Text-to-Image
MLX
MiniMax H3
Safetensors
English
apple-silicon
macos
image-to-image
image-editing
image-generation
Instructions to use devin-lai/MiniMax-H3-Image-mlx with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use devin-lai/MiniMax-H3-Image-mlx with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] hf download devin-lai/MiniMax-H3-Image-mlx --local-dir MiniMax-H3-Image-mlx
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
Download reports/operators.json from devin-lai/MiniMax-H3-Image-mlx: direct link, hf CLI and curl.
- Browser
- Download file 36.3 kB
-
https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/resolve/main/reports/operators.json
- Command line
-
hf download hf://devin-lai/MiniMax-H3-Image-mlx/reports/operators.json
-
curl -L -o operators.json https://huggingface.co/devin-lai/MiniMax-H3-Image-mlx/resolve/main/reports/operators.json
36.3 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"
},
"precision_environment": {
"MLX_ENABLE_TF32": "0",
"MLX_CONV_WINOGRAD": "0"
}
},
"warmup": 2,
"repeats": 7,
"scope": "Synthetic operator microbenchmarks; excludes loading/quantization. Not image latency.",
"trials": [
{
"operator": "rms_adaln",
"variant": "native_rounding",
"shape": [
256,
5376
],
"median_ms": 1.173374941572547,
"min_ms": 0.616749981418252,
"max_ms": 1.2301669921725988,
"samples_ms": [
0.616749981418252,
1.2266251724213362,
1.0637501254677773,
1.173374941572547,
1.2301669921725988,
1.195583026856184,
1.1052920017391443
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "rms_adaln",
"variant": "metal_tg128",
"shape": [
256,
5376
],
"median_ms": 0.5940420087426901,
"min_ms": 0.5400839727371931,
"max_ms": 0.7085409015417099,
"samples_ms": [
0.7085409015417099,
0.5697500891983509,
0.5824591498821974,
0.5400839727371931,
0.6797090172767639,
0.5940420087426901,
0.609041890129447
],
"max_abs": 0.015625,
"relative_rmse": 1.7477590517955832e-05,
"equal_fraction": 0.9999935030937195
},
{
"operator": "rms_adaln",
"variant": "metal_tg256",
"shape": [
256,
5376
],
"median_ms": 0.5604161415249109,
"min_ms": 0.2424169797450304,
"max_ms": 0.6824578158557415,
"samples_ms": [
0.6682500243186951,
0.5432919133454561,
0.6824578158557415,
0.6537078879773617,
0.5389170255511999,
0.5604161415249109,
0.2424169797450304
],
"max_abs": 0.015625,
"relative_rmse": 2.209606464020908e-05,
"equal_fraction": 0.9999912977218628
},
{
"operator": "rms_adaln",
"variant": "metal_tg512",
"shape": [
256,
5376
],
"median_ms": 0.8023330010473728,
"min_ms": 0.19370880909264088,
"max_ms": 0.8323339279741049,
"samples_ms": [
0.8323339279741049,
0.8083749562501907,
0.8090410847216845,
0.7428331300616264,
0.8023330010473728,
0.19370880909264088,
0.7230001501739025
],
"max_abs": 0.015625,
"relative_rmse": 1.8376344087300822e-05,
"equal_fraction": 0.9999912977218628
},
{
"operator": "qk_norm_rope",
"variant": "native",
"shape": [
256,
56,
128
],
"median_ms": 1.1400829534977674,
"min_ms": 0.6993331480771303,
"max_ms": 1.366166863590479,
"samples_ms": [
0.7345420308411121,
0.7162499241530895,
0.6993331480771303,
1.2012079823762178,
1.366166863590479,
1.1400829534977674,
1.2222910299897194
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "qk_norm_rope",
"variant": "metal",
"shape": [
256,
56,
128
],
"median_ms": 0.26829191483557224,
"min_ms": 0.24679210036993027,
"max_ms": 0.9639160707592964,
"samples_ms": [
0.9639160707592964,
0.848833005875349,
0.8924168068915606,
0.26829191483557224,
0.24679210036993027,
0.264165922999382,
