887e35d728998bf48083f998710c06ff8721","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.006073833,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.005813625,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.005653083,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}}]},{"outer":1,"implementation":"baseline","profile":{"device":"mps","case":"tiny","batch_size":1,"width":128,"layers":3},"phase":"forced","capture":"round1-baseline-forced-mps-tiny-b1-w128-l3.stdout","capture_sha256":"56f9bb5ab5366aff3ecab8b660148381df2758df78ea756aa3dee775fc4fcf39","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.063770042,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.068284167,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.063149416,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"967b9b77f73828682532120b9b8818a79002d65f1db3e0c3663d67e1020e6638","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.008265084,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.007595958,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.007582667,"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":"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":"06e17f03182c6318081dbc3a8a5555dd8d013097dc42862d566e7e77524a19e1","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.051361959,"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":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.054054625,"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":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.05850975,"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":111616,"device_peak_increment_bytes":111616,"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":"forced","capture":"round1-candidate-forced-mps-tiny-b8-w128-l3.stdout","capture_sha256":"3a02f2d8a124e9f07814208a4051b1780625db5103cd557728840f8ba44fd722","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.007188542,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.007229917,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.006879125,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}}]},{"outer":1,"implementation":"baseline","profile":{"device":"mps","case":"tiny","batch_size":8,"width":128,"layers":3},"phase":"forced","capture":"round1-baseline-forced-mps-tiny-b8-w128-l3.stdout","capture_sha256":"8e41c8874be973f74032fabd3378b381a14acc8a48a92f091104c7a8251cfcbc","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.332666042,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.337962083,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.356680584,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"a3d1272039f2016ae017fc5ca6ca649f6edcebddacb5c86246afa9fade147ec7","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.006630459,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.0051405,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.005528625,"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":"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":"9caba5811b346ca368f6c5ed7c58f844f7de54bf8f02d75c1fedd763eced4a7c","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.348025416,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.366388666,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.357229708,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"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":"forced","capture":"round1-candidate-forced-cpu-tiny-b1-w32-l2.stdout","capture_sha256":"3bd2424abdbb370fbaef15f47ca2967c7ed11a1ca1af437592d4b4f9ece92d48","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000665917,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000665208,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000663792,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":1,"implementation":"baseline","profile":{"device":"cpu","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round1-baseline-forced-cpu-tiny-b1-w32-l2.stdout","capture_sha256":"3f96de052e7a48fff865a7710358acb894f88be54ca025e7a93a2c15d4c0cc3e","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000724083,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000774,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000753375,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"1c222e4a6b35acabbc7eb222c23298e23e2a3dac81aa0a23d7ba14d83d81b817","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000721959,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000730458,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000764292,"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":"2274f16dd5b6aa5268d538cba01c4687c6a00331bebe0863aa07bc19fdf3cb7a","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000698042,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.00070825,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000695083,"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":"forced","capture":"round1-candidate-forced-mps-tiny-b1-w32-l2.stdout","capture_sha256":"fa9ad0a15246c81cb8c66052e3014d2d34dbccb645eb6d317d49356999a6728b","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.006956834,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.005377166,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.005432958,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}}]},{"outer":1,"implementation":"baseline","profile":{"device":"mps","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round1-baseline-forced-mps-tiny-b1-w32-l2.stdout","capture_sha256":"0e443a41dba25a5be0f6f5127d04c0e8f0121e16aecfb684b3a9d1a85d407537","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.035197334,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.033160458,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.030818083,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":2,"implementation":"candidate","profile":{"device":"mps","case":"tiny","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round2-candidate-forced-mps-tiny-b8-w32-l2.stdout","capture_sha256":"d89cc8f723a066e7a90c33f5b26369ef25cd31a8a21b9ce4ebba5d032e21fe1b","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.006049209,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.004699375,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.004828041,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}}]},{"outer":2,"implementation":"baseline","profile":{"device":"mps","case":"tiny","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round2-baseline-forced-mps-tiny-b8-w32-l2.stdout","capture_sha256":"205fead0865226263397a2fba7cabcbd85a06dee2034818eb0844c806177d7ba","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.232736084,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.231402625,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.24427875,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"8f6b9