| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Test ModelOpt quantization method setup and weight loading. |
| 4 | |
| 5 | Run `pytest tests/quantization/test_modelopt.py`. |
| 6 | """ |
| 7 | |
| 8 | import os |
| 9 | from typing import Any, NoReturn |
| 10 | from unittest.mock import MagicMock, Mock, patch |
| 11 | |
| 12 | import pytest |
| 13 | import torch |
| 14 | |
| 15 | from tests.quantization.utils import ( |
| 16 | is_quant_method_supported, |
| 17 | load_model_without_vllm_runner, |
| 18 | ) |
| 19 | from vllm.config import VllmConfig, set_current_vllm_config |
| 20 | from vllm.config.model import ModelConfig |
| 21 | from vllm.forward_context import set_forward_context |
| 22 | from vllm.model_executor.kernels.linear import ( |
| 23 | FlashInferCuteDslNvFp4W4A16LinearKernel, |
| 24 | HummingNvFp4LinearKernel, |
| 25 | MarlinNvFp4LinearKernel, |
| 26 | ) |
| 27 | from vllm.model_executor.layers.attention import Attention |
| 28 | from vllm.model_executor.layers.fused_moe import RoutedExperts |
| 29 | from vllm.model_executor.layers.linear import UnquantizedLinearMethod |
| 30 | from vllm.model_executor.layers.quantization.modelopt import ( |
| 31 | LINEAR_ALGOS, |
| 32 | ModelOptFp8Config, |
| 33 | ModelOptLinearMethod, |
| 34 | ModelOptMixedPrecisionConfig, |
| 35 | ModelOptMxFp8Config, |
| 36 | ModelOptNvFp4Config, |
| 37 | ) |
| 38 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 39 | kFp8StaticTensorSym, |
| 40 | kMxfp8Dynamic, |
| 41 | kMxfp8Static, |
| 42 | kNvfp4Dynamic, |
| 43 | kNvfp4Static, |
| 44 | ) |
| 45 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 46 | ParallelLMHead, |
| 47 | VocabParallelEmbedding, |
| 48 | ) |
| 49 | from vllm.platforms import current_platform |
| 50 | |
| 51 | |
| 52 | @pytest.fixture(scope="function", autouse=True) |
| 53 | def enable_pickle(monkeypatch): |
| 54 | """`LLM.apply_model` requires pickling a function.""" |
| 55 | monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1") |
| 56 | |
| 57 | |
| 58 | def _skip(msg: str) -> NoReturn: |
| 59 | pytest.skip(msg) |
| 60 | raise RuntimeError(msg) |
| 61 | |
| 62 | |
| 63 | def _snapshot_download_or_skip(model_id: str) -> str: |
| 64 | try: |
| 65 | from vllm.transformers_utils.repo_utils import hf_api |
| 66 | except Exception as e: # pragma: no cover |
| 67 | _skip(f"huggingface_hub is required to download {model_id}: {e}") |
| 68 | |
| 69 | try: |
| 70 | return hf_api().snapshot_download( |
| 71 | repo_id=model_id, |
| 72 | repo_type="model", |
| 73 | # These checkpoints are already small; download full repo for simplicity. |
| 74 | allow_patterns=["*"], |
| 75 | ) |
| 76 | except Exception as e: |
| 77 | _skip(f"Failed to download {model_id} from the HF Hub: {e}") |
| 78 | |
| 79 | |
| 80 | def _mock_lm_head() -> Mock: |
| 81 | lm_head = Mock(spec=ParallelLMHead) |
| 82 | lm_head.__class__ = ParallelLMHead |
| 83 | return lm_head |
| 84 | |
| 85 | |
| 86 | def _mixed_precision_config(quantized_layers: dict) -> ModelOptMixedPrecisionConfig: |
| 87 | return ModelOptMixedPrecisionConfig( |
| 88 | kv_cache_quant_method=None, |
| 89 | exclude_modules=[], |
| 90 | quantized_layers=quantized_layers, |
| 91 | fp8_config=ModelOptFp8Config( |
| 92 | quant_method="FP8", |
| 93 | is_checkpoint_fp8_serialized=True, |
| 94 | kv_cache_quant_method=None, |
| 95 | exclude_modules=[], |
| 96 | ), |
| 97 | nvfp4_config=ModelOptNvFp4Config( |
| 98 | is_checkpoint_nvfp4_serialized=True, |
| 99 | kv_cache_quant_algo=None, |
| 100 | exclude_modules=[], |
| 101 | ), |
| 102 | w4a16_nvfp4_config=ModelOptNvFp4Config( |
| 103 | quant_method="W4A16_NVFP4", |
| 104 | is_checkpoint_nvfp4_serialized=True, |
| 105 | kv_cache_quant_algo=None, |
| 106 | exclude_modules=[], |
| 107 | ), |
| 108 | mxfp8_config=ModelOptMxFp8Config( |
| 109 | is_checkpoint_mxfp8_serialized=True, |
| 110 | kv_cache_quant_algo=None, |
| 111 | exclude_modules=[], |
| 112 | ), |
| 113 | ) |
| 114 | |
| 115 | |
| 116 | def test_modelopt_nvfp4_quantizes_parallel_lm_head(): |
| 117 | config = ModelOptNvFp4Config( |
| 118 | is_checkpoint_nvfp4_serialized=True, |
| 119 | kv_cache_quant_algo=None, |
| 120 | exclude_modules=[], |
| 121 | ) |
| 122 | |
| 123 | method = config.get_quant_method(_mock_lm_head(), prefix="lm_head") |
| 124 | |
| 125 | assert isinstance(method, ModelOptLinearMethod) |
| 126 | assert method.spec.weight is kNvfp4Static |
| 127 | assert method.spec.activation is kNvfp4Dynamic |
| 128 | |
| 129 | |
| 130 | def test_modelopt_mxfp8_preserves_per_row_checkpoint_scales(dist_init, monkeypatch): |
| 131 | """Standard MXFP8 checkpoints already have one scale row per weight row.""" |
| 132 | from vllm.model_executor.layers.linear import ReplicatedLinear |
| 133 | |
| 134 | kernel = Mock() |
| 135 | kernel.input_quant_key.return_value = None |
| 136 | monkeypatch.setattr( |
| 137 | "vllm.model_executor.layers.quantization.modelopt.init_mxfp8_linear_kernel", |
| 138 | lambda **kwargs: kernel, |
| 139 | ) |