0.2524999435991049
],
"max_abs": 0.00390625,
"relative_rmse": 3.6376052321429597e-06,
"equal_fraction": 0.9999983906745911
},
{
"operator": "gemm",
"variant": "bf16",
"shape": [
256,
5376,
21504
],
"median_ms": 2.5721671991050243,
"min_ms": 2.5016670115292072,
"max_ms": 2.6360841002315283,
"samples_ms": [
2.5721671991050243,
2.580333035439253,
2.587541937828064,
2.5619592051953077,
2.550124889239669,
2.6360841002315283,
2.5016670115292072
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "gemm",
"variant": "q8_g32",
"shape": [
256,
5376,
21504
],
"median_ms": 3.0226248782128096,
"min_ms": 3.0050419736653566,
"max_ms": 3.1408751383423805,
"samples_ms": [
3.018749877810478,
3.040374955162406,
3.0050419736653566,
3.1321251299232244,
3.1408751383423805,
3.0226248782128096,
3.0074999667704105
],
"max_abs": 0.0625,
"relative_rmse": 0.006850533653050661,
"equal_fraction": 0.24900999665260315
},
{
"operator": "gemm",
"variant": "q8_g64",
"shape": [
256,
5376,
21504
],
"median_ms": 2.9889577999711037,
"min_ms": 2.945707878097892,
"max_ms": 3.2797919120639563,
"samples_ms": [
3.110166173428297,
3.0175000429153442,
3.2797919120639563,
2.9889577999711037,
2.970875008031726,
2.945707878097892,
2.9832920990884304
],
"max_abs": 0.0625,
"relative_rmse": 0.007352763321250677,
"equal_fraction": 0.23354558646678925
},
{
"operator": "gemm",
"variant": "q8_g128",
"shape": [
256,
5376,
21504
],
"median_ms": 2.979417098686099,
"min_ms": 2.963416976854205,
"max_ms": 3.222458064556122,
"samples_ms": [
2.9804580844938755,
3.222458064556122,
3.1075000297278166,
2.969249850139022,
2.963416976854205,
2.979417098686099,
2.9705839697271585
],
"max_abs": 0.0625,
"relative_rmse": 0.007834218442440033,
"equal_fraction": 0.22005045413970947
},
{
"operator": "gemm",
"variant": "q4_g32",
"shape": [
256,
5376,
21504
],
"median_ms": 2.7746250852942467,
"min_ms": 2.7560831513255835,
"max_ms": 3.058916889131069,
"samples_ms": [
3.058916889131069,
2.7560831513255835,
2.7758749201893806,
2.8548329137265682,
2.7711251750588417,
2.7608750388026237,
2.7746250852942467
],
"max_abs": 0.64453125,
"relative_rmse": 0.08106927573680878,
"equal_fraction": 0.0220609400421381
},
{
"operator": "gemm",
"variant": "q4_g64",
"shape": [
256,
5376,
21504
],
"median_ms": 2.7238749898970127,
"min_ms": 2.707334002479911,
"max_ms": 2.966000000014901,
"samples_ms": [
2.966000000014901,
2.7539171278476715,
2.7230000123381615,
2.7238749898970127,
2.711624838411808,
2.7322079986333847,
2.707334002479911
],
"max_abs": 0.6953125,
"relative_rmse": 0.0909692719578743,
"equal_fraction": 0.01975395530462265
},
{
"operator": "gemm",
"variant": "q4_g128",
"shape": [
256,
5376,
21504
],
"median_ms": 2.727207960560918,
"min_ms": 2.6830830611288548,
"max_ms": 2.914041979238391,
"samples_ms": [
2.9036658816039562,
2.914041979238391,
2.7150840032845736,
2.8310411144047976,
2.727207960560918,
2.6830830611288548,
2.6853750459849834
],
"max_abs": 0.7509765625,
"relative_rmse": 0.09998571127653122,
"equal_fraction": 0.017951058223843575
},
{
"operator": "gemm",
"variant": "bf16",
"shape": [
256,
5376,
28672
],
"median_ms": 3.347999881953001,
"min_ms": 3.28362500295043,
"max_ms": 3.5925419069826603,
"samples_ms": [
3.347999881953001,
3.2983338460326195,
3.5925419069826603,
3.4004999324679375,
3.321249969303608,
3.3945830073207617,
3.28362500295043
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "gemm",
"variant": "q8_g32",
"shape": [
256,
5376,
28672
],
"median_ms": 3.924708114936948,