efa3fbf27c6fd5fc46071d769905855607e3e9d041502d3a532a3fc3e95","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.005267459,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.00368175,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.003064,"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":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":"5f665498fe909326de31a080e18c4c0f177d5fb23e868c86e2589abdcab45e07","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.244100167,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.235761125,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.241736166,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}}]},{"outer":2,"implementation":"candidate","profile":{"device":"mps","case":"long","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round2-candidate-forced-mps-long-b1-w32-l2.stdout","capture_sha256":"dc16c3ef252bcfcf81408697ecabe44a4a6e07baf6db83d87a2731773528efb9","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.022195125,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.021923125,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.021722042,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}}]},{"outer":2,"implementation":"baseline","profile":{"device":"mps","case":"long","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round2-baseline-forced-mps-long-b1-w32-l2.stdout","capture_sha256":"fe6dafff814ce91c4adbd341ca9c30dcb00c55f184bc5a65d8e18728bf81ef7f","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.23683025,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.23565175,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.238474125,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"9b7d8cb7cd78fb2e94e6b314755fd0cdd1ebd8e9f0f6dcd9777c0e15a9625425","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.021331,"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-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.021321542,"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-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.022195083,"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":"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":"07e7685e7d323c04a0213068c8fd09c21b15a8e20acb64d8bfb93ff858e2d33b","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.238316042,"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":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.246487,"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":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.240249542,"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":52736,"device_peak_increment_bytes":52736,"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":"forced","capture":"round2-candidate-forced-mps-long-b8-w32-l2.stdout","capture_sha256":"3cb10eccbc418851bb3e4d68c97cf463c263f5e9aa759e07390bfcfc5bc882af","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.023809417,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.021975584,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.022621125,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}}]},{"outer":2,"implementation":"baseline","profile":{"device":"mps","case":"long","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round2-baseline-forced-mps-long-b8-w32-l2.stdout","capture_sha256":"693ce24b51002b3f31dd600493880ed6503a11c45c6b345a81257043ae19cb3a","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":1.943686083,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":1.933485042,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":1.905253125,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"97367a2bcc9e18752b26c8def08b341402353667f8b7f650ca8b98ccd1d3c5eb","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.026189166,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.024024875,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.023817125,"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":"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":"64617d338e5afb9cdcea3ce1b3ff8a6d96e5e2521bf926b5c06f5c8669a3dffa","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":1.951989083,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":1.91864075,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":1.8940625,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"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":"forced","capture":"round2-candidate-forced-mps-ragged-b8-w32-l2.stdout","capture_sha256":"2cccef9f0b7bd8f672e3d9a53a86e2de3939abd92b2ffda4dac2195e034a9200","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.011422625,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.011696416,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.011498792,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}}]},{"outer":2,"implementation":"baseline","profile":{"device":"mps","case":"ragged","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round2-baseline-forced-mps-ragged-b8-w32-l2.stdout","capture_sha256":"e97b33b701eae377cf142a0adb2ad16e2c774c80ebe8663f03d1cc91b7fddd2b","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.327299541,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.329518125,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.34433275,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"d83e86edb3efc66cc5e5454f1070f7beaafcba959511de40a1650f85069a847d","records":[{"schema":"aner.metal-decoder-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,"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":0.016217583,"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-decoder-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,"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":0.014868083,"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-decoder-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,"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":0.012971959,"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":"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":"97c35ae78880dea904819bfdfbb456a0fa7240a48d9c417c8cd4231dd28005ca","records":[{"schema":"aner.metal-decoder-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,"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":0.33877325,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"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":0.343376833,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"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":0.360259,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"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":"forced","capture":"round2-candidate-forced-mps-tiny-b1-w128-l3.stdout","capture_sha256":"3d58e8ee0931cb41d52cc42302ec76f72bbb43ea1ea501a49c5222f197248c84","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.006180833,