| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Test model set-up and weight loading for quark-quantized models. |
| 4 | |
| 5 | Run `pytest tests/quantization/test_quark.py`. |
| 6 | |
| 7 | See also `tests/kernels/moe/test_ocp_mx_moe.py`. |
| 8 | """ |
| 9 | |
| 10 | import importlib.metadata |
| 11 | from dataclasses import dataclass |
| 12 | from importlib.util import find_spec |
| 13 | from types import SimpleNamespace |
| 14 | from unittest.mock import MagicMock, Mock, patch |
| 15 | |
| 16 | import huggingface_hub |
| 17 | import lm_eval |
| 18 | import pytest |
| 19 | import torch |
| 20 | from packaging import version |
| 21 | |
| 22 | from tests.quantization.utils import load_model_without_vllm_runner |
| 23 | from vllm._aiter_ops import is_aiter_found_and_supported, rocm_aiter_ops |
| 24 | from vllm.config import VllmConfig, set_current_vllm_config |
| 25 | from vllm.config.cache import CacheConfig, CacheDType |
| 26 | from vllm.forward_context import set_forward_context |
| 27 | from vllm.model_executor import parameter |
| 28 | from vllm.model_executor.kernels.linear.scaled_mm.aiter import ( |
| 29 | AiterHipbMMPerTokenFp8ScaledMMLinearKernel, |
| 30 | AiterPerTokenFp8ScaledMMLinearKernel, |
| 31 | AiterPreshuffledPerTokenFp8ScaledMMLinearKernel, |
| 32 | ) |
| 33 | from vllm.model_executor.kernels.linear.scaled_mm.ScaledMMLinearKernel import ( |
| 34 | FP8ScaledMMLinearLayerConfig, |
| 35 | ) |
| 36 | from vllm.model_executor.layers.attention import Attention |
| 37 | from vllm.model_executor.layers.fused_moe import ( |
| 38 | FusedMoeWeightScaleSupported, |
| 39 | RoutedExperts, |
| 40 | UnquantizedFusedMoEMethod, |
| 41 | ) |
| 42 | from vllm.model_executor.layers.fused_moe.activation import MoEActivation |
| 43 | from vllm.model_executor.layers.fused_moe.config import ( |
| 44 | FusedMoEConfig, |
| 45 | FusedMoEParallelConfig, |
| 46 | RoutingMethodType, |
| 47 | ) |
| 48 | from vllm.model_executor.layers.linear import LinearBase, UnquantizedLinearMethod |
| 49 | from vllm.model_executor.layers.quantization.quark.quark import ( # noqa: E501 |
| 50 | QuarkConfig, |
| 51 | QuarkLinearMethod, |
| 52 | QuarkNVFP4, |
| 53 | QuarkOCP_MX, |
| 54 | QuarkW8A8Fp8, |
| 55 | QuarkW8A8Fp8PerBlock, |
| 56 | QuarkW8A8Int8, |
| 57 | ) |
| 58 | from vllm.model_executor.layers.quantization.quark.quark_moe import ( # noqa: E501 |
| 59 | QuarkMoEMethod, |
| 60 | QuarkW4A8Fp8MoEMethod, |
| 61 | QuarkW4A16Int4MoEMethod, |
| 62 | QuarkW8A8Fp8MoEMethod, |
| 63 | QuarkW8A8Int8MoEMethod, |
| 64 | ) |
| 65 | from vllm.model_executor.layers.quantization.quark.schemes import ( |
| 66 | QuarkScheme, |
| 67 | QuarkW4A16Int4, |
| 68 | ) |
| 69 | from vllm.model_executor.layers.quantization.quark.utils import ( |
| 70 | QuarkQTensorHint, |
| 71 | canonicalize_quark_packed_int4, |
| 72 | should_ignore_layer, |
| 73 | ) |
| 74 | from vllm.model_executor.layers.quantization.utils.mxfp4_utils import ( |
| 75 | quant_dequant_mxfp4, |
| 76 | ) |
| 77 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 78 | QuantKey, |
| 79 | is_layer_skipped, |
| 80 | kFp8Dynamic128Sym, |
| 81 | kFp8DynamicTensorSym, |
| 82 | kFp8DynamicTokenSym, |
| 83 | kFp8Static128BlockE8M0Sym, |
| 84 | kFp8Static128BlockSym, |
| 85 | kFp8StaticChannelSym, |
| 86 | kFp8StaticTensorSym, |
| 87 | kInt4W4A8StaticChannelSym, |
| 88 | kInt8DynamicTensorAsym, |
| 89 | kInt8DynamicTensorSym, |
| 90 | kInt8DynamicTokenAsym, |
| 91 | kInt8DynamicTokenSym, |
| 92 | kInt8StaticChannelSym, |
| 93 | kInt8StaticTensorAsym, |
| 94 | kInt8StaticTensorSym, |
| 95 | kMxfp4Dynamic, |
| 96 | kMxfp4Static, |
| 97 | kMxfp6E2M3Dynamic, |
| 98 | kMxfp6E2M3Static, |
| 99 | kMxfp6E3M2Dynamic, |
| 100 | kMxfp6E3M2Static, |
| 101 | kNvfp4Dynamic, |
| 102 | kNvfp4Static, |
| 103 | ) |
| 104 | from vllm.model_executor.models.llama import LlamaForCausalLM |
| 105 | from vllm.model_executor.models.utils import WeightsMapper |
| 106 | from vllm.platforms import current_platform |
| 107 | from vllm.transformers_utils.repo_utils import hf_api |
| 108 | |
| 109 | if current_platform.is_rocm(): |
| 110 | from vllm.platforms.rocm import on_gfx942, on_gfx950 |
| 111 | else: |
| 112 | |
| 113 | def on_gfx942() -> bool: |
| 114 | return False |
| 115 | |
| 116 | def on_gfx950() -> bool: |
| 117 | return False |
| 118 | |
| 119 | |
| 120 | from .reference_mxfp4 import dq_mxfp4_torch, qdq_mxfp4_torch |
| 121 | |
| 122 | # Minimum amd-quark version for MXFP4/OCP_MX tests (single source of truth). |
| 123 | QUARK_MXFP4_MIN_VERSION = "0.12" |
| 124 | |
| 125 | QUARK_MXFP4_AVAILABLE = find_spec("quark") is not None and version.parse( |
| 126 | importlib.metadata.version("amd-quark") |
| 127 | ) >= version.parse(QUARK_MXFP4_MIN_VERSION) |
| 128 | |
| 129 | AITER_AVAILABLE = is_aiter_found_and_supported() |
| 130 | |
| 131 | AITER_PTPC_KERNELS = ( |
| 132 | AiterHipbMMPerTokenFp8ScaledMMLinearKernel, |
| 133 | AiterPreshuffledPerTokenFp8ScaledMMLinearKernel, |
| 134 | AiterPerTokenFp8ScaledMMLinearKernel, |
| 135 | ) |
| 136 | |
| 137 | DEVICE_TYPE = current_platform.device_type |
| 138 | |
| 139 | |
| 140 | @dataclass(frozen=True) |
| 141 | class QTensorConfig: |
| 142 | name: str |
| 143 | weight: QuarkQTensorHint |
| 144 | input_tensors: QuarkQTensorHint |
| 145 | weight_quant_key: QuantKey | None = None |
| 146 | act_quant_key: QuantKey | None = None |
| 147 | dispatch_cls: type[QuarkScheme] | type[QuarkMoEMethod] | None = None |
| 148 | expected_error: tuple[type[Exception], str] | None = None |
| 149 | |
| 150 | |
| 151 | QTENSOR_CONFIGS = [ |
| 152 | QTensorConfig( |
| 153 | name="fp8_w8a8_static_tensor", |
| 154 | weight={"dtype": "fp8_e4m3", "qscheme": "per_tensor", "is_dynamic": False}, |
| 155 | input_tensors={ |
| 156 | "dtype": "fp8_e4m3", |
| 157 | "qscheme": "per_tensor", |
| 158 | "is_dynamic": False, |
| 159 | }, |
| 160 | weight_quant_key=kFp8StaticTensorSym, |
| 161 | act_quant_key=kFp8StaticTensorSym, |
| 162 | dispatch_cls=QuarkW8A8Fp8, |
| 163 | ), |
| 164 | QTensorConfig( |
| 165 | name="fp8_w8a8_static_tensor_single_entry_lists", |
| 166 | weight=[ |
| 167 | { |
| 168 | "dtype": "fp8_e4m3", |
| 169 | "qscheme": "per_tensor", |
| 170 | "is_dynamic": False, |
| 171 | } |
| 172 | ], |
| 173 | input_tensors=[ |
| 174 | { |
| 175 | "dtype": "fp8_e4m3", |
| 176 | "qscheme": "per_tensor", |
| 177 | "is_dynamic": False, |
| 178 | } |
| 179 | ], |
| 180 | weight_quant_key=kFp8StaticTensorSym, |
| 181 | act_quant_key=kFp8StaticTensorSym, |
| 182 | dispatch_cls=QuarkW8A8Fp8, |
| 183 | ), |
| 184 | QTensorConfig( |
| 185 | name="fp8_w8a8_dynamic_tensor", |
| 186 | weight={"dtype": "fp8_e4m3", "qscheme": "per_tensor", "is_dynamic": False}, |
| 187 | input_tensors={ |
| 188 | "dtype": "fp8_e4m3", |
| 189 | "qscheme": "per_tensor", |
| 190 | "is_dynamic": True, |
| 191 | }, |
| 192 | weight_quant_key=kFp8StaticTensorSym, |
| 193 | act_quant_key=kFp8DynamicTensorSym, |
| 194 | dispatch_cls=QuarkW8A8Fp8, |
| 195 | ), |
| 196 | QTensorConfig( |
| 197 | name="fp8_w8a8_dynamic_token", |
| 198 | weight={"dtype": "fp8_e4m3", "qscheme": "per_channel", "is_dynamic": False}, |
| 199 | input_tensors={ |
| 200 | "dtype": "fp8_e4m3", |
| 201 | "qscheme": "per_channel", |
| 202 | "is_dynamic": True, |
| 203 | }, |
| 204 | weight_quant_key=kFp8StaticChannelSym, |
| 205 | act_quant_key=kFp8DynamicTokenSym, |
| 206 | dispatch_cls=QuarkW8A8Fp8, |
| 207 | ), |
| 208 | QTensorConfig( |
| 209 | name="fp8_w8a8_channel_static_tensor", |
| 210 | weight={"dtype": "fp8_e4m3", "qscheme": "per_channel", "is_dynamic": False}, |
| 211 | input_tensors={ |
| 212 | "dtype": "fp8_e4m3", |
| 213 | "qscheme": "per_tensor", |
| 214 | "is_dynamic": False, |
| 215 | }, |
| 216 | weight_quant_key=kFp8StaticChannelSym, |
| 217 | act_quant_key=kFp8StaticTensorSym, |
| 218 | dispatch_cls=QuarkW8A8Fp8, |
| 219 | ), |
| 220 | QTensorConfig( |
| 221 | name="fp8_w8a8_tensor_dynamic_token", |
| 222 | weight={"dtype": "fp8_e4m3", "qscheme": "per_tensor", "is_dynamic": False}, |
| 223 | input_tensors={ |
| 224 | "dtype": "fp8_e4m3", |
| 225 | "qscheme": "per_channel", |
| 226 | "is_dynamic": True, |
| 227 | }, |
| 228 | weight_quant_key=kFp8StaticTensorSym, |
| 229 | act_quant_key=kFp8DynamicTokenSym, |
| 230 | dispatch_cls=QuarkW8A8Fp8, |
| 231 | ), |
| 232 | QTensorConfig( |
| 233 | name="fp8_w8a8_dynamic_block_fp32", |
| 234 | weight={ |
| 235 | "dtype": "fp8_e4m3", |
| 236 | "qscheme": "per_block", |
| 237 | "is_dynamic": False, |
| 238 | "block_size": [128, 128], |
| 239 | "symmetric": True, |
| 240 | }, |
| 241 | input_tensors={ |
| 242 | "dtype": "fp8_e4m3", |
| 243 | "qscheme": "per_group", |
| 244 | "is_dynamic": True, |
| 245 | "group_size": 128, |
| 246 | "symmetric": True, |
| 247 | }, |
| 248 | weight_quant_key=kFp8Static128BlockSym, |
| 249 | act_quant_key=kFp8Dynamic128Sym, |
| 250 | dispatch_cls=QuarkW8A8Fp8PerBlock, |
| 251 | ), |
| 252 | QTensorConfig( |
| 253 | name="fp8_w8a8_dynamic_block_e8m0", |
| 254 | weight={ |
| 255 | "dtype": "fp8_e4m3", |
| 256 | "qscheme": "per_block", |
| 257 | "is_dynamic": False, |
| 258 | "block_size": [128, 128], |
| 259 | "symmetric": True, |
| 260 | "scale_type": "float8_e8m0fnu", |
| 261 | }, |
| 262 | input_tensors={ |
| 263 | "dtype": "fp8_e4m3", |
| 264 | "qscheme": "per_group", |
| 265 | "is_dynamic": True, |
| 266 | "group_size": 128, |
| 267 | "symmetric": True, |
| 268 | }, |
| 269 | weight_quant_key=kFp8Static128BlockE8M0Sym, |
| 270 | act_quant_key=kFp8Dynamic128Sym, |
| 271 | dispatch_cls=QuarkW8A8Fp8PerBlock, |
| 272 | ), |
| 273 | QTensorConfig( |
| 274 | name="fp8_w8a8_dynamic_block_fp32_moe", |
| 275 | weight={ |
| 276 | "dtype": "fp8_e4m3", |
| 277 | "qscheme": "per_block", |
| 278 | "is_dynamic": False, |
| 279 | "block_size": [128, 128], |
| 280 | "symmetric": True, |
| 281 | }, |
| 282 | input_tensors={ |
| 283 | "dtype": "fp8_e4m3", |
| 284 | "qscheme": "per_group", |
| 285 | "is_dynamic": True, |
| 286 | "group_size": 128, |
| 287 | "symmetric": True, |
| 288 | }, |
| 289 | weight_quant_key=kFp8Static128BlockSym, |
| 290 | act_quant_key=kFp8Dynamic128Sym, |
| 291 | dispatch_cls=QuarkW8A8Fp8MoEMethod, |
| 292 | ), |
| 293 | QTensorConfig( |
| 294 | name="fp8_w8a8_block_static_input", |
| 295 | weight={ |
| 296 | "dtype": "fp8_e4m3", |
| 297 | "qscheme": "per_block", |
| 298 | "is_dynamic": False, |
| 299 | "block_size": [128, 128], |
| 300 | "symmetric": True, |
| 301 | }, |
| 302 | input_tensors={ |
| 303 | "dtype": "fp8_e4m3", |
| 304 | "qscheme": "per_group", |
| 305 | "is_dynamic": False, |
| 306 | "group_size": 128, |
| 307 | "symmetric": True, |
| 308 | }, |
| 309 | expected_error=(NotImplementedError, "No quark compatible scheme"), |
| 310 | ), |
| 311 | QTensorConfig( |
| 312 | name="fp8_w8a8_block_group_size_mismatch", |
| 313 | weight={ |
| 314 | "dtype": "fp8_e4m3", |
| 315 | "qscheme": "per_block", |
| 316 | "is_dynamic": False, |
| 317 | "block_size": [128, 128], |
| 318 | "symmetric": True, |
| 319 | }, |
| 320 | input_tensors={ |
| 321 | "dtype": "fp8_e4m3", |
| 322 | "qscheme": "per_group", |
| 323 | "is_dynamic": True, |
| 324 | "group_size": 64, |
| 325 | "symmetric": True, |
| 326 | }, |
| 327 | expected_error=(NotImplementedError, "No quark compatible scheme"), |
| 328 | ), |
| 329 | QTensorConfig( |
| 330 | name="fp8_w8a8_block_missing_block_size", |
| 331 | weight={ |
| 332 | "dtype": "fp8_e4m3", |
| 333 | "qscheme": "per_block", |
| 334 | "is_dynamic": False, |
| 335 | "symmetric": True, |
| 336 | }, |
| 337 | input_tensors={ |
| 338 | "dtype": "fp8_e4m3", |
| 339 | "qscheme": "per_group", |
| 340 | "is_dynamic": True, |
| 341 | "group_size": 128, |
| 342 | "symmetric": True, |
| 343 | }, |
| 344 | expected_error=(ValueError, "requires `block_size`"), |
| 345 | ), |
| 346 | QTensorConfig( |
| 347 | name="int8_w8a8_static_symmetric", |
| 348 | weight={ |
| 349 | "dtype": "int8", |
| 350 | "qscheme": "per_tensor", |
| 351 | "is_dynamic": False, |
| 352 | "symmetric": True, |
| 353 | }, |
| 354 | input_tensors={ |
| 355 | "dtype": "int8", |
| 356 | "qscheme": "per_tensor", |
| 357 | "is_dynamic": False, |
| 358 | "symmetric": True, |
| 359 | }, |