| 140 | config = ModelOptMxFp8Config.from_config({"quant_method": "mxfp8"}) |
| 141 | linear = ReplicatedLinear(64, 64, bias=False, quant_config=config) |
| 142 | scales = torch.arange(128, dtype=torch.uint8).reshape(64, 2) |
| 143 | linear.weight_scale.weight_loader(linear.weight_scale, scales) |
| 144 | assert torch.equal(linear.weight_scale, scales) |
| 145 | |
| 146 | |
| 147 | def test_modelopt_mxfp8_pre_processed_weights_follow_kernel(monkeypatch): |
| 148 | """The weight cache daemon can only serve MXFP8 layers whose kernel needs |
| 149 | no post-load state beyond the parameters it exports.""" |
| 150 | from vllm.config.quantization import QuantSpec |
| 151 | from vllm.model_executor.kernels.linear.mxfp8 import Mxfp8LinearKernel |
| 152 | |
| 153 | method = ModelOptLinearMethod.__new__(ModelOptLinearMethod) |
| 154 | method.spec = QuantSpec(weight=kMxfp8Static, activation=kMxfp8Dynamic) |
| 155 | method.kernel = None |
| 156 | assert not method.supports_pre_processed_weights |
| 157 | |
| 158 | method.kernel = Mock(spec=Mxfp8LinearKernel) |
| 159 | method.kernel.supports_pre_processed_weights = False |
| 160 | assert not method.supports_pre_processed_weights |
| 161 | method.kernel.supports_pre_processed_weights = True |
| 162 | assert method.supports_pre_processed_weights |
| 163 | |
| 164 | monkeypatch.setattr( |
| 165 | "vllm.model_executor.layers.quantization.modelopt.is_weights_pre_processed", |
| 166 | lambda: True, |
| 167 | ) |
| 168 | layer = torch.nn.Module() |
| 169 | method.process_weights_after_loading(layer) |
| 170 | method.kernel.process_weights_after_loading.assert_not_called() |
| 171 | |
| 172 | method.kernel.supports_pre_processed_weights = False |
| 173 | with pytest.raises(RuntimeError, match="pre-processed"): |
| 174 | method.process_weights_after_loading(layer) |
| 175 | |
| 176 | |
| 177 | def test_modelopt_fp8_updates_weight_dims_after_transpose(): |
| 178 | """Humming reads weight.input_dim/output_dim. Swapping the |
| 179 | ModelWeightParameter for a plain Parameter drops them, so the per-tensor |
| 180 | FP8 scheme must restore them for the transposed [in, out] layout. |
| 181 | """ |
| 182 | from vllm.config.quantization import QuantSpec |
| 183 | from vllm.model_executor.layers.quantization.modelopt import ( |
| 184 | SCHEME_FOR, |
| 185 | CkptCtx, |
| 186 | FormatScheme, |
| 187 | ) |
| 188 | |
| 189 | layer = torch.nn.Module() |
| 190 | layer.register_parameter( |
| 191 | "weight", torch.nn.Parameter(torch.empty(3, 2), requires_grad=False) |
| 192 | ) |
| 193 | layer.register_parameter( |
| 194 | "weight_scale", torch.nn.Parameter(torch.ones(1), requires_grad=False) |
| 195 | ) |
| 196 | layer.register_parameter( |
| 197 | "input_scale", torch.nn.Parameter(torch.ones(1), requires_grad=False) |
| 198 | ) |
| 199 | layer.logical_widths = [3] |
| 200 | |
| 201 | method = ModelOptLinearMethod.__new__(ModelOptLinearMethod) |
| 202 | method.spec = QuantSpec(weight=kFp8StaticTensorSym, activation=kFp8StaticTensorSym) |
| 203 | method.ctx = CkptCtx() |
| 204 | method.fmt = FormatScheme() |
| 205 | method.wkey = SCHEME_FOR[kFp8StaticTensorSym] |
| 206 | method.akey = SCHEME_FOR[kFp8StaticTensorSym] |
| 207 | method.kernel = Mock() |
| 208 | method.process_weights_after_loading(layer) |
| 209 | |
| 210 | assert layer.weight.shape == (2, 3) |
| 211 | assert layer.weight.input_dim == 0 |
| 212 | assert layer.weight.output_dim == 1 |
| 213 | method.kernel.process_weights_after_loading.assert_called_once_with(layer) |
| 214 | |
| 215 | |
| 216 | def test_modelopt_linear_algos_table_matches_resolve(): |
| 217 | """LINEAR_ALGOS is the single source of truth for supported linear algos. |
| 218 | |
| 219 | Every entry must be dispatchable by resolve(), and every config's |
| 220 | validation list must be derived from it -- so adding a format is one row |
| 221 | here plus one row in resolve(), with nothing else to keep in sync. |
| 222 | """ |
| 223 | from vllm.model_executor.layers.quantization.modelopt import ( |
| 224 | QUANT_ALGOS, |
| 225 | algos_owned_by, |
| 226 | resolve, |
| 227 | ) |
| 228 | |
| 229 | class _Cfg: |
| 230 | group_size = 16 |
| 231 | |
| 232 | for algo in LINEAR_ALGOS: |
| 233 | spec, _, _ = resolve(algo, _Cfg(), "layer") |
| 234 | assert spec.weight is not None, algo |
| 235 | |
| 236 | assert list(QUANT_ALGOS) == [*LINEAR_ALGOS, "MIXED_PRECISION"] |
| 237 | assert algos_owned_by("modelopt") == ( |
| 238 | "FP8", |
| 239 | "FP8_PER_CHANNEL_PER_TOKEN", |
| 240 | "FP8_PB_WO", |
| 241 | ) |
| 242 | assert algos_owned_by("modelopt_fp4") == ("NVFP4", "W4A16_NVFP4") |
| 243 | assert algos_owned_by("modelopt_mxfp8") == ("MXFP8",) |
| 244 | |
| 245 | |
| 246 | @pytest.mark.parametrize("algo", list(LINEAR_ALGOS)) |
| 247 | def test_modelopt_mixed_precision_dispatches_every_linear_algo(algo): |
| 248 | """Mixed precision must route every algo in LINEAR_ALGOS through the |
| 249 | generic method. FP8_PER_CHANNEL_PER_TOKEN and FP8_PB_WO used to fall |
| 250 | through to UnquantizedLinearMethod, which loses the checkpoint's scales. |
| 251 | """ |
| 252 | from vllm.model_executor.layers.linear import LinearBase |
| 253 | from vllm.model_executor.layers.quantization import modelopt as m |
| 254 | |
| 255 | config = m.ModelOptMixedPrecisionConfig.from_config( |
| 256 | { |
| 257 | "quantization": { |
| 258 | "quant_algo": "MIXED_PRECISION", |
| 259 | "kv_cache_quant_algo": None, |
| 260 | "exclude_modules": [], |
| 261 | "group_size": 16, |
| 262 | "quantized_layers": { |
| 263 | "model.layers.0.mlp.down_proj": {"quant_algo": algo} |
| 264 | }, |
| 265 | } |
| 266 | } |
| 267 | ) |
| 268 | method = config.get_quant_method( |
| 269 | MagicMock(spec=LinearBase), "model.layers.0.mlp.down_proj" |
| 270 | ) |
| 271 | |
| 272 | assert isinstance(method, ModelOptLinearMethod), (algo, type(method).__name__) |
| 273 | |
| 274 | |
| 275 | @pytest.mark.parametrize("algo", ["FP8_PB_WO", "FP8_BLOCK_SCALES"]) |
| 276 | def test_modelopt_mixed_precision_dispatches_block_fp8_moe(algo): |
| 277 | config = ModelOptMixedPrecisionConfig.from_config( |
| 278 | { |
| 279 | "quantization": { |
| 280 | "quant_algo": "MIXED_PRECISION", |
| 281 | "quantized_layers": { |
| 282 | "mtp.layers.48.mlp.experts": { |
| 283 | "quant_algo": algo, |
| 284 | "group_size": 128, |
| 285 | } |
| 286 | }, |
| 287 | } |
| 288 | } |
| 289 | ) |
| 290 | layer = MagicMock(spec=RoutedExperts) |
| 291 | expected = object() |
| 292 | |
| 293 | with patch( |
| 294 | "vllm.model_executor.layers.quantization.modelopt.Fp8MoEMethod", |
| 295 | return_value=expected, |
| 296 | ) as method_cls: |
| 297 | method = config.get_quant_method(layer, "mtp.layers.48.mlp.experts") |
| 298 | |
| 299 | assert method is expected |
| 300 | fp8_config, called_layer = method_cls.call_args.args |
| 301 | assert called_layer is layer |
| 302 | assert fp8_config.weight_block_size == [128, 128] |
| 303 | assert fp8_config.activation_scheme == "dynamic" |
| 304 | |
| 305 | |
| 306 | def test_modelopt_nvfp4_leaves_excluded_parallel_lm_head_unquantized(): |
| 307 | config = ModelOptNvFp4Config( |
| 308 | is_checkpoint_nvfp4_serialized=True, |
| 309 | kv_cache_quant_algo=None, |
| 310 | exclude_modules=["lm_head"], |
| 311 | ) |
| 312 | |
| 313 | method = config.get_quant_method(_mock_lm_head(), prefix="lm_head") |
| 314 | |
| 315 | assert isinstance(method, UnquantizedLinearMethod) |
| 316 | |
| 317 | |
| 318 | def test_modelopt_mixed_precision_quantizes_parallel_lm_head(): |
| 319 | config = _mixed_precision_config( |
| 320 | {"lm_head": {"quant_algo": "NVFP4", "group_size": 16}} |
| 321 | ) |
| 322 | |
| 323 | method = config.get_quant_method(_mock_lm_head(), prefix="lm_head") |
| 324 | |
| 325 | assert isinstance(method, ModelOptLinearMethod) |
| 326 | assert method.spec.weight is kNvfp4Static |
| 327 | assert method.spec.activation is kNvfp4Dynamic |
| 328 | |
| 329 | |
| 330 | def test_modelopt_mixed_precision_resolves_declared_packed_projection(): |
| 331 | config = _mixed_precision_config( |
| 332 | { |
| 333 | "model.layers.0.self_attn.q_proj": {"quant_algo": "MXFP8"}, |
| 334 | "model.layers.0.self_attn.k_proj": {"quant_algo": "MXFP8"}, |
| 335 | "model.layers.0.self_attn.v_proj": {"quant_algo": "MXFP8"}, |
| 336 | } |
| 337 | ) |
| 338 | config.packed_modules_mapping = {"qkv_proj": ["q_proj", "k_proj", "v_proj"]} |
| 339 | |
| 340 | assert config._resolve_quant_algo("model.layers.0.self_attn.qkv_proj") == "MXFP8" |
| 341 | |
| 342 | |
| 343 | def test_modelopt_mixed_precision_does_not_quantize_unlisted_fused_sibling(): |
| 344 | config = _mixed_precision_config( |
| 345 | { |
| 346 | "model.layers.0.linear_attn.in_proj_qkv": {"quant_algo": "FP8"}, |
| 347 | "model.layers.0.linear_attn.in_proj_z": {"quant_algo": "FP8"}, |
| 348 | "model.layers.0.linear_attn.out_proj": {"quant_algo": "FP8"}, |
| 349 | } |
| 350 | ) |
| 351 | config.packed_modules_mapping = { |
| 352 | "in_proj_qkvz": ["in_proj_qkv", "in_proj_z"], |
| 353 | "in_proj_ba": ["in_proj_b", "in_proj_a"], |
| 354 | } |
| 355 | |
| 356 | assert ( |
| 357 | config._resolve_quant_algo("model.layers.0.linear_attn.in_proj_qkvz") == "FP8" |
| 358 | ) |
| 359 | assert config._resolve_quant_algo("model.layers.0.linear_attn.in_proj_ba") is None |
| 360 | |
| 361 | |
| 362 | def test_modelopt_mixed_precision_composes_gemma4_mappers(): |
| 363 | from vllm.model_executor.models.gemma4 import Gemma4ForCausalLM |
| 364 | from vllm.model_executor.models.gemma4_mm import ( |
| 365 | Gemma4ForConditionalGeneration, |
| 366 | ) |
| 367 | |
| 368 | config = _mixed_precision_config( |
| 369 | { |
| 370 | "model.language_model.layers.0.experts": { |
| 371 | "quant_algo": "NVFP4", |
| 372 | "group_size": 16, |
| 373 | }, |
| 374 | "model.language_model.layers.1.experts.gate_up_proj": { |
| 375 | "quant_algo": "NVFP4", |
| 376 | "group_size": 16, |
| 377 | }, |
| 378 | } |
| 379 | ) |
| 380 | |
| 381 | config.apply_vllm_mapper( |
