Instructions to use water1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use water1234/MiniMax-H3-MLX-Argus-Calibrated-INT8 with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir MiniMax-H3-MLX-Argus-Calibrated-INT8 water1234/MiniMax-H3-MLX-Argus-Calibrated-INT8
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
File size: 11,004 Bytes
70505b0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 | {
"schema_version": 3,
"profile": "m4-pro-calibrated-int8",
"algorithm": {
"family": "activation-aware diagonal-Hessian affine weight-only INT8",
"objective": "minimize the diagonal-Hessian approximation to ||WX-Q(W)X||_F^2",
"optimizer": "three alternating weighted least-squares affine refits on Torch/CUDA",
"fallback": "retain the plain MLX RTN group whenever a fitted group is not better",
"packing": "four consecutive uint8 codes per uint32, first code in least-significant bits"
},
"bits": 8,
"group_size": 32,
"mode": "affine",
"quantize_adaln": true,
"adaln_bits": 8,
"quantized_layers": {
"8": 254
},
"bf16_layers": [
"blocks.14.adaln_proj.linear",
"blocks.46.mlp.fc2",
"blocks.47.mlp.fc2",
"blocks.48.mlp.fc2"
],
"bf16_exceptions": [
{
"path": "blocks.47.mlp.fc2",
"parameters": 77070336,
"rtn_diagonal_hessian_relative_error": 0.006842044550077006,
"full_dit_parameter_fraction": 0.002326792637428219,
"selection_reason": "highest calibration diagonal-Hessian RTN sensitivity within the preregistered BF16 parameter budget"
},
{
"path": "blocks.14.adaln_proj.linear",
"parameters": 260112384,
"rtn_diagonal_hessian_relative_error": 0.006829738475327525,
"full_dit_parameter_fraction": 0.00785292515132024,
"selection_reason": "highest calibration diagonal-Hessian RTN sensitivity within the preregistered BF16 parameter budget"
},
{
"path": "blocks.46.mlp.fc2",
"parameters": 77070336,
"rtn_diagonal_hessian_relative_error": 0.006813439715784915,
"full_dit_parameter_fraction": 0.002326792637428219,
"selection_reason": "highest calibration diagonal-Hessian RTN sensitivity within the preregistered BF16 parameter budget"
},
{
"path": "blocks.48.mlp.fc2",
"parameters": 77070336,
"rtn_diagonal_hessian_relative_error": 0.006765240564342134,
"full_dit_parameter_fraction": 0.002326792637428219,
"selection_reason": "highest calibration diagonal-Hessian RTN sensitivity within the preregistered BF16 parameter budget"
}
],
"bf16_exception_parameters": 491323392,
"bf16_exception_full_dit_parameter_fraction": 0.014833303063604896,
"protected_source_precision_modules": [
"video_patch_proj",
"audio_patch_proj",
"condition_proj",
"time_embedder",
"final_layer",
"norms",
"linear_biases"
],
"calibration": {
"path": "/root/.argus-skill-latest/workspaces/s-4e0f9076/research/minimax-h3-int8-calibration-cuda-fp32-r939557dc.npz",
"sha256": "0ff2892e605979bd7a2fe0c2807bdd05d9c2b1c54af1bcd53bd3f00b046d18dc",
"manifest": {
"algorithm": "diagonal-hessian-activation-aware-affine-int8",
"bits": 8,
"cases": [
{
"height": 256,
"prompt": "A red fox leaps over a mossy log in a misty forest at dawn.",
"seed": 101,
"split": "calibration",
"step_index": 0,
"width": 256
},
{
"height": 192,
"prompt": "Waves crash against black volcanic rocks under a grey sky.",
"seed": 202,
"split": "calibration",
"step_index": 2,
"width": 320
},
{
"height": 320,
"prompt": "A street musician plays saxophone on a rainy neon-lit corner.",
"seed": 303,
"split": "calibration",
"step_index": 4,
"width": 192
},
{
"height": 192,
"prompt": "Steam rises from a bowl of noodles on a wooden table.",
"seed": 404,
"split": "holdout",
"step_index": 1,
"width": 256
},
{
"height": 256,
"prompt": "An origami robot walks across a sunlit library desk.",
"seed": 505,
"split": "holdout",
"step_index": 3,
"width": 192
}
],
"collector": {
"cuda": "13.1",
"cuda_visible_devices": "1",
"dit_load": {
"assigned_parameter_tensors": 638,
"compute_dtype": "float32",
"dropped_source_keys": [
"rope.inv_freq"
],
"load_seconds": 39.6426787157543,
"missing_keys": [],
"peak_cuda_memory_bytes": 132720563200,
"source_tensors": 535,
"unexpected_keys": []
},
"framework": "PyTorch/CUDA FP32 DiT teacher with one-time BF16 conditioner",
"gpu": "NVIDIA B200",
"script": "scripts/calibrate_int8_torch.py",
"text_encoder": {
"encode_seconds": 1.0854376470670104,
"parity": {
"comparison": null,
"passed": null,
"performed": false,
"prompts": {},
"tolerance": {
"cosine": 0.997,
"relative_l2": 0.08
}
},
"peak_cuda_memory_bytes": 50327620608,
"token_ids_exact_between_source_assets": true,
"torch_loader": {
"full_layers": 64,