| 360 | weight_quant_key=kInt8StaticTensorSym, |
| 361 | act_quant_key=kInt8StaticTensorSym, |
| 362 | dispatch_cls=QuarkW8A8Int8, |
| 363 | ), |
| 364 | QTensorConfig( |
| 365 | name="int8_w8a8_static_asymmetric", |
| 366 | weight={ |
| 367 | "dtype": "int8", |
| 368 | "qscheme": "per_tensor", |
| 369 | "is_dynamic": False, |
| 370 | "symmetric": True, |
| 371 | }, |
| 372 | input_tensors={ |
| 373 | "dtype": "int8", |
| 374 | "qscheme": "per_tensor", |
| 375 | "is_dynamic": False, |
| 376 | "symmetric": False, |
| 377 | }, |
| 378 | weight_quant_key=kInt8StaticTensorSym, |
| 379 | act_quant_key=kInt8StaticTensorAsym, |
| 380 | dispatch_cls=QuarkW8A8Int8, |
| 381 | ), |
| 382 | QTensorConfig( |
| 383 | name="int8_w8a8_channel_static_symmetric", |
| 384 | weight={ |
| 385 | "dtype": "int8", |
| 386 | "qscheme": "per_channel", |
| 387 | "is_dynamic": False, |
| 388 | "symmetric": True, |
| 389 | }, |
| 390 | input_tensors={ |
| 391 | "dtype": "int8", |
| 392 | "qscheme": "per_tensor", |
| 393 | "is_dynamic": False, |
| 394 | "symmetric": True, |
| 395 | }, |
| 396 | weight_quant_key=kInt8StaticChannelSym, |
| 397 | act_quant_key=kInt8StaticTensorSym, |
| 398 | dispatch_cls=QuarkW8A8Int8, |
| 399 | ), |
| 400 | QTensorConfig( |
| 401 | name="int8_w8a8_channel_static_asymmetric", |
| 402 | weight={ |
| 403 | "dtype": "int8", |
| 404 | "qscheme": "per_channel", |
| 405 | "is_dynamic": False, |
| 406 | "symmetric": True, |
| 407 | }, |
| 408 | input_tensors={ |
| 409 | "dtype": "int8", |
| 410 | "qscheme": "per_tensor", |
| 411 | "is_dynamic": False, |
| 412 | "symmetric": False, |
| 413 | }, |
| 414 | weight_quant_key=kInt8StaticChannelSym, |
| 415 | act_quant_key=kInt8StaticTensorAsym, |
| 416 | dispatch_cls=QuarkW8A8Int8, |
| 417 | ), |
| 418 | QTensorConfig( |
| 419 | name="int8_w8a8_dynamic_tensor_symmetric", |
| 420 | weight={ |
| 421 | "dtype": "int8", |
| 422 | "qscheme": "per_tensor", |
| 423 | "is_dynamic": False, |
| 424 | "symmetric": True, |
| 425 | }, |
| 426 | input_tensors={ |
| 427 | "dtype": "int8", |
| 428 | "qscheme": "per_channel", |
| 429 | "is_dynamic": True, |
| 430 | "symmetric": True, |
| 431 | }, |
| 432 | weight_quant_key=kInt8StaticTensorSym, |
| 433 | act_quant_key=kInt8DynamicTensorSym, |
| 434 | dispatch_cls=QuarkW8A8Int8, |
| 435 | ), |
| 436 | QTensorConfig( |
| 437 | name="int8_w8a8_dynamic_tensor_asymmetric", |
| 438 | weight={ |
| 439 | "dtype": "int8", |
| 440 | "qscheme": "per_tensor", |
| 441 | "is_dynamic": False, |
| 442 | "symmetric": True, |
| 443 | }, |
| 444 | input_tensors={ |
| 445 | "dtype": "int8", |
| 446 | "qscheme": "per_channel", |
| 447 | "is_dynamic": True, |
| 448 | "symmetric": False, |
| 449 | }, |
| 450 | weight_quant_key=kInt8StaticTensorSym, |
| 451 | act_quant_key=kInt8DynamicTensorAsym, |
| 452 | dispatch_cls=QuarkW8A8Int8, |
| 453 | ), |
| 454 | QTensorConfig( |
| 455 | name="int8_w8a8_dynamic_token", |
| 456 | weight={ |
| 457 | "dtype": "int8", |
| 458 | "qscheme": "per_channel", |
| 459 | "is_dynamic": False, |
| 460 | "symmetric": True, |
| 461 | }, |
| 462 | input_tensors={ |
| 463 | "dtype": "int8", |
| 464 | "qscheme": "per_channel", |
| 465 | "is_dynamic": True, |
| 466 | "symmetric": True, |
| 467 | }, |
| 468 | weight_quant_key=kInt8StaticChannelSym, |
| 469 | act_quant_key=kInt8DynamicTokenSym, |
| 470 | dispatch_cls=QuarkW8A8Int8, |
| 471 | ), |
| 472 | QTensorConfig( |
| 473 | name="int8_w8a8_dynamic_token_asymmetric", |
| 474 | weight={ |
| 475 | "dtype": "int8", |
| 476 | "qscheme": "per_channel", |
| 477 | "is_dynamic": False, |
| 478 | "symmetric": True, |
| 479 | }, |
| 480 | input_tensors={ |
| 481 | "dtype": "int8", |
| 482 | "qscheme": "per_channel", |
| 483 | "is_dynamic": True, |
| 484 | "symmetric": False, |
| 485 | }, |
| 486 | weight_quant_key=kInt8StaticChannelSym, |
| 487 | act_quant_key=kInt8DynamicTokenAsym, |
| 488 | dispatch_cls=QuarkW8A8Int8, |
| 489 | ), |
| 490 | QTensorConfig( |
| 491 | name="ocp_mx_mxfp4_weight_only", |
| 492 | weight={ |
| 493 | "dtype": "fp4", |
| 494 | "qscheme": "per_group", |
| 495 | "group_size": 32, |
| 496 | "scale_format": "e8m0", |
| 497 | "is_dynamic": False, |
| 498 | }, |
| 499 | input_tensors=None, |
| 500 | weight_quant_key=kMxfp4Static, |
| 501 | act_quant_key=None, |
| 502 | dispatch_cls=QuarkOCP_MX, |
| 503 | ), |
| 504 | QTensorConfig( |
| 505 | name="ocp_mx_mxfp4_activation", |
| 506 | weight={ |
| 507 | "dtype": "fp4", |
| 508 | "qscheme": "per_group", |
| 509 | "group_size": 32, |
| 510 | "scale_format": "e8m0", |
| 511 | "is_dynamic": False, |
| 512 | }, |
| 513 | input_tensors={ |
| 514 | "dtype": "fp4", |
| 515 | "qscheme": "per_group", |
| 516 | "group_size": 32, |
| 517 | "scale_format": "e8m0", |
| 518 | "is_dynamic": True, |
| 519 | }, |
| 520 | weight_quant_key=kMxfp4Static, |
| 521 | act_quant_key=kMxfp4Dynamic, |
| 522 | dispatch_cls=QuarkOCP_MX, |
| 523 | ), |
| 524 | QTensorConfig( |
| 525 | name="ocp_mx_mxfp6_e3m2", |
| 526 | weight={ |
| 527 | "dtype": "fp6_e3m2", |
| 528 | "qscheme": "per_group", |
| 529 | "group_size": 32, |
| 530 | "scale_format": "e8m0", |
| 531 | "is_dynamic": False, |
| 532 | }, |
| 533 | input_tensors={ |
| 534 | "dtype": "fp6_e3m2", |
| 535 | "qscheme": "per_group", |
| 536 | "group_size": 32, |
| 537 | "scale_format": "e8m0", |
| 538 | "is_dynamic": True, |
| 539 | }, |
| 540 | weight_quant_key=kMxfp6E3M2Static, |
| 541 | act_quant_key=kMxfp6E3M2Dynamic, |
| 542 | dispatch_cls=QuarkOCP_MX, |
| 543 | ), |
| 544 | QTensorConfig( |
| 545 | name="ocp_mx_mxfp4_mxfp6_e3m2_activation", |
| 546 | weight={ |
| 547 | "dtype": "fp4", |
| 548 | "qscheme": "per_group", |
| 549 | "group_size": 32, |
| 550 | "scale_format": "e8m0", |
| 551 | "is_dynamic": False, |
| 552 | }, |
| 553 | input_tensors={ |
| 554 | "dtype": "fp6_e3m2", |
| 555 | "qscheme": "per_group", |
| 556 | "group_size": 32, |
| 557 | "scale_format": "e8m0", |
| 558 | "is_dynamic": True, |
| 559 | }, |
| 560 | weight_quant_key=kMxfp4Static, |
| 561 | act_quant_key=kMxfp6E3M2Dynamic, |
| 562 | dispatch_cls=QuarkOCP_MX, |
| 563 | ), |
| 564 | QTensorConfig( |
| 565 | name="ocp_mx_mxfp4_mxfp6_e2m3_activation", |
| 566 | weight={ |
| 567 | "dtype": "fp4", |
| 568 | "qscheme": "per_group", |
| 569 | "group_size": 32, |
| 570 | "scale_format": "e8m0", |
| 571 | "is_dynamic": False, |
| 572 | }, |
| 573 | input_tensors={ |
| 574 | "dtype": "fp6_e2m3", |
| 575 | "qscheme": "per_group", |
| 576 | "group_size": 32, |
| 577 | "scale_format": "e8m0", |
| 578 | "is_dynamic": True, |
| 579 | }, |
| 580 | weight_quant_key=kMxfp4Static, |
| 581 | act_quant_key=kMxfp6E2M3Dynamic, |
| 582 | dispatch_cls=QuarkOCP_MX, |
| 583 | ), |
| 584 | QTensorConfig( |
| 585 | name="ocp_mx_mxfp6_e2m3", |
| 586 | weight={ |
| 587 | "dtype": "fp6_e2m3", |
| 588 | "qscheme": "per_group", |
| 589 | "group_size": 32, |
| 590 | "scale_format": "e8m0", |
| 591 | "is_dynamic": False, |
| 592 | }, |
| 593 | input_tensors={ |
| 594 | "dtype": "fp6_e2m3", |
| 595 | "qscheme": "per_group", |
| 596 | "group_size": 32, |
| 597 | "scale_format": "e8m0", |
| 598 | "is_dynamic": True, |
| 599 | }, |
| 600 | weight_quant_key=kMxfp6E2M3Static, |
| 601 | act_quant_key=kMxfp6E2M3Dynamic, |
| 602 | dispatch_cls=QuarkOCP_MX, |
| 603 | ), |
| 604 | QTensorConfig( |
| 605 | name="nvfp4", |
| 606 | weight=[ |
| 607 | { |
| 608 | "dtype": "fp4", |
| 609 | "qscheme": "per_group", |
| 610 | "group_size": 16, |
| 611 | "is_dynamic": False, |
| 612 | }, |
| 613 | {"dtype": "fp8_e4m3", "qscheme": "per_tensor", "is_dynamic": False}, |
| 614 | ], |
| 615 | input_tensors=[ |
| 616 | { |
| 617 | "dtype": "fp4", |
| 618 | "qscheme": "per_group", |
| 619 | "group_size": 16, |
| 620 | "is_dynamic": True, |
| 621 | }, |
| 622 | {"dtype": "fp8_e4m3", "qscheme": "per_tensor", "is_dynamic": False}, |
| 623 | ], |
| 624 | weight_quant_key=kNvfp4Static, |
| 625 | act_quant_key=kNvfp4Dynamic, |
| 626 | dispatch_cls=QuarkNVFP4, |
| 627 | ), |
| 628 | QTensorConfig( |
| 629 | name="w4a8_fp8_static", |
| 630 | weight=[ |
| 631 | {"dtype": "fp8_e4m3", "qscheme": "per_tensor", "is_dynamic": False}, |
| 632 | { |
| 633 | "dtype": "int4", |
| 634 | "qscheme": "per_channel", |
| 635 | "is_dynamic": False, |
| 636 | "symmetric": True, |
| 637 | "ch_axis": 0, |
| 638 | }, |
| 639 | ], |
| 640 | input_tensors={ |
| 641 | "dtype": "fp8_e4m3", |
| 642 | "qscheme": "per_tensor", |
| 643 | "is_dynamic": False, |
| 644 | }, |
| 645 | weight_quant_key=kInt4W4A8StaticChannelSym, |
| 646 | act_quant_key=kFp8StaticTensorSym, |
| 647 | dispatch_cls=QuarkW4A8Fp8MoEMethod, |
| 648 | ), |
| 649 | QTensorConfig( |
| 650 | name="w4a8_fp8_dynamic", |
| 651 | weight=[ |
| 652 | {"dtype": "fp8_e4m3", "qscheme": "per_tensor", "is_dynamic": False}, |
| 653 | { |
| 654 | "dtype": "int4", |
| 655 | "qscheme": "per_channel", |
| 656 | "is_dynamic": False, |
| 657 | "symmetric": True, |
| 658 | "ch_axis": 0, |
| 659 | }, |
| 660 | ], |
| 661 | input_tensors={ |
| 662 | "dtype": "fp8_e4m3", |
| 663 | "qscheme": "per_channel", |
| 664 | "is_dynamic": True, |
| 665 | }, |
| 666 | weight_quant_key=kInt4W4A8StaticChannelSym, |
| 667 | act_quant_key=kFp8DynamicTokenSym, |
| 668 | dispatch_cls=QuarkW4A8Fp8MoEMethod, |
| 669 | ), |
| 670 | QTensorConfig( |
| 671 | name="w4a8_fp8_static_single_entry_input", |
| 672 | weight=[ |
| 673 | {"dtype": "fp8_e4m3", "qscheme": "per_tensor", "is_dynamic": False}, |
| 674 | { |
| 675 | "dtype": "int4", |
| 676 | "qscheme": "per_channel", |
| 677 | "is_dynamic": False, |
| 678 | "symmetric": True, |
| 679 | "ch_axis": 0, |
| 680 | }, |
| 681 | ], |
| 682 | input_tensors=[ |
| 683 | { |
| 684 | "dtype": "fp8_e4m3", |
| 685 | "qscheme": "per_tensor", |
| 686 | "is_dynamic": False, |
| 687 | } |
| 688 | ], |
| 689 | weight_quant_key=kInt4W4A8StaticChannelSym, |
| 690 | act_quant_key=kFp8StaticTensorSym, |
| 691 | dispatch_cls=QuarkW4A8Fp8MoEMethod, |
| 692 | ), |
| 693 | ] |
| 694 | |
| 695 | |
| 696 | def _make_qtensor_config( |
| 697 | weight: QuarkQTensorHint, |
| 698 | input_tensors: QuarkQTensorHint, |
| 699 | exclude: list[str] | None = None, |
| 700 | ) -> QuarkConfig: |
| 701 | return QuarkConfig( |
| 702 | { |
| 703 | "global_quant_config": { |
| 704 | "weight": weight, |
| 705 | "input_tensors": input_tensors, |
| 706 | }, |
| 707 | "layer_type_quant_config": {}, |
| 708 | "exclude": exclude or [], |
| 709 | } |
| 710 | ) |
| 711 | |
| 712 | |
| 713 | def _make_test_moe_config() -> FusedMoEConfig: |
| 714 | return FusedMoEConfig( |
| 715 | num_experts=8, |
| 716 | experts_per_token=2, |
| 717 | hidden_dim=256, |
| 718 | intermediate_size=256, |
| 719 | num_local_experts=8, |
| 720 | num_logical_experts=8, |
| 721 | activation=MoEActivation.SILU, |
| 722 | device=current_platform.device_type, |
| 723 | routing_method=RoutingMethodType.Renormalize, |
| 724 | moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(), |
| 725 | in_dtype=torch.bfloat16, |
| 726 | ) |
| 727 | |
| 728 | |
| 729 | if QUARK_MXFP4_AVAILABLE: |
| 730 | from quark.torch.export.nn.modules.realquantizer import StaticScaledRealQuantizer |
| 731 | from quark.torch.kernel import mx as mx_kernel |
| 732 | from quark.torch.quantization.config.config import FP4PerGroupSpec |
| 733 | |
| 734 | try: |
| 735 | hf_api().list_repo_refs( |
| 736 | "amd/Llama-3.3-70B-Instruct-WMXFP4-AMXFP4-KVFP8-Scale-UINT8-SQ" |
| 737 | ) |
| 738 | HF_HUB_AMD_ORG_ACCESS = True |
| 739 | except huggingface_hub.errors.RepositoryNotFoundError: |
| 740 | HF_HUB_AMD_ORG_ACCESS = False |
| 741 | |
| 742 | |
| 743 | @pytest.fixture(scope="function", autouse=True) |
| 744 | def enable_pickle(monkeypatch): |
| 745 | """`LLM.apply_model` requires pickling a function.""" |
| 746 | monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1") |
| 747 | |
| 748 | |
| 749 | def test_quark_w8a8_fp8_per_block_registers_weight_scale(monkeypatch): |
| 750 | from vllm.model_executor.layers.quantization.utils.fp8_utils import ( |
| 751 | get_fp8_block_weight_scale, |
| 752 | ) |
| 753 | |
| 754 | monkeypatch.setattr( |
| 755 | "vllm.model_executor.layers.quantization.quark.schemes." |
| 756 | "quark_w8a8_fp8.get_current_vllm_config", |
