| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import torch |
| 5 | |
| 6 | from vllm.logger import init_logger |
| 7 | from vllm.model_executor.layers.attention import Attention |
| 8 | from vllm.model_executor.layers.fused_moe import ( |
| 9 | FusedMoEConfig, |
| 10 | FusedMoEMethodBase, |
| 11 | FusedMoEParallelConfig, |
| 12 | FusedMoEQuantConfig, |
| 13 | RoutedExperts, |
| 14 | SharedExperts, |
| 15 | ) |
| 16 | from vllm.model_executor.layers.fused_moe import modular_kernel as mk |
| 17 | from vllm.model_executor.layers.fused_moe.config import ( |
| 18 | mxfp4_w4a16_moe_quant_config, |
| 19 | ) |
| 20 | from vllm.model_executor.layers.fused_moe.moe_output import UnfinalizedMoEOutput |
| 21 | from vllm.model_executor.layers.fused_moe.oracle.mxfp4 import ( |
| 22 | TRITON_BACKENDS, |
| 23 | Mxfp4MoeBackend, |
| 24 | convert_gpt_oss_weight_to_mxfp4_moe_kernel_format, |
| 25 | convert_weight_to_mxfp4_moe_kernel_format, |
| 26 | make_mxfp4_moe_kernel, |
| 27 | make_mxfp4_moe_quant_config, |
| 28 | mxfp4_round_up_hidden_size_and_intermediate_size, |
| 29 | select_deepseek_v4_mxfp4_moe_backend, |
| 30 | select_mxfp4_moe_backend, |
| 31 | ) |
| 32 | from vllm.model_executor.layers.linear import LinearBase, UnquantizedLinearMethod |
| 33 | from vllm.model_executor.layers.quantization import QuantizationMethods |
| 34 | from vllm.model_executor.layers.quantization.base_config import ( |
| 35 | QuantizationConfig, |
| 36 | QuantizeMethodBase, |
| 37 | ) |
| 38 | from vllm.model_executor.layers.quantization.utils.quant_utils import is_layer_skipped |
| 39 | from vllm.model_executor.utils import ( |
| 40 | is_weights_pre_processed, |
| 41 | replace_parameter, |
| 42 | set_weight_attrs, |
| 43 | ) |
| 44 | from vllm.platforms import current_platform |
| 45 | |
| 46 | logger = init_logger(__name__) |
| 47 | |
| 48 | |
| 49 | class Mxfp4Config(QuantizationConfig): |
| 50 | """Canonical base config for MXFP4 quantization. |
| 51 | |
| 52 | Subclasses override get_name() and override_quantization_method() to |
| 53 | register themselves as the handler for a specific checkpoint format. |
| 54 | """ |
| 55 | |
| 56 | def __init__(self, ignored_layers: list[str] | None = None): |
| 57 | super().__init__() |
| 58 | self.ignored_layers = ignored_layers |
| 59 | |
| 60 | @classmethod |
| 61 | def from_config(cls, config): |
| 62 | return cls() |
| 63 | |
| 64 | @classmethod |
| 65 | def get_min_capability(cls) -> int: |
| 66 | return 80 |
| 67 | |
| 68 | @classmethod |
| 69 | def get_name(cls) -> QuantizationMethods: |
| 70 | return "mxfp4" |
| 71 | |
| 72 | @classmethod |
| 73 | def get_supported_act_dtypes(cls) -> list[torch.dtype]: |
| 74 | return [torch.bfloat16] |
| 75 | |
| 76 | @classmethod |
| 77 | def get_config_filenames(cls) -> list[str]: |
| 78 | return [] |
| 79 | |
| 80 | def _make_moe_method(self, moe: FusedMoEConfig) -> FusedMoEMethodBase: |
| 81 | """MoE method for RoutedExperts. Subclasses override to pick a |
| 82 | checkpoint-specific kernel family.""" |
| 83 | return Mxfp4MoEMethod(moe) |
| 84 | |
| 85 | def get_quant_method( |
| 86 | self, layer: torch.nn.Module, prefix: str |
| 87 | ) -> "QuantizeMethodBase | None": |
| 88 | if isinstance(layer, LinearBase): |
| 89 | if self.ignored_layers and is_layer_skipped( |
| 90 | prefix=prefix, |
| 91 | ignored_layers=self.ignored_layers, |
| 92 | fused_mapping=self.packed_modules_mapping, |
| 93 | ): |
| 94 | return UnquantizedLinearMethod() |
| 95 | logger.debug_once( |
| 96 | "MXFP4 linear layer is not implemented - falling back to " |
| 97 | "UnquantizedLinearMethod.", |
| 98 | ) |
| 99 | return UnquantizedLinearMethod() |
| 100 | elif isinstance(layer, RoutedExperts): |
| 101 | return self._make_moe_method(layer.moe_config) |
| 102 | elif isinstance(layer, Attention): |
| 103 | logger.debug_once( |
| 104 | "MXFP4 attention layer is not implemented. " |
