devin-lai's picture
Release v0.4.0: selective MLX model downloads
3da88e3 verified
Raw History Blame Contribute Delete
22.4 kB
{
"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": 7,
"scope": "Original block-0 weights; synthetic BF16 activations. Includes per-call GPU weight reconstruction, excludes initial packing.",
"trials": [
{
"operator": "attn.qkv_proj",
"variant": "bf16",
"shape": [
4181,
5376,
21504
],
"median_ms": 39.11329107359052,
"min_ms": 38.95195806398988,
"max_ms": 39.387915981933475,
"samples_ms": [
39.11329107359052,
39.00787513703108,
39.387915981933475,
39.19341694563627,
39.008582942187786,
39.19645817950368,
38.95195806398988
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "attn.qkv_proj",
"variant": "packed12",
"shape": [
4181,
5376,
21504
],
"pack_s": 0.2350818330887705,
"weight_bytes": 176202178,
"ratio": 0.7620838623738884,
"weight_bits_identical": true,
"escape_fraction": 0.004027948357891334,
"median_ms": 41.564124869182706,
"min_ms": 41.3626660592854,
"max_ms": 41.81537497788668,
"samples_ms": [
41.81537497788668,
41.564124869182706,
41.52962495572865,
41.72691679559648,
41.444499976933,
41.617542039602995,
41.3626660592854
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "attn.qkv_proj",
"variant": "packed12_compiled",
"shape": [
4181,
5376,
21504
],
"median_ms": 41.82958393357694,
"min_ms": 41.48062504827976,
"max_ms": 41.99729207903147,
"samples_ms": [
41.67441697791219,
41.8610000051558,
41.99729207903147,
41.48062504827976,
41.56637494452298,
41.889374842867255,
41.82958393357694
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "attn.qkv_proj",
"variant": "q8_fused",
"shape": [
4181,
5376,
21504
],
"median_ms": 45.41745805181563,
"min_ms": 45.24320783093572,
"max_ms": 45.74962495826185,
"samples_ms": [
45.27574987150729,
45.68949993699789,
45.24320783093572,
45.253500109538436,
45.41745805181563,
45.58362509123981,
45.74962495826185
],
"max_abs": 0.5,
"relative_rmse": 0.007152453996241093,
"equal_fraction": 0.23806069791316986
},
{
"operator": "attn.qkv_proj",
"variant": "q8_dequant_gemm",
"shape": [
4181,
5376,
21504
],
"median_ms": 40.749500039964914,
"min_ms": 40.459583047777414,
"max_ms": 40.915916906669736,
"samples_ms": [
40.459583047777414,
40.79162492416799,
40.66333407536149,
40.915916906669736,
40.749500039964914,
40.86683294735849,
40.717124938964844
],
"max_abs": 0.5,
"relative_rmse": 0.007152453996241093,
"equal_fraction": 0.23806069791316986
},
{
"operator": "attn.qkv_proj",
"variant": "q6_fused",
"shape": [
4181,
5376,
21504
],
"median_ms": 46.02895793505013,
"min_ms": 45.826958026736975,
"max_ms": 46.16720788180828,
"samples_ms": [
45.86966591887176,
46.16720788180828,
46.087624970823526,
45.917625073343515,
45.826958026736975,
46.09595891088247,
46.02895793505013
],
"max_abs": 1.40625,
"relative_rmse": 0.020706482231616974,
"equal_fraction": 0.0854361429810524
},
{
"operator": "attn.qkv_proj",
"variant": "q6_dequant_gemm",
"shape": [
4181,
5376,
21504
],
"median_ms": 40.31316703185439,
"min_ms": 40.03754095174372,
"max_ms": 40.45050009153783,
"samples_ms": [
40.25179194286466,
40.24758283048868,
40.03754095174372,
40.38529214449227,
40.31316703185439,
40.44608282856643,
40.45050009153783
],
"max_abs": 1.40625,
"relative_rmse": 0.020706482231616974,
"equal_fraction": 0.0854361429810524
},
{
"operator": "attn.qkv_proj",
"variant": "q4_fused",
"shape": [
4181,
5376,
21504
],
"median_ms": 42.86429099738598,
"min_ms": 42.5539577845484,
"max_ms": 42.981999926269054,
"samples_ms": [
42.86429099738598,
42.5539577845484,
42.57708298973739,
42.72995796054602,
42.90879191830754,
42.91945789009333,
42.981999926269054
],
"max_abs": 5.60546875,
"relative_rmse": 0.08441592752933502,
"equal_fraction": 0.021116377785801888
},
{
"operator": "attn.qkv_proj",
"variant": "q4_dequant_gemm",
"shape": [
4181,
5376,
21504
],
"median_ms": 40.29725003056228,
"min_ms": 40.13475007377565,
"max_ms": 40.35129211843014,
"samples_ms": [
40.13475007377565,
40.339750004932284,
40.31404200941324,
40.24154180660844,
40.35129211843014,
40.22395913489163,
40.29725003056228
],
"max_abs": 5.60546875,
"relative_rmse": 0.08441592752933502,
"equal_fraction": 0.021116377785801888
},
{
"operator": "attn.out_proj",
"variant": "bf16",
"shape": [
4181,
7168,
5376
],
"median_ms": 13.413249980658293,
"min_ms": 13.186333002522588,
"max_ms": 14.14366695098579,
"samples_ms": [
13.424291042611003,
13.681416166946292,
14.14366695098579,
13.386917067691684,
13.186333002522588,
13.283791951835155,
13.413249980658293
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "attn.out_proj",
"variant": "packed12",
"shape": [
4181,
7168,
5376
],
"pack_s": 0.06456262501887977,
"weight_bytes": 58028596,
"ratio": 0.75293036220836,
"weight_bits_identical": true,
"escape_fraction": 0.0009767701025722789,
"median_ms": 14.004125026986003,
"min_ms": 13.932083966210485,
"max_ms": 14.128583017736673,
"samples_ms": [
14.128583017736673,
13.964999932795763,
14.004125026986003,
13.947749976068735,
13.932083966210485,
14.098458923399448,
14.062334084883332
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "attn.out_proj",
"variant": "packed12_compiled",
"shape": [
4181,
7168,
5376
],
"median_ms": 14.036833075806499,
"min_ms": 13.777584070339799,
"max_ms": 14.199041994288564,
"samples_ms": [
14.036833075806499,
14.028500067070127,
14.091833960264921,
14.036166947335005,
13.777584070339799,
14.133417047560215,
14.199041994288564
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "attn.out_proj",
"variant": "q8_fused",
"shape": [
4181,
7168,
5376
],
"median_ms": 15.811874996870756,
"min_ms": 15.60616702772677,
"max_ms": 15.950917033478618,
"samples_ms": [
15.950917033478618,
15.882750041782856,
15.60616702772677,
15.812082914635539,
15.811874996870756,
15.758000081405044,
15.778542030602694
],
"max_abs": 0.5,
"relative_rmse": 0.006857909727841616,
"equal_fraction": 0.24853673577308655
},
{
"operator": "attn.out_proj",
"variant": "q8_dequant_gemm",
"shape": [
4181,
7168,
5376
],
"median_ms": 13.878457946702838,
"min_ms": 13.743874849751592,
"max_ms": 14.05216590501368,
"samples_ms": [
13.942958088591695,
13.81287514232099,
13.743874849751592,
13.878457946702838,
14.05216590501368,
13.864542124792933,
13.975624926388264
],
"max_abs": 0.5,