"load_seconds": 10.279082759749144,
"loaded_layers": 50,
"loaded_parameter_tensors": 552,
"missing_keys": [],
"unexpected_keys": []
}
},
"torch": "2.11.0a0+eb65b36914.nv26.02",
"transformer_mapping": "official convert_transformer_key + reorder_interleaved_qkv; PyTorch hooks mapped back to original MLX module paths"
},
"created_at": "2026-08-11T16:27:54.492972+00:00",
"group_size": 32,
"implementation": {
"commit": "db1d7651847c2b382beae2094d28214712ff811e",
"dirty": true,
"repository": "https://github.com/Argus-AiTeam/minimax-h3-mac"
},
"mode": "affine",
"sampling": {
"adaln_selection": "real video/audio scheduler timesteps at preregistered case indices, recorded separately from the full per-forward timestep table",
"adaln_split_timesteps": {
"calibration": [
0.0,
0.05263161659240723,
0.1818181872367859,
0.25,
0.5714285373687744
],
"holdout": [
0.020408153533935547,
0.07692301273345947,
0.1111111044883728,
0.3333333134651184
]
},
"duration_seconds": 0.2,
"max_rows_per_call": 8,
"max_rows_per_layer": 64,
"selection": "evenly spaced rows from each real layer input",
"steps": 6
},
"schema_version": 1,
"source": {
"model": "MiniMaxAI/MiniMax-H3",
"revision": "939557dc319dd91227e30195a763f272ba7f8765",
"transformer_config_sha256": "f619093a231fcfbcc3d035bec26c50ad864e7331a500d5c519f5045dc1e50458"
},
"split_contract": "calibration and holdout prompts, seeds, and recorded AdaLN timestep inputs are disjoint",
"token_ids": {
"A red fox leaps over a mossy log in a misty forest at dawn.": [
32,
2518,
38835,
83458,
916,
264,
77243,
88,
1487,
304,
264,
8600,
88,
13638,
518,
38393,
13
],
"A street musician plays saxophone on a rainy neon-lit corner.": [
32,
8592,
38744,
11088,
63008,
77203,
389,
264,
62757,
46652,
2852,
275,
9131,
13
],
"An origami robot walks across a sunlit library desk.": [
2082,
2713,
10606,
12305,
22479,
3941,
264,
7015,
31635,
6733,
18010,
13
],
"Steam rises from a bowl of noodles on a wooden table.": [
41644,
37168,
504,
264,
19212,
315,
60160,
389,
264,
22360,
1965,
13
],
"Waves crash against black volcanic rocks under a grey sky.": [
54,
4693,
9920,
2348,
3691,
66064,
23035,
1212,
264,
19780,
12884,
13
]
}
}
},
"calibration_objective": {
"plain_rtn_diagonal_hessian_sse": 4572725.832609705,
"candidate_diagonal_hessian_sse": 2366819.81746003,
"relative_improvement": 0.4824050458959487
},
"optimizer": {
"framework": "Torch/CUDA",
"device": "cuda:0",
"gpu": "NVIDIA B200",
"torch": "2.11.0a0+eb65b36914.nv26.02",
"cuda": "13.1",
"calls": 512,
"seconds": 958.1897342111915,
"peak_memory_bytes": 16988678144,
"layer_residency": "one dense source layer and its FP32 calibration rows are uploaded at a time; CUDA tensors are released after MLX uint32/scales/biases are materialized"
},
"source": {
"model": "MiniMaxAI/MiniMax-H3",
"revision": "939557dc319dd91227e30195a763f272ba7f8765",
"subfolder": "FL2VA/transformer",
"config_sha256": "f619093a231fcfbcc3d035bec26c50ad864e7331a500d5c519f5045dc1e50458"
},
"builder": {
"repository": "https://github.com/Argus-AiTeam/minimax-h3-mac",
"repository_commit": "db1d7651847c2b382beae2094d28214712ff811e",
"repository_dirty": true,
"mlx_version": "0.32.0",
"python": "3.12.3",
"platform": "Linux-6.8.0-1043-nvidia-x86_64-with-glibc2.39",
"implementation_sha256": {
"scripts/calibrate_int8.py": "d29123d6982475574b01bca1ceffa15770f507ea44d6bdb1779bb53ce745faa2",
"scripts/calibrate_int8_torch.py": "693a2b12af17ccc75124db649334af7f1dff6d51afec47ed6b9aacd0207c9892",
"scripts/probe_real_cuda_parity.py": "5798dcb3fed3e69ceed8f0157fa6bb7ce8ffad8eda4590947e5819b431919ad3",
"scripts/activation_quant_torch.py": "ab7ec7def56b5cb442743732dd16e9d2d74ebe3228261e85c72c427c2adcc64a",
"scripts/build_calibrated_int8.py": "49108bf521b40e11477d4e6f908a877810d75487990b5cad54b063a1c362ff70",
"scripts/eval_calibrated_int8.py": "81205559dd460dfaa4eb09eac796907b1fb4f7cafb70a43009f8391472ce5bf1",
"scripts/build_quant.py": "a0d1a552286f30e3219596bd5f06e5b99fb0eeefdf3b952202152f3ffcdac106",
"minimax_h3_mlx/activation_quant.py": "37621b33250fa2cb497e5cd71fc63bd124a050336d27de9e3b4154df9bc77b10",
"minimax_h3_mlx/quantize.py": "c908638eeec83c127a1b90f93aef607e5d12563f78011702a78465a2010ba275",
"minimax_h3_mlx/load.py": "4a87519b5795f92c08e5f8daf31e023b28f6491a32744a48afe804a211cbc92d",
"minimax_h3_mlx/streaming.py": "dbf76e338bd4c4f9a01b2f8517039b74a0f0953a3f20bebd4ae39eac43fef2c3",
"minimax_h3_mlx/dit.py": "a75791c95244ae496490f7ae0275e3ac42007506bec7b467ad74c56ca8f55c46"
}
},
"gb_on_disk": 37.797,
"gb_resident_after_adaln_drop": 22.928
} |