"min_ms": 3.8868330884724855,
"max_ms": 3.9580839220434427,
"samples_ms": [
3.924708114936948,
3.9580839220434427,
3.9263751823455095,
3.929541911929846,
3.8885409012436867,
3.8868330884724855,
3.892374923452735
],
"max_abs": 0.0625,
"relative_rmse": 0.0068595209158957005,
"equal_fraction": 0.24881459772586823
},
{
"operator": "gemm",
"variant": "q8_g64",
"shape": [
256,
5376,
28672
],
"median_ms": 3.9282499346882105,
"min_ms": 3.88133293017745,
"max_ms": 4.043041029945016,
"samples_ms": [
3.9579591248184443,
3.8992909248918295,
3.886499907821417,
4.043041029945016,
3.963541006669402,
3.88133293017745,
3.9282499346882105
],
"max_abs": 0.0625,
"relative_rmse": 0.007344856392592192,
"equal_fraction": 0.23325484991073608
},
{
"operator": "gemm",
"variant": "q8_g128",
"shape": [
256,
5376,
28672
],
"median_ms": 3.890999825671315,
"min_ms": 3.86029202491045,
"max_ms": 4.0629159193485975,
"samples_ms": [
4.019708139821887,
3.86029202491045,
3.8872910663485527,
3.864959115162492,
4.0629159193485975,
3.9462910499423742,
3.890999825671315
],
"max_abs": 0.0625,
"relative_rmse": 0.007839743979275227,
"equal_fraction": 0.21977084875106812
},
{
"operator": "gemm",
"variant": "q4_g32",
"shape": [
256,
5376,
28672
],
"median_ms": 3.6355829797685146,
"min_ms": 3.614082932472229,
"max_ms": 3.8034170866012573,
"samples_ms": [
3.6440829280763865,
3.622000105679035,
3.663749899715185,
3.6355829797685146,
3.6264171358197927,
3.8034170866012573,
3.614082932472229
],
"max_abs": 0.63214111328125,
"relative_rmse": 0.08107927441596985,
"equal_fraction": 0.02212129347026348
},
{
"operator": "gemm",
"variant": "q4_g64",
"shape": [
256,
5376,
28672
],
"median_ms": 3.5808749962598085,
"min_ms": 3.555125091224909,
"max_ms": 3.729666117578745,
"samples_ms": [
3.5808749962598085,
3.5678748972713947,
3.729666117578745,
3.6466249730437994,
3.5678751301020384,
3.6204170901328325,
3.555125091224909
],
"max_abs": 0.7109375,
"relative_rmse": 0.09099964797496796,
"equal_fraction": 0.01962498389184475
},
{
"operator": "gemm",
"variant": "q4_g128",
"shape": [
256,
5376,
28672
],
"median_ms": 3.5667498596012592,
"min_ms": 3.5494170151650906,
"max_ms": 3.6764999385923147,
"samples_ms": [
3.6764999385923147,
3.553125075995922,
3.5667498596012592,
3.5774579737335443,
3.5494170151650906,
3.554082941263914,
3.568042069673538
],
"max_abs": 0.84375,
"relative_rmse": 0.10004469007253647,
"equal_fraction": 0.017870495095849037
},
{
"operator": "gemm",
"variant": "bf16",
"shape": [
256,
14336,
5376
],
"median_ms": 2.098832977935672,
"min_ms": 2.0805420354008675,
"max_ms": 2.1304998081177473,
"samples_ms": [
2.113000024110079,
2.098832977935672,
2.1304998081177473,
2.094374969601631,
2.0852498710155487,
2.1142091136425734,
2.0805420354008675
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "gemm",
"variant": "q8_g32",
"shape": [
256,
14336,
5376
],
"median_ms": 2.3393749725073576,
"min_ms": 2.302875043824315,
"max_ms": 2.5676661171019077,
"samples_ms": [
2.3640000727027655,
2.5676661171019077,
2.556499792262912,
2.3393749725073576,
2.338541904464364,
2.302875043824315,
2.3251669481396675
],
"max_abs": 0.0859375,
"relative_rmse": 0.006851653102785349,
"equal_fraction": 0.2485344409942627
},
{
"operator": "gemm",
"variant": "q8_g64",
"shape": [
256,
14336,
5376
],
"median_ms": 2.398208947852254,
"min_ms": 2.242250135168433,
"max_ms": 2.6454171165823936,
"samples_ms": [
2.398208947852254,
2.2548330016434193,
2.5238750968128443,
2.6454171165823936,
2.398416865617037,