"relative_rmse": 0.006857909727841616,
"equal_fraction": 0.24853673577308655
},
{
"operator": "attn.out_proj",
"variant": "q6_fused",
"shape": [
4181,
7168,
5376
],
"median_ms": 15.862250002101064,
"min_ms": 15.767458127811551,
"max_ms": 16.03629207238555,
"samples_ms": [
15.862250002101064,
15.81766689196229,
15.960833989083767,
15.948832966387272,
15.767458127811551,
15.833416022360325,
16.03629207238555
],
"max_abs": 1.25,
"relative_rmse": 0.019887777045369148,
"equal_fraction": 0.08924407511949539
},
{
"operator": "attn.out_proj",
"variant": "q6_dequant_gemm",
"shape": [
4181,
7168,
5376
],
"median_ms": 13.771500205621123,
"min_ms": 13.648458989337087,
"max_ms": 13.843791093677282,
"samples_ms": [
13.771500205621123,
13.843791093677282,
13.777583139017224,
13.748375000432134,
13.73504102230072,
13.648458989337087,
13.793708989396691
],
"max_abs": 1.25,
"relative_rmse": 0.019887777045369148,
"equal_fraction": 0.08924407511949539
},
{
"operator": "attn.out_proj",
"variant": "q4_fused",
"shape": [
4181,
7168,
5376
],
"median_ms": 14.759707963094115,
"min_ms": 14.564833836629987,
"max_ms": 14.91695805452764,
"samples_ms": [
14.81087482534349,
14.86175018362701,
14.627082971855998,
14.91695805452764,
14.759707963094115,
14.564833836629987,
14.668332878500223
],
"max_abs": 5.78125,
"relative_rmse": 0.08107094466686249,
"equal_fraction": 0.022073309868574142
},
{
"operator": "attn.out_proj",
"variant": "q4_dequant_gemm",
"shape": [
4181,
7168,
5376
],
"median_ms": 13.788500102236867,
"min_ms": 13.597874902188778,
"max_ms": 13.935040915384889,
"samples_ms": [
13.892709044739604,
13.923333026468754,
13.597874902188778,
13.788500102236867,
13.668040977790952,
13.742083916440606,
13.935040915384889
],
"max_abs": 5.78125,
"relative_rmse": 0.08107094466686249,
"equal_fraction": 0.022073309868574142
},
{
"operator": "mlp.fc1",
"variant": "bf16",
"shape": [
4181,
5376,
28672
],
"median_ms": 52.165666827932,
"min_ms": 51.81224993430078,
"max_ms": 52.24533309228718,
"samples_ms": [
52.16937488876283,
52.165666827932,
52.24479106254876,
51.92012502811849,
51.90241592936218,
52.24533309228718,
51.81224993430078
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "mlp.fc1",
"variant": "packed12",
"shape": [
4181,
5376,
28672
],
"pack_s": 0.28152816696092486,
"weight_bytes": 233553616,
"ratio": 0.7575989288537681,
"weight_bits_identical": true,
"escape_fraction": 0.0025329719595357674,
"median_ms": 54.94029214605689,
"min_ms": 54.67533296905458,
"max_ms": 55.15158316120505,
"samples_ms": [
54.93524996563792,
54.94029214605689,
55.1342919934541,
55.15158316120505,
54.67533296905458,
54.85250009223819,
54.94895810261369
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "mlp.fc1",
"variant": "packed12_compiled",
"shape": [
4181,
5376,
28672
],
"median_ms": 54.986333940178156,
"min_ms": 54.71341707743704,
"max_ms": 55.26600009761751,
"samples_ms": [
55.13437488116324,
55.26600009761751,
54.84370794147253,
54.986333940178156,
55.15583395026624,
54.784000152722,
54.71341707743704
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "mlp.fc1",
"variant": "q8_fused",
"shape": [
4181,
5376,
28672
],
"median_ms": 60.483790934085846,
"min_ms": 60.2590839844197,
"max_ms": 60.80966698937118,
"samples_ms": [
60.30862499028444,
60.483790934085846,
60.80966698937118,
60.6498341076076,
60.41408306919038,
60.2590839844197,
60.61466713435948
],
"max_abs": 0.6875,
"relative_rmse": 0.00697753531858325,
"equal_fraction": 0.24016006290912628
},
{
"operator": "mlp.fc1",
"variant": "q8_dequant_gemm",
"shape": [
4181,
5376,
28672
],
"median_ms": 54.15787501260638,
"min_ms": 54.041167022660375,
"max_ms": 54.48079202324152,
"samples_ms": [
54.08804211765528,
54.15787501260638,
54.06516697257757,
54.041167022660375,
54.163584019988775,
54.43433299660683,
54.48079202324152
],
"max_abs": 0.6875,
"relative_rmse": 0.00697753531858325,
"equal_fraction": 0.24016006290912628
},
{
"operator": "mlp.fc1",
"variant": "q6_fused",
"shape": [
4181,
5376,
28672
],
"median_ms": 61.271291924640536,
"min_ms": 60.923040844500065,
"max_ms": 61.29683298058808,
"samples_ms": [
61.28195789642632,
61.00579211488366,
61.13933282904327,
61.28733418881893,
60.923040844500065,
61.29683298058808,
61.271291924640536
],
"max_abs": 1.625,
"relative_rmse": 0.020274344831705093,
"equal_fraction": 0.08617563545703888
},
{
"operator": "mlp.fc1",
"variant": "q6_dequant_gemm",
"shape": [
4181,
5376,
28672
],
"median_ms": 53.775500040501356,
"min_ms": 53.66216693073511,
"max_ms": 54.12320885807276,
"samples_ms": [
53.74274984933436,
53.71450004167855,
53.775500040501356,
54.12320885807276,
53.901457926258445,
53.84779185988009,
53.66216693073511
],
"max_abs": 1.625,
"relative_rmse": 0.020274344831705093,
"equal_fraction": 0.08617563545703888
},
{
"operator": "mlp.fc1",
"variant": "q4_fused",
"shape": [
4181,
5376,
28672
],
"median_ms": 57.11166700348258,
"min_ms": 56.81737489067018,
"max_ms": 57.686541927978396,
"samples_ms": [
56.81737489067018,
57.11166700348258,
56.91687506623566,
57.09958402439952,
57.40487505681813,
57.686541927978396,
57.62820807285607
],
"max_abs": 7.21875,
"relative_rmse": 0.08270496875047684,
"equal_fraction": 0.021269498392939568
},
{
"operator": "mlp.fc1",
"variant": "q4_dequant_gemm",
"shape": [
4181,
5376,
28672
],
"median_ms": 54.25675003789365,
"min_ms": 53.82283288054168,
"max_ms": 54.87637501209974,
"samples_ms": [
53.82283288054168,
54.06649992801249,
54.25675003789365,
53.96083300001919,
54.30883285589516,
54.32837479747832,
54.87637501209974
],
"max_abs": 7.21875,
"relative_rmse": 0.08270496875047684,
"equal_fraction": 0.021269498392939568
},
{
"operator": "mlp.fc2",
"variant": "bf16",
"shape": [
4181,
14336,
5376
],
"median_ms": 32.67379105091095,
"min_ms": 32.297125086188316,
"max_ms": 33.28745788894594,
"samples_ms": [
33.28745788894594,
32.67379105091095,
32.43308397941291,
32.297125086188316,
33.098625019192696,
32.57412509992719,
32.885042019188404
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "mlp.fc2",
"variant": "packed12",
"shape": [
4181,
14336,
5376
],
"pack_s": 0.128432875033468,
"weight_bytes": 116549002,
"ratio": 0.7561210191168752,
"weight_bits_identical": true,
"escape_fraction": 0.0020403310555179104,
"median_ms": 34.36666703782976,
"min_ms": 33.84170797653496,