| 105 | "Skipping quantization for this layer.", |
| 106 | ) |
| 107 | return None |
| 108 | |
| 109 | |
| 110 | class GptOssMxfp4Config(Mxfp4Config): |
| 111 | """MXFP4 config for GPT-OSS checkpoints. |
| 112 | |
| 113 | Checkpoints carry ``"quant_method": "mxfp4"`` in their JSON config. |
| 114 | override_quantization_method() maps that to the canonical internal name |
| 115 | so that the rest of the loading path uses "gpt_oss_mxfp4" consistently. |
| 116 | """ |
| 117 | |
| 118 | @classmethod |
| 119 | def get_name(cls) -> QuantizationMethods: |
| 120 | return "gpt_oss_mxfp4" |
| 121 | |
| 122 | @classmethod |
| 123 | def override_quantization_method( |
| 124 | cls, hf_quant_cfg, user_quant, hf_config=None |
| 125 | ) -> QuantizationMethods | None: |
| 126 | # Match both "mxfp4" (original checkpoint value) and "gpt_oss_mxfp4" |
| 127 | # (already normalized by verify_and_update_model_config) so that |
| 128 | # explicit --quantization mxfp4 from the user doesn't cause a mismatch. |
| 129 | if not ( |
| 130 | isinstance(hf_quant_cfg, dict) |
| 131 | and hf_quant_cfg.get("quant_method") in ("mxfp4", "gpt_oss_mxfp4") |
| 132 | ): |
| 133 | return None |
| 134 | # Require explicit confirmation that this is a GPT-OSS model. |
| 135 | # Do NOT fall back to returning the override when hf_config is None, |
| 136 | # as that would silently claim all mxfp4 checkpoints. |
| 137 | model_type = getattr(hf_config, "model_type", None) |
| 138 | if model_type != "gpt_oss": |
| 139 | return None |
| 140 | return "gpt_oss_mxfp4" |
| 141 | |
| 142 | def _make_moe_method(self, moe: FusedMoEConfig) -> FusedMoEMethodBase: |
| 143 | return GptOssMxfp4MoEMethod(moe) |
| 144 | |
| 145 | |
| 146 | class GptOssMxfp4MoEMethod(FusedMoEMethodBase): |
| 147 | """MXFP4 MoE quantization method.""" |
| 148 | |
| 149 | def __init__(self, moe: FusedMoEConfig): |
| 150 | super().__init__(moe) |
| 151 | self.weight_dtype = "gpt_oss_mxfp4" |
| 152 | self.mxfp4_backend, self.experts_cls = select_mxfp4_moe_backend(moe) |
| 153 | |
| 154 | self.max_capture_size = moe.max_capture_size |
| 155 | |
| 156 | self._cache_permute_indices: dict[torch.Size, torch.Tensor] = {} |
| 157 | self.moe_kernel: mk.FusedMoEKernel | None = None |
| 158 | |
| 159 | # Used for triton kernel precision configs |
| 160 | self.w13_precision_config = None |
| 161 | self.w2_precision_config = None |
| 162 | |
| 163 | @property |
| 164 | def skip_forward_padding(self) -> bool: |
| 165 | # SM100_FI_MXFP4_MXFP8_TRTLLM supports padding with mxfp8 quant |
| 166 | # so can skip the padding in the forward before applying the moe method |
| 167 | return self.mxfp4_backend == Mxfp4MoeBackend.FLASHINFER_TRTLLM_MXFP4_MXFP8 |
| 168 | |
| 169 | # TODO(bnell): move to MK/expert_class? |
| 170 | @property |
| 171 | def has_unpadded_output(self) -> bool: |
| 172 | return self.mxfp4_backend in [ |
| 173 | Mxfp4MoeBackend.FLASHINFER_TRTLLM_MXFP4_MXFP8, |
| 174 | Mxfp4MoeBackend.FLASHINFER_TRTLLM_MXFP4_BF16, |
| 175 | ] |
| 176 | |
| 177 | def maybe_roundup_sizes( |
| 178 | self, |
| 179 | hidden_size: int, |
| 180 | intermediate_size_per_partition: int, |
| 181 | act_dtype: torch.dtype, |
| 182 | moe_parallel_config: FusedMoEParallelConfig, |
| 183 | ) -> tuple[int, int]: |
| 184 | hidden_size, intermediate_size_per_partition = super().maybe_roundup_sizes( |
| 185 | hidden_size=hidden_size, |
| 186 | intermediate_size_per_partition=intermediate_size_per_partition, |
| 187 | act_dtype=act_dtype, |
| 188 | moe_parallel_config=moe_parallel_config, |
| 189 | ) |
| 190 | return mxfp4_round_up_hidden_size_and_intermediate_size( |
| 191 | self.mxfp4_backend, hidden_size, intermediate_size_per_partition |
| 192 | ) |
| 193 | |
| 194 | def create_weights( |
| 195 | self, |
| 196 | layer: RoutedExperts, |
| 197 | num_experts: int, |
| 198 | hidden_size: int, |
| 199 | intermediate_size_per_partition: int, |