2.37491587176919,
2.242250135168433
],
"max_abs": 0.09375,
"relative_rmse": 0.007338221650570631,
"equal_fraction": 0.2334180623292923
},
{
"operator": "gemm",
"variant": "q8_g128",
"shape": [
256,
14336,
5376
],
"median_ms": 2.2256250958889723,
"min_ms": 2.209833823144436,
"max_ms": 2.2738338448107243,
"samples_ms": [
2.2117909975349903,
2.214916981756687,
2.209833823144436,
2.263291971758008,
2.2330000065267086,
2.2256250958889723,
2.2738338448107243
],
"max_abs": 0.09375,
"relative_rmse": 0.007827789522707462,
"equal_fraction": 0.2200208455324173
},
{
"operator": "gemm",
"variant": "q4_g32",
"shape": [
256,
14336,
5376
],
"median_ms": 2.159459050744772,
"min_ms": 2.059374935925007,
"max_ms": 2.193290973082185,
"samples_ms": [
2.059374935925007,
2.18175002373755,
2.159459050744772,
2.075666096061468,
2.193290973082185,
2.122832927852869,
2.159666968509555
],
"max_abs": 0.9296875,
"relative_rmse": 0.08106718212366104,
"equal_fraction": 0.02191670797765255
},
{
"operator": "gemm",
"variant": "q4_g64",
"shape": [
256,
14336,
5376
],
"median_ms": 2.067500026896596,
"min_ms": 1.9949160050600767,
"max_ms": 2.128707943484187,
"samples_ms": [
2.067500026896596,
2.0210829097777605,
2.096999902278185,
2.128707943484187,
1.9949160050600767,
2.0117908716201782,
2.0698329899460077
],
"max_abs": 1.0390625,
"relative_rmse": 0.09107401221990585,
"equal_fraction": 0.0195334292948246
},
{
"operator": "gemm",
"variant": "q4_g128",
"shape": [
256,
14336,
5376
],
"median_ms": 2.079749945551157,
"min_ms": 2.0619581919163465,
"max_ms": 2.380209043622017,
"samples_ms": [
2.0619581919163465,
2.0739170722663403,
2.1086670458316803,
2.079749945551157,
2.380209043622017,
2.083666855469346,
2.062832936644554
],
"max_abs": 1.171875,
"relative_rmse": 0.10009276866912842,
"equal_fraction": 0.01786004938185215
},
{
"operator": "sdpa",
"variant": "query_chunk_0",
"shape": [
1,
256,
56,
128
],
"median_ms": 0.32275007106363773,
"min_ms": 0.31208316795527935,
"max_ms": 0.3370409831404686,
"samples_ms": [
0.3233749885112047,
0.32275007106363773,
0.31208316795527935,
0.31349994242191315,
0.33125001937150955,
0.3370409831404686,
0.31766598112881184
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "sdpa",
"variant": "query_chunk_128",
"shape": [
1,
256,
56,
128
],
"median_ms": 0.6250408478081226,
"min_ms": 0.5880000535398722,
"max_ms": 0.684875063598156,
"samples_ms": [
0.684875063598156,
0.6232080049812794,
0.6250408478081226,
0.6210829596966505,
0.631541945040226,
0.632166862487793,
0.5880000535398722
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "sdpa",
"variant": "query_chunk_512",
"shape": [
1,
256,
56,
128
],
"median_ms": 0.28816587291657925,
"min_ms": 0.25708298198878765,
"max_ms": 0.2957081887871027,
"samples_ms": [
0.2957081887871027,
0.29191700741648674,
0.28900010511279106,
0.27449987828731537,
0.28816587291657925,
0.2626250497996807,
0.25708298198878765
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "sdpa",
"variant": "query_chunk_1024",
"shape": [
1,
256,
56,
128
],
"median_ms": 0.2906669396907091,
"min_ms": 0.2794580068439245,
"max_ms": 0.3147500101476908,
"samples_ms": [
0.27962494641542435,
0.2935829106718302,
0.29037497006356716,
0.3147500101476908,
0.2794580068439245,
0.2906669396907091,
0.2928751055151224
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "rms_adaln",
"variant": "native_rounding",
"shape": [
4096,
5376
],
"median_ms": 3.5581670235842466,
"min_ms": 3.5174579825252295,
"max_ms": 3.779958002269268,
"samples_ms": [