| 757 | lambda: SimpleNamespace(model_config=SimpleNamespace(dtype=torch.bfloat16)), |
| 758 | ) |
| 759 | scheme = QuarkW8A8Fp8PerBlock(kFp8Static128BlockSym, kFp8Dynamic128Sym) |
| 760 | |
| 761 | layer = torch.nn.Module() |
| 762 | layer.weight_scale = torch.tensor([2.0]) |
| 763 | assert get_fp8_block_weight_scale(layer) is None |
| 764 | layer.scheme = scheme |
| 765 | assert get_fp8_block_weight_scale(layer) is layer.weight_scale |
| 766 | layer.weight_scale_inv = torch.tensor([3.0]) |
| 767 | assert get_fp8_block_weight_scale(layer) is layer.weight_scale |
| 768 | layer.scheme = None |
| 769 | assert get_fp8_block_weight_scale(layer) is layer.weight_scale_inv |
| 770 | |
| 771 | loaded = torch.nn.Module() |
| 772 | |
| 773 | def weight_loader(param, loaded_weight): |
| 774 | return None |
| 775 | |
| 776 | dummy_param = torch.nn.Parameter(torch.empty(1), requires_grad=False) |
| 777 | with ( |
| 778 | patch( |
| 779 | "vllm.model_executor.layers.quantization.quark.schemes.quark_w8a8_fp8." |
| 780 | "validate_fp8_block_shape" |
| 781 | ), |
| 782 | patch( |
| 783 | "vllm.model_executor.layers.quantization.quark.schemes.quark_w8a8_fp8." |
| 784 | "create_fp8_weight_parameter", |
| 785 | return_value=dummy_param, |
| 786 | ), |
| 787 | patch( |
| 788 | "vllm.model_executor.layers.quantization.quark.schemes.quark_w8a8_fp8." |
| 789 | "create_fp8_scale_parameter", |
| 790 | return_value=dummy_param, |
| 791 | ), |
| 792 | patch( |
| 793 | "vllm.model_executor.layers.quantization.quark.schemes.quark_w8a8_fp8." |
| 794 | "init_fp8_linear_kernel", |
| 795 | return_value=MagicMock(), |
| 796 | ), |
| 797 | ): |
| 798 | scheme.create_weights( |
| 799 | loaded, |
| 800 | output_partition_sizes=[256], |
| 801 | input_size_per_partition=256, |
| 802 | params_dtype=torch.bfloat16, |
| 803 | weight_loader=weight_loader, |
| 804 | input_size=256, |
| 805 | output_size=256, |
| 806 | ) |
| 807 | assert hasattr(loaded, "weight_scale") |
| 808 | assert not hasattr(loaded, "weight_scale_inv") |
| 809 | |
| 810 | |
| 811 | def test_quark_config_has_no_model_specific_fused_mappings(): |
| 812 | config = QuarkConfig({}) |
| 813 | |
| 814 | assert "gate_up_proj" not in config.packed_modules_mapping |
| 815 | assert "fused_wqa_wkv" not in config.packed_modules_mapping |
| 816 | |
| 817 | |
| 818 | def test_quark_config_preserves_existing_packed_modules_mapping(): |
| 819 | class CustomQuarkConfig(QuarkConfig): |
| 820 | packed_modules_mapping = {"custom_proj": ["a", "b"]} |
| 821 | |
| 822 | config = CustomQuarkConfig({}) |
| 823 | |
| 824 | assert config.packed_modules_mapping["custom_proj"] == ["a", "b"] |
| 825 | |
| 826 | |
| 827 | def test_quant_method_dispatch_ignored(default_vllm_config): |
| 828 | config = _make_qtensor_config(None, None, exclude=["linear", "experts"]) |
| 829 | |
| 830 | class TestLinear(LinearBase): |
| 831 | def __init__(self): |
| 832 | torch.nn.Module.__init__(self) |
| 833 | |
| 834 | class TestRoutedExperts(RoutedExperts): |
| 835 | def __init__(self): |
| 836 | torch.nn.Module.__init__(self) |
| 837 | self.moe_config = _make_test_moe_config() |
| 838 | |
| 839 | assert config.get_quant_method_target("linear", LinearBase) == ( |
| 840 | None, |
| 841 | None, |
| 842 | UnquantizedLinearMethod, |
| 843 | ) |
| 844 | assert isinstance( |
| 845 | config.get_quant_method(TestLinear(), "linear"), UnquantizedLinearMethod |
| 846 | ) |
| 847 | |
| 848 | assert config.get_quant_method_target("experts", RoutedExperts) == ( |
| 849 | None, |
| 850 | None, |
| 851 | UnquantizedFusedMoEMethod, |
| 852 | ) |
| 853 | assert isinstance( |
| 854 | config.get_quant_method(TestRoutedExperts(), "experts"), |
| 855 | UnquantizedFusedMoEMethod, |
| 856 | ) |
| 857 | |
| 858 | mxfp4_config = _make_qtensor_config( |
| 859 | { |
| 860 | "dtype": "fp4", |
| 861 | "qscheme": "per_group", |
| 862 | "group_size": 32, |
| 863 | "scale_format": "e8m0", |
| 864 | "is_dynamic": False, |
| 865 | }, |
| 866 | None, |
| 867 | exclude=["self_attn.q_proj", "mlp.down_proj"], |
| 868 | ) |
| 869 | assert mxfp4_config.get_quant_method_target("self_attn.q_proj", LinearBase) == ( |
| 870 | None, |
| 871 | None, |
| 872 | UnquantizedLinearMethod, |
| 873 | ) |
| 874 | assert isinstance( |
| 875 | mxfp4_config.get_quant_method(TestLinear(), "self_attn.q_proj"), |
| 876 | UnquantizedLinearMethod, |
| 877 | ) |
| 878 | |
| 879 | assert mxfp4_config.get_quant_method_target("mlp.down_proj", LinearBase) == ( |
| 880 | None, |
| 881 | None, |
| 882 | UnquantizedLinearMethod, |
| 883 | ) |
| 884 | assert isinstance( |
| 885 | mxfp4_config.get_quant_method(TestLinear(), "mlp.down_proj"), |
| 886 | UnquantizedLinearMethod, |
| 887 | ) |
| 888 | |
| 889 | |
| 890 | @pytest.mark.parametrize("case", QTENSOR_CONFIGS, ids=lambda case: case.name) |
| 891 | def test_quant_method_dispatch_target(case): |
| 892 | config = _make_qtensor_config(case.weight, case.input_tensors) |
| 893 | if case.expected_error is not None: |
| 894 | error_type, error_message = case.expected_error |
| 895 | with pytest.raises(error_type, match=error_message): |
| 896 | config.get_quant_method_target("linear", LinearBase) |
| 897 | return |
| 898 | |
| 899 | assert case.dispatch_cls is not None |
| 900 | is_linear = issubclass(case.dispatch_cls, QuarkScheme) |
| 901 | |
| 902 | weight_quant_key, act_quant_key, method_cls = config.get_quant_method_target( |
| 903 | "linear" if is_linear else "experts", |
| 904 | LinearBase if is_linear else RoutedExperts, |
| 905 | ) |
| 906 | |
| 907 | assert weight_quant_key == case.weight_quant_key |
| 908 | assert act_quant_key == case.act_quant_key |
| 909 | assert method_cls is (QuarkLinearMethod if is_linear else case.dispatch_cls) |
| 910 | |
| 911 | |
| 912 | def test_quant_method_dispatch_mxfp8_2d_block(default_vllm_config): |
| 913 | """Shape B (DeepSeek-V4.1): 32x32 per-block e8m0 routes to QuarkOCP_MX. |
| 914 | |
| 915 | Verifies that get_quant_method and get_quant_method_target agree (the |
| 916 | short-circuit that previously caused them to diverge has been removed). |
| 917 | """ |
| 918 | from vllm.model_executor.layers.quantization.quark.schemes import QuarkOCP_MX |
| 919 | from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp8Static |
| 920 | |
| 921 | default_vllm_config.model_config = SimpleNamespace(dtype=torch.bfloat16) |
| 922 | mxfp8_spec = { |
| 923 | "weight": { |
| 924 | "dtype": "fp8_e4m3", |
| 925 | "qscheme": "per_block", |
| 926 | "block_size": [32, 32], |
| 927 | "symmetric": True, |
| 928 | "is_dynamic": False, |
| 929 | "scale_type": "float8_e8m0fnu", |
| 930 | }, |
| 931 | "input_tensors": { |
| 932 | "dtype": "fp8_e4m3", |
| 933 | "qscheme": "per_group", |
| 934 | "group_size": 32, |
| 935 | "symmetric": True, |
| 936 | "is_dynamic": True, |
| 937 | }, |
| 938 | } |
| 939 | config = QuarkConfig( |
| 940 | { |
| 941 | "global_quant_config": { |
| 942 | "weight": { |
| 943 | "dtype": "fp4", |
| 944 | "qscheme": "per_group", |
| 945 | "group_size": 32, |
| 946 | "scale_format": "e8m0", |
| 947 | "is_dynamic": False, |
| 948 | }, |
| 949 | "input_tensors": { |
| 950 | "dtype": "fp4", |
| 951 | "qscheme": "per_group", |
| 952 | "group_size": 32, |
| 953 | "scale_format": "e8m0", |
| 954 | "is_dynamic": True, |
| 955 | }, |
| 956 | }, |
| 957 | "layer_type_quant_config": {}, |
| 958 | "layer_quant_config": {"layers.0.attn.wkv": mxfp8_spec}, |
| 959 | "exclude": [], |
| 960 | } |
| 961 | ) |
| 962 | |
| 963 | class TestLinear(LinearBase): |
| 964 | def __init__(self): |
| 965 | torch.nn.Module.__init__(self) |
| 966 | |
| 967 | # get_quant_method_target now routes MXFP8 through the matcher chain. |
| 968 | wk, ak, mcls = config.get_quant_method_target("layers.0.attn.wkv", LinearBase) |
| 969 | assert wk == kMxfp8Static |
| 970 | assert mcls is QuarkLinearMethod |
| 971 | |
| 972 | linear = TestLinear() |
| 973 | method = config.get_quant_method(linear, "layers.0.attn.wkv") |
| 974 | assert isinstance(method, QuarkLinearMethod) |
| 975 | assert isinstance(linear.scheme, QuarkOCP_MX) |
| 976 | assert linear.scheme.weight_quant_key == kMxfp8Static |
| 977 | assert linear.scheme.scale_block_rows == 32 |
| 978 | |
| 979 | # Experts still fall through to the global MXFP4 spec. |
| 980 | assert ( |
| 981 | config.get_quant_method_target("layers.0.ffn.experts", RoutedExperts)[2] |
| 982 | is not QuarkLinearMethod |
| 983 | ) |
| 984 | |
| 985 | # 128x128 per-block FP8 (DeepSeek V4) keeps its existing Quark scheme. |
| 986 | v4_spec = { |
| 987 | "weight": {**mxfp8_spec["weight"], "block_size": [128, 128]}, |
| 988 | "input_tensors": {**mxfp8_spec["input_tensors"], "group_size": 128}, |
| 989 | } |
| 990 | v4_config = _make_qtensor_config(v4_spec["weight"], v4_spec["input_tensors"]) |
| 991 | assert v4_config.get_quant_method_target("linear", LinearBase)[2] is ( |
| 992 | QuarkLinearMethod |
| 993 | ) |
| 994 | linear = TestLinear() |
| 995 | assert isinstance(v4_config.get_quant_method(linear, "linear"), QuarkLinearMethod) |
| 996 | assert isinstance(linear.scheme, QuarkW8A8Fp8PerBlock) |
| 997 | |
| 998 | |
| 999 | def test_quant_method_dispatch_mxfp8_canonical(default_vllm_config): |
| 1000 | """Shape A (canonical MX): per_group/group_size=32/e8m0 FP8 routes |
| 1001 | to QuarkOCP_MX with scale_block_rows == 1. |
| 1002 | |
| 1003 | This covers the canonical MXFP8 spelling (no 2-D block). Validated by |
| 1004 | synthetic dispatch test only — no canonical-MXFP8 checkpoint on hand. |
| 1005 | """ |
| 1006 | from vllm.model_executor.layers.quantization.quark.schemes import QuarkOCP_MX |
| 1007 | from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp8Static |
| 1008 | |
| 1009 | default_vllm_config.model_config = SimpleNamespace(dtype=torch.bfloat16) |
| 1010 | mxfp8_canonical_spec = { |
| 1011 | "weight": { |
| 1012 | "dtype": "fp8_e4m3", |
| 1013 | "qscheme": "per_group", |
| 1014 | "group_size": 32, |
| 1015 | "scale_format": "e8m0", |
| 1016 | "is_dynamic": False, |
| 1017 | }, |
| 1018 | "input_tensors": { |
| 1019 | "dtype": "fp8_e4m3", |
| 1020 | "qscheme": "per_group", |
| 1021 | "group_size": 32, |
| 1022 | "scale_format": "e8m0", |
| 1023 | "is_dynamic": True, |
| 1024 | }, |
| 1025 | } |
| 1026 | config = QuarkConfig( |
| 1027 | { |
| 1028 | "global_quant_config": mxfp8_canonical_spec, |
| 1029 | "layer_type_quant_config": {}, |
| 1030 | "layer_quant_config": {}, |
| 1031 | "exclude": [], |
| 1032 | } |
| 1033 | ) |
| 1034 | |
| 1035 | wk, ak, mcls = config.get_quant_method_target("linear", LinearBase) |
| 1036 | assert wk == kMxfp8Static |
| 1037 | assert mcls is QuarkLinearMethod |
| 1038 | |
| 1039 | class TestLinear(LinearBase): |
| 1040 | def __init__(self): |
| 1041 | torch.nn.Module.__init__(self) |
| 1042 | |
| 1043 | linear = TestLinear() |
| 1044 | method = config.get_quant_method(linear, "linear") |
| 1045 | assert isinstance(method, QuarkLinearMethod) |
| 1046 | assert isinstance(linear.scheme, QuarkOCP_MX) |
| 1047 | assert linear.scheme.weight_quant_key == kMxfp8Static |
| 1048 | assert linear.scheme.scale_block_rows == 1 |
| 1049 | |
| 1050 | |
| 1051 | def test_quant_method_dispatch_mxfp8_moe_raises(default_vllm_config): |
| 1052 | """MXFP8 in a MoE config raises ValueError — experts are unsupported.""" |
| 1053 | from vllm.model_executor.layers.quantization.quark.quark_moe import ( |
| 1054 | QuarkOCP_MX_MoEMethod, |
| 1055 | ) |
| 1056 | |
| 1057 | default_vllm_config.model_config = SimpleNamespace(dtype=torch.bfloat16) |
| 1058 | mxfp8_spec = { |
| 1059 | "weight": { |
| 1060 | "dtype": "fp8_e4m3", |
| 1061 | "qscheme": "per_group", |
| 1062 | "group_size": 32, |
| 1063 | "scale_format": "e8m0", |
| 1064 | "is_dynamic": False, |
| 1065 | }, |
| 1066 | "input_tensors": { |
| 1067 | "dtype": "fp8_e4m3", |
| 1068 | "qscheme": "per_group", |
| 1069 | "group_size": 32, |
| 1070 | "scale_format": "e8m0", |
| 1071 | "is_dynamic": True, |
| 1072 | }, |
| 1073 | } |
| 1074 | config = QuarkConfig( |
| 1075 | { |
| 1076 | "global_quant_config": mxfp8_spec, |
| 1077 | "layer_type_quant_config": {}, |
| 1078 | "layer_quant_config": {}, |
| 1079 | "exclude": [], |
| 1080 | } |
| 1081 | ) |