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.005150583,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.005298833,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}}]},{"outer":2,"implementation":"baseline","profile":{"device":"mps","case":"tiny","batch_size":1,"width":128,"layers":3},"phase":"forced","capture":"round2-baseline-forced-mps-tiny-b1-w128-l3.stdout","capture_sha256":"c7af53d804d60f2bcc3bd2df326d5ca3fdfe8c71c2613d405488193455155052","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.04238625,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.042205167,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.042059542,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"e421bf9b542de0851610ef299aa7c642587d7f4c2b688c65486141a54d0ac27a","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.006678333,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.005713709,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.00470675,"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":"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":"7869180229cc85af89b491cf8a856c8addce0e8659b65e6563e8c6c0cbddabcb","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.045186667,"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":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.045982959,"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":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.043911292,"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":111616,"device_peak_increment_bytes":111616,"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":"forced","capture":"round2-candidate-forced-mps-tiny-b8-w128-l3.stdout","capture_sha256":"ed8b7c6fa2eefc96c8bce5ce64eacfc695be4d35efe6a9beead13fcb989de8ab","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.006968541,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.006008125,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.005004541,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}}]},{"outer":2,"implementation":"baseline","profile":{"device":"mps","case":"tiny","batch_size":8,"width":128,"layers":3},"phase":"forced","capture":"round2-baseline-forced-mps-tiny-b8-w128-l3.stdout","capture_sha256":"f7b5a6c0f01f04f215cb080939fba728cda8bef0678d7318e237392be971bbd1","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.337073625,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.343515416,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.350593292,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"e8c13fd7ac00c14eb746a0c4ab1da41bc2044f3b352bbaa8f1000609bda76f31","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.006592167,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.006491959,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.005946,"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":"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":"a5d70c4926184f0dbd15ffeabd8dccc8831a625523a5ce46f3d1cda99916d19d","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.331690958,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.335246291,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.339811708,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"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":"forced","capture":"round2-candidate-forced-cpu-tiny-b1-w32-l2.stdout","capture_sha256":"d3a772325143fbd5e36b8cca93364d78e02cf3a6add8499e0111704bf50f66d8","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000725375,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000713,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000679375,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":2,"implementation":"baseline","profile":{"device":"cpu","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round2-baseline-forced-cpu-tiny-b1-w32-l2.stdout","capture_sha256":"116c2b4f13456e37989e006679c8ca0db53b62e1462ec73203a837e202b30d73","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000718667,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000724416,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000720666,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"d9a95773763a0fafa59aa758f8e1d98c6dbd7edf628aa5e9b2abac4de031ef16","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000687625,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000679875,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.00074575,"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":"7fdad4568447d52960a24ce2430ee7861d0991f8572a3ac5d776fcb548cda704","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000737834,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000732667,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.00076775,"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":"forced","capture":"round2-candidate-forced-mps-tiny-b1-w32-l2.stdout","capture_sha256":"6baebaf9cbd43ab56c4b87dd71731f1a21ea60054f61da9549419c50c3939fcc","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.007122625,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.004816708,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.004307833,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}}]},{"outer":2,"implementation":"baseline","profile":{"device":"mps","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round2-baseline-forced-mps-tiny-b1-w32-l2.stdout","capture_sha256":"6e754a89fb899e88003ca7f6225030273fcd810a355ff9e868ebabbcea0886cf","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.032620625,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.031771,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.030540125,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"ad8539e5c7861c23df28bc2d6655810ee30a619f66a76f0f14e070f87cd079f0","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.006187375,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.005028792,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.004814041,"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":"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":"39d20ac36545bb893a1bf065780ce6dc46977563794be99d37201811a32575a0","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.030037167,"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":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.030532375,"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":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.028955,"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":24064,"device_peak_increment_bytes":24064,"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},"phase":"natural","capture":"round3-baseline-natural-mps-tiny-b8-w32-l2.stdout","capture_sha256":"850eda3dba48a6b1d19cff8aa81495014d23fb41f2d75347302f5c3cd7ce56da","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.244296667,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