3.5195830278098583,
3.706582821905613,
3.779958002269268,
3.5999580286443233,
3.5581670235842466,
3.530999878421426,
3.5174579825252295
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "rms_adaln",
"variant": "metal_tg128",
"shape": [
4096,
5376
],
"median_ms": 0.5828330758959055,
"min_ms": 0.5625830963253975,
"max_ms": 0.7374170236289501,
"samples_ms": [
0.5748749244958162,
0.6655419711023569,
0.5828330758959055,
0.5625830963253975,
0.7374170236289501,
0.5845420528203249,
0.5679160822182894
],
"max_abs": 0.03125,
"relative_rmse": 1.081336540664779e-05,
"equal_fraction": 0.9999977350234985
},
{
"operator": "rms_adaln",
"variant": "metal_tg256",
"shape": [
4096,
5376
],
"median_ms": 0.542541965842247,
"min_ms": 0.5144169554114342,
"max_ms": 0.5596249829977751,
"samples_ms": [
0.547375064343214,
0.5328329280018806,
0.5336250178515911,
0.5144169554114342,
0.5596249829977751,
0.542541965842247,
0.5430420860648155
],
"max_abs": 0.015625,
"relative_rmse": 9.978849448089022e-06,
"equal_fraction": 0.9999976754188538
},
{
"operator": "rms_adaln",
"variant": "metal_tg512",
"shape": [
4096,
5376
],
"median_ms": 0.5379160866141319,
"min_ms": 0.5112909711897373,
"max_ms": 0.5407920107245445,
"samples_ms": [
0.5379160866141319,
0.5390408914536238,
0.5112909711897373,
0.5176248960196972,
0.5320410709828138,
0.5397088825702667,
0.5407920107245445
],
"max_abs": 0.03125,
"relative_rmse": 1.3310983376868535e-05,
"equal_fraction": 0.9999973177909851
},
{
"operator": "qk_norm_rope",
"variant": "native",
"shape": [
4096,
56,
128
],
"median_ms": 10.960625018924475,
"min_ms": 10.198707925155759,
"max_ms": 13.119417009875178,
"samples_ms": [
13.119417009875178,
12.153874849900603,
11.330457869917154,
10.960625018924475,
10.477333795279264,
10.198707925155759,
10.489040985703468
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "qk_norm_rope",
"variant": "metal",
"shape": [
4096,
56,
128
],
"median_ms": 2.057416830211878,
"min_ms": 2.032166812568903,
"max_ms": 2.1125830244272947,
"samples_ms": [
2.1125830244272947,
2.032166812568903,
2.0543329883366823,
2.108125016093254,
2.0634999964386225,
2.057416830211878,
2.04420811496675
],
"max_abs": 0.015625,
"relative_rmse": 9.562711966282222e-06,
"equal_fraction": 0.9999966621398926
},
{
"operator": "gemm",
"variant": "bf16",
"shape": [
4096,
5376,
21504
],
"median_ms": 37.58908296003938,
"min_ms": 37.561249919235706,
"max_ms": 37.69441717304289,
"samples_ms": [
37.66870801337063,
37.616125075146556,
37.69441717304289,
37.58908296003938,
37.573833018541336,
37.58049989119172,
37.561249919235706
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "gemm",
"variant": "q8_g32",
"shape": [
4096,
5376,
21504
],
"median_ms": 43.576583033427596,
"min_ms": 43.46295795403421,
"max_ms": 43.9285000320524,
"samples_ms": [
43.9285000320524,
43.55879197828472,
43.46295795403421,
43.5891249217093,
43.576583033427596,
43.54733298532665,
43.7540418934077
],
"max_abs": 0.0625,
"relative_rmse": 0.006848405115306377,
"equal_fraction": 0.24890799820423126
},
{
"operator": "gemm",
"variant": "q8_g64",
"shape": [
4096,
5376,
21504
],
"median_ms": 43.335665948688984,
"min_ms": 42.99841704778373,
"max_ms": 43.70733303949237,
"samples_ms": [
43.32091589458287,
42.99841704778373,
43.70733303949237,
43.48420910537243,
43.335665948688984,
43.42833301052451,
43.065707897767425
],
"max_abs": 0.0625,
"relative_rmse": 0.007345186080783606,
"equal_fraction": 0.2333000749349594
},
{
"operator": "gemm",
"variant": "q8_g128",
"shape": [
4096,
5376,
21504
],
"median_ms": 43.26391592621803,
"min_ms": 42.786624981090426,
"max_ms": 43.67449996061623,
"samples_ms": [
43.431750033050776,
43.67449996061623,
43.26391592621803,
42.962417006492615,
42.786624981090426,
43.42845897190273,
43.2283750269562
],
"max_abs": 0.0703125,
"relative_rmse": 0.00781948957592249,
"equal_fraction": 0.22026407718658447
},
{
"operator": "gemm",
"variant": "q4_g32",
"shape": [
4096,
5376,
21504
],
"median_ms": 41.20029206387699,
"min_ms": 41.03312501683831,
"max_ms": 41.38141707517207,
"samples_ms": [
41.1425840575248,
41.38141707517207,
41.20029206387699,
41.150250006467104,
41.20558290742338,
41.33666702546179,
41.03312501683831
],
"max_abs": 0.71875,
"relative_rmse": 0.0810740664601326,
"equal_fraction": 0.022041741758584976
},
{
"operator": "gemm",
"variant": "q4_g64",
"shape": [
4096,
5376,
21504
],
"median_ms": 40.671000024303794,
"min_ms": 40.518667083233595,
"max_ms": 40.88425007648766,
"samples_ms": [
40.66379205323756,
40.611709002405405,
40.518667083233595,
40.88425007648766,
40.671000024303794,
40.70158419199288,
40.81641719676554
],
"max_abs": 0.763671875,
"relative_rmse": 0.09098212420940399,
"equal_fraction": 0.019682418555021286
},
{
"operator": "gemm",
"variant": "q4_g128",
"shape": [
4096,
5376,
21504
],
"median_ms": 40.219750022515655,
"min_ms": 40.07566696964204,
"max_ms": 40.326499845832586,
"samples_ms": [
40.313333040103316,
40.326499845832586,
40.219750022515655,
40.076584089547396,
40.18954187631607,
40.07566696964204,
40.235500084236264
],
"max_abs": 0.861328125,
"relative_rmse": 0.10006927698850632,
"equal_fraction": 0.01790381595492363
},
{
"operator": "gemm",
"variant": "bf16",
"shape": [
4096,
5376,
28672
],
"median_ms": 50.183625193312764,
"min_ms": 50.01025018282235,
"max_ms": 50.473124952986836,
"samples_ms": [
50.073042046278715,
50.473124952986836,
50.01025018282235,
50.22287485189736,
50.05783401429653,
50.183625193312764,
50.22233398631215
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "gemm",
"variant": "q8_g32",
"shape": [
4096,
5376,
28672
],
"median_ms": 58.16854094155133,
"min_ms": 57.83437518402934,
"max_ms": 58.548124972730875,
"samples_ms": [
58.37570782750845,
58.548124972730875,
58.050625026226044,
58.243790874257684,
58.11225017532706,
58.16854094155133,
57.83437518402934
],
"max_abs": 0.0625,
"relative_rmse": 0.0068477122113108635,
"equal_fraction": 0.2489926517009735
},
{
"operator": "gemm",
"variant": "q8_g64",
"shape": [
4096,
5376,
28672
],
"median_ms": 57.96945793554187,
"min_ms": 57.74970795027912,
"max_ms": 58.27745795249939,
"samples_ms": [
58.05641692131758,
57.74970795027912,
57.96945793554187,
57.95050016604364,
58.27745795249939,
57.87858390249312,
57.978666853159666
],
"max_abs": 0.0625,
"relative_rmse": 0.007344013080000877,
"equal_fraction": 0.23339642584323883
},
{
"operator": "gemm",
"variant": "q8_g128",
"shape": [
4096,
5376,
28672
],
"median_ms": 58.31704195588827,
"min_ms": 57.85416602157056,
"max_ms": 58.81566694006324,
"samples_ms": [
57.94695811346173,
57.85416602157056,
58.45699994824827,
58.159250067546964,
58.31704195588827,
58.81566694006324,
58.56541683897376
],
"max_abs": 0.06640625,
"relative_rmse": 0.00781780295073986,
"equal_fraction": 0.2203386276960373
},
{
"operator": "gemm",
"variant": "q4_g32",
"shape": [
4096,
5376,
28672
],
"median_ms": 55.75945810414851,
"min_ms": 54.988875053822994,
"max_ms": 56.057958863675594,
"samples_ms": [
54.988875053822994,
55.60045805759728,