"max_ms": 34.62899988517165,
"samples_ms": [
34.23387510702014,
34.372332971543074,
34.62899988517165,
34.4923329539597,
33.84170797653496,
34.36666703782976,
33.9440000243485
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "mlp.fc2",
"variant": "packed12_compiled",
"shape": [
4181,
14336,
5376
],
"median_ms": 34.0739160310477,
"min_ms": 33.992790849879384,
"max_ms": 34.83095811679959,
"samples_ms": [
34.31983385235071,
34.83095811679959,
34.06149987131357,
33.992790849879384,
34.204334020614624,
34.0739160310477,
34.02716713026166
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "mlp.fc2",
"variant": "q8_fused",
"shape": [
4181,
14336,
5376
],
"median_ms": 34.68745783902705,
"min_ms": 34.13212485611439,
"max_ms": 35.295374924317,
"samples_ms": [
34.59445899352431,
34.13212485611439,
35.0765420589596,
34.826042130589485,
34.68745783902705,
35.295374924317,
34.50066689401865
],
"max_abs": 1.125,
"relative_rmse": 0.007229790091514587,
"equal_fraction": 0.23662720620632172
},
{
"operator": "mlp.fc2",
"variant": "q8_dequant_gemm",
"shape": [
4181,
14336,
5376
],
"median_ms": 33.93920813687146,
"min_ms": 33.23116712272167,
"max_ms": 34.14066694676876,
"samples_ms": [
34.02454196475446,
33.23883400298655,
34.019374987110496,
33.81558391265571,
33.93920813687146,
33.23116712272167,
34.14066694676876
],
"max_abs": 1.125,
"relative_rmse": 0.007229790091514587,
"equal_fraction": 0.23662720620632172
},
{
"operator": "mlp.fc2",
"variant": "q6_fused",
"shape": [
4181,
14336,
5376
],
"median_ms": 34.468917176127434,
"min_ms": 34.02270912192762,
"max_ms": 34.86333298496902,
"samples_ms": [
34.86333298496902,
34.098458010703325,
34.02270912192762,
34.199790796265006,
34.74954189732671,
34.468917176127434,
34.50925019569695
],
"max_abs": 3.5,
"relative_rmse": 0.021220030263066292,
"equal_fraction": 0.08387544006109238
},
{
"operator": "mlp.fc2",
"variant": "q6_dequant_gemm",
"shape": [
4181,
14336,
5376
],
"median_ms": 33.64562499336898,
"min_ms": 33.47887517884374,
"max_ms": 34.06262514181435,
"samples_ms": [
33.79079094156623,
34.06262514181435,
33.64562499336898,
33.51258300244808,
33.47887517884374,
33.56566606089473,
33.88137510046363
],
"max_abs": 3.5,
"relative_rmse": 0.021220030263066292,
"equal_fraction": 0.08387544006109238
},
{
"operator": "mlp.fc2",
"variant": "q4_fused",
"shape": [
4181,
14336,
5376
],
"median_ms": 31.817874871194363,
"min_ms": 31.516709132120013,
"max_ms": 32.1920420974493,
"samples_ms": [
31.516709132120013,
32.19029190950096,
31.671958975493908,
31.74712508916855,
32.1920420974493,
32.01295807957649,
31.817874871194363
],
"max_abs": 13.5,
"relative_rmse": 0.0866343304514885,
"equal_fraction": 0.020664894953370094
},
{
"operator": "mlp.fc2",
"variant": "q4_dequant_gemm",
"shape": [
4181,
14336,
5376
],
"median_ms": 33.572624903172255,
"min_ms": 33.22533308528364,
"max_ms": 33.93350006081164,
"samples_ms": [
33.38700020685792,
33.81758299656212,
33.42291712760925,
33.93350006081164,
33.572624903172255,
33.611832885071635,
33.22533308528364
],
"max_abs": 13.5,
"relative_rmse": 0.0866343304514885,
"equal_fraction": 0.020664894953370094
}
]
}