| 1082 | wk, ak, mcls = config.get_quant_method_target("experts", RoutedExperts) |
| 1083 | assert mcls is QuarkOCP_MX_MoEMethod |
| 1084 | assert wk is not None |
| 1085 | # The OCP MX MoE constructor should fail loudly for MXFP8. |
| 1086 | fake_moe_config = MagicMock() |
| 1087 | with pytest.raises(ValueError, match="MXFP8 experts are not supported"): |
| 1088 | QuarkOCP_MX_MoEMethod(fake_moe_config, wk, ak) |
| 1089 | |
| 1090 | |
| 1091 | @pytest.mark.parametrize( |
| 1092 | ("weight", "input_tensors"), |
| 1093 | [ |
| 1094 | pytest.param( |
| 1095 | { |
| 1096 | "dtype": "int8", |
| 1097 | "qscheme": "per_group", |
| 1098 | "is_dynamic": False, |
| 1099 | "symmetric": True, |
| 1100 | }, |
| 1101 | { |
| 1102 | "dtype": "int8", |
| 1103 | "qscheme": "per_tensor", |
| 1104 | "is_dynamic": False, |
| 1105 | "symmetric": True, |
| 1106 | }, |
| 1107 | id="single_entry", |
| 1108 | ), |
| 1109 | pytest.param( |
| 1110 | [ |
| 1111 | {"dtype": "int8", "qscheme": "per_tensor"}, |
| 1112 | {"dtype": "int8", "qscheme": "per_tensor"}, |
| 1113 | ], |
| 1114 | [ |
| 1115 | {"dtype": "int8", "qscheme": "per_tensor"}, |
| 1116 | {"dtype": "int8", "qscheme": "per_tensor"}, |
| 1117 | ], |
| 1118 | id="multi_entry", |
| 1119 | ), |
| 1120 | ], |
| 1121 | ) |
| 1122 | def test_quant_method_dispatch_unsupported(weight, input_tensors): |
| 1123 | config = _make_qtensor_config(weight, input_tensors) |
| 1124 | |
| 1125 | class TestRoutedExperts(RoutedExperts): |
| 1126 | def __init__(self): |
| 1127 | torch.nn.Module.__init__(self) |
| 1128 | |
| 1129 | with pytest.raises(RuntimeError, match="^Unsupported FusedMoe scheme$"): |
| 1130 | config.get_quant_method_target("experts", RoutedExperts) |
| 1131 | |
| 1132 | with pytest.raises(RuntimeError, match="^Unsupported FusedMoe scheme$"): |
| 1133 | config.get_quant_method(TestRoutedExperts(), "experts") |
| 1134 | |
| 1135 | |
| 1136 | @pytest.mark.parametrize( |
| 1137 | "case", |
| 1138 | [case for case in QTENSOR_CONFIGS if case.expected_error is None], |
| 1139 | ids=lambda case: case.name, |
| 1140 | ) |
| 1141 | def test_quant_method_dispatch_instantiation(case, monkeypatch, default_vllm_config): |
| 1142 | config = _make_qtensor_config(case.weight, case.input_tensors) |
| 1143 | assert case.dispatch_cls is not None |
| 1144 | if issubclass(case.dispatch_cls, QuarkScheme): |
| 1145 | |
| 1146 | class TestLinear(LinearBase): |
| 1147 | def __init__(self): |
| 1148 | torch.nn.Module.__init__(self) |
| 1149 | |
| 1150 | monkeypatch.setattr( |
| 1151 | "vllm.model_executor.layers.quantization.quark.schemes." |
| 1152 | "quark_w8a8_fp8.get_current_vllm_config", |
| 1153 | lambda: SimpleNamespace(model_config=SimpleNamespace(dtype=torch.bfloat16)), |
| 1154 | ) |
| 1155 | layer = TestLinear() |
| 1156 | method = config.get_quant_method(layer, "linear") |
| 1157 | |
| 1158 | assert isinstance(method, QuarkLinearMethod) |
| 1159 | assert isinstance(layer.scheme, case.dispatch_cls) |
| 1160 | if case.weight_quant_key == kFp8Static128BlockE8M0Sym: |
| 1161 | # TODO: Remove once E8M0 quant key is properly handled in oracle |
| 1162 | assert layer.scheme.weight_quant_key == kFp8Static128BlockSym |
| 1163 | else: |
| 1164 | assert layer.scheme.weight_quant_key == case.weight_quant_key |
| 1165 | assert layer.scheme.activation_quant_key == case.act_quant_key |
| 1166 | else: |
| 1167 | |
| 1168 | class TestRoutedExperts(RoutedExperts): |
| 1169 | def __init__(self): |
| 1170 | torch.nn.Module.__init__(self) |
| 1171 | self.moe_config = _make_test_moe_config() |
| 1172 | |
| 1173 | for target in ( |
| 1174 | "select_fp8_moe_backend", |
| 1175 | "select_int8_moe_backend", |
| 1176 | "select_mxfp4_moe_backend", |
| 1177 | "backend_to_kernel_cls", |
| 1178 | "select_nvfp4_moe_backend", |
| 1179 | ): |
| 1180 | monkeypatch.setattr( |
| 1181 | f"vllm.model_executor.layers.quantization.quark.quark_moe.{target}", |
| 1182 | lambda *args, **kwargs: (object(), object()), |
| 1183 | ) |
| 1184 | |
| 1185 | # AssertionError: W4A8 FP8 MoE requires ROCm AITER fused MoE support |
| 1186 | monkeypatch.setattr( |
| 1187 | "vllm.model_executor.layers.quantization.quark.quark_moe." |
| 1188 | "rocm_aiter_ops.is_fused_moe_enabled", |
| 1189 | lambda: True, |
| 1190 | ) |
| 1191 | |
| 1192 | # default_vllm_config carries no model, but the FP8 and OCP MX methods |
| 1193 | # read the model type off the HF config. |
| 1194 | monkeypatch.setattr( |
| 1195 | "vllm.model_executor.layers.quantization.quark.quark_moe." |
| 1196 | "get_current_vllm_config", |
| 1197 | lambda: SimpleNamespace( |
| 1198 | model_config=SimpleNamespace(hf_config=SimpleNamespace()) |
| 1199 | ), |
| 1200 | ) |
| 1201 | |
| 1202 | layer = TestRoutedExperts() |
| 1203 | method = config.get_quant_method(layer, "experts") |
| 1204 | |
| 1205 | assert isinstance(method, case.dispatch_cls) |
| 1206 | |
| 1207 | |
| 1208 | QUARK_MOE_MODULE = "vllm.model_executor.layers.quantization.quark.quark_moe" |
| 1209 | # validate_fp8_block_shape_moe imports this from vllm.distributed when called, |
| 1210 | # so the patch has to target the source module rather than a local binding. |
| 1211 | TP_WORLD_SIZE = "vllm.distributed.get_tensor_model_parallel_world_size" |
| 1212 | |
| 1213 | |
| 1214 | def _make_per_block_fp8_moe_method( |
| 1215 | activation_quant_key: QuantKey = kFp8Dynamic128Sym, |
| 1216 | ) -> QuarkW8A8Fp8MoEMethod: |
| 1217 | return QuarkW8A8Fp8MoEMethod( |
| 1218 | _make_test_moe_config(), |
| 1219 | kFp8Static128BlockSym, |
| 1220 | activation_quant_key, |
| 1221 | ) |
| 1222 | |
| 1223 | |
| 1224 | def test_quark_w8a8_fp8_moe_per_block_requires_dynamic_group_input(): |
| 1225 | with ( |
| 1226 | patch( |
| 1227 | f"{QUARK_MOE_MODULE}.select_fp8_moe_backend", |
| 1228 | return_value=(Mock(), Mock()), |
| 1229 | ), |
| 1230 | patch(f"{QUARK_MOE_MODULE}.get_current_vllm_config"), |
| 1231 | pytest.raises(ValueError, match="per-block scales"), |
| 1232 | ): |
| 1233 | _make_per_block_fp8_moe_method(kFp8StaticTensorSym) |
| 1234 | |
| 1235 | |
| 1236 | def test_quark_w8a8_fp8_moe_per_block_weight_shapes(): |
| 1237 | with ( |
| 1238 | patch( |
| 1239 | f"{QUARK_MOE_MODULE}.select_fp8_moe_backend", |
| 1240 | return_value=(Mock(), Mock()), |
| 1241 | ), |
| 1242 | patch(f"{QUARK_MOE_MODULE}.get_current_vllm_config"), |
| 1243 | patch(TP_WORLD_SIZE, return_value=1), |
| 1244 | ): |
| 1245 | method = _make_per_block_fp8_moe_method() |
| 1246 | assert method.block_quant |
| 1247 | assert method.weight_block_size == [128, 128] |
| 1248 | |
| 1249 | layer = torch.nn.Module() |
| 1250 | method.create_weights( |
| 1251 | layer, |
| 1252 | num_experts=4, |
| 1253 | hidden_size=512, |
| 1254 | intermediate_size_per_partition=256, |
| 1255 | params_dtype=torch.bfloat16, |
| 1256 | ) |
| 1257 | |
| 1258 | w13_num_shards = method.moe.w13_num_shards |
| 1259 | assert layer.weight_block_size == [128, 128] |
| 1260 | assert layer.w13_weight.shape == (4, w13_num_shards * 256, 512) |
| 1261 | assert layer.w2_weight.shape == (4, 512, 256) |
| 1262 | # Quark exports block scales as `weight_scale`, like the per-tensor and |
| 1263 | # per-channel schemes, so all schemes register the same parameter name. |
| 1264 | assert not hasattr(layer, "w13_weight_scale_inv") |
| 1265 | assert not hasattr(layer, "w2_weight_scale_inv") |
| 1266 | # One scale per 128x128 tile of each expert's weight. |
| 1267 | assert layer.w13_weight_scale.shape == ( |
| 1268 | 4, |
| 1269 | w13_num_shards * (256 // 128), |
| 1270 | 512 // 128, |
| 1271 | ) |
| 1272 | assert layer.w2_weight_scale.shape == (4, 512 // 128, 256 // 128) |
| 1273 | # The loader shards block scales on the block grid, not per row. |
| 1274 | for scale in (layer.w13_weight_scale, layer.w2_weight_scale): |
| 1275 | assert scale.quant_method == FusedMoeWeightScaleSupported.BLOCK.value |
| 1276 | |
| 1277 | |
| 1278 | def test_quark_w8a8_fp8_moe_per_block_rejects_misaligned_partition(): |
| 1279 | with ( |
| 1280 | patch( |
| 1281 | f"{QUARK_MOE_MODULE}.select_fp8_moe_backend", |
| 1282 | return_value=(Mock(), Mock()), |
| 1283 | ), |
| 1284 | patch(f"{QUARK_MOE_MODULE}.get_current_vllm_config"), |
| 1285 | patch(TP_WORLD_SIZE, return_value=1), |
| 1286 | pytest.raises(ValueError, match="not divisible by"), |
| 1287 | ): |
| 1288 | _make_per_block_fp8_moe_method().create_weights( |
| 1289 | torch.nn.Module(), |
| 1290 | num_experts=4, |
| 1291 | hidden_size=512, |
| 1292 | intermediate_size_per_partition=192, |
| 1293 | params_dtype=torch.bfloat16, |
| 1294 | ) |
| 1295 | |
| 1296 | |
| 1297 | def test_quark_fp8_ptpc_exposes_kernel_input_quant_key(monkeypatch): |
| 1298 | """QuarkW8A8Fp8 must advertise the key its kernel consumes pre-quantized. |
| 1299 | |
| 1300 | Only kernel selection runs, no GEMM. The shape matters: AITER PTPC |
| 1301 | kernels decline untuned shapes, leaving a torch kernel that has no key. |
| 1302 | """ |
| 1303 | # Llama-3.1-70B qkv_proj at TP1 (aiter-tuned N, K on gfx950). |
| 1304 | N, K = 10240, 8192 |
| 1305 | dtype = torch.bfloat16 |
| 1306 | kernel_config = FP8ScaledMMLinearLayerConfig( |
| 1307 | weight_quant_key=kFp8StaticChannelSym, |
| 1308 | activation_quant_key=kFp8DynamicTokenSym, |
| 1309 | weight_shape=(N, K), |
| 1310 | input_dtype=dtype, |
| 1311 | out_dtype=dtype, |
| 1312 | ) |
| 1313 | if not any( |
| 1314 | cls.is_supported()[0] and cls.can_implement(kernel_config)[0] |
| 1315 | for cls in AITER_PTPC_KERNELS |
| 1316 | ): |
| 1317 | pytest.skip("no AITER PTPC kernel is usable for this shape and environment") |
| 1318 | |
| 1319 | monkeypatch.setattr(parameter, "get_tensor_model_parallel_rank", lambda: 0) |
| 1320 | monkeypatch.setattr(parameter, "get_tensor_model_parallel_world_size", lambda: 1) |
| 1321 | |
| 1322 | scheme = QuarkW8A8Fp8.__new__(QuarkW8A8Fp8) |
| 1323 | scheme.weight_qscheme = "per_channel" |
| 1324 | scheme.is_static_input_scheme = False |
| 1325 | scheme.activation_quant_key = kFp8DynamicTokenSym |
| 1326 | scheme.weight_quant_key = kFp8StaticChannelSym |
| 1327 | scheme.out_dtype = dtype |
| 1328 | scheme.input_dtype = dtype |
| 1329 | |
| 1330 | layer = torch.nn.Module() |
| 1331 | with set_current_vllm_config(VllmConfig()): |
| 1332 | scheme.create_weights( |
| 1333 | layer, |
| 1334 | output_partition_sizes=[N], |
| 1335 | input_size_per_partition=K, |
| 1336 | params_dtype=dtype, |
| 1337 | weight_loader=lambda *args, **kwargs: None, |
| 1338 | ) |
| 1339 | |
| 1340 | assert isinstance(scheme.fp8_linear, AITER_PTPC_KERNELS) |
| 1341 | assert layer.input_quant_key == kFp8DynamicTokenSym |
| 1342 | |
| 1343 | |
| 1344 | @pytest.mark.parametrize("kv_cache_dtype", ["auto", "fp8"]) |
| 1345 | def test_quark_fp8_w_per_tensor_a_per_tensor( |
| 1346 | kv_cache_dtype: CacheDType, monkeypatch, dist_init, workspace_init |
| 1347 | ): |
| 1348 | model_path = "amd/Llama-3.1-8B-Instruct-FP8-KV-Quark-test" |
| 1349 | checkpoint_scales = {} |
| 1350 | scale_names = { |
| 1351 | "model.layers.0.self_attn.k_proj.output_scale", |
| 1352 | "model.layers.0.self_attn.v_proj.output_scale", |
| 1353 | } |
| 1354 | original_load_weights = LlamaForCausalLM.load_weights |
| 1355 | |
| 1356 | def load_weights(self, weights): |
| 1357 | def capture_scales(): |
| 1358 | for name, weight in weights: |
| 1359 | if name in scale_names: |
| 1360 | checkpoint_scales[name] = weight.detach().cpu() |
| 1361 | yield name, weight |
| 1362 | |
| 1363 | return original_load_weights(self, capture_scales()) |
| 1364 | |
| 1365 | monkeypatch.setattr(LlamaForCausalLM, "load_weights", load_weights) |
| 1366 | model, vllm_config = load_model_without_vllm_runner( |
| 1367 | model_path, |
| 1368 | model_config_kwargs={"hf_overrides": {"num_hidden_layers": 3}}, |
| 1369 | vllm_config_kwargs={"cache_config": CacheConfig(cache_dtype=kv_cache_dtype)}, |
| 1370 | ) |
| 1371 | |
| 1372 | qkv_proj = model.model.layers[0].self_attn.qkv_proj |
| 1373 | assert isinstance(qkv_proj.quant_method, QuarkLinearMethod) |
| 1374 | assert isinstance(qkv_proj.scheme, QuarkW8A8Fp8) |
| 1375 | assert len(qkv_proj.input_scale.shape) == 0 |
| 1376 | assert qkv_proj.weight.dtype is current_platform.fp8_dtype() |
| 1377 | assert len(qkv_proj.weight_scale.shape) == 0 |
| 1378 | |
| 1379 | attn = model.model.layers[0].self_attn.attn |
| 1380 | if kv_cache_dtype == "fp8": |
| 1381 | assert checkpoint_scales.keys() == scale_names |
| 1382 | scale_multiplier = 2 if current_platform.is_fp8_fnuz() else 1 |
| 1383 | assert attn._k_scale_float == ( |
| 1384 | checkpoint_scales["model.layers.0.self_attn.k_proj.output_scale"].item() |
| 1385 | * scale_multiplier |
| 1386 | ) |
| 1387 | assert attn._v_scale_float == ( |
| 1388 | checkpoint_scales["model.layers.0.self_attn.v_proj.output_scale"].item() |
| 1389 | * scale_multiplier |
| 1390 | ) |
| 1391 | else: |
| 1392 | assert attn._k_scale_float == 1.0 |
| 1393 | assert attn._v_scale_float == 1.0 |
| 1394 | |
| 1395 | monkeypatch.setattr(Attention, "forward", lambda _, q, k, v: q.contiguous()) |
| 1396 | input_ids = torch.tensor([1, 2, 3, 4], device=DEVICE_TYPE) |
| 1397 | positions = torch.arange(input_ids.numel(), device=DEVICE_TYPE) |
| 1398 | with ( |
| 1399 | set_current_vllm_config(vllm_config), |
| 1400 | set_forward_context(None, vllm_config, num_tokens=input_ids.numel()), |
| 1401 | ): |
| 1402 | hidden_states = model(input_ids, positions, None) |
| 1403 | logits = model.compute_logits(hidden_states) |
| 1404 | assert torch.isfinite(logits).all() |
| 1405 | |
| 1406 | |
| 1407 | def test_quark_fp8_w_per_channel_a_per_token(monkeypatch, dist_init, workspace_init): |
| 1408 | model_path = "amd/Qwen2.5-1.5B-Instruct-ptpc-Quark-ts" |
| 1409 | model, vllm_config = load_model_without_vllm_runner( |
| 1410 | model_path, |
| 1411 | model_config_kwargs={"hf_overrides": {"num_hidden_layers": 3}}, |
| 1412 | ) |
| 1413 | |
| 1414 | qkv_proj = model.model.layers[0].self_attn.qkv_proj |
| 1415 | assert isinstance(qkv_proj.quant_method, QuarkLinearMethod) |
| 1416 | assert isinstance(qkv_proj.scheme, QuarkW8A8Fp8) |
| 1417 | assert qkv_proj.weight.dtype is current_platform.fp8_dtype() |
| 1418 | assert qkv_proj.weight_scale.shape[0] == qkv_proj.weight.shape[1] |
| 1419 | assert qkv_proj.weight_scale.shape[1] == 1 |
| 1420 | |
| 1421 | monkeypatch.setattr(Attention, "forward", lambda _, q, k, v: q.contiguous()) |
| 1422 | input_ids = torch.tensor([1, 2, 3, 4], device=DEVICE_TYPE) |
| 1423 | positions = torch.arange(input_ids.numel(), device=DEVICE_TYPE) |
| 1424 | with ( |
| 1425 | set_current_vllm_config(vllm_config), |
| 1426 | set_forward_context(None, vllm_config, num_tokens=input_ids.numel()), |
| 1427 | ): |
| 1428 | hidden_states = model(input_ids, positions, None) |
| 1429 | logits = model.compute_logits(hidden_states) |
| 1430 | assert torch.isfinite(logits).all() |
| 1431 | |
| 1432 | |
| 1433 | def test_quark_int8_w_per_tensor_a_per_tensor(monkeypatch, dist_init, workspace_init): |
| 1434 | model_path = "amd/Llama-3.1-8B-Instruct-w-int8-a-int8-sym-test" |
| 1435 | model, vllm_config = load_model_without_vllm_runner( |
| 1436 | model_path, |
| 1437 | model_config_kwargs={"hf_overrides": {"num_hidden_layers": 3}}, |
| 1438 | ) |
| 1439 | with set_current_vllm_config(vllm_config): |
| 1440 | qkv_proj = model.model.layers[0].self_attn.qkv_proj |
| 1441 | assert isinstance(qkv_proj.quant_method, QuarkLinearMethod) |
| 1442 | assert isinstance(qkv_proj.scheme, QuarkW8A8Int8) |
| 1443 | |
| 1444 | monkeypatch.setattr(Attention, "forward", lambda _, q, k, v: q.contiguous()) |
| 1445 | input_ids = torch.tensor([1, 2, 3, 4], device=DEVICE_TYPE) |
| 1446 | positions = torch.arange(input_ids.numel(), device=DEVICE_TYPE) |
| 1447 | with set_forward_context(None, vllm_config, num_tokens=input_ids.numel()): |
| 1448 | hidden_states = model(input_ids, positions, None) |
| 1449 | logits = model.compute_logits(hidden_states) |
| 1450 | assert torch.isfinite(logits).all() |
| 1451 | |
| 1452 | |
| 1453 | @pytest.mark.parametrize("tp", [1]) |
| 1454 | def test_quark_int8_w8a8_moe(vllm_runner, tp): |
| 1455 | """Test W8A8 INT8 MoE quantization with a tiny Qwen3 MoE model.""" |
| 1456 | model_path = "amd/tiny-qwen3-moe-w8a8-int8" |
| 1457 | with vllm_runner( |
| 1458 | model_path, |
| 1459 | enforce_eager=True, |
| 1460 | tensor_parallel_size=tp, |
| 1461 | gpu_memory_utilization=0.1, |
| 1462 | ) as llm: |
| 1463 | |
| 1464 | def check_model(model): |
| 1465 | layer = model.model.layers[0] |
| 1466 | # MoE experts should use QuarkW8A8Int8MoEMethod |
| 1467 | moe = layer.mlp.experts |
| 1468 | assert isinstance(moe._quant_method, QuarkW8A8Int8MoEMethod), ( |
| 1469 | f"Expected QuarkW8A8Int8MoEMethod, got {type(moe._quant_method)}" |
| 1470 | ) |
| 1471 | # Non-MoE linear layers should use QuarkW8A8Int8 |
| 1472 | qkv_proj = layer.self_attn.qkv_proj |
| 1473 | assert isinstance(qkv_proj.scheme, QuarkW8A8Int8) |
| 1474 | |
| 1475 | llm.apply_model(check_model) |
| 1476 | |
| 1477 | output = llm.generate_greedy("Hello", max_tokens=4) |
| 1478 | assert output |
| 1479 | |
| 1480 | |
| 1481 | @pytest.mark.parametrize("tp", [1]) |
| 1482 | def test_quark_fp8_w8a8_per_block_moe(vllm_runner, tp): |
| 1483 | """Test per-block (128x128) FP8 MoE quantization with a tiny Qwen3 MoE model.""" |
| 1484 | model_path = "Adamji/tiny-qwen3-moe-fp8-per-block" |
| 1485 | with vllm_runner( |
| 1486 | model_path, |
| 1487 | enforce_eager=True, |
| 1488 | tensor_parallel_size=tp, |
| 1489 | gpu_memory_utilization=0.1, |
| 1490 | ) as llm: |
| 1491 | |
| 1492 | def check_model(model): |
| 1493 | experts = model.model.layers[0].mlp.experts |
| 1494 | method = experts._quant_method |
| 1495 | assert isinstance(method, QuarkW8A8Fp8MoEMethod), ( |
| 1496 | f"Expected QuarkW8A8Fp8MoEMethod, got {type(method)}" |
| 1497 | ) |
| 1498 | assert method.weight_qscheme == "per_block" |
| 1499 | assert method.weight_block_size == [128, 128] |
| 1500 | |
| 1501 | # hidden_size=128, moe_intermediate_size=256 and 4 experts, so one |
| 1502 | # scale per 128x128 tile, with w13 stacking gate on top of up. |
| 1503 | routed_experts = experts.routed_experts |
| 1504 | assert routed_experts.w13_weight_scale.shape == (4, 4, 1) |
| 1505 | assert routed_experts.w2_weight_scale.shape == (4, 1, 2) |
| 1506 | # Quark exports the block scales under the same name as the |
| 1507 | # per-tensor and per-channel schemes. |
| 1508 | assert not hasattr(routed_experts, "w13_weight_scale_inv") |
| 1509 | assert not hasattr(routed_experts, "w2_weight_scale_inv") |
| 1510 | |
| 1511 | llm.apply_model(check_model) |
| 1512 | |
| 1513 | output = llm.generate_greedy("Hello", max_tokens=4) |
| 1514 | assert output |
| 1515 | |
| 1516 | |
| 1517 | @pytest.mark.skipif( |
| 1518 | not (on_gfx950() or on_gfx942()), |
| 1519 | reason="Quark W4A8 (INT4-FP8) MoE requires the AITER kernel on gfx942/gfx950", |
| 1520 | ) |
| 1521 | def test_quark_w4a8_fp8_moe(monkeypatch, dist_init, workspace_init): |
| 1522 | """Test W4A8 (INT4 weight + FP8 activation) MoE with a tiny Qwen3 MoE model. |
| 1523 | |
| 1524 | W4A8 dispatches through the AITER fused MoE kernel, so AITER must be on. |
| 1525 | """ |
| 1526 | monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1") |
| 1527 | monkeypatch.setenv("VLLM_ROCM_USE_AITER_MOE", "1") |
| 1528 | rocm_aiter_ops.refresh_env_variables() |
| 1529 | |
| 1530 | model_path = "amd/tiny-qwen3-moe-w4a8" |
| 1531 | model, vllm_config = load_model_without_vllm_runner( |
| 1532 | model_path, |
| 1533 | ) |
| 1534 | with set_current_vllm_config(vllm_config): |
| 1535 | moe = model.model.layers[0].mlp.experts |
| 1536 | assert isinstance(moe._quant_method, QuarkW4A8Fp8MoEMethod), ( |
| 1537 | f"Expected QuarkW4A8Fp8MoEMethod, got {type(moe._quant_method)}" |
| 1538 | ) |
| 1539 | |
| 1540 | monkeypatch.setattr(Attention, "forward", lambda _, q, k, v: q.contiguous()) |
| 1541 | input_ids = torch.tensor([1, 2, 3, 4], device=DEVICE_TYPE) |
| 1542 | positions = torch.arange(input_ids.numel(), device=DEVICE_TYPE) |
| 1543 | with set_forward_context(None, vllm_config, num_tokens=input_ids.numel()): |
| 1544 | hidden_states = model(input_ids, positions, None) |
| 1545 | logits = model.compute_logits(hidden_states) |
| 1546 | assert torch.isfinite(logits).all() |
| 1547 | |
| 1548 | |
| 1549 | def test_quark_fp8_parity(dist_init, workspace_init): |
| 1550 | quark_model_id = "amd-quark/llama-tiny-fp8-quark-quant-method" |
| 1551 | fp8_model_id = "amd-quark/llama-tiny-fp8-quant-method" |
| 1552 | |
| 1553 | def load_state_dict(model_id: str) -> dict[str, torch.Tensor]: |
| 1554 | model, _ = load_model_without_vllm_runner(model_id) |
| 1555 | return {k: v.cpu() for k, v in model.state_dict().items()} |
| 1556 | |
| 1557 | quark_state_dict = load_state_dict(quark_model_id) |
| 1558 | fp8_state_dict = load_state_dict(fp8_model_id) |
| 1559 | |
| 1560 | assert fp8_state_dict.keys() == quark_state_dict.keys() |
| 1561 | |
| 1562 | for key in fp8_state_dict: |
| 1563 | assert torch.equal(fp8_state_dict[key], quark_state_dict[key]) |
| 1564 | |
| 1565 | |
| 1566 | @dataclass |
| 1567 | class AccuracyTestConfig: |
| 1568 | model_name: str |
| 1569 | excepted_value: float |
| 1570 | |
| 1571 | def get_model_args( |
| 1572 | self, |
| 1573 | tp_size: int, |
| 1574 | model_max_len: int | None = None, |
| 1575 | kwargs: dict | None = None, |
| 1576 | ) -> dict: |
| 1577 | if kwargs is None: |
| 1578 | kwargs = {} |
| 1579 | |
| 1580 | model_args = { |
| 1581 | "pretrained": self.model_name, |
| 1582 | "dtype": "auto", |
| 1583 | "add_bos_token": True, |
| 1584 | "tensor_parallel_size": tp_size, |
| 1585 | "gpu_memory_utilization": 0.7, |
| 1586 | **kwargs, |
| 1587 | } |
| 1588 | if model_max_len is not None: |
| 1589 | model_args["max_model_len"] = model_max_len |
| 1590 | |
| 1591 | return model_args |
| 1592 | |
| 1593 | |
| 1594 | WIKITEXT_ACCURACY_CONFIGS = [ |
| 1595 | AccuracyTestConfig( |
| 1596 | model_name="fxmarty/qwen1.5_moe_a2.7b_chat_w_fp4_a_fp6_e2m3", |
| 1597 | excepted_value=11.3, |
| 1598 | ), |
| 1599 | AccuracyTestConfig( |
| 1600 | model_name="fxmarty/qwen1.5_moe_a2.7b_chat_w_fp6_e3m2_a_fp6_e3m2", |
| 1601 | excepted_value=10.6, |
| 1602 | ), |
| 1603 | ] |
| 1604 | |
| 1605 | |
| 1606 | @pytest.mark.skipif( |
| 1607 | not QUARK_MXFP4_AVAILABLE, |
| 1608 | reason=f"amd-quark>={QUARK_MXFP4_MIN_VERSION} is not available", |
| 1609 | ) |
| 1610 | @pytest.mark.parametrize( |
| 1611 | "config", WIKITEXT_ACCURACY_CONFIGS, ids=lambda config: config.model_name |
| 1612 | ) |
| 1613 | @pytest.mark.parametrize("tp_size", [1, 2]) |
| 1614 | def test_ocp_mx_wikitext_correctness(config: AccuracyTestConfig, tp_size: int): |
| 1615 | device_count = torch.accelerator.device_count() |
| 1616 | if device_count < tp_size: |
| 1617 | pytest.skip(f"This test requires >={tp_size} gpus, got only {device_count}") |
| 1618 | |
| 1619 | results = lm_eval.simple_evaluate( |
| 1620 | model="vllm", |
| 1621 | model_args=config.get_model_args( |
| 1622 | tp_size=tp_size, kwargs={"cudagraph_capture_sizes": [16]} |
| 1623 | ), |
| 1624 | tasks="wikitext", |
| 1625 | batch_size=64, |
| 1626 | ) |
| 1627 | |
| 1628 | measured_value = results["results"]["wikitext"]["word_perplexity,none"] |
| 1629 | assert measured_value == pytest.approx(config.excepted_value, abs=0.1) |
| 1630 | |
| 1631 | |
| 1632 | GSM8K_ACCURACY_CONFIGS = [ |
| 1633 | # Private model. |
| 1634 | AccuracyTestConfig( |
| 1635 | model_name="amd/DeepSeek-R1-WMXFP4-AMXFP4-Scale-UINT8-MoE-Quant", |
| 1636 | excepted_value=0.96, |
| 1637 | ), |
| 1638 | ] |
| 1639 | |
| 1640 | |
| 1641 | @pytest.mark.parametrize("config", GSM8K_ACCURACY_CONFIGS) |
| 1642 | @pytest.mark.skipif( |
| 1643 | not QUARK_MXFP4_AVAILABLE, |
| 1644 | reason=f"amd-quark>={QUARK_MXFP4_MIN_VERSION} is not available", |
| 1645 | ) |
| 1646 | @pytest.mark.skipif( |
| 1647 | not HF_HUB_AMD_ORG_ACCESS, |
| 1648 | reason="Read access to huggingface.co/amd is required for this test.", |
| 1649 | ) |
| 1650 | def test_mxfp4_gsm8k_correctness(config: AccuracyTestConfig): |
| 1651 | device_count = torch.accelerator.device_count() |
| 1652 | if device_count < 8: |
| 1653 | pytest.skip(f"This test requires >=8 gpus, got only {device_count}") |
| 1654 | |
| 1655 | task = "gsm8k" |
| 1656 | rtol = 0.03 |
| 1657 | |
| 1658 | results = lm_eval.simple_evaluate( |
| 1659 | model="vllm", |
| 1660 | model_args=config.get_model_args(tp_size=8, model_max_len=38768), |
| 1661 | tasks=task, |
| 1662 | batch_size=64, |
| 1663 | num_fewshot=8, |
| 1664 | ) |
| 1665 | |
| 1666 | EXPECTED_VALUE = config.excepted_value |
| 1667 | measured_value = results["results"][task]["exact_match,strict-match"] |
| 1668 | assert ( |
| 1669 | measured_value - rtol < EXPECTED_VALUE |
| 1670 | and measured_value + rtol > EXPECTED_VALUE |
| 1671 | ), f"Expected: {EXPECTED_VALUE} | Measured: {measured_value}" |
| 1672 | |
| 1673 | |
| 1674 | @pytest.mark.skipif( |
| 1675 | not QUARK_MXFP4_AVAILABLE, |
| 1676 | reason=f"amd-quark>={QUARK_MXFP4_MIN_VERSION} is not available", |
| 1677 | ) |
| 1678 | @pytest.mark.parametrize("float_dtype", [torch.bfloat16, torch.float16]) |
| 1679 | @pytest.mark.parametrize("scalings", [[2.3, 0.03, 7.3, 0.1, 0.004, 17.3, 1e4, 1e-4]]) |
| 1680 | def test_mxfp4_fused_qdq_match_quark(float_dtype: torch.dtype, scalings: list[int]): |
| 1681 | torch.manual_seed(0) |
| 1682 | |
| 1683 | hidden_size = 64 * 32 |
| 1684 | inp = (torch.rand(1, hidden_size, dtype=float_dtype, device=DEVICE_TYPE) - 0.5) * 2 |
| 1685 | for i in range(hidden_size // 32): |
| 1686 | inp[:, i * 32 : (i + 1) * 32] = ( |
| 1687 | inp[:, i * 32 : (i + 1) * 32] * scalings[i % len(scalings)] |
| 1688 | ) |
| 1689 | |
| 1690 | inp_kernel = inp.clone() |
| 1691 | inp_kernel_clone = inp_kernel.clone() |
| 1692 | |
| 1693 | res_hip = mx_kernel.qdq_mxfp4_hip(inp_kernel_clone, "even") |
| 1694 | res_torch = qdq_mxfp4_torch(inp_kernel, "even") |
| 1695 | |
| 1696 | for i in range(hidden_size // 32): |
| 1697 | assert torch.all(torch.isfinite(res_hip[:, i * 32 : (i + 1) * 32])) |
| 1698 | assert torch.all(torch.isfinite(res_torch[:, i * 32 : (i + 1) * 32])) |
| 1699 | |
| 1700 | torch.testing.assert_close( |
| 1701 | res_hip[:, i * 32 : (i + 1) * 32], res_torch[:, i * 32 : (i + 1) * 32] |
| 1702 | ) |
| 1703 | |
| 1704 | |
| 1705 | @pytest.mark.skipif( |
| 1706 | not QUARK_MXFP4_AVAILABLE, |
| 1707 | reason=f"amd-quark>={QUARK_MXFP4_MIN_VERSION} is not available", |
| 1708 | ) |
| 1709 | @pytest.mark.parametrize("float_dtype", [torch.bfloat16, torch.float16]) |
| 1710 | @pytest.mark.parametrize("scalings", [[2.3, 0.03, 7.3, 0.1, 0.004, 17.3, 1e4, 1e-4]]) |
| 1711 | def test_mxfp4_dequant_kernel_match_quark( |
| 1712 | float_dtype: torch.dtype, scalings: list[int] |
| 1713 | ): |
| 1714 | qspec = FP4PerGroupSpec( |
| 1715 | ch_axis=-1, |
| 1716 | group_size=32, |
| 1717 | scale_format="e8m0", |
| 1718 | scale_calculation_mode="even", |
| 1719 | is_dynamic=False, |
| 1720 | ).to_quantization_spec() |
| 1721 | |
| 1722 | weight_quantizer = StaticScaledRealQuantizer( |
| 1723 | qspec=qspec, |
| 1724 | quantizer=None, |
| 1725 | reorder=False, |
| 1726 | real_quantized=True, |
| 1727 | float_dtype=float_dtype, |
| 1728 | device=DEVICE_TYPE, |
| 1729 | ) |
| 1730 | |
| 1731 | observer = qspec.observer_cls(qspec, device=DEVICE_TYPE) |
| 1732 | |
| 1733 | hidden_size = 512 |
| 1734 | shape = (11008, hidden_size) |
| 1735 | |
| 1736 | w = (torch.rand(shape, device=DEVICE_TYPE, dtype=float_dtype) - 0.5) * 2 |
| 1737 | |
| 1738 | # Make it so that different groups have different scales. |
| 1739 | for i in range(hidden_size // 32): |
| 1740 | w[:, i * 32 : (i + 1) * 32] = ( |
| 1741 | w[:, i * 32 : (i + 1) * 32] * scalings[i % len(scalings)] |
| 1742 | ) |
| 1743 | |
| 1744 | observer(w) |
| 1745 | scale, _ = observer._calculate_qparams() |
| 1746 | weight_quantizer.scale = scale |
| 1747 | |
| 1748 | w_mxfp4 = weight_quantizer.to_real_quantize_params(w).to(DEVICE_TYPE) |
| 1749 | weight_quantizer.maybe_convert_and_transpose_scale() |
| 1750 | |
| 1751 | scale = weight_quantizer.scale |
| 1752 | |
| 1753 | out_hip = mx_kernel.dq_mxfp4_hip(w_mxfp4, scale, float_dtype) |
| 1754 | |
| 1755 | out_torch = dq_mxfp4_torch(w_mxfp4, scale, float_dtype) |
| 1756 | |
| 1757 | assert torch.equal(out_hip, out_torch) |
| 1758 | |
| 1759 | |
| 1760 | @pytest.mark.skipif( |
| 1761 | not QUARK_MXFP4_AVAILABLE, |
| 1762 | reason=f"amd-quark>={QUARK_MXFP4_MIN_VERSION} is not available", |
| 1763 | ) |
| 1764 | @pytest.mark.skipif( |
| 1765 | not AITER_AVAILABLE, |
| 1766 | reason="AITER is not found or not supported on the current platform", |
| 1767 | ) |
| 1768 | @pytest.mark.parametrize("float_dtype", [torch.bfloat16, torch.float16]) |
| 1769 | @pytest.mark.parametrize("scalings", [[2.3, 0.03, 7.3, 0.1, 0.004, 17.3, 1e4, 1e-4]]) |
| 1770 | def test_mxfp4_dynamic_quant_match_quark( |
| 1771 | float_dtype: torch.dtype, scalings: list[float] |
| 1772 | ): |
| 1773 | """`AiterMxfp4LinearKernel` quantizes weights dynamically through AITER's |
| 1774 | `dynamic_mxfp4_quant`, while the emulation path quantizes/dequantizes |
| 1775 | through Quark's `qdq_mxfp4`. Check that both agree on the same input. |
| 1776 | """ |
| 1777 | from aiter.ops.triton.quant import dynamic_mxfp4_quant |
| 1778 | |
| 1779 | torch.manual_seed(0) |
| 1780 | |
| 1781 | hidden_size = 32 * 64 |
| 1782 | inp = (torch.rand(48, hidden_size, dtype=float_dtype, device=DEVICE_TYPE) - 0.5) * 2 |
| 1783 | for i in range(hidden_size // 32): |
| 1784 | inp[:, i * 32 : (i + 1) * 32] = ( |
| 1785 | inp[:, i * 32 : (i + 1) * 32] * scalings[i % len(scalings)] |
| 1786 | ) |
| 1787 | |
| 1788 | x_q, x_s = dynamic_mxfp4_quant(inp) |
| 1789 | out_dynamic_quant = dq_mxfp4_torch(x_q, x_s, float_dtype) |
| 1790 | |
| 1791 | out_quark_qdq = quant_dequant_mxfp4(inp) |
| 1792 | |
| 1793 | assert torch.equal(out_dynamic_quant, out_quark_qdq) |
| 1794 | |
| 1795 | |
| 1796 | # Unit tests for ``is_layer_skipped`` fused-name handling. |
| 1797 | |
| 1798 | FUSED_MAPPING = { |
| 1799 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 1800 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 1801 | } |
| 1802 | |
| 1803 | |
| 1804 | def test_quark_should_ignore_layer_checks_children(): |
| 1805 | assert should_ignore_layer( |
| 1806 | "model.layers.78.mlp.experts", |
| 1807 | ["model.layers.78.mlp.experts.0.down_proj"], |
| 1808 | check_children=True, |
| 1809 | ) |
| 1810 | |
| 1811 | |
| 1812 | def test_quark_should_ignore_layer_rejects_partial_fused_matches(): |
| 1813 | with pytest.raises(ValueError, match="different quantization schemes"): |
| 1814 | should_ignore_layer( |
| 1815 | "model.layers.0.self_attn.qkv_proj", |
| 1816 | ["model.layers.0.self_attn.q_proj"], |
| 1817 | FUSED_MAPPING, |
| 1818 | ) |
| 1819 | |
| 1820 | |
| 1821 | def test_fused_name_listed_directly_is_skipped(): |
| 1822 | # Regression for Step-3.5-Flash-FP8: the checkpoint lists the fused |
| 1823 | # name (``qkv_proj``) directly in ``modules_to_not_convert``. When a |
| 1824 | # ``packed_modules_mapping`` is registered on the model, the fused |
| 1825 | # match must still win over per-shard expansion. |
| 1826 | ignored = ["model.layers.0.self_attn.qkv_proj"] |
| 1827 | assert is_layer_skipped( |
| 1828 | prefix="model.layers.0.self_attn.qkv_proj", |
| 1829 | ignored_layers=ignored, |
| 1830 | fused_mapping=FUSED_MAPPING, |
| 1831 | ) |
| 1832 | assert is_layer_skipped( |
| 1833 | prefix="model.layers.0.mlp.gate_up_proj", |
| 1834 | ignored_layers=["model.layers.0.mlp.gate_up_proj"], |
| 1835 | fused_mapping=FUSED_MAPPING, |
| 1836 | ) |
| 1837 | |
| 1838 | |
| 1839 | def test_unfused_shards_listed_is_skipped(): |
| 1840 | # Quark INT8 style: per-shard names listed; all shards present means |
| 1841 | # the fused layer is skipped via expansion. |
| 1842 | ignored = [ |
| 1843 | "model.layers.0.self_attn.q_proj", |
| 1844 | "model.layers.0.self_attn.k_proj", |
| 1845 | "model.layers.0.self_attn.v_proj", |
| 1846 | ] |
| 1847 | assert is_layer_skipped( |
| 1848 | prefix="model.layers.0.self_attn.qkv_proj", |
| 1849 | ignored_layers=ignored, |
| 1850 | fused_mapping=FUSED_MAPPING, |
| 1851 | ) |
| 1852 | |
| 1853 | |
| 1854 | def test_partial_shards_raises(): |
| 1855 | # Only some shards listed -> ambiguous, must raise. Fused name is |
| 1856 | # not in ignored_layers, so we fall through to per-shard expansion. |
| 1857 | ignored = ["model.layers.0.self_attn.q_proj"] |
| 1858 | with pytest.raises(ValueError): |
| 1859 | is_layer_skipped( |
| 1860 | prefix="model.layers.0.self_attn.qkv_proj", |
| 1861 | ignored_layers=ignored, |
| 1862 | fused_mapping=FUSED_MAPPING, |
| 1863 | ) |
| 1864 | |
| 1865 | |
| 1866 | def test_not_skipped_when_nothing_listed(): |
| 1867 | assert not is_layer_skipped( |
| 1868 | prefix="model.layers.0.self_attn.qkv_proj", |
| 1869 | ignored_layers=["model.layers.0.mlp.gate_up_proj"], |
| 1870 | fused_mapping=FUSED_MAPPING, |
| 1871 | ) |
| 1872 | |
| 1873 | |
| 1874 | def test_non_fused_layer_unaffected(): |
| 1875 | assert is_layer_skipped( |
| 1876 | prefix="model.layers.0.self_attn.o_proj", |
| 1877 | ignored_layers=["model.layers.0.self_attn.o_proj"], |
| 1878 | fused_mapping=FUSED_MAPPING, |
| 1879 | ) |
| 1880 | assert not is_layer_skipped( |
| 1881 | prefix="model.layers.0.self_attn.o_proj", |
| 1882 | ignored_layers=["model.layers.1.self_attn.o_proj"], |
| 1883 | fused_mapping=FUSED_MAPPING, |
| 1884 | ) |
| 1885 | |
| 1886 | |
| 1887 | def test_substr_match_on_fused_name(): |
| 1888 | # Substring matching: a fused-name match should also |
| 1889 | # short-circuit before shard expansion. |
| 1890 | assert is_layer_skipped( |
| 1891 | prefix="model.layers.0.self_attn.qkv_proj", |
| 1892 | ignored_layers=["self_attn.qkv_proj"], |
| 1893 | fused_mapping=FUSED_MAPPING, |
| 1894 | match_mode="substring", |
| 1895 | ) |
| 1896 | |
| 1897 | |
| 1898 | @pytest.mark.parametrize( |
| 1899 | ("prefix", "ignored_layer", "expected"), |
| 1900 | [ |
| 1901 | ("model.layers.0.self_attn.b_proj", "b_proj", True), |
| 1902 | ("model.layers.0.self_attn.q_b_proj", "b_proj", False), |
| 1903 | ("model.layers.0.self_attn.kv_b_proj", "b_proj", False), |
| 1904 | ("model.layers.5.self_attn.g_proj", "5.self_attn.g_proj", True), |
| 1905 | ("model.layers.6.self_attn.g_proj", "5.self_attn.g_proj", False), |
| 1906 | ], |
| 1907 | ) |
| 1908 | def test_suffix_match_at_module_boundary(prefix, ignored_layer, expected): |
| 1909 | assert ( |
| 1910 | is_layer_skipped( |
| 1911 | prefix=prefix, |
| 1912 | ignored_layers=[ignored_layer], |
| 1913 | match_mode="suffix", |
| 1914 | ) |
| 1915 | is expected |
| 1916 | ) |
| 1917 | |
| 1918 | |
| 1919 | _GLM5_MXFP4_WEIGHT = { |
| 1920 | "dtype": "fp4", |
| 1921 | "qscheme": "per_group", |
| 1922 | "group_size": 32, |
| 1923 | "scale_format": "e8m0", |
| 1924 | "is_dynamic": False, |
| 1925 | } |
| 1926 | _GLM5_BLOCK_FP8_WEIGHT = { |
| 1927 | "dtype": "fp8_e4m3", |
| 1928 | "qscheme": "per_block", |
| 1929 | "is_dynamic": False, |
| 1930 | "block_size": [128, 128], |
| 1931 | "symmetric": True, |
| 1932 | } |
| 1933 | _GLM5_BLOCK_FP8_INPUT = { |
| 1934 | "dtype": "fp8_e4m3", |
| 1935 | "qscheme": "per_group", |
| 1936 | "is_dynamic": True, |
| 1937 | "group_size": 128, |
| 1938 | "symmetric": True, |
| 1939 | } |
| 1940 | _GLM5_GATE_UP = "language_model.model.layers.0.mlp.gate_up_proj" |
| 1941 | |
| 1942 | |
| 1943 | def _glm5_mixed_precision_config() -> QuarkConfig: |
| 1944 | """Global MXFP4 with the dense-MLP gate/up shards marked block-FP8.""" |
| 1945 | fp8 = {"weight": _GLM5_BLOCK_FP8_WEIGHT, "input_tensors": _GLM5_BLOCK_FP8_INPUT} |
| 1946 | return QuarkConfig( |
| 1947 | { |
| 1948 | "global_quant_config": { |
| 1949 | "weight": _GLM5_MXFP4_WEIGHT, |
| 1950 | "input_tensors": None, |
| 1951 | }, |
| 1952 | "layer_type_quant_config": {}, |
| 1953 | "layer_quant_config": {"*mlp.gate_proj": fp8, "*mlp.up_proj": fp8}, |
| 1954 | "exclude": [], |
| 1955 | } |
| 1956 | ) |
| 1957 | |
| 1958 | |
| 1959 | class _GLM5RecordingParam: |
| 1960 | """Stand-in for a BF16 target param that records what gets loaded.""" |
| 1961 | |
| 1962 | def __init__(self): |
| 1963 | self.loaded_weight: torch.Tensor | None = None |
| 1964 | self.loaded_shard: object = "unset" |
| 1965 | |
| 1966 | def weight_loader(self, param, loaded_weight, shard_id=None): |
| 1967 | self.loaded_weight = loaded_weight |
| 1968 | self.loaded_shard = shard_id |
| 1969 | |
| 1970 | |
| 1971 | def _glm5_block_fp8(out_dim: int, in_dim: int, block: int = 128): |
| 1972 | """Return a (fp8 weight, f32 per-block scale) pair.""" |
| 1973 | weight = (torch.randn(out_dim, in_dim) * 0.1).to(torch.float8_e4m3fn) |
| 1974 | scale = torch.rand(out_dim // block, in_dim // block, dtype=torch.float32) + 0.5 |
| 1975 | return weight, scale |
| 1976 | |
| 1977 | |
| 1978 | def test_glm5next_gate_up_proj_mapping_is_genuine_fusion(): |
| 1979 | from vllm.models.glm5next.common.model import Glm5NextForConditionalGeneration |
| 1980 | |
| 1981 | mapping = Glm5NextForConditionalGeneration.packed_modules_mapping |
| 1982 | assert mapping["gate_up_proj"] == ["gate_proj", "up_proj"] |
| 1983 | |
| 1984 | |
| 1985 | def test_glm5next_genuine_fusion_resolves_gate_up_to_block_fp8(): |
| 1986 | # The fused module must expand to its real shard names so each resolves to |
| 1987 | # the per-layer block-FP8 entry (rather than the global MXFP4 scheme). |
| 1988 | config = _glm5_mixed_precision_config() |
| 1989 | config.packed_modules_mapping = {"gate_up_proj": ["gate_proj", "up_proj"]} |
| 1990 | _, _, scheme_cls = config.get_scheme_cls(LinearBase, _GLM5_GATE_UP) |
| 1991 | assert scheme_cls is QuarkW8A8Fp8PerBlock |
| 1992 | |
| 1993 | |
| 1994 | def test_glm5next_attn_loader_accepts_quark_weight_scale(): |
| 1995 | # Regression for KeyError on '...kv_a_proj_with_mqa.weight_scale': the fused |
| 1996 | # q_a/kv_a projection is kept BF16 and dequantized on load; the Quark scale |
| 1997 | # name must be recognized and routed to the correct fused shard. |
| 1998 | from vllm.models.glm5next.common.model import ( |
| 1999 | _dequant_fp8_block, |
| 2000 | _try_load_fp8_attn_proj, |
| 2001 | ) |
| 2002 | |
| 2003 | prefix = "layers.3.self_attn" |
| 2004 | target = _GLM5RecordingParam() |
| 2005 | params_dict = {f"{prefix}.fused_qkv_a_proj.weight": target} |
| 2006 | buf: dict = {} |
| 2007 | loaded: set = set() |
| 2008 | |
| 2009 | weight, scale = _glm5_block_fp8(256, 256) |
| 2010 | # fp8 weight arrives first -> buffered, nothing loaded yet. |
| 2011 | assert ( |
| 2012 | _try_load_fp8_attn_proj( |
| 2013 | f"{prefix}.kv_a_proj_with_mqa.weight", weight, buf, params_dict, loaded, 0 |
| 2014 | ) |
| 2015 | is True |
| 2016 | ) |
| 2017 | assert target.loaded_weight is None |
| 2018 | # Quark-named scale completes the pair -> dequantize + load. |
| 2019 | assert ( |
| 2020 | _try_load_fp8_attn_proj( |
| 2021 | f"{prefix}.kv_a_proj_with_mqa.weight_scale", |
| 2022 | scale, |
| 2023 | buf, |
| 2024 | params_dict, |
| 2025 | loaded, |
| 2026 | 0, |
| 2027 | ) |
| 2028 | is True |
| 2029 | ) |
| 2030 | assert target.loaded_weight is not None |
| 2031 | assert target.loaded_weight.dtype == torch.bfloat16 |
| 2032 | assert target.loaded_shard == 1 # kv_a is shard 1 of fused_qkv_a_proj |
| 2033 | assert torch.equal(target.loaded_weight, _dequant_fp8_block(weight, scale, 128)) |
| 2034 | assert f"{prefix}.fused_qkv_a_proj.weight" in loaded |
| 2035 | |
| 2036 | |
| 2037 | def test_glm5next_attn_loader_accepts_deepseek_weight_scale_inv(): |
| 2038 | # DeepSeek scale name keeps working unchanged. |
| 2039 | from vllm.models.glm5next.common.model import ( |
| 2040 | _dequant_fp8_block, |
| 2041 | _try_load_fp8_attn_proj, |
| 2042 | ) |
| 2043 | |
| 2044 | prefix = "layers.7.self_attn" |
| 2045 | target = _GLM5RecordingParam() |
| 2046 | params_dict = {f"{prefix}.o_proj.weight": target} |
| 2047 | buf: dict = {} |
| 2048 | loaded: set = set() |
| 2049 | |
| 2050 | weight, scale = _glm5_block_fp8(128, 256) |
| 2051 | _try_load_fp8_attn_proj( |
| 2052 | f"{prefix}.o_proj.weight", weight, buf, params_dict, loaded, 0 |
| 2053 | ) |
| 2054 | assert ( |
| 2055 | _try_load_fp8_attn_proj( |
| 2056 | f"{prefix}.o_proj.weight_scale_inv", scale, buf, params_dict, loaded, 0 |
| 2057 | ) |
| 2058 | is True |
| 2059 | ) |
| 2060 | assert target.loaded_shard is None # o_proj is a direct (non-fused) proj |
| 2061 | assert torch.equal(target.loaded_weight, _dequant_fp8_block(weight, scale, 128)) |
| 2062 | |
| 2063 | |
| 2064 | _REVERSE_AWQ_PACK_ORDER = [0, 4, 1, 5, 2, 6, 3, 7] |
| 2065 | |
| 2066 | |
| 2067 | def _quark_int4_config( |
| 2068 | *, |
| 2069 | pack_method: str = "reorder", |
| 2070 | symmetric: bool = True, |
| 2071 | exclude: list[str] | None = None, |
| 2072 | ) -> dict: |
| 2073 | return { |
| 2074 | "quant_method": "quark", |
| 2075 | "export": {"pack_method": pack_method, "kv_cache_group": []}, |
| 2076 | "global_quant_config": { |
| 2077 | "weight": { |
| 2078 | "dtype": "int4", |
| 2079 | "group_size": 128, |
| 2080 | "symmetric": symmetric, |
| 2081 | } |
| 2082 | }, |
| 2083 | "exclude": exclude or [], |
| 2084 | } |
| 2085 | |
| 2086 | |
| 2087 | def _sign_extend_int4_nibbles(t: torch.Tensor) -> torch.Tensor: |
| 2088 | mask = (t & 0x8).bool() |
| 2089 | t = t.clone() |
| 2090 | t[mask] = t[mask] | 0xF0 |
| 2091 | return t |
| 2092 | |
| 2093 | |
| 2094 | def _dequantize_quark_signed_awq_torch( |
| 2095 | qweight: torch.Tensor, |
| 2096 | scales: torch.Tensor, |
| 2097 | qzeros: torch.Tensor, |
| 2098 | group_size: int, |
| 2099 | *, |
| 2100 | pack_reorder: bool = True, |
| 2101 | ) -> torch.Tensor: |
| 2102 | bits = 4 |
| 2103 | shifts = torch.arange(0, 32, bits, device=qweight.device) |
| 2104 | iweights = ((qweight[:, :, None] >> shifts[None, None, :]) & 0xF).to(torch.int8) |
| 2105 | iweights = iweights.view(qweight.shape[0], -1) |
| 2106 | zeros = ((qzeros[:, :, None] >> shifts[None, None, :]) & 0xF).to(torch.int8) |
| 2107 | zeros = zeros.view(qzeros.shape[0], -1) |
| 2108 | |
| 2109 | if pack_reorder: |
| 2110 | order = torch.tensor(_REVERSE_AWQ_PACK_ORDER, device=qweight.device) |
| 2111 | else: |
| 2112 | order = torch.arange(8, device=qweight.device) |
| 2113 | iweights = iweights.view(qweight.shape[0], -1, 8)[:, :, order].reshape( |
| 2114 | qweight.shape[0], -1 |
| 2115 | ) |
| 2116 | zeros = zeros.view(qzeros.shape[0], -1, 8)[:, :, order].reshape(qzeros.shape[0], -1) |
| 2117 | iweights = _sign_extend_int4_nibbles(iweights & 0xF) |
| 2118 | zeros = _sign_extend_int4_nibbles(zeros & 0xF) |
| 2119 | |
| 2120 | scales = scales.repeat_interleave(group_size, dim=0) |
| 2121 | zeros = zeros.repeat_interleave(group_size, dim=0) |
| 2122 | return (iweights - zeros) * scales |
| 2123 | |
| 2124 | |
| 2125 | def _dequantize_awq_unsigned_torch( |
| 2126 | qweight: torch.Tensor, |
| 2127 | scales: torch.Tensor, |
| 2128 | qzeros: torch.Tensor, |
| 2129 | group_size: int, |
| 2130 | *, |
| 2131 | pack_reorder: bool = True, |
| 2132 | ) -> torch.Tensor: |
| 2133 | bits = 4 |
| 2134 | shifts = torch.arange(0, 32, bits, device=qweight.device) |
| 2135 | iweights = ((qweight[:, :, None] >> shifts[None, None, :]) & 0xF).to(torch.int8) |
| 2136 | iweights = iweights.view(qweight.shape[0], -1) |
| 2137 | zeros = ((qzeros[:, :, None] >> shifts[None, None, :]) & 0xF).to(torch.int8) |
| 2138 | zeros = zeros.view(qzeros.shape[0], -1) |
| 2139 | |
| 2140 | if pack_reorder: |
| 2141 | order = torch.tensor(_REVERSE_AWQ_PACK_ORDER, device=qweight.device) |
| 2142 | else: |
| 2143 | order = torch.arange(8, device=qweight.device) |
| 2144 | iweights = iweights.view(qweight.shape[0], -1, 8)[:, :, order].reshape( |
| 2145 | qweight.shape[0], -1 |
| 2146 | ) |
| 2147 | zeros = zeros.view(qzeros.shape[0], -1, 8)[:, :, order].reshape(qzeros.shape[0], -1) |
| 2148 | |
| 2149 | scales = scales.repeat_interleave(group_size, dim=0) |
| 2150 | zeros = zeros.repeat_interleave(group_size, dim=0) |
| 2151 | return (iweights - zeros) * scales |
| 2152 | |
| 2153 | |
| 2154 | def _pack_int4_nibbles(nibbles: torch.Tensor, *, pack_reorder: bool) -> torch.Tensor: |
| 2155 | pack_order = ( |
| 2156 | torch.tensor(_REVERSE_AWQ_PACK_ORDER, device=nibbles.device) |
| 2157 | if pack_reorder |
| 2158 | else torch.arange(8, device=nibbles.device) |
| 2159 | ) |
| 2160 | shifts = pack_order * 4 |
| 2161 | return ((nibbles.to(torch.int64) & 0xF) << shifts).sum(dim=-1).to(torch.int32) |
| 2162 | |
| 2163 | |
| 2164 | class TestQuarkInt4Format: |
| 2165 | """Tests for Quark INT4 export format compatibility.""" |
| 2166 | |
| 2167 | def test_quark_order_pack_method_uses_native_int4_scheme(self): |
| 2168 | quant_config = QuarkConfig.from_config(_quark_int4_config(pack_method="order")) |
| 2169 | weight_config = quant_config.quant_config["global_quant_config"] |
| 2170 | weight_key, act_key, scheme_cls = quant_config._get_scheme_cls_from_config( |
| 2171 | weight_config |
| 2172 | ) |
| 2173 | assert scheme_cls is QuarkW4A16Int4 |
| 2174 | |
| 2175 | scheme = quant_config.init_scheme( |
| 2176 | scheme_cls, |
| 2177 | weight_quant_key=weight_key, |
| 2178 | activation_quant_key=act_key, |
| 2179 | weight_config=weight_config["weight"], |
| 2180 | ) |
| 2181 | assert isinstance(scheme, QuarkW4A16Int4) |
| 2182 | assert not scheme.pack_reorder |
| 2183 | |
| 2184 | def test_quark_int4_scheme_supports_asymmetric_weights(self): |
| 2185 | quant_config = QuarkConfig.from_config(_quark_int4_config(symmetric=False)) |
| 2186 | weight_config = quant_config.quant_config["global_quant_config"] |
| 2187 | weight_key, act_key, scheme_cls = quant_config._get_scheme_cls_from_config( |
| 2188 | weight_config |
| 2189 | ) |
| 2190 | assert scheme_cls is QuarkW4A16Int4 |
| 2191 | |
| 2192 | scheme = quant_config.init_scheme( |
| 2193 | scheme_cls, |
| 2194 | weight_quant_key=weight_key, |
| 2195 | activation_quant_key=act_key, |
| 2196 | weight_config=weight_config["weight"], |
| 2197 | ) |
| 2198 | assert isinstance(scheme, QuarkW4A16Int4) |
| 2199 | assert not scheme.is_symmetric |
| 2200 | |
| 2201 | @pytest.mark.parametrize("missing_field", ["group_size", "symmetric"]) |
| 2202 | def test_quark_int4_scheme_requires_weight_config_fields(self, missing_field): |
| 2203 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 2204 | kInt4Static, |
| 2205 | ) |
| 2206 | |
| 2207 | weight_config = { |
| 2208 | "dtype": "int4", |
| 2209 | "group_size": 128, |
| 2210 | "symmetric": True, |
| 2211 | } |
| 2212 | weight_config.pop(missing_field) |
| 2213 | quant_config = QuarkConfig.from_config( |
| 2214 | { |
| 2215 | "quant_method": "quark", |
| 2216 | "export": {"pack_method": "reorder", "kv_cache_group": []}, |
| 2217 | "global_quant_config": {"weight": weight_config}, |
| 2218 | "exclude": [], |
| 2219 | } |
| 2220 | ) |
| 2221 | |
| 2222 | with pytest.raises(ValueError, match=missing_field): |
| 2223 | quant_config.init_scheme( |
| 2224 | QuarkW4A16Int4, |
| 2225 | weight_quant_key=kInt4Static, |
| 2226 | activation_quant_key=None, |
| 2227 | weight_config=weight_config, |
| 2228 | ) |
| 2229 | |
| 2230 | def test_quark_int4_moe_uses_native_moe_method(self): |
| 2231 | from unittest.mock import MagicMock, patch |
| 2232 | |
| 2233 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 2234 | kInt4StaticAsym, |
| 2235 | ) |
| 2236 | |
| 2237 | quant_config = QuarkConfig.from_config(_quark_int4_config(symmetric=False)) |
| 2238 | moe_config = type("MoeConfig", (), {})() |
| 2239 | moe_config.has_bias = False |
| 2240 | |
| 2241 | mock_backend = MagicMock() |
| 2242 | mock_experts_cls = MagicMock() |
| 2243 | with patch( |
| 2244 | "vllm.model_executor.layers.quantization.quark.quark_moe" |
| 2245 | ".select_wna16_moe_backend", |
| 2246 | return_value=(mock_backend, mock_experts_cls), |
| 2247 | ): |
| 2248 | method = QuarkW4A16Int4MoEMethod( |
| 2249 | kInt4StaticAsym, |
| 2250 | None, |
| 2251 | quant_config.quant_config["global_quant_config"]["weight"], |
| 2252 | quant_config.pack_method, |
| 2253 | moe_config, |
| 2254 | ) |
| 2255 | |
| 2256 | assert method.group_size == 128 |
| 2257 | assert method.pack_reorder |
| 2258 | assert method.use_wna16_backend |
| 2259 | |
| 2260 | def test_triton_wna16_experts_supports_asymmetric_int4(self): |
| 2261 | from vllm.model_executor.layers.fused_moe.experts.triton_moe import ( |
| 2262 | TritonWNA16Experts, |
| 2263 | ) |
| 2264 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 2265 | kInt4Static, |
| 2266 | kInt4Static32, |
| 2267 | kInt4Static32Asym, |
| 2268 | kInt4StaticAsym, |
| 2269 | ) |
| 2270 | |
| 2271 | for key in [kInt4Static, kInt4Static32, kInt4StaticAsym, kInt4Static32Asym]: |
| 2272 | assert TritonWNA16Experts._supports_quant_scheme(key, None), ( |
| 2273 | f"{key} should be supported by TritonWNA16Experts" |
| 2274 | ) |
| 2275 | |
| 2276 | def test_quark_moe_loader_shards_w2_zero_point_across_tp_ranks(self): |
| 2277 | """w2 zero-points are sharded along the intermediate dim per TP rank. |
| 2278 | |
| 2279 | Loading the same Quark checkpoint tensor at tensor_parallel_size=2 for |
| 2280 | both ranks must jointly reconstruct the tensor_parallel_size=1 (full) |
| 2281 | result, with no overlap or gap. This guards the custom loader's TP |
| 2282 | split for the packed zero-point layout. |
| 2283 | """ |
| 2284 | |
| 2285 | class _FakeMoEConfig: |
| 2286 | has_bias = False |
| 2287 | |
| 2288 | def __init__(self, tp_size, tp_rank): |
| 2289 | self.tp_size = tp_size |
| 2290 | self.tp_rank = tp_rank |
| 2291 | |
| 2292 | class _FakeLayer: |
| 2293 | def __init__(self, moe_config): |
| 2294 | self.moe_config = moe_config |
| 2295 | self.intermediate_size_per_partition = 64 |
| 2296 | self.group_size_div_factor = 1 |
| 2297 | |
| 2298 | def load_w2_zero_point(raw_zp, tp_size, tp_rank): |
| 2299 | from unittest.mock import MagicMock, patch |
| 2300 | |
| 2301 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 2302 | kInt4StaticAsym, |
| 2303 | ) |
| 2304 | |
| 2305 | quant_config = QuarkConfig.from_config(_quark_int4_config(symmetric=False)) |
| 2306 | moe_config = _FakeMoEConfig(tp_size, tp_rank) |
| 2307 | with patch( |
| 2308 | "vllm.model_executor.layers.quantization.quark.quark_moe" |
| 2309 | ".select_wna16_moe_backend", |
| 2310 | return_value=(MagicMock(), MagicMock()), |
| 2311 | ): |
| 2312 | method = QuarkW4A16Int4MoEMethod( |
| 2313 | kInt4StaticAsym, |
| 2314 | None, |
| 2315 | quant_config.quant_config["global_quant_config"]["weight"], |
| 2316 | quant_config.pack_method, |
| 2317 | moe_config, |
| 2318 | ) |
| 2319 | |
| 2320 | layer = _FakeLayer(moe_config) |
| 2321 | # The loader converts the packed zero-point, then writes the |
| 2322 | # per-rank slice into param.data[expert_id]. A single expert dim is |
| 2323 | # enough; the last dim shrinks by tp_size after sharding. |
| 2324 | param = torch.nn.Parameter( |
| 2325 | torch.zeros(1, 16, 8 // tp_size, dtype=torch.uint8), |
| 2326 | requires_grad=False, |
| 2327 | ) |
| 2328 | loader = method.get_weight_loader( |
| 2329 | layer, |
| 2330 | weight_loader=None, |
| 2331 | ) |
| 2332 | loader( |
| 2333 | param, |
| 2334 | raw_zp.clone(), |
| 2335 | weight_name="w2_weight_zero_point", |
| 2336 | shard_id="w2", |
| 2337 | expert_id=0, |
| 2338 | ) |
| 2339 | return param.data[0] |
| 2340 | |
| 2341 | # Quark-packed zero-point for one expert: rows = hidden // pack_factor, |
| 2342 | # cols carry the (packed) intermediate dim, split evenly across ranks. |
| 2343 | raw_zp = torch.randint(0, 256, (8, 16), dtype=torch.uint8) |
| 2344 | |
| 2345 | full = load_w2_zero_point(raw_zp, tp_size=1, tp_rank=0) |
| 2346 | shard0 = load_w2_zero_point(raw_zp, tp_size=2, tp_rank=0) |
| 2347 | shard1 = load_w2_zero_point(raw_zp, tp_size=2, tp_rank=1) |
| 2348 | |
| 2349 | expected = full.view(full.size(0), 2, -1) |
| 2350 | torch.testing.assert_close(shard0, expected[:, 0]) |
| 2351 | torch.testing.assert_close(shard1, expected[:, 1]) |
| 2352 | |
| 2353 | def test_quark_shared_expert_gate_keeps_quantized_tensors(self): |
| 2354 | quant_config = QuarkConfig.from_config(_quark_int4_config()) |
| 2355 | mapper = quant_config.get_cache_scale_mapper() |
| 2356 | |
| 2357 | output_names = { |
| 2358 | name |
| 2359 | for name, _ in mapper.apply( |
| 2360 | [ |
| 2361 | ( |
| 2362 | "model.language_model.layers.0.mlp.shared_expert_gate.weight", |
| 2363 | torch.zeros(1), |
| 2364 | ), |
| 2365 | ( |
| 2366 | "model.language_model.layers.0.mlp.shared_expert_gate" |
| 2367 | ".weight_scale", |
| 2368 | torch.zeros(1), |
| 2369 | ), |
| 2370 | ( |
| 2371 | "model.language_model.layers.0.mlp.shared_expert_gate" |
| 2372 | ".weight_zero_point", |
| 2373 | torch.zeros(1), |
| 2374 | ), |
| 2375 | ] |
| 2376 | ) |
| 2377 | } |
| 2378 | |
| 2379 | assert ( |
| 2380 | "model.language_model.layers.0.mlp.shared_expert_gate.weight" |
| 2381 | in output_names |
| 2382 | ) |
| 2383 | assert ( |
| 2384 | "model.language_model.layers.0.mlp.shared_expert_gate.weight_scale" |
| 2385 | in output_names |
| 2386 | ) |
| 2387 | assert ( |
| 2388 | "model.language_model.layers.0.mlp.shared_expert_gate.weight_zero_point" |
| 2389 | in output_names |
| 2390 | ) |
| 2391 | |
| 2392 | def test_quark_apply_mapper_updates_exclude_and_layer_quant_config(self): |
| 2393 | quant_config = QuarkConfig.from_config( |
| 2394 | { |
| 2395 | **_quark_int4_config(), |
| 2396 | "exclude": ["lm_head"], |
| 2397 | "layer_quant_config": { |
| 2398 | "model.language_model.layers.0.mlp.gate": { |
| 2399 | "weight": {"dtype": "float16"}, |
| 2400 | }, |
| 2401 | }, |
| 2402 | } |
| 2403 | ) |
| 2404 | quant_config.apply_vllm_mapper( |
| 2405 | WeightsMapper( |
| 2406 | orig_to_new_prefix={ |
| 2407 | "lm_head": "language_model.lm_head", |
| 2408 | "model.language_model.": "language_model.model.", |
| 2409 | }, |
| 2410 | ) |
| 2411 | ) |
| 2412 | |
| 2413 | layer_quant_config = quant_config.quant_config["layer_quant_config"] |
| 2414 | assert "language_model.model.layers.0.mlp.gate" in layer_quant_config |
| 2415 | assert "model.language_model.layers.0.mlp.gate" not in layer_quant_config |
| 2416 | assert quant_config.quant_config["exclude"] == [ |
| 2417 | "language_model.lm_head", |
| 2418 | ] |
| 2419 | |
| 2420 | def test_quark_apply_mapper_ignores_non_string_list_entries(self): |
| 2421 | quant_config = QuarkConfig.from_config( |
| 2422 | { |
| 2423 | **_quark_int4_config(), |
| 2424 | "algo_config": [ |
| 2425 | {"name": "qronos", "inside_layer_modules": ["self_attn.q_proj"]} |
| 2426 | ], |
| 2427 | } |
| 2428 | ) |
| 2429 | quant_config.apply_vllm_mapper(WeightsMapper()) |
| 2430 | |
| 2431 | assert quant_config.quant_config["algo_config"] == [ |
| 2432 | {"name": "qronos", "inside_layer_modules": ["self_attn.q_proj"]} |
| 2433 | ] |
| 2434 | |
| 2435 | def test_quark_exclude_matches_via_mapper_prefix(self): |
| 2436 | quant_config = QuarkConfig.from_config(_quark_int4_config(exclude=["lm_head"])) |
| 2437 | quant_config.apply_vllm_mapper( |
| 2438 | WeightsMapper(orig_to_new_prefix={"lm_head": "language_model.lm_head"}) |
| 2439 | ) |
| 2440 | |
| 2441 | exclude_layers = quant_config.quant_config["exclude"] |
| 2442 | assert should_ignore_layer("language_model.lm_head", ignore=exclude_layers) |
| 2443 | assert not should_ignore_layer( |
| 2444 | "language_model.model.layers.0.mlp.gate", ignore=exclude_layers |
| 2445 | ) |
| 2446 | |
| 2447 | |
| 2448 | @pytest.mark.parametrize("symmetric", [False, True]) |
| 2449 | @pytest.mark.parametrize("pack_method", ["order", "reorder"]) |
| 2450 | def test_quark_int4_canonicalizes_pack_for_kernel_layout(pack_method, symmetric): |
| 2451 | """Quark pack order is normalized to the layout expected by awq_* ops. |
| 2452 | |
| 2453 | This reuses the existing AWQ dequant/gemm kernels for compute only; loading |
| 2454 | still goes through the native Quark quantization path, not AutoAWQ. |
| 2455 | """ |
| 2456 | pack_reorder = pack_method == "reorder" |
| 2457 | group_size = 2 |
| 2458 | packed_values = torch.tensor( |
| 2459 | [ |
| 2460 | [0, 1, 7, 8, 9, 15, 2, 14], |
| 2461 | [15, 8, 0, 3, 12, 7, 1, 9], |
| 2462 | ], |
| 2463 | dtype=torch.int32, |
| 2464 | ) |
| 2465 | packed_zeros = torch.zeros((1, 8), dtype=torch.int32) |
| 2466 | qweight = _pack_int4_nibbles(packed_values, pack_reorder=pack_reorder).view(2, 1) |
| 2467 | qzeros = _pack_int4_nibbles(packed_zeros, pack_reorder=pack_reorder).view(1, 1) |
| 2468 | scales = torch.ones((1, 8), dtype=torch.float16) |
| 2469 | |
| 2470 | dequantize = ( |
| 2471 | _dequantize_quark_signed_awq_torch |
| 2472 | if symmetric |
| 2473 | else _dequantize_awq_unsigned_torch |
| 2474 | ) |
| 2475 | expected = dequantize( |
| 2476 | qweight, |
| 2477 | scales, |
| 2478 | qzeros, |
| 2479 | group_size, |
| 2480 | pack_reorder=pack_reorder, |
| 2481 | ) |
| 2482 | canonical_weight = canonicalize_quark_packed_int4( |
| 2483 | qweight, |
| 2484 | pack_reorder=pack_reorder, |
| 2485 | is_symmetric=symmetric, |
| 2486 | ) |
| 2487 | canonical_zero = canonicalize_quark_packed_int4( |
| 2488 | qzeros, |
| 2489 | pack_reorder=pack_reorder, |
| 2490 | is_symmetric=symmetric, |
| 2491 | ) |
| 2492 | actual = _dequantize_awq_unsigned_torch( |
| 2493 | canonical_weight, scales, canonical_zero, group_size |
| 2494 | ) |
| 2495 | |
| 2496 | assert torch.equal(actual, expected) |
| 2497 | |
| 2498 | |
| 2499 | # --------------------------------------------------------------------------- |
| 2500 | # override_quantization_method: Quark always delegates to QuarkConfig |
| 2501 | # --------------------------------------------------------------------------- |
| 2502 | |
| 2503 | _QUARK_MXFP4_CFG = { |
| 2504 | "quant_method": "quark", |
| 2505 | "global_quant_config": { |
| 2506 | "weight": {"dtype": "fp4", "qscheme": "per_group", "group_size": 32}, |
| 2507 | }, |
| 2508 | } |
| 2509 | |
| 2510 | _QUARK_MXFP8_CFG = { |
| 2511 | "quant_method": "quark", |
| 2512 | "global_quant_config": { |
| 2513 | "weight": { |
| 2514 | "dtype": "fp8_e4m3", |
| 2515 | "qscheme": "per_block", |
| 2516 | "block_size": [32, 32], |
| 2517 | "scale_type": "float8_e8m0fnu", |
| 2518 | "symmetric": True, |
| 2519 | }, |
| 2520 | }, |
| 2521 | } |
| 2522 | |
| 2523 | _QUARK_MIXED_CFG = { |
| 2524 | "quant_method": "quark", |
| 2525 | "global_quant_config": { |
| 2526 | "weight": {"dtype": "fp4", "qscheme": "per_group", "group_size": 32}, |
| 2527 | }, |
| 2528 | "layer_quant_config": {"layers.0.attn": {"weight": {"dtype": "fp8_e4m3"}}}, |
| 2529 | } |
| 2530 | |
| 2531 | _FP8_CFG = {"quant_method": "fp8"} |
| 2532 | |
| 2533 | _DSV4_FP8_CFG = {"quant_method": "deepseek_v4_fp8"} |
| 2534 | |
| 2535 | |
| 2536 | @pytest.mark.parametrize( |
| 2537 | "quant_config_cls", |
| 2538 | [ |
| 2539 | "deepseek_v41", |
| 2540 | "deepseek_v4", |
| 2541 | ], |
| 2542 | ) |
| 2543 | @pytest.mark.parametrize( |
| 2544 | "hf_quant_cfg, expect_none", |
| 2545 | [ |
| 2546 | (_QUARK_MXFP4_CFG, True), |
| 2547 | (_QUARK_MXFP8_CFG, True), |
| 2548 | (_QUARK_MIXED_CFG, True), |
| 2549 | (_FP8_CFG, False), |
| 2550 | (_DSV4_FP8_CFG, False), |
| 2551 | ], |
| 2552 | ids=[ |
| 2553 | "quark-mxfp4", |
| 2554 | "quark-mxfp8", |
| 2555 | "quark-mixed", |
| 2556 | "native-fp8", |
| 2557 | "deepseek-v4-fp8", |
| 2558 | ], |
| 2559 | ) |
| 2560 | def test_quark_override_delegates_to_quark_config( |
| 2561 | quant_config_cls, hf_quant_cfg, expect_none |
| 2562 | ): |
| 2563 | """Any quant_method=='quark' must return None (delegate to QuarkConfig). |
| 2564 | |
| 2565 | Native FP8 configs should still be claimed by the model-specific config. |
| 2566 | """ |
| 2567 | import importlib |
| 2568 | |
| 2569 | if quant_config_cls == "deepseek_v41": |
| 2570 | # Module was renamed from deepseek_v4_1 to deepseek_v41 upstream. |
| 2571 | for mod_name in ( |
| 2572 | "vllm.models.deepseek_v41.quant_config", |
| 2573 | "vllm.models.deepseek_v4_1.quant_config", |
| 2574 | ): |
| 2575 | try: |
| 2576 | mod = importlib.import_module(mod_name) |
| 2577 | break |
| 2578 | except ModuleNotFoundError: |
| 2579 | continue |
| 2580 | else: |
| 2581 | pytest.skip("deepseek_v41 quant_config not found") |
| 2582 | model_type = "deepseek_v41" |
| 2583 | else: |
| 2584 | mod = importlib.import_module("vllm.models.deepseek_v4.quant_config") |
| 2585 | model_type = "deepseek_v4" |
| 2586 | |
| 2587 | config_cls = mod.DeepseekV4FP8Config |
| 2588 | hf_config = SimpleNamespace(model_type=model_type) |
| 2589 | result = config_cls.override_quantization_method( |
| 2590 | hf_quant_cfg, None, hf_config=hf_config |
| 2591 | ) |
| 2592 | |
| 2593 | if expect_none: |
| 2594 | assert result is None, ( |
| 2595 | f"Quark config should delegate to QuarkConfig (return None), got {result!r}" |
| 2596 | ) |
| 2597 | else: |
| 2598 | assert result == "deepseek_v4_fp8", ( |
| 2599 | f"Native FP8 should be claimed by DeepseekV4FP8Config, got {result!r}" |
| 2600 | ) |