55.75945810414851,
55.441749980673194,
56.057958863675594,
55.944374995306134,
55.823208997026086
],
"max_abs": 0.6640625,
"relative_rmse": 0.08107572048902512,
"equal_fraction": 0.022053532302379608
},
{
"operator": "gemm",
"variant": "q4_g64",
"shape": [
4096,
5376,
28672
],
"median_ms": 55.38108292967081,
"min_ms": 54.57354104146361,
"max_ms": 55.79287512227893,
"samples_ms": [
54.57354104146361,
55.38108292967081,
55.60295912437141,
55.10041699744761,
55.79287512227893,
55.68016692996025,
55.32279098406434
],
"max_abs": 0.763916015625,
"relative_rmse": 0.09099233150482178,
"equal_fraction": 0.019668912515044212
},
{
"operator": "gemm",
"variant": "q4_g128",
"shape": [
4096,
5376,
28672
],
"median_ms": 55.30212516896427,
"min_ms": 54.67195808887482,
"max_ms": 55.79587514512241,
"samples_ms": [
54.813416907563806,
54.67195808887482,
55.29566598124802,
55.30212516896427,
55.365709122270346,
55.59012503363192,
55.79587514512241
],
"max_abs": 0.853515625,
"relative_rmse": 0.10007014870643616,
"equal_fraction": 0.017912209033966064
},
{
"operator": "gemm",
"variant": "bf16",
"shape": [
4096,
14336,
5376
],
"median_ms": 29.900042107328773,
"min_ms": 28.89745798893273,
"max_ms": 30.47887491993606,
"samples_ms": [
30.47887491993606,
29.908499913290143,
29.61533283814788,
28.89745798893273,
29.382749926298857,
29.900042107328773,
30.169625068083405
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "gemm",
"variant": "q8_g32",
"shape": [
4096,
14336,
5376
],
"median_ms": 32.60995796881616,
"min_ms": 32.3443750385195,
"max_ms": 33.24191691353917,
"samples_ms": [
32.51500008627772,
32.3443750385195,
32.59654110297561,
32.8078749589622,
32.94933307915926,
32.60995796881616,
33.24191691353917
],
"max_abs": 0.09375,
"relative_rmse": 0.006849921308457851,
"equal_fraction": 0.24884523451328278
},
{
"operator": "gemm",
"variant": "q8_g64",
"shape": [
4096,
14336,
5376
],
"median_ms": 32.50916698016226,
"min_ms": 31.81283315643668,
"max_ms": 33.11962494626641,
"samples_ms": [
32.13399997912347,
31.956125050783157,
31.81283315643668,
32.66279213130474,
32.50916698016226,
33.11962494626641,
32.73437498137355
],
"max_abs": 0.09375,
"relative_rmse": 0.0073441751301288605,
"equal_fraction": 0.23339390754699707
},
{
"operator": "gemm",
"variant": "q8_g128",
"shape": [
4096,
14336,
5376
],
"median_ms": 32.53870899789035,
"min_ms": 31.987624941393733,
"max_ms": 32.83950011245906,
"samples_ms": [
32.462208066135645,
32.62395807541907,
32.83950011245906,
32.55154099315405,
32.53870899789035,
32.40154217928648,
31.987624941393733
],
"max_abs": 0.125,
"relative_rmse": 0.007824343629181385,
"equal_fraction": 0.22009055316448212
},
{
"operator": "gemm",
"variant": "q4_g32",
"shape": [
4096,
14336,
5376
],
"median_ms": 30.538500053808093,
"min_ms": 29.95408303104341,
"max_ms": 30.80308297649026,
"samples_ms": [
30.08712502196431,
29.95408303104341,
30.80308297649026,
30.737000051885843,
30.45966592617333,
30.538500053808093,
30.643749982118607
],
"max_abs": 1.1328125,
"relative_rmse": 0.08104117959737778,
"equal_fraction": 0.02208491787314415
},
{
"operator": "gemm",
"variant": "q4_g64",
"shape": [
4096,
14336,
5376
],
"median_ms": 29.812082881107926,
"min_ms": 29.65637482702732,
"max_ms": 30.311124864965677,
"samples_ms": [
29.684999957680702,
29.65637482702732,
29.743541963398457,
29.951457865536213,
29.812082881107926,
30.01641691662371,
30.311124864965677
],
"max_abs": 1.18359375,
"relative_rmse": 0.09096471965312958,
"equal_fraction": 0.019711222499608994
},
{
"operator": "gemm",
"variant": "q4_g128",
"shape": [
4096,
14336,
5376
],
"median_ms": 29.769500019028783,
"min_ms": 29.42449995316565,
"max_ms": 29.965666122734547,
"samples_ms": [
29.884499963372946,
29.42449995316565,
29.77370796725154,
29.965666122734547,
29.769500019028783,
29.60404218174517,
29.57754209637642
],
"max_abs": 1.32421875,
"relative_rmse": 0.10005950182676315,
"equal_fraction": 0.01792476326227188
},
{
"operator": "sdpa",
"variant": "query_chunk_0",
"shape": [
1,
4096,
56,
128
],
"median_ms": 24.146165931597352,
"min_ms": 23.585624992847443,
"max_ms": 24.465749971568584,
"samples_ms": [
23.585624992847443,
23.93633290193975,
24.196083890274167,
24.06216599047184,
24.146165931597352,
24.465749971568584,
24.197667138651013
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "sdpa",
"variant": "query_chunk_128",
"shape": [
1,
4096,
56,
128
],
"median_ms": 41.21925006620586,
"min_ms": 41.048833169043064,
"max_ms": 41.29937500692904,
"samples_ms": [
41.06008284725249,
41.21925006620586,
41.29687510430813,
41.25925013795495,
41.29937500692904,
41.048833169043064,
41.09108285047114
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "sdpa",
"variant": "query_chunk_512",
"shape": [
1,
4096,
56,
128
],
"median_ms": 26.095834095031023,
"min_ms": 25.888917036354542,
"max_ms": 26.268708053976297,
"samples_ms": [
25.888917036354542,
26.00116585381329,
25.939542101696134,
26.095834095031023,
26.154958875849843,
26.268708053976297,
26.120957918465137
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "sdpa",
"variant": "query_chunk_1024",
"shape": [
1,
4096,
56,
128
],
"median_ms": 25.337791070342064,
"min_ms": 25.203832890838385,
"max_ms": 25.563500123098493,
"samples_ms": [
25.29804199002683,
25.253874948248267,
25.337791070342064,
25.203832890838385,
25.563500123098493,
25.462209014222026,
25.39620897732675
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
}
]
}
{
"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"
},
"precision_environment": {
"MLX_ENABLE_TF32": "0",
"MLX_CONV_WINOGRAD": "0"
}
},
"warmup": 2,
"repeats": 7,
"scope": "Synthetic operator microbenchmarks; excludes loading/quantization. Not image latency.",
"trials": [
{
"operator": "rms_adaln",
"variant": "native_rounding",
"shape": [
256,
5376
],
"median_ms": 1.173374941572547,
"min_ms": 0.616749981418252,
"max_ms": 1.2301669921725988,
"samples_ms": [
0.616749981418252,
1.2266251724213362,
1.0637501254677773,
1.173374941572547,
1.2301669921725988,
1.195583026856184,
1.1052920017391443
],
"max_abs": 0.0,
"relative_rmse": 0.0,
"equal_fraction": 1.0
},
{
"operator": "rms_adaln",
"variant": "metal_tg128",
"shape": [
256,
5376
],
"median_ms": 0.5940420087426901,
"min_ms": 0.5400839727371931,
"max_ms": 0.7085409015417099,
"samples_ms": [
0.7085409015417099,
0.5697500891983509,
0.5824591498821974,
0.5400839727371931,
0.6797090172767639,
0.5940420087426901,
0.609041890129447
],
"max_abs": 0.015625,
"relative_rmse": 1.7477590517955832e-05,
"equal_fraction": 0.9999935030937195
},
{
"operator": "rms_adaln",
"variant": "metal_tg256",
"shape": [
256,
5376
],
"median_ms": 0.5604161415249109,
"min_ms": 0.2424169797450304,
"max_ms": 0.6824578158557415,
"samples_ms": [
0.6682500243186951,
0.5432919133454561,
0.6824578158557415,
0.6537078879773617,
0.5389170255511999,
0.5604161415249109,
0.2424169797450304
],
"max_abs": 0.015625,
"relative_rmse": 2.209606464020908e-05,
"equal_fraction": 0.9999912977218628
},
{
"operator": "rms_adaln",