{ "environment": { "platform": "macOS-27.0.1-arm64-arm-64bit", "machine": "arm64", "metal": { "device_name": "Apple M5 Pro", "max_recommended_working_set_size": 55662788608, "memory_size": 68719476736, "architecture": "applegpu_g17s", "max_buffer_length": 41747087360, "resource_limit": 499000 }, "versions": { "mlx": "0.32.4", "mlx-metal": "0.32.4", "numpy": "2.5.4", "safetensors": "0.8.0", "transformers": "4.57.6", "torch": "2.15.0", "Pillow": "13.0.0", "scikit-image": "0.26.0" }, "precision_environment": { "MLX_ENABLE_TF32": "0", "MLX_CONV_WINOGRAD": "0" } }, "warmup": 2, "repeats": 5, "scope": "Original projection weights, synthetic BF16 inputs. Custom Metal 4 TensorOps with FP32 accumulation and BF16 bias epilogue.", "trials": [ { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 0, 0, 0 ], "simdgroups": 0, "walk_m": 0, "median_ms": 39.11950020119548, "min_ms": 39.022000040858984, "max_ms": 39.206292014569044, "samples_ms": [ 39.13562512025237, 39.022000040858984, 39.206292014569044, 39.11950020119548, 39.02737516909838 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 32, 64, 128 ], "simdgroups": 2, "walk_m": 0, "median_ms": 144.08591715618968, "min_ms": 143.81708297878504, "max_ms": 144.4397501181811, "samples_ms": [ 144.4397501181811, 143.81708297878504, 144.08591715618968, 144.16858297772706, 143.81849998608232 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 32, 64, 128 ], "simdgroups": 2, "walk_m": 8, "median_ms": 42.84683312289417, "min_ms": 42.720959056168795, "max_ms": 42.899500112980604, "samples_ms": [ 42.84683312289417, 42.899500112980604, 42.720959056168795, 42.8632500115782, 42.73395799100399 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 32, 64, 256 ], "simdgroups": 2, "walk_m": 0, "median_ms": 144.37312516383827, "min_ms": 143.9994580578059, "max_ms": 144.90454108454287, "samples_ms": [ 144.90454108454287, 143.9994580578059, 144.37312516383827, 144.53170797787607, 144.06895893625915 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 32, 64, 256 ], "simdgroups": 2, "walk_m": 8, "median_ms": 41.73733294010162, "min_ms": 41.50099982507527, "max_ms": 41.90316703170538, "samples_ms": [ 41.73975018784404, 41.50099982507527, 41.73733294010162, 41.690500220283866, 41.90316703170538 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 64, 128 ], "simdgroups": 4, "walk_m": 0, "median_ms": 76.95283298380673, "min_ms": 76.59599999897182, "max_ms": 77.17770803719759, "samples_ms": [ 76.95283298380673, 76.84358395636082, 77.12462497875094, 76.59599999897182, 77.17770803719759 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 64, 128 ], "simdgroups": 4, "walk_m": 8, "median_ms": 42.57095814682543, "min_ms": 42.45987511239946, "max_ms": 42.72100003436208, "samples_ms": [ 42.58745792321861, 42.45987511239946, 42.57095814682543, 42.53150010481477, 42.72100003436208 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 64, 256 ], "simdgroups": 4, "walk_m": 0, "median_ms": 76.77087490446866, "min_ms": 76.64783322252333, "max_ms": 77.23387493751943, "samples_ms": [ 76.7487499397248, 77.23387493751943, 76.77087490446866, 76.64783322252333, 77.14020903222263 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 64, 256 ], "simdgroups": 4, "walk_m": 8, "median_ms": 41.06487496756017, "min_ms": 40.93854199163616, "max_ms": 41.63033398799598, "samples_ms": [ 40.93854199163616, 41.015249909833074, 41.06487496756017, 41.276958072558045, 41.63033398799598 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 128, 128 ], "simdgroups": 4, "walk_m": 0, "median_ms": 75.16441587358713, "min_ms": 74.81487491168082, "max_ms": 75.63787512481213, "samples_ms": [ 75.63787512481213, 74.81487491168082, 75.01008291728795, 75.16441587358713, 75.40129194967449 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 128, 128 ], "simdgroups": 4, "walk_m": 8, "median_ms": 40.33562494441867, "min_ms": 40.23950011469424, "max_ms": 40.57650011964142, "samples_ms": [ 40.57650011964142, 40.33562494441867, 40.23950011469424, 40.33204214647412, 40.38066603243351 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 128, 256 ], "simdgroups": 4, "walk_m": 0, "median_ms": 74.62074980139732, "min_ms": 74.47524997405708, "max_ms": 75.09104209020734, "samples_ms": [ 74.62074980139732, 75.09104209020734, 74.54266701824963, 74.79133293963969, 74.47524997405708 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 128, 256 ], "simdgroups": 4, "walk_m": 8, "median_ms": 39.65325001627207, "min_ms": 39.51312508434057, "max_ms": 39.91074999794364, "samples_ms": [ 39.51312508434057, 39.65325001627207, 39.56470894627273, 39.91074999794364, 39.74874992854893 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 64, 128 ], "simdgroups": 4, "walk_m": 0, "median_ms": 46.420250087976456, "min_ms": 46.2697499897331, "max_ms": 46.63291689939797, "samples_ms": [ 46.382749918848276, 46.63291689939797, 46.420250087976456, 46.2697499897331, 46.44833295606077 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 64, 128 ], "simdgroups": 4, "walk_m": 8, "median_ms": 40.789457969367504, "min_ms": 40.339959086850286, "max_ms": 41.80283402092755, "samples_ms": [ 40.38770799525082, 40.339959086850286, 40.789457969367504, 41.80283402092755, 41.13591695204377 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 64, 256 ], "simdgroups": 4, "walk_m": 0, "median_ms": 45.47000001184642, "min_ms": 45.36420805379748, "max_ms": 45.94416683539748, "samples_ms": [ 45.36420805379748, 45.94416683539748, 45.47000001184642, 45.78987509012222, 45.43199995532632 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 64, 256 ], "simdgroups": 4, "walk_m": 8, "median_ms": 40.022207889705896, "min_ms": 39.64875009842217, "max_ms": 40.58245802298188, "samples_ms": [ 39.695374900475144, 39.64875009842217, 40.022207889705896, 40.58245802298188, 40.487749967724085 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 128, 128 ], "simdgroups": 4, "walk_m": 0, "median_ms": 55.45079102739692, "min_ms": 55.025249952450395, "max_ms": 55.971749825403094, "samples_ms": [ 55.2547920960933, 55.45079102739692, 55.025249952450395, 55.93649996444583, 55.971749825403094 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 128, 128 ], "simdgroups": 4, "walk_m": 8, "median_ms": 54.957584012299776, "min_ms": 54.046041099354625, "max_ms": 55.90949999168515, "samples_ms": [ 54.046041099354625, 54.527875036001205, 54.957584012299776, 55.12870894744992, 55.90949999168515 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 128, 256 ], "simdgroups": 4, "walk_m": 0, "median_ms": 51.92316696047783, "min_ms": 50.93825003132224, "max_ms": 52.514916053041816, "samples_ms": [ 50.93825003132224, 51.41841690056026, 51.92316696047783, 52.130750147625804, 52.514916053041816 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 128, 256 ], "simdgroups": 4, "walk_m": 8, "median_ms": 50.32787495292723, "min_ms": 49.495709128677845, "max_ms": 51.822875160723925, "samples_ms": [ 49.495709128677845, 49.761957954615355, 50.36799982190132, 50.32787495292723, 51.822875160723925 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 128, 128 ], "simdgroups": 8, "walk_m": 0, "median_ms": 50.32845796085894, "min_ms": 50.06470903754234, "max_ms": 50.51270918920636, "samples_ms": [ 50.457790959626436, 50.06470903754234, 50.51270918920636, 50.19462504424155, 50.32845796085894 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 128, 128 ], "simdgroups": 8, "walk_m": 8, "median_ms": 43.097833869978786, "min_ms": 42.822916992008686, "max_ms": 44.24499999731779, "samples_ms": [ 42.822916992008686, 43.097833869978786, 43.96050004288554, 43.00912516191602, 44.24499999731779 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 128, 256 ], "simdgroups": 8, "walk_m": 0, "median_ms": 48.885042080655694, "min_ms": 48.5747498460114, "max_ms": 49.46512496098876, "samples_ms": [ 48.5747498460114, 48.93749998882413, 48.885042080655694, 49.46512496098876, 48.775041941553354 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 128, 128, 256 ], "simdgroups": 8, "walk_m": 8, "median_ms": 42.02324990183115, "min_ms": 40.935083059594035, "max_ms": 42.17162495478988, "samples_ms": [ 40.935083059594035, 41.284624952822924, 42.02324990183115, 42.12037520483136, 42.17162495478988 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 128, 128 ], "simdgroups": 8, "walk_m": 0, "median_ms": 79.10916698165238, "min_ms": 78.5035421140492, "max_ms": 79.44233296439052, "samples_ms": [ 79.21604183502495, 79.44233296439052, 79.10916698165238, 78.5035421140492, 78.95054202526808 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 128, 128 ], "simdgroups": 8, "walk_m": 8, "median_ms": 44.0463749691844, "min_ms": 42.65099996700883, "max_ms": 44.993416871875525, "samples_ms": [ 42.65099996700883, 43.315166840329766, 44.0463749691844, 44.62029202841222, 44.993416871875525 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 128, 256 ], "simdgroups": 8, "walk_m": 0, "median_ms": 77.51595904119313, "min_ms": 76.97137491777539, "max_ms": 77.76708295568824, "samples_ms": [ 77.76708295568824, 77.75079202838242, 76.97137491777539, 77.29237503372133, 77.51595904119313 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 }, { "operator": "attn.qkv_proj", "tokens": 4181, "tile": [ 64, 128, 256 ], "simdgroups": 8, "walk_m": 8, "median_ms": 42.69133391790092, "min_ms": 41.10400006175041, "max_ms": 43.06004196405411, "samples_ms": [ 41.10400006175041, 41.98370920494199, 42.75595908984542, 42.69133391790092, 43.06004196405411 ], "max_abs": 0.0, "relative_rmse": 0.0, "equal_fraction": 1.0 } ] }