| 382 | Gemma4ForConditionalGeneration.hf_to_vllm_mapper.get_rename_mapper() |
| 383 | ) |
| 384 | config.apply_vllm_mapper(Gemma4ForCausalLM.hf_to_vllm_mapper.get_rename_mapper()) |
| 385 | |
| 386 | expected_prefix = "language_model.model.layers.0.experts" |
| 387 | assert set(config.quantized_layers) == { |
| 388 | expected_prefix, |
| 389 | "language_model.model.layers.1.experts.gate_up_proj", |
| 390 | } |
| 391 | assert config._resolve_quant_algo(expected_prefix) == "NVFP4" |
| 392 | |
| 393 | |
| 394 | def test_modelopt_mixed_precision_infers_fused_gate_up_projection(): |
| 395 | from vllm.model_executor.layers.linear import LinearBase |
| 396 | |
| 397 | config = _mixed_precision_config( |
| 398 | { |
| 399 | "model.layers.0.mlp.gate_proj": {"quant_algo": "NVFP4"}, |
| 400 | "model.layers.0.mlp.up_proj": {"quant_algo": "NVFP4"}, |
| 401 | } |
| 402 | ) |
| 403 | |
| 404 | fake_layer = MagicMock(spec=LinearBase) |
| 405 | method = config.get_quant_method(fake_layer, "model.layers.0.mlp.gate_up_proj") |
| 406 | |
| 407 | assert isinstance(method, ModelOptLinearMethod) |
| 408 | assert method.spec.weight is kNvfp4Static |
| 409 | assert method.spec.activation is kNvfp4Dynamic |
| 410 | |
| 411 | |
| 412 | @pytest.mark.parametrize( |
| 413 | ("quantized_prefix", "missing_prefix"), |
| 414 | [ |
| 415 | ("model.layers.0.mlp.gate_proj", "model.layers.0.mlp.down_proj"), |
| 416 | ("model.layers.0.self_attn.o_proj", "model.layers.0.self_attn.qkv_proj"), |
| 417 | ], |
| 418 | ) |
| 419 | def test_modelopt_mixed_precision_does_not_infer_missing_sibling_linear( |
| 420 | quantized_prefix, missing_prefix |
| 421 | ): |
| 422 | from vllm.model_executor.layers.linear import LinearBase |
| 423 | |
| 424 | config = _mixed_precision_config( |
| 425 | { |
| 426 | quantized_prefix: {"quant_algo": "NVFP4"}, |
| 427 | } |
| 428 | ) |
| 429 | |
| 430 | fake_layer = MagicMock(spec=LinearBase) |
| 431 | method = config.get_quant_method(fake_layer, missing_prefix) |
| 432 | |
| 433 | assert isinstance(method, UnquantizedLinearMethod) |
| 434 | |
| 435 | |
| 436 | def test_vocab_parallel_embedding_weight_loader_accepts_scalar_scale(): |
| 437 | holder = Mock() |
| 438 | scale = torch.nn.Parameter(torch.empty(1)) |
| 439 | loaded_scale = torch.tensor(2.0) |
| 440 | |
| 441 | VocabParallelEmbedding.weight_loader(holder, scale, loaded_scale) |
| 442 | |
| 443 | assert torch.equal(scale, loaded_scale.reshape(1)) |
| 444 | |
| 445 | |
| 446 | @pytest.mark.skipif( |
| 447 | not is_quant_method_supported("modelopt"), |
| 448 | reason="ModelOpt FP8 is not supported on this GPU type.", |
| 449 | ) |
| 450 | def test_modelopt_fp8_checkpoint_setup(default_vllm_config, vllm_runner): |
| 451 | """Test ModelOpt FP8 checkpoint loading and structure validation.""" |
| 452 | # TODO: provide a small publicly available test checkpoint |
| 453 | model_path = ( |
| 454 | "/home/scratch.omniml_data_1/zhiyu/ckpts/test_ckpts/" |
| 455 | "TinyLlama-1.1B-Chat-v1.0-fp8-0710" |
| 456 | ) |
| 457 | |
| 458 | # Skip test if checkpoint doesn't exist |
| 459 | if not os.path.exists(model_path): |
| 460 | pytest.skip( |
| 461 | f"Test checkpoint not found at {model_path}. " |
| 462 | "This test requires a local ModelOpt FP8 checkpoint." |
| 463 | ) |
| 464 | |
| 465 | # Set model config as model_config.dtype is required in ModelOptLinearMethod. |
| 466 | default_vllm_config.model_config = ModelConfig() |
| 467 | with vllm_runner(model_path, quantization="modelopt", enforce_eager=True) as llm: |
| 468 | |
| 469 | def check_model(model): |
| 470 | layer = model.model.layers[0] |
| 471 | |
| 472 | qkv_proj = layer.self_attn.qkv_proj |
| 473 | o_proj = layer.self_attn.o_proj |
| 474 | gate_up_proj = layer.mlp.gate_up_proj |
| 475 | down_proj = layer.mlp.down_proj |
| 476 | |
| 477 | # Check that ModelOpt quantization method is properly applied |
| 478 | assert isinstance(qkv_proj.quant_method, ModelOptLinearMethod) |
| 479 | assert isinstance(o_proj.quant_method, ModelOptLinearMethod) |
| 480 | assert isinstance(gate_up_proj.quant_method, ModelOptLinearMethod) |
| 481 | assert isinstance(down_proj.quant_method, ModelOptLinearMethod) |
| 482 | |
| 483 | # Check weight dtype is FP8 |
| 484 | assert qkv_proj.weight.dtype == torch.float8_e4m3fn |
| 485 | assert o_proj.weight.dtype == torch.float8_e4m3fn |
| 486 | assert gate_up_proj.weight.dtype == torch.float8_e4m3fn |
| 487 | assert down_proj.weight.dtype == torch.float8_e4m3fn |
| 488 | |
| 489 | # Check scales are present and have correct dtype |
| 490 | assert hasattr(qkv_proj, "weight_scale") |
| 491 | assert hasattr(qkv_proj, "input_scale") |
| 492 | assert qkv_proj.weight_scale.dtype == torch.float32 |
| 493 | assert qkv_proj.input_scale.dtype == torch.float32 |
| 494 | |
| 495 | assert hasattr(o_proj, "weight_scale") |
| 496 | assert hasattr(o_proj, "input_scale") |
| 497 | assert o_proj.weight_scale.dtype == torch.float32 |
| 498 | assert o_proj.input_scale.dtype == torch.float32 |
| 499 | |
| 500 | assert hasattr(gate_up_proj, "weight_scale") |
| 501 | assert hasattr(gate_up_proj, "input_scale") |
| 502 | assert gate_up_proj.weight_scale.dtype == torch.float32 |
| 503 | assert gate_up_proj.input_scale.dtype == torch.float32 |
| 504 | |
| 505 | assert hasattr(down_proj, "weight_scale") |
| 506 | assert hasattr(down_proj, "input_scale") |
| 507 | assert down_proj.weight_scale.dtype == torch.float32 |
| 508 | assert down_proj.input_scale.dtype == torch.float32 |
| 509 | |
| 510 | llm.apply_model(check_model) |
| 511 | |
| 512 | # Run a simple generation test to ensure the model works |
| 513 | output = llm.generate_greedy(["Hello my name is"], max_tokens=4) |
| 514 | assert output |
| 515 | print(f"ModelOpt FP8 output: {output}") |
| 516 | |
| 517 | |
| 518 | @pytest.mark.skipif( |
| 519 | not is_quant_method_supported("modelopt"), |
| 520 | reason="ModelOpt FP8 is not supported on this GPU type.", |
| 521 | ) |
| 522 | def test_modelopt_fp8_pc_pt_checkpoint_setup(monkeypatch, dist_init, workspace_init): |
| 523 | """Test ModelOpt FP8_PER_CHANNEL_PER_TOKEN checkpoint setup.""" |
| 524 | model_id = "CedricHwang/qwen2.5-0.5b-modelopt-fp8-pc-pt" |
| 525 | model_path = _snapshot_download_or_skip(model_id) |
| 526 | |
| 527 | model, vllm_config = load_model_without_vllm_runner( |
| 528 | model_path, |
| 529 | quantization="modelopt", |
| 530 | ) |
| 531 | layer = model.model.layers[0] |
| 532 | |
| 533 | qkv_proj = layer.self_attn.qkv_proj |
| 534 | o_proj = layer.self_attn.o_proj |
| 535 | gate_up_proj = layer.mlp.gate_up_proj |
| 536 | down_proj = layer.mlp.down_proj |
| 537 | |
| 538 | assert isinstance(qkv_proj.quant_method, ModelOptLinearMethod) |
| 539 | assert isinstance(o_proj.quant_method, ModelOptLinearMethod) |
| 540 | assert isinstance(gate_up_proj.quant_method, ModelOptLinearMethod) |
| 541 | assert isinstance(down_proj.quant_method, ModelOptLinearMethod) |
| 542 | |
| 543 | fp8_dtype = current_platform.fp8_dtype() |
| 544 | assert qkv_proj.weight.dtype == fp8_dtype |
| 545 | assert o_proj.weight.dtype == fp8_dtype |
| 546 | assert gate_up_proj.weight.dtype == fp8_dtype |
| 547 | assert down_proj.weight.dtype == fp8_dtype |
| 548 | |
| 549 | # Per-channel scales; activations are dynamically scaled per token. |
| 550 | for projection in (qkv_proj, o_proj, gate_up_proj, down_proj): |
| 551 | assert hasattr(projection, "weight_scale") |
| 552 | assert projection.weight_scale.dtype == torch.float32 |
| 553 | assert projection.weight_scale.dim() == 1 |
| 554 | assert not hasattr(projection, "input_scale") |
| 555 | |
| 556 | monkeypatch.setattr(Attention, "forward", lambda _, q, k, v: q.contiguous()) |
| 557 | input_ids = torch.tensor([1, 2, 3, 4], device=current_platform.device_type) |
| 558 | positions = torch.arange(input_ids.numel(), device=current_platform.device_type) |
| 559 | with ( |
| 560 | set_current_vllm_config(vllm_config), |
| 561 | set_forward_context(None, vllm_config, num_tokens=input_ids.numel()), |
| 562 | ): |
| 563 | hidden_states = model(input_ids, positions, None) |
| 564 | logits = model.compute_logits(hidden_states) |
| 565 | assert torch.isfinite(logits).all() |
| 566 | |
| 567 | |
| 568 | @pytest.mark.skipif( |
| 569 | not is_quant_method_supported("modelopt"), |
| 570 | reason="ModelOpt FP8 is not supported on this GPU type.", |
| 571 | ) |
| 572 | def test_modelopt_fp8_pb_wo_checkpoint_setup(monkeypatch, dist_init, workspace_init): |
| 573 | """Test ModelOpt FP8_PB_WO checkpoint setup.""" |
| 574 | model_id = "CedricHwang/qwen2.5-0.5b-modelopt-fp8-pb-wo" |
| 575 | model_path = _snapshot_download_or_skip(model_id) |
| 576 | |
| 577 | model, vllm_config = load_model_without_vllm_runner( |
| 578 | model_path, |
| 579 | quantization="modelopt", |
| 580 | ) |
| 581 | layer = model.model.layers[0] |
| 582 | |
| 583 | qkv_proj = layer.self_attn.qkv_proj |
| 584 | o_proj = layer.self_attn.o_proj |
| 585 | gate_up_proj = layer.mlp.gate_up_proj |
| 586 | down_proj = layer.mlp.down_proj |
| 587 | |
| 588 | assert isinstance(qkv_proj.quant_method, ModelOptLinearMethod) |
| 589 | assert isinstance(o_proj.quant_method, ModelOptLinearMethod) |
| 590 | assert isinstance(gate_up_proj.quant_method, ModelOptLinearMethod) |
| 591 | assert isinstance(down_proj.quant_method, ModelOptLinearMethod) |
| 592 | |
| 593 | fp8_dtype = current_platform.fp8_dtype() |
| 594 | assert qkv_proj.weight.dtype == fp8_dtype |
| 595 | assert o_proj.weight.dtype == fp8_dtype |
| 596 | assert gate_up_proj.weight.dtype == fp8_dtype |
| 597 | assert down_proj.weight.dtype == fp8_dtype |
| 598 | |
| 599 | # Block scales are materialized as a 2D [out_blk, in_blk] tensor. |
| 600 | for projection in (qkv_proj, o_proj, gate_up_proj, down_proj): |
| 601 | assert hasattr(projection, "weight_scale") |
| 602 | assert projection.weight_scale.dtype == torch.float32 |
| 603 | assert projection.weight_scale.dim() == 2 |
| 604 | |
| 605 | monkeypatch.setattr(Attention, "forward", lambda _, q, k, v: q.contiguous()) |
| 606 | input_ids = torch.tensor([1, 2, 3, 4], device=current_platform.device_type) |
| 607 | positions = torch.arange(input_ids.numel(), device=current_platform.device_type) |
| 608 | with ( |
| 609 | set_current_vllm_config(vllm_config), |
| 610 | set_forward_context(None, vllm_config, num_tokens=input_ids.numel()), |
| 611 | ): |
| 612 | hidden_states = model(input_ids, positions, None) |
| 613 | logits = model.compute_logits(hidden_states) |
| 614 | assert torch.isfinite(logits).all() |
| 615 | |
| 616 | |
| 617 | def test_modelopt_nvfp4_config_dispatches_w4a4_method(): |
| 618 | """``quant_method="NVFP4"`` (W4A4) resolves to a |
| 619 | ``(kNvfp4Static, kNvfp4Dynamic)`` QuantSpec under the generic |
| 620 | ``ModelOptLinearMethod``.""" |
| 621 | from vllm.model_executor.layers.linear import LinearBase |
| 622 | |
| 623 | config = ModelOptNvFp4Config( |
| 624 | quant_method="NVFP4", |
| 625 | is_checkpoint_nvfp4_serialized=True, |
| 626 | kv_cache_quant_algo=None, |
| 627 | exclude_modules=[], |
| 628 | ) |
| 629 | assert config.quant_method == "NVFP4" |
| 630 | |
| 631 | method = config.get_quant_method( |
| 632 | MagicMock(spec=LinearBase), "model.layers.0.fake_proj" |
| 633 | ) |
| 634 | assert isinstance(method, ModelOptLinearMethod) |
| 635 | assert method.spec.weight is kNvfp4Static |
| 636 | assert method.spec.activation is kNvfp4Dynamic |
| 637 | |
| 638 | |
| 639 | def test_modelopt_nvfp4_config_dispatches_w4a16_method(): |
| 640 | """``quant_method="W4A16_NVFP4"`` resolves to a weight-only QuantSpec |
| 641 | (``activation=None``) — distinct from the W4A4 sibling. |
| 642 | |
| 643 | A regression here would mean a W4A16 NVFP4 checkpoint silently loaded |
| 644 | with a dynamic fp4 activation key, registering an ``input_scale`` and |
| 645 | routing to the cutlass W4A4 NVFP4 GEMM instead of FP4 Marlin. |
| 646 | """ |
| 647 | from vllm.model_executor.layers.linear import LinearBase |
| 648 | |
| 649 | config = ModelOptNvFp4Config( |
| 650 | quant_method="W4A16_NVFP4", |
| 651 | is_checkpoint_nvfp4_serialized=True, |
| 652 | kv_cache_quant_algo=None, |
| 653 | exclude_modules=[], |
| 654 | ) |
| 655 | assert config.quant_method == "W4A16_NVFP4" |
| 656 | |
| 657 | method = config.get_quant_method( |
| 658 | MagicMock(spec=LinearBase), "model.layers.0.fake_proj" |
| 659 | ) |
| 660 | assert isinstance(method, ModelOptLinearMethod) |
| 661 | assert method.spec.weight is kNvfp4Static |
| 662 | assert method.spec.activation is None |
| 663 | |
| 664 | |
| 665 | def test_modelopt_linear_method_builder_registry_override(monkeypatch): |
| 666 | """The bespoke-method escape hatch: a format registered in |
| 667 | ``LINEAR_METHOD_BUILDERS`` routes that algo to its own method instead of the |
| 668 | generic ``ModelOptLinearMethod``. This is how a format that cannot be a |
| 669 | ``(weight, activation)`` key pair plugs into dispatch.""" |
| 670 | from vllm.model_executor.layers.linear import LinearBase |
| 671 | from vllm.model_executor.layers.quantization import modelopt as m |
| 672 | |
| 673 | sentinel = object() |
| 674 | monkeypatch.setitem(m.LINEAR_METHOD_BUILDERS, "NVFP4", lambda cfg, prefix: sentinel) |
| 675 | |
| 676 | config = ModelOptNvFp4Config( |
| 677 | quant_method="NVFP4", |
| 678 | is_checkpoint_nvfp4_serialized=True, |
| 679 | kv_cache_quant_algo=None, |
| 680 | exclude_modules=[], |
| 681 | ) |
| 682 | method = config.get_quant_method( |
| 683 | MagicMock(spec=LinearBase), "model.layers.0.fake_proj" |
| 684 | ) |
| 685 | assert method is sentinel # bespoke builder wins over the generic path |
| 686 | |
| 687 | |
| 688 | @pytest.mark.parametrize( |
| 689 | ("linear_backend", "kernel_cls"), |
| 690 | [ |
| 691 | ("auto", MarlinNvFp4LinearKernel), |
| 692 | ("humming", HummingNvFp4LinearKernel), |
| 693 | ("flashinfer_cutedsl", FlashInferCuteDslNvFp4W4A16LinearKernel), |
| 694 | ], |
| 695 | ) |
| 696 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA only") |
| 697 | def test_modelopt_w4a16_respects_linear_backend(linear_backend, kernel_cls): |
| 698 | """W4A16 (`activation=None`) kernel selection honors ``--linear-backend``: |
| 699 | ``use_a16=True`` defaults to Marlin, but an explicit backend wins. The |
| 700 | generic method routes this through ``select_linear_kernel``.""" |
| 701 | from vllm.config.quantization import QuantSpec |
| 702 | from vllm.model_executor.layers.quantization.modelopt import ( |
| 703 | RuntimeDtypes, |
| 704 | select_linear_kernel, |
| 705 | ) |
| 706 | |
| 707 | if linear_backend != "auto": |
| 708 | is_supported, reason = kernel_cls.is_supported() |
| 709 | if not is_supported: |
| 710 | pytest.skip(reason) |
| 711 | |
| 712 | vllm_config = VllmConfig() |
| 713 | vllm_config.kernel_config.linear_backend = linear_backend |
| 714 | spec = QuantSpec(weight=kNvfp4Static, activation=None) |
| 715 | rt = RuntimeDtypes(torch.bfloat16, torch.bfloat16) |
| 716 | with set_current_vllm_config(vllm_config): |