| 200 | params_dtype: torch.dtype, |
| 201 | **extra_weight_attrs, |
| 202 | ): |
| 203 | self.num_experts = num_experts |
| 204 | weight_dtype = torch.uint8 |
| 205 | scale_dtype = torch.uint8 |
| 206 | mxfp4_block = 32 |
| 207 | |
| 208 | layer.params_dtype = params_dtype |
| 209 | layer.num_experts = num_experts |
| 210 | self.intermediate_size = intermediate_size_per_partition |
| 211 | self.hidden_size = hidden_size |
| 212 | |
| 213 | # Fused gate_up_proj (column parallel) |
| 214 | w13_weight = torch.nn.Parameter( |
| 215 | torch.zeros( |
| 216 | num_experts, |
| 217 | self.moe.w13_num_shards * intermediate_size_per_partition, |
| 218 | hidden_size // 2, |
| 219 | dtype=weight_dtype, |
| 220 | ), |
| 221 | requires_grad=False, |
| 222 | ) |
| 223 | layer.register_parameter("w13_weight", w13_weight) |
| 224 | set_weight_attrs(w13_weight, extra_weight_attrs) |
| 225 | |
| 226 | w13_weight_scale = torch.nn.Parameter( |
| 227 | torch.zeros( |
| 228 | num_experts, |
| 229 | self.moe.w13_num_shards * intermediate_size_per_partition, |
| 230 | hidden_size // mxfp4_block, |
| 231 | dtype=scale_dtype, |
| 232 | ), |
| 233 | requires_grad=False, |
| 234 | ) |
| 235 | layer.register_parameter("w13_weight_scale", w13_weight_scale) |
| 236 | set_weight_attrs(w13_weight_scale, extra_weight_attrs) |
| 237 | w13_weight_scale.quant_method = "block" |
| 238 | |
| 239 | # down_proj (row parallel) |
| 240 | w2_weight = torch.nn.Parameter( |
| 241 | torch.zeros( |
| 242 | num_experts, |
| 243 | hidden_size, |
| 244 | intermediate_size_per_partition // 2, |
| 245 | dtype=weight_dtype, |
| 246 | ), |
| 247 | requires_grad=False, |
| 248 | ) |
| 249 | layer.register_parameter("w2_weight", w2_weight) |
| 250 | set_weight_attrs(w2_weight, extra_weight_attrs) |
| 251 | |
| 252 | w2_weight_scale = torch.nn.Parameter( |
| 253 | torch.zeros( |
| 254 | num_experts, |
| 255 | hidden_size, |
| 256 | intermediate_size_per_partition // mxfp4_block, |
| 257 | dtype=scale_dtype, |
| 258 | ), |
| 259 | requires_grad=False, |
| 260 | ) |
| 261 | layer.register_parameter("w2_weight_scale", w2_weight_scale) |
| 262 | set_weight_attrs(w2_weight_scale, extra_weight_attrs) |
| 263 | w2_weight_scale.quant_method = "block" |
| 264 | |
| 265 | if self.moe.has_bias: |
| 266 | w13_bias = torch.nn.Parameter( |
| 267 | torch.zeros( |
| 268 | num_experts, |
| 269 | self.moe.w13_num_shards * intermediate_size_per_partition, |
| 270 | dtype=torch.bfloat16, |
| 271 | ), |
| 272 | requires_grad=False, |
| 273 | ) |
| 274 | layer.register_parameter("w13_bias", w13_bias) |
| 275 | set_weight_attrs(w13_bias, extra_weight_attrs) |
| 276 | |
| 277 | w2_bias = torch.nn.Parameter( |
| 278 | torch.zeros( |
| 279 | num_experts, |
| 280 | hidden_size, |
| 281 | dtype=torch.bfloat16, |
| 282 | ), |
| 283 | requires_grad=False, |
| 284 | ) |
| 285 | layer.register_parameter("w2_bias", w2_bias) |
| 286 | set_weight_attrs(w2_bias, extra_weight_attrs) |
| 287 | |
| 288 | def _setup_kernel( |
| 289 | self, |
| 290 | layer: RoutedExperts, |
| 291 | w13: torch.Tensor, |
| 292 | w2: torch.Tensor, |
| 293 | w13_scale: torch.Tensor, |
| 294 | w2_scale: torch.Tensor, |
| 295 | w13_bias: torch.Tensor | None = None, |
| 296 | w2_bias: torch.Tensor | None = None, |
| 297 | ) -> None: |
| 298 | num_experts = self.num_experts |
| 299 | intermediate_size = self.intermediate_size |
| 300 | hidden_size = self.hidden_size |
| 301 | sf_block_size = 32 |
| 302 | |
| 303 | # Shape assertions |
| 304 | assert ( |
| 305 | w13.dim() == 3 |
| 306 | and w13.shape[0] == num_experts |
| 307 | and w13.shape[1] == intermediate_size * self.moe.w13_num_shards |
| 308 | and w13.shape[2] == hidden_size // 2 |
| 309 | ) |
| 310 | assert ( |
| 311 | w13_scale.dim() == 3 |
| 312 | and w13_scale.shape[0] == num_experts |
| 313 | and w13_scale.shape[1] == intermediate_size * self.moe.w13_num_shards |