.249522959,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.24617575,"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":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"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":"d4fef660b2f2f0a3dbb225f9acfb1a17b85a0145496cb628f0e5db3a5f3fca65","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.006139375,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.006234125,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.004274083,"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":1,"width":32,"layers":2},"phase":"forced","capture":"round3-baseline-forced-mps-long-b1-w32-l2.stdout","capture_sha256":"6e4406689784ba55803b5e8546fee02099325e38302cb6891844c09324097d5f","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.236169875,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.243877875,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.232304959,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":3,"implementation":"candidate","profile":{"device":"mps","case":"long","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round3-candidate-forced-mps-long-b1-w32-l2.stdout","capture_sha256":"685c2594c7d35374ad964d540911a1a01bd42624f320dd4cca736fd2eda9c6f1","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.022345667,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.020939833,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.021348416,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":36352,"kv_elements":9088,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":43636,"device_peak_increment_bytes":43636,"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":1088,"packed_output_rows":64}}]},{"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":"5ec195b2120a27b9a499514369b54c796c484eb75f820a81edde0bcadb307b6e","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.228350042,"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":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.234039375,"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":52736,"device_peak_increment_bytes":52736,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.234058916,"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":52736,"device_peak_increment_bytes":52736,"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":"4f5435b0b573de63931e43a0bf04197882413c31ebba77fad0535a5c89932c38","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":0,"seconds":0.021737625,"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-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":1,"seconds":0.021545959,"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-decoder-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,"prompt_lengths":[8],"budgets":[64],"row_seeds":[7301],"iteration":2,"seconds":0.021268417,"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":"baseline","profile":{"device":"mps","case":"long","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round3-baseline-forced-mps-long-b8-w32-l2.stdout","capture_sha256":"64f20e8df456fbe208ad6b6069adabad175795aca8a9b6e40432f8a05a31effc","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":1.902845375,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":1.908818209,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":1.957197459,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":3,"implementation":"candidate","profile":{"device":"mps","case":"long","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round3-candidate-forced-mps-long-b8-w32-l2.stdout","capture_sha256":"473114fb39bbd9cd75632eb623f2022f1874c242d815a837eb48ae4f9c0d610c","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.032314083,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.026680042,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"long","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.024002375,"timed_logit_readback":false,"forced_next_logit_rows":512,"kv_bytes":290816,"kv_elements":72704,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":349060,"device_peak_increment_bytes":349060,"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":1088,"packed_output_rows":512}}]},{"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":"b209d145637e419681f463c9857f32fcf904de0155bb73bf10092b8be38393f6","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":1.884974958,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":1.920978,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":1.908085209,"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":348832,"device_peak_increment_bytes":307200,"host_peak_increment_bytes":58016,"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":"eb857d0e409b278bf5f05238ed4c32f77f29ca80b9067f08b1db72ca4377dafb","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.022532292,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.022243209,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[64,64,64,64,64,64,64,64],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.022300792,"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":"baseline","profile":{"device":"mps","case":"ragged","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round3-baseline-forced-mps-ragged-b8-w32-l2.stdout","capture_sha256":"7ef103fe4b61359141a1d75c031ac829bb04fcb5e0c8d8f88cf8c4429a70ffd9","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.3445675,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.34090825,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.348212417,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":3,"implementation":"candidate","profile":{"device":"mps","case":"ragged","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round3-candidate-forced-mps-ragged-b8-w32-l2.stdout","capture_sha256":"e282dfc5a4e1a56d2d50f72727167aadc49fd535e46212ce05518b17a3cde444","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.020506541,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.017186875,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"ragged","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"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":0.014273333,"timed_logit_readback":false,"forced_next_logit_rows":90,"kv_bytes":97792,"kv_elements":24448,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":191136,"device_peak_increment_bytes":191136,"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":544,"packed_output_rows":90}}]},{"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":"9d15e968f2c01069f1fb6f6e1296285bc3f3256a7f0418bb1166ec21e3db4a57","records":[{"schema":"aner.metal-decoder-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,"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":0.334903,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"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":0.323350041,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"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":0.33395125,"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":194048,"device_peak_increment_bytes":194048,"host_peak_increment_bytes":58016,"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":"5f0d7b8851e3fd07e0716dedfd45c2a014c4b7598b8024015a4b01696889786f","records":[{"schema":"aner.metal-decoder-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,"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":0.013683833,"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-decoder-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,"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":0.01350575,"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-decoder-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,"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":0.012811834,"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":"baseline","profile":{"device":"mps","case":"tiny","batch_size":1,"width":128,"layers":3},"phase":"forced","capture":"round3-baseline-forced-mps-tiny-b1-w128-l3.stdout","capture_sha256":"484c7c7a11b699f823f3ecc03efd76919d2aba963f54f19bdbbe0d1000a3094b","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.041082875,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.04082325,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.041846083,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":3,"implementation":"candidate","profile":{"device":"mps","case":"tiny","batch_size":1,"width":128,"layers":3},"phase":"forced","capture":"round3-candidate-forced-mps-tiny-b1-w128-l3.stdout","capture_sha256":"193491fb04c5ab5e0a68fe093378ff3a5122d85dbd983aeb0a6cc426ea6d1f31","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.00694125,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.005894292,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.005494667,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":46080,"kv_elements":11520,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":71796,"device_peak_increment_bytes":71796,"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":192,"packed_output_rows":8}}]},{"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":"ebc5815801993a1351ee8333e9cf4b8bdb9ca9b63f9bfa621a14eb1e1291a79e","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.041024417,"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":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.041450792,"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":111616,"device_peak_increment_bytes":111616,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.045233958,"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":111616,"device_peak_increment_bytes":111616,"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":"e55ab883f7cf6fe880b5ce614dceff1dab1801afa774f127477f1bb0f68a63a3","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.005481458,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.005208709,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.004850459,"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":"baseline","profile":{"device":"mps","case":"tiny","batch_size":8,"width":128,"layers":3},"phase":"forced","capture":"round3-baseline-forced-mps-tiny-b8-w128-l3.stdout","capture_sha256":"a4c57a478cbcbcd577d63ae17fc289a4db986883a6cde4afc7d67e8ca9334b0e","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.3428755,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.34132775,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.338919583,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":3,"implementation":"candidate","profile":{"device":"mps","case":"tiny","batch_size":8,"width":128,"layers":3},"phase":"forced","capture":"round3-candidate-forced-mps-tiny-b8-w128-l3.stdout","capture_sha256":"b6b341cc5d1c7798db3e8643afcc9131f843e144b73c87c4fff15941d80d735e","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.0088015,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.005534,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":128,"layers":3,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.005116541,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":368640,"kv_elements":92160,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":599048,"device_peak_increment_bytes":599048,"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":192,"packed_output_rows":64}}]},{"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":"7ea7a2177f8c88c887aea714c73a288ef4e8b1cca63d25ef8c684fe39e676c6a","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.343027292,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.335254041,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":16576},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.402531125,"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":434176,"device_peak_increment_bytes":434176,"host_peak_increment_bytes":58016,"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":"ba635e261e7c9ea072169b532f35d37609047493aa9726888d8bf70998212473","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.0064525,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.006018791,"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-decoder-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,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.005067834,"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":"baseline","profile":{"device":"cpu","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round3-baseline-forced-cpu-tiny-b1-w32-l2.stdout","capture_sha256":"62d1e9788a4960421735e4aa3ec737813828a7d0291e27163386e964b9529ccf","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000771792,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000808541,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000743083,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":3,"implementation":"candidate","profile":{"device":"cpu","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round3-candidate-forced-cpu-tiny-b1-w32-l2.stdout","capture_sha256":"e07feaeda37969aa6daded4aaf5d46c31ba9954d18bb9161338a27d9b1822ba8","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000728542,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000730459,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"cpu","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.0007195,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"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,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"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":"f4b47fc374a795fa9d97a6ff60f8bde84ef703ea4c872b36c557f5f633b835a7","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000673958,"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-decoder-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,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000694084,"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":"f5f132066fb356c87e4a8cf2fb38c3b5331c5f269b7c7fbc3a155371de10a96d","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.000672375,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.000675458,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.000667666,"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":"forced","capture":"round3-baseline-forced-mps-tiny-b1-w32-l2.stdout","capture_sha256":"a31932b188c826bc797b2d188943bfdcade1c7f955b1c7c7e3ec6ee01ec912fb","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.032702875,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.034467417,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.032514916,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":0,"readback_peak_increment_bytes":0},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":3,"implementation":"candidate","profile":{"device":"mps","case":"tiny","batch_size":1,"width":32,"layers":2},"phase":"forced","capture":"round3-candidate-forced-mps-tiny-b1-w32-l2.stdout","capture_sha256":"b4d182815a051e70e5ae4a342f8b72cfa8989df0fe039fbe7157ab32f63ddbf9","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.005904416,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.004834,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":1,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.005319042,"timed_logit_readback":false,"forced_next_logit_rows":8,"kv_bytes":7680,"kv_elements":1920,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":14964,"device_peak_increment_bytes":14964,"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":136,"packed_output_rows":8}}]},{"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":"6ddc5038d8b9ba126b3e2ff8c2654179aa297b632cfb31a57812e717feb8e1b1","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.031417042,"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":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.032091209,"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":24064,"device_peak_increment_bytes":24064,"host_peak_increment_bytes":2072,"readback_peak_increment_bytes":2072},"statistics":{"available":false}},{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.030122042,"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":24064,"device_peak_increment_bytes":24064,"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":"8c600d37b6cf9fd3e25ae59ab2f659c847d1eecd7b35e3a0f823dd2a611b8a0e","records":[{"schema":"aner.metal-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":0,"seconds":0.006475417,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":1,"seconds":0.004845416,"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-decoder-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,"prompt_lengths":[8],"budgets":[8],"row_seeds":[7301],"iteration":2,"seconds":0.004890583,"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":"baseline","profile":{"device":"mps","case":"tiny","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round3-baseline-forced-mps-tiny-b8-w32-l2.stdout","capture_sha256":"6814f37c49c763492054db964781a351cd54fe316cf6cdf167b31c55d26a2527","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.238206,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.236664917,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.237070416,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":8,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119456,"device_peak_increment_bytes":77824,"host_peak_increment_bytes":58016,"readback_peak_increment_bytes":2072},"statistics":{"available":false,"forward_calls":0,"command_buffers":0,"status_reads":0,"gemm_calls":0,"compute_encoders":0,"packed_output_rows":0}}]},{"outer":3,"implementation":"candidate","profile":{"device":"mps","case":"tiny","batch_size":8,"width":32,"layers":2},"phase":"forced","capture":"round3-candidate-forced-mps-tiny-b8-w32-l2.stdout","capture_sha256":"e2b3460a7d0d11306f534aa3f3e8220a04c1a572ebf1ae021a522e402f2e8332","records":[{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":0,"seconds":0.005178458,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":1,"seconds":0.003048584,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}},{"schema":"aner.metal-decoder-benchmark/1","phase":"forced","device":"mps","dtype":"Float32","case":"tiny","batch_size":8,"width":32,"layers":2,"heads":4,"context":96,"vocabulary":259,"model_seed":20261015,"dropout":0,"prompt_lengths":[8,8,8,8,8,8,8,8],"budgets":[8,8,8,8,8,8,8,8],"row_seeds":[7301,7402,7503,7604,7705,7806,7907,8008],"iteration":2,"seconds":0.003309125,"timed_logit_readback":false,"forced_next_logit_rows":64,"kv_bytes":61440,"kv_elements":15360,"kv_allocations":1,"memory":{"retained_increment_bytes":0,"all_peak_increment_bytes":119684,"device_peak_increment_bytes":119684,"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":136,"packed_output_rows":64}}]}],"qualification":{"focused_actual_mps_groups":{"whole_plan":6604129,"single_cache":2392539,"batched_cache":9335843,"gpu_tensor":256,"generation_device":19251,"sequence_device":9098},"check_count_scope":"Assertion counts overlap between groups; they are not independent timing replicates.","address_and_undefined_behavior_sanitizer":"Metal whole-plan and GPU integration plus CPU Float32 cache passed; platform leak detection disabled.","metal_validation":{"api":"enabled; all whole-plan contracts passed","shader":"enabled; all whole-plan contracts passed","environment":{"MTL_DEBUG_LAYER":"1","MTL_DEBUG_LAYER_ERROR_MODE":"assert","MTL_SHADER_VALIDATION":"1","MTL_SHADER_VALIDATION_REPORT_TO_STDERR":"1","MTL_SHADER_VALIDATION_ABORT_ON_FAULT":"1"},"documentation":["https://developer.apple.com/documentation/xcode/validating-your-apps-metal-api-usage","https://developer.apple.com/documentation/xcode/validating-your-apps-metal-shader-usage"]},"rounding_regression":"The original Metal normalization affine expression is retained; near-tie GPT2 greedy/sampled parity passes.","normal_regressions":{"registered":289,"executed_or_accounted_normal_groups":283,"passed":274,"platform_or_live_service_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"],"failed":0,"research_campaigns_excluded":6}},"interpretation":["One command buffer/status read applies to decoder forward. Model setup, cache growth copies and host token selection can add separate work.","Natural generation remains a host-driven loop; this is not a device-only multi-step recurrence.","Native bounded decoder and supported GPT2 inference only; pretrained LibTorch training and embedded llama.cpp inference are unchanged.","Measurements cover these reference models and this Mac; they do not establish Gemma 2B or PyTorch superiority."]}