| 717 | kernel = select_linear_kernel(spec, MagicMock(), rt) |
| 718 | assert isinstance(kernel, kernel_cls) |
| 719 | |
| 720 | |
| 721 | @pytest.mark.parametrize( |
| 722 | "quant_method, expected_use_a16, act_key_is_none", |
| 723 | [ |
| 724 | ("NVFP4", False, False), # W4A4 default |
| 725 | ("W4A16_NVFP4", True, True), # native W4A16 ckpt |
| 726 | ], |
| 727 | ) |
| 728 | def test_modelopt_nvfp4_moe_dispatches_to_marlin_when_w4a16( |
| 729 | quant_method, expected_use_a16, act_key_is_none |
| 730 | ): |
| 731 | """``ModelOptNvFp4FusedMoE``: when the ckpt's ``quant_method`` is |
| 732 | ``W4A16_NVFP4``, the MoE class must pass ``activation_key=None`` to |
| 733 | ``select_nvfp4_moe_backend``. That filters out every W4A4 backend |
| 734 | (their ``_supports_quant_scheme`` requires |
| 735 | ``(kNvfp4Static, kNvfp4Dynamic)`` exactly); Marlin survives because |
| 736 | it only checks ``weight_key``. A regression here would mean a W4A16 |
| 737 | ckpt silently went to the cutlass W4A4 path. |
| 738 | """ |
| 739 | from vllm.model_executor.layers.quantization.modelopt import ( |
| 740 | ModelOptNvFp4Config, |
| 741 | ModelOptNvFp4FusedMoE, |
| 742 | ) |
| 743 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 744 | kNvfp4Dynamic, |
| 745 | kNvfp4Static, |
| 746 | ) |
| 747 | |
| 748 | config = ModelOptNvFp4Config( |
| 749 | quant_method=quant_method, |
| 750 | is_checkpoint_nvfp4_serialized=True, |
| 751 | kv_cache_quant_algo=None, |
| 752 | exclude_modules=[], |
| 753 | group_size=16, |
| 754 | ) |
| 755 | |
| 756 | mock_select = MagicMock(return_value=(MagicMock(), MagicMock())) |
| 757 | with ( |
| 758 | patch( |
| 759 | "vllm.model_executor.layers.quantization.modelopt.select_nvfp4_moe_backend", |
| 760 | mock_select, |
| 761 | ), |
| 762 | patch( |
| 763 | "vllm.model_executor.layers.quantization.modelopt." |
| 764 | "is_global_sf_supported_for_nvfp4_backend", |
| 765 | return_value=False, |
| 766 | ), |
| 767 | ): |
| 768 | moe = ModelOptNvFp4FusedMoE(config, MagicMock()) |
| 769 | |
| 770 | assert moe.use_a16 is expected_use_a16 |
| 771 | _, kwargs = mock_select.call_args |
| 772 | assert kwargs["weight_key"] is kNvfp4Static |
| 773 | if act_key_is_none: |
| 774 | assert kwargs["activation_key"] is None |
| 775 | else: |
| 776 | assert kwargs["activation_key"] is kNvfp4Dynamic |
| 777 | |
| 778 | |
| 779 | @pytest.mark.parametrize( |
| 780 | "per_layer_algo, expected_weight, expected_activation", |
| 781 | [ |
| 782 | ("NVFP4", kNvfp4Static, kNvfp4Dynamic), |
| 783 | ("W4A16_NVFP4", kNvfp4Static, None), |
| 784 | ("FP8", kFp8StaticTensorSym, kFp8StaticTensorSym), |
| 785 | ("MXFP8", kMxfp8Static, kMxfp8Dynamic), |
| 786 | ], |
| 787 | ) |
| 788 | def test_modelopt_mixed_precision_dispatches_linear_layer( |
| 789 | per_layer_algo, expected_weight, expected_activation |
| 790 | ): |
| 791 | """``ModelOptMixedPrecisionConfig.get_quant_method`` routes a Linear layer |
| 792 | to the generic ``ModelOptLinearMethod`` with the ``QuantSpec`` resolved |
| 793 | from its per-layer ``quant_algo`` entry in ``quantized_layers``. A |
| 794 | regression here would mean a layer got the wrong ``(weight, activation)`` |
| 795 | key pair or fell through to ``UnquantizedLinearMethod`` — e.g. a W4A16 |
| 796 | layer picking up a dynamic fp4 activation key (cutlass W4A4 path) instead |
| 797 | of the weight-only Marlin path. |
| 798 | """ |
| 799 | from vllm.model_executor.layers.linear import LinearBase |
| 800 | from vllm.model_executor.layers.quantization import modelopt as m |
| 801 | |
| 802 | hf_quant_config: dict[str, Any] = { |
| 803 | "quantization": { |
| 804 | "quant_algo": "MIXED_PRECISION", |
| 805 | "kv_cache_quant_algo": None, |
| 806 | "exclude_modules": [], |
| 807 | "group_size": 16, |
| 808 | "quantized_layers": { |
| 809 | "model.layers.0.fake_proj": {"quant_algo": per_layer_algo}, |
| 810 | }, |
| 811 | } |
| 812 | } |
| 813 | config = m.ModelOptMixedPrecisionConfig.from_config(hf_quant_config) |
| 814 | |
| 815 | fake_layer = MagicMock(spec=LinearBase) |
| 816 | method = config.get_quant_method(fake_layer, "model.layers.0.fake_proj") |
| 817 | |
| 818 | assert isinstance(method, m.ModelOptLinearMethod) |
| 819 | assert method.spec.weight is expected_weight |
| 820 | assert method.spec.activation is expected_activation |
| 821 | |
| 822 | |
| 823 | def test_modelopt_mixed_precision_builds_w4a16_sibling_config(): |
| 824 | """Sanity: ``ModelOptMixedPrecisionConfig._from_config`` builds **two** |
| 825 | NVFP4 sub-configs — one for W4A4 (default) and one tagged |
| 826 | ``quant_method='W4A16_NVFP4'`` — so per-layer dispatch can hand |
| 827 | Marlin-bound layers the right config without re-instantiating it on |
| 828 | every call. |
| 829 | """ |
| 830 | from vllm.model_executor.layers.quantization import modelopt as m |
| 831 | |
| 832 | hf_quant_config: dict[str, Any] = { |
| 833 | "quantization": { |
| 834 | "quant_algo": "MIXED_PRECISION", |
| 835 | "kv_cache_quant_algo": None, |
| 836 | "exclude_modules": [], |
| 837 | "group_size": 16, |
| 838 | "quantized_layers": { |
| 839 | "model.layers.0.a": {"quant_algo": "NVFP4"}, |
| 840 | "model.layers.0.b": {"quant_algo": "W4A16_NVFP4"}, |
| 841 | }, |
| 842 | } |
| 843 | } |
| 844 | config = m.ModelOptMixedPrecisionConfig.from_config(hf_quant_config) |
| 845 | |
| 846 | assert isinstance(config, m.ModelOptMixedPrecisionConfig) |
| 847 | assert config.nvfp4_config.quant_method == "NVFP4" |
| 848 | assert config.w4a16_nvfp4_config.quant_method == "W4A16_NVFP4" |
| 849 | |
| 850 | |
| 851 | def test_modelopt_fp8_pb_wo_hides_output_padding(monkeypatch): |
| 852 | """FP8_PB_WO output width that is not a multiple of 128 (a partial trailing |
| 853 | block) is padded up to a block boundary before the kernel post-load, the |
| 854 | GEMM runs on the padded weight, and the output is trimmed back to the |
| 855 | logical width with bias added after. Faithful port of wei-zhao #53132's |
| 856 | test_modelopt_fp8_pb_wo_hides_output_padding for the generic method + |
| 857 | _Fp8PbWoPartialBlock FormatScheme. |
| 858 | |
| 859 | Width is the real motivating case -- GLM's fused qkv_a_proj, q_a 2048 + |
| 860 | kv_a 576 = 2624 (replicated, so no TP degree makes it a 128-multiple), |
| 861 | padded to 2688 = 21 * 128. |
| 862 | """ |
| 863 | from vllm.config.quantization import QuantSpec |
| 864 | from vllm.model_executor.layers.quantization import modelopt as mo |
| 865 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 866 | kFp8Dynamic128Sym, |
| 867 | kFp8Static128BlockSym, |
| 868 | ) |
| 869 | |
| 870 | kernel = Mock() |
| 871 | init_fp8_linear_kernel = Mock(return_value=kernel) |
| 872 | monkeypatch.setattr(mo, "init_fp8_linear_kernel", init_fp8_linear_kernel) |
| 873 | monkeypatch.setattr(mo, "expose_input_quant_key", lambda layer, k: None) |
| 874 | |
| 875 | method = ModelOptLinearMethod.__new__(ModelOptLinearMethod) |
| 876 | method.spec = QuantSpec(weight=kFp8Static128BlockSym, activation=kFp8Dynamic128Sym) |
| 877 | method.ctx = mo.CkptCtx() |
| 878 | method.fmt = mo._PB_WO_PARTIAL_BLOCK |
| 879 | method.wkey = mo.SCHEME_FOR[kFp8Static128BlockSym] |
| 880 | method.akey = mo.SCHEME_FOR[kFp8Dynamic128Sym] |
| 881 | method.input_dtype = method.out_dtype = torch.bfloat16 |
| 882 | method.marlin_input_dtype = None |
| 883 | |
| 884 | layer = torch.nn.Module() |
| 885 | with ( |
| 886 | patch( |
| 887 | "vllm.model_executor.parameter.get_tensor_model_parallel_rank", |
| 888 | return_value=0, |
| 889 | ), |
| 890 | patch( |
| 891 | "vllm.model_executor.parameter.get_tensor_model_parallel_world_size", |
| 892 | return_value=1, |
| 893 | ), |
| 894 | ): |
| 895 | # output 2624 = 2048 + 576, input 128 |
| 896 | method.create_weights( |
| 897 | layer, 128, [2048, 576], 128, 2624, torch.bfloat16, weight_loader=Mock() |
| 898 | ) |
| 899 | |
| 900 | # loaded at logical size; scale is cdiv(2624, 128) = 21 block rows |
| 901 | assert layer.weight.shape == (2624, 128) |
| 902 | assert layer.weight_scale.shape == (21, 1, 1, 1) |
| 903 | assert init_fp8_linear_kernel.call_args.kwargs["weight_shape"] == (2688, 128) |
| 904 | |
| 905 | layer.weight.data.fill_(1) |
| 906 | method.process_weights_after_loading(layer) |
| 907 | |
| 908 | # weight padded to the block boundary; the pad rows are zero |
| 909 | assert layer.weight.shape == (2688, 128) |
| 910 | assert torch.count_nonzero(layer.weight[2624:].float()) == 0 |
| 911 | kernel.process_weights_after_loading.assert_called_once_with(layer) |
| 912 | |
| 913 | # apply: GEMM on padded weight (bias=None), output trimmed + bias added |
| 914 | physical_output = torch.randn(4, 2688, dtype=torch.bfloat16) |
| 915 | kernel.apply_weights.return_value = physical_output |
| 916 | bias = torch.randn(2624, dtype=torch.bfloat16) |
| 917 | output = method.apply(layer, torch.randn(4, 128), bias) |
| 918 | |
| 919 | torch.testing.assert_close(output, physical_output[:, :2624] + bias) |
| 920 | assert output.shape == (4, 2624) |
| 921 | assert output.is_contiguous() |
| 922 | kernel.apply_weights.assert_called_once() |
| 923 | assert kernel.apply_weights.call_args.kwargs["bias"] is None |
| 924 | |
| 925 | |
| 926 | def test_modelopt_fp8_pb_wo_rejects_non_128_input(): |
| 927 | """Input width must still be a multiple of 128 (same as #53132, which only |
| 928 | pads the output). A partial input block is refused loudly rather than |
| 929 | silently loading wrong scales.""" |
| 930 | from vllm.model_executor.layers.quantization import modelopt as mo |
| 931 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 932 | kFp8Static128BlockSym, |
| 933 | ) |
| 934 | |
| 935 | scheme = mo.SCHEME_FOR[kFp8Static128BlockSym] |
| 936 | shapes = mo.Shapes([128], 100, torch.bfloat16) # input 100 not divisible by 128 |
| 937 | with pytest.raises(ValueError, match="in divisible by 128"): |
| 938 | scheme.create_weights( |
| 939 | torch.nn.Module(), mo.WEIGHT, mo.CkptCtx(), shapes, Mock() |
| 940 | ) |