| 314 | and w13_scale.shape[2] == hidden_size // sf_block_size |
| 315 | ) |
| 316 | assert ( |
| 317 | w2.dim() == 3 |
| 318 | and w2.shape[0] == num_experts |
| 319 | and w2.shape[1] == hidden_size |
| 320 | and w2.shape[2] == intermediate_size // 2 |
| 321 | ) |
| 322 | assert ( |
| 323 | w2_scale.dim() == 3 |
| 324 | and w2_scale.shape[1] == hidden_size |
| 325 | and w2_scale.shape[2] == intermediate_size // sf_block_size |
| 326 | ) |
| 327 | if w13_bias is not None: |
| 328 | assert ( |
| 329 | w13_bias.dim() == 2 |
| 330 | and w13_bias.shape[0] == num_experts |
| 331 | and w13_bias.shape[1] == intermediate_size * self.moe.w13_num_shards |
| 332 | ) |
| 333 | if w2_bias is not None: |
| 334 | assert ( |
| 335 | w2_bias.dim() == 2 |
| 336 | and w2_bias.shape[0] == num_experts |
| 337 | and w2_bias.shape[1] == hidden_size |
| 338 | ) |
| 339 | |
| 340 | # Convert weights to kernel format |
| 341 | w13, w2, w13_scale, w2_scale, w13_bias, w2_bias = ( |
| 342 | convert_gpt_oss_weight_to_mxfp4_moe_kernel_format( |
| 343 | mxfp4_backend=self.mxfp4_backend, |
| 344 | layer=layer, |
| 345 | w13_weight=w13, |
| 346 | w2_weight=w2, |
| 347 | w13_weight_scale=w13_scale, |
| 348 | w2_weight_scale=w2_scale, |
| 349 | w13_bias=w13_bias, |
| 350 | w2_bias=w2_bias, |
| 351 | _cache_permute_indices=self._cache_permute_indices, |
| 352 | ) |
| 353 | ) |
| 354 | |
| 355 | # For TRITON backends, weights are wrapped tensors from triton_kernels |
| 356 | # that don't support .detach(). Manually assign parameters. |
| 357 | if self.mxfp4_backend not in TRITON_BACKENDS: |
| 358 | replace_parameter(layer, "w13_weight", w13) |
| 359 | replace_parameter(layer, "w2_weight", w2) |
| 360 | replace_parameter(layer, "w13_weight_scale", w13_scale) |
| 361 | replace_parameter(layer, "w2_weight_scale", w2_scale) |
| 362 | else: |
| 363 | layer.w13_weight = w13 |
| 364 | layer.w2_weight = w2 |
| 365 | self.w13_precision_config = w13_scale |
| 366 | self.w2_precision_config = w2_scale |
| 367 | |
| 368 | if w13_bias is not None and w2_bias is not None: |
| 369 | replace_parameter(layer, "w13_bias", w13_bias) |
| 370 | replace_parameter(layer, "w2_bias", w2_bias) |
| 371 | |
| 372 | # Build quant config |
| 373 | self.moe_quant_config = self.get_fused_moe_quant_config(layer) |
| 374 | |
| 375 | # Build kernel (modular or monolithic) |
| 376 | if self.moe_quant_config is not None and self.experts_cls is not None: |
| 377 | self.moe_kernel = make_mxfp4_moe_kernel( |
| 378 | moe_quant_config=self.moe_quant_config, |
| 379 | moe_config=self.moe, |
| 380 | mxfp4_backend=self.mxfp4_backend, |
| 381 | experts_cls=self.experts_cls, |
| 382 | routing_tables=layer._expert_routing_tables(), |
| 383 | ) |
| 384 | self.moe_kernel.fused_experts.process_weights_after_loading(layer) |
| 385 | |
| 386 | def process_weights_after_loading(self, layer: RoutedExperts) -> None: |
| 387 | w13 = layer.w13_weight |
| 388 | w2 = layer.w2_weight |
| 389 | w13_scale = layer.w13_weight_scale |
| 390 | w2_scale = layer.w2_weight_scale |
| 391 | w13_bias = getattr(layer, "w13_bias", None) |
| 392 | w2_bias = getattr(layer, "w2_bias", None) |
| 393 | |
| 394 | if self.mxfp4_backend == Mxfp4MoeBackend.NONE: |
| 395 | return |
| 396 | |
| 397 | self._setup_kernel(layer, w13, w2, w13_scale, w2_scale, w13_bias, w2_bias) |
| 398 | |
| 399 | def get_fused_moe_quant_config( |
| 400 | self, layer: RoutedExperts |
| 401 | ) -> FusedMoEQuantConfig | None: |
| 402 | w1_bias = getattr(layer, "w13_bias", None) |
| 403 | w2_bias = getattr(layer, "w2_bias", None) |
| 404 | |
| 405 | if self.mxfp4_backend in TRITON_BACKENDS: |
| 406 | # TRITON backends free w13/w2_weight_scale after swizzling; the |
| 407 | # swizzled scales live inside the precision configs instead. |
| 408 | assert self.w13_precision_config is not None |
| 409 | assert self.w2_precision_config is not None |
| 410 | w1_scale = self.w13_precision_config |
| 411 | w2_scale = self.w2_precision_config |
| 412 | else: |
| 413 | w1_scale = layer.w13_weight_scale |
| 414 | w2_scale = layer.w2_weight_scale |
| 415 | |
| 416 | return make_mxfp4_moe_quant_config( |
| 417 | mxfp4_backend=self.mxfp4_backend, |
| 418 | w1_scale=w1_scale, |
| 419 | w2_scale=w2_scale, |
| 420 | w1_bias=w1_bias, |
| 421 | w2_bias=w2_bias, |
| 422 | gemm1_alpha=1.702, |
| 423 | gemm1_beta=1.0, |
| 424 | swiglu_limit=7.0, |
| 425 | layer=layer, |
| 426 | ) |
| 427 | |
| 428 | def apply( |
| 429 | self, |
| 430 | layer: RoutedExperts, |
| 431 | x: torch.Tensor, |
| 432 | topk_weights: torch.Tensor, |
| 433 | topk_ids: torch.Tensor, |
| 434 | shared_experts: SharedExperts | None, |
| 435 | shared_experts_input: torch.Tensor | None, |
| 436 | ) -> torch.Tensor | UnfinalizedMoEOutput: |
| 437 | assert not self.is_monolithic |
| 438 | assert self.moe_kernel is not None |
| 439 | return self.moe_kernel.apply( |
| 440 | hidden_states=x, |
| 441 | w1=layer.w13_weight, |
| 442 | w2=layer.w2_weight, |
| 443 | topk_weights=topk_weights, |
| 444 | topk_ids=topk_ids, |
| 445 | activation=layer.activation, |
| 446 | global_num_experts=layer.global_num_experts, |
| 447 | apply_router_weight_on_input=layer.apply_router_weight_on_input, |
| 448 | expert_map=layer.expert_map, |
| 449 | shared_experts=shared_experts, |
| 450 | shared_experts_input=shared_experts_input, |
| 451 | ) |
| 452 | |
| 453 | def apply_monolithic( |
| 454 | self, |
| 455 | layer: RoutedExperts, |
| 456 | x: torch.Tensor, |
| 457 | router_logits: torch.Tensor, |
| 458 | input_ids: torch.Tensor | None = None, |
| 459 | ) -> torch.Tensor | UnfinalizedMoEOutput: |
| 460 | assert self.is_monolithic |
| 461 | assert self.moe_kernel is not None |
| 462 | return self.moe_kernel.apply_monolithic( |
| 463 | hidden_states=x, |
| 464 | w1=layer.w13_weight, |
| 465 | w2=layer.w2_weight, |
| 466 | router_logits=router_logits, |
| 467 | activation=layer.activation, |
| 468 | global_num_experts=layer.global_num_experts, |
| 469 | expert_map=layer.expert_map, |
| 470 | apply_router_weight_on_input=layer.apply_router_weight_on_input, |
| 471 | num_expert_group=layer.num_expert_group, |
| 472 | topk_group=layer.topk_group, |
| 473 | e_score_correction_bias=layer.e_score_correction_bias, |
| 474 | routed_scaling_factor=layer.routed_scaling_factor, |
| 475 | ) |
| 476 | |
| 477 | |
| 478 | class Mxfp4MoEMethod(FusedMoEMethodBase): |
| 479 | """MXFP4 MoE quantization method.""" |
| 480 | |
| 481 | supports_pre_processed_weights = True |
| 482 | |
| 483 | def __init__(self, moe: FusedMoEConfig): |
| 484 | super().__init__(moe) |
| 485 | |
| 486 | self.weight_dtype = "mxfp4" |
| 487 | self.mxfp4_backend, self.experts_cls = select_deepseek_v4_mxfp4_moe_backend(moe) |
| 488 | |
| 489 | self.max_capture_size = moe.max_capture_size |
| 490 | |
| 491 | self._cache_permute_indices: dict[torch.Size, torch.Tensor] = {} |
| 492 | self.moe_kernel: mk.FusedMoEKernel | None = None |
| 493 | |
| 494 | # Used for triton kernel precision configs |
| 495 | self.w13_precision_config = None |
| 496 | self.w2_precision_config = None |
| 497 | |
| 498 | @property |
| 499 | def supports_eplb(self) -> bool: |
| 500 | return True |
| 501 | |
| 502 | @property |
| 503 | def skip_forward_padding(self) -> bool: |
| 504 | # SM100_FI_MXFP4_MXFP8_TRTLLM supports padding with mxfp8 quant |
| 505 | # so can skip the padding in the forward before applying the moe method |
| 506 | return self.mxfp4_backend == Mxfp4MoeBackend.FLASHINFER_TRTLLM_MXFP4_MXFP8 |
| 507 | |
| 508 | # TODO(bnell): move to MK/expert_class? |
| 509 | @property |
| 510 | def has_unpadded_output(self) -> bool: |
| 511 | return self.mxfp4_backend in [ |
| 512 | Mxfp4MoeBackend.FLASHINFER_TRTLLM_MXFP4_MXFP8, |
| 513 | Mxfp4MoeBackend.FLASHINFER_TRTLLM_MXFP4_BF16, |
| 514 | ] |
| 515 | |
| 516 | def maybe_roundup_sizes( |
| 517 | self, |
| 518 | hidden_size: int, |
| 519 | intermediate_size_per_partition: int, |
| 520 | act_dtype: torch.dtype, |
| 521 | moe_parallel_config: FusedMoEParallelConfig, |
| 522 | ) -> tuple[int, int]: |
| 523 | hidden_size, intermediate_size_per_partition = super().maybe_roundup_sizes( |
| 524 | hidden_size=hidden_size, |
| 525 | intermediate_size_per_partition=intermediate_size_per_partition, |
| 526 | act_dtype=act_dtype, |
| 527 | moe_parallel_config=moe_parallel_config, |
| 528 | ) |
| 529 | return mxfp4_round_up_hidden_size_and_intermediate_size( |
| 530 | self.mxfp4_backend, |
| 531 | hidden_size, |
| 532 | intermediate_size_per_partition, |
| 533 | activation=self.moe.activation, |
| 534 | ) |
| 535 | |
| 536 | @staticmethod |
| 537 | def _encode_mxfp4_weight_scale(loaded_weight: torch.Tensor) -> torch.Tensor: |
| 538 | if loaded_weight.dtype == torch.uint8: |
| 539 | return loaded_weight |
| 540 | if loaded_weight.dtype == torch.float8_e8m0fnu: |
| 541 | return loaded_weight.view(torch.uint8) |
| 542 | if loaded_weight.is_floating_point(): |
| 543 | return loaded_weight.to(torch.float8_e8m0fnu).view(torch.uint8) |
| 544 | return loaded_weight |
| 545 | |
| 546 | @staticmethod |
| 547 | def get_scale_weight_loader(weight_loader): |
| 548 | def mxfp4_weight_loader( |
| 549 | param: torch.nn.Parameter, |
| 550 | loaded_weight: torch.Tensor, |
| 551 | weight_name: str, |
| 552 | shard_id: str, |
| 553 | expert_id: int, |
| 554 | return_success: bool = False, |
| 555 | ) -> bool | None: |
| 556 | loaded_weight = Mxfp4MoEMethod._encode_mxfp4_weight_scale(loaded_weight) |
| 557 | return weight_loader( |
| 558 | param, |
| 559 | loaded_weight, |
| 560 | weight_name, |
| 561 | shard_id, |
| 562 | expert_id, |
| 563 | return_success=return_success, |
| 564 | ) |
| 565 | |
| 566 | return mxfp4_weight_loader |
| 567 | |
| 568 | def create_weights( |
| 569 | self, |
| 570 | layer: RoutedExperts, |
| 571 | num_experts: int, |
| 572 | hidden_size: int, |
| 573 | intermediate_size_per_partition: int, |
| 574 | params_dtype: torch.dtype, |
| 575 | **extra_weight_attrs, |
| 576 | ): |
| 577 | self.num_experts = num_experts |
| 578 | weight_dtype = torch.uint8 |
| 579 | scale_dtype = torch.uint8 |
| 580 | mxfp4_block = 32 |
| 581 | |
| 582 | layer.params_dtype = params_dtype |
| 583 | layer.num_experts = num_experts |
| 584 | self.intermediate_size = intermediate_size_per_partition |
| 585 | self.hidden_size = hidden_size |
| 586 | weight_loader = extra_weight_attrs.pop("weight_loader") |
| 587 | scale_weight_loader = Mxfp4MoEMethod.get_scale_weight_loader(weight_loader) |
| 588 | |
| 589 | # Fused gate_up_proj (column parallel) |
| 590 | w13_weight = torch.nn.Parameter( |
| 591 | torch.zeros( |
| 592 | num_experts, |
| 593 | self.moe.w13_num_shards * intermediate_size_per_partition, |
| 594 | hidden_size // 2, |
| 595 | dtype=weight_dtype, |
| 596 | ), |
| 597 | requires_grad=False, |
| 598 | ) |
| 599 | layer.register_parameter("w13_weight", w13_weight) |
| 600 | set_weight_attrs(w13_weight, extra_weight_attrs) |
| 601 | set_weight_attrs(w13_weight, {"weight_loader": weight_loader}) |
| 602 | |
| 603 | w13_weight_scale = torch.nn.Parameter( |
| 604 | torch.zeros( |
| 605 | num_experts, |
| 606 | self.moe.w13_num_shards * intermediate_size_per_partition, |
| 607 | hidden_size // mxfp4_block, |
| 608 | dtype=scale_dtype, |
| 609 | ), |
| 610 | requires_grad=False, |
| 611 | ) |
| 612 | layer.register_parameter("w13_weight_scale", w13_weight_scale) |
| 613 | set_weight_attrs(w13_weight_scale, extra_weight_attrs) |
| 614 | set_weight_attrs(w13_weight_scale, {"weight_loader": scale_weight_loader}) |
| 615 | w13_weight_scale.quant_method = "block" |
| 616 | |
| 617 | # down_proj (row parallel) |
| 618 | w2_weight = torch.nn.Parameter( |
| 619 | torch.zeros( |
| 620 | num_experts, |
| 621 | hidden_size, |
| 622 | intermediate_size_per_partition // 2, |
| 623 | dtype=weight_dtype, |
| 624 | ), |
| 625 | requires_grad=False, |
| 626 | ) |
| 627 | layer.register_parameter("w2_weight", w2_weight) |
| 628 | set_weight_attrs(w2_weight, extra_weight_attrs) |
| 629 | set_weight_attrs(w2_weight, {"weight_loader": weight_loader}) |
| 630 | |
| 631 | w2_weight_scale = torch.nn.Parameter( |
| 632 | torch.zeros( |
| 633 | num_experts, |
| 634 | hidden_size, |
| 635 | intermediate_size_per_partition // mxfp4_block, |
| 636 | dtype=scale_dtype, |
| 637 | ), |
| 638 | requires_grad=False, |
| 639 | ) |
| 640 | layer.register_parameter("w2_weight_scale", w2_weight_scale) |
| 641 | set_weight_attrs(w2_weight_scale, extra_weight_attrs) |
| 642 | set_weight_attrs(w2_weight_scale, {"weight_loader": scale_weight_loader}) |
| 643 | w2_weight_scale.quant_method = "block" |
| 644 | |
| 645 | if self.moe.has_bias: |
| 646 | w13_bias = torch.nn.Parameter( |
| 647 | torch.zeros( |
| 648 | num_experts, |
| 649 | self.moe.w13_num_shards * intermediate_size_per_partition, |
| 650 | dtype=torch.bfloat16, |
| 651 | ), |
| 652 | requires_grad=False, |
| 653 | ) |
| 654 | layer.register_parameter("w13_bias", w13_bias) |
| 655 | set_weight_attrs(w13_bias, extra_weight_attrs) |
| 656 | set_weight_attrs(w13_bias, {"weight_loader": weight_loader}) |
| 657 | |
| 658 | w2_bias = torch.nn.Parameter( |
| 659 | torch.zeros( |
| 660 | num_experts, |
| 661 | hidden_size, |
| 662 | dtype=torch.bfloat16, |
| 663 | ), |
| 664 | requires_grad=False, |
| 665 | ) |
| 666 | layer.register_parameter("w2_bias", w2_bias) |
| 667 | set_weight_attrs(w2_bias, extra_weight_attrs) |
| 668 | set_weight_attrs(w2_bias, {"weight_loader": weight_loader}) |
| 669 | |
| 670 | def _setup_kernel( |
| 671 | self, |
| 672 | layer: RoutedExperts, |
| 673 | w13: torch.Tensor, |
| 674 | w2: torch.Tensor, |
| 675 | w13_scale: torch.Tensor, |
| 676 | w2_scale: torch.Tensor, |
| 677 | w13_bias: torch.Tensor | None = None, |
| 678 | w2_bias: torch.Tensor | None = None, |
| 679 | ) -> None: |
| 680 | num_experts = self.num_experts |
| 681 | intermediate_size = self.intermediate_size |
| 682 | hidden_size = self.hidden_size |
| 683 | sf_block_size = 32 |
| 684 | |
| 685 | # Shape assertions — skipped for SITU since its kernel handles native |
| 686 | # (non-256-aligned) intermediate sizes without prior round-up. |
| 687 | from vllm.model_executor.layers.fused_moe.activation import MoEActivation |
| 688 | |
| 689 | if self.moe.activation != MoEActivation.SITU: |
| 690 | assert ( |
| 691 | w13.dim() == 3 |
| 692 | and w13.shape[0] == num_experts |
| 693 | and w13.shape[1] == intermediate_size * self.moe.w13_num_shards |
| 694 | and w13.shape[2] == hidden_size // 2 |
| 695 | ) |
| 696 | assert ( |
| 697 | w13_scale.dim() == 3 |
| 698 | and w13_scale.shape[0] == num_experts |
| 699 | and w13_scale.shape[1] == intermediate_size * self.moe.w13_num_shards |
| 700 | and w13_scale.shape[2] == hidden_size // sf_block_size |
| 701 | ) |
| 702 | assert ( |
| 703 | w2.dim() == 3 |
| 704 | and w2.shape[0] == num_experts |
| 705 | and w2.shape[1] == hidden_size |
| 706 | and w2.shape[2] == intermediate_size // 2 |
| 707 | ) |
| 708 | assert ( |
| 709 | w2_scale.dim() == 3 |
| 710 | and w2_scale.shape[1] == hidden_size |
| 711 | and w2_scale.shape[2] == intermediate_size // sf_block_size |
| 712 | ) |
| 713 | if w13_bias is not None: |
| 714 | assert ( |
| 715 | w13_bias.dim() == 2 |
| 716 | and w13_bias.shape[0] == num_experts |
| 717 | and w13_bias.shape[1] == intermediate_size * self.moe.w13_num_shards |
| 718 | ) |
| 719 | if w2_bias is not None: |
| 720 | assert ( |
| 721 | w2_bias.dim() == 2 |
| 722 | and w2_bias.shape[0] == num_experts |
| 723 | and w2_bias.shape[1] == hidden_size |
| 724 | ) |
| 725 | |
| 726 | # Convert weights to kernel format |
| 727 | w13, w2, w13_scale, w2_scale, w13_bias, w2_bias = ( |
| 728 | convert_weight_to_mxfp4_moe_kernel_format( |
| 729 | mxfp4_backend=self.mxfp4_backend, |
| 730 | layer=layer, |
| 731 | w13_weight=w13, |
| 732 | w2_weight=w2, |
| 733 | w13_weight_scale=w13_scale, |
| 734 | w2_weight_scale=w2_scale, |
| 735 | w13_bias=w13_bias, |
| 736 | w2_bias=w2_bias, |
| 737 | _cache_permute_indices=self._cache_permute_indices, |
| 738 | activation=self.moe.activation, |
| 739 | use_separated_a4w4=self.moe.use_mxfp4_w4a4_dsv4, |
| 740 | ) |
| 741 | ) |
| 742 | |
| 743 | # For TRITON backends, weights are wrapped tensors from triton_kernels |
| 744 | # that don't support .detach(). Manually assign parameters. |
| 745 | is_gfx1250 = False |
| 746 | if current_platform.is_rocm(): |
| 747 | from vllm.platforms.rocm import on_gfx1250 |
| 748 | |
| 749 | is_gfx1250 = on_gfx1250() |
| 750 | |
| 751 | uses_triton_weight_format = self.mxfp4_backend in TRITON_BACKENDS or ( |
| 752 | self.mxfp4_backend == Mxfp4MoeBackend.AITER_MXFP4_BF16 and is_gfx1250 |
| 753 | ) |
| 754 | if not uses_triton_weight_format: |
| 755 | replace_parameter(layer, "w13_weight", w13) |
| 756 | replace_parameter(layer, "w2_weight", w2) |
| 757 | replace_parameter(layer, "w13_weight_scale", w13_scale) |
| 758 | replace_parameter(layer, "w2_weight_scale", w2_scale) |
| 759 | else: |
| 760 | layer.w13_weight = w13 |
| 761 | layer.w2_weight = w2 |
| 762 | self.w13_precision_config = w13_scale |
| 763 | self.w2_precision_config = w2_scale |
| 764 | |
| 765 | if w13_bias is not None and w2_bias is not None: |
| 766 | replace_parameter(layer, "w13_bias", w13_bias) |
| 767 | replace_parameter(layer, "w2_bias", w2_bias) |
| 768 | |
| 769 | # Build quant config |
| 770 | self._build_moe_kernel(layer) |
| 771 | |
| 772 | def _build_moe_kernel(self, layer: RoutedExperts) -> None: |
| 773 | """Build the modular MoE kernel from the (already in-format) weights.""" |
| 774 | self.moe_quant_config = self.get_fused_moe_quant_config(layer) |
| 775 | if self.moe_quant_config is not None and self.experts_cls is not None: |
| 776 | self.moe_kernel = make_mxfp4_moe_kernel( |
| 777 | moe_quant_config=self.moe_quant_config, |
| 778 | moe_config=self.moe, |
| 779 | mxfp4_backend=self.mxfp4_backend, |
| 780 | experts_cls=self.experts_cls, |
| 781 | routing_tables=layer._expert_routing_tables(), |
| 782 | ) |
| 783 | self.moe_kernel.fused_experts.process_weights_after_loading(layer) |
| 784 | |
| 785 | def process_weights_after_loading(self, layer): |
| 786 | if self.mxfp4_backend == Mxfp4MoeBackend.NONE: |
| 787 | return |
| 788 | |
| 789 | if is_weights_pre_processed(): |
| 790 | if self.mxfp4_backend != Mxfp4MoeBackend.FLASHINFER_TRTLLM_MXFP4_MXFP8: |
| 791 | raise RuntimeError( |
| 792 | "pre-processed weights require FLASHINFER_TRTLLM_MXFP4_MXFP8 " |
| 793 | f"moe backend, got {self.mxfp4_backend}" |
| 794 | ) |
| 795 | self._build_moe_kernel(layer) |
| 796 | return |
| 797 | |
| 798 | w13 = layer.w13_weight |
| 799 | w2 = layer.w2_weight |
| 800 | w13_scale = layer.w13_weight_scale |
| 801 | w2_scale = layer.w2_weight_scale |
| 802 | w13_bias = getattr(layer, "w13_bias", None) |
| 803 | w2_bias = getattr(layer, "w2_bias", None) |
| 804 | |
| 805 | self._setup_kernel(layer, w13, w2, w13_scale, w2_scale, w13_bias, w2_bias) |
| 806 | |
| 807 | def get_fused_moe_quant_config( |
| 808 | self, |
| 809 | layer: RoutedExperts, |
| 810 | ) -> FusedMoEQuantConfig | None: |
| 811 | w1_bias = getattr(layer, "w13_bias", None) |
| 812 | w2_bias = getattr(layer, "w2_bias", None) |
| 813 | swiglu_limit = getattr(layer, "swiglu_limit", None) |
| 814 | |
| 815 | is_gfx1250 = False |
| 816 | if current_platform.is_rocm(): |
| 817 | from vllm.platforms.rocm import on_gfx1250 |
| 818 | |
| 819 | is_gfx1250 = on_gfx1250() |
| 820 | |
| 821 | if self.mxfp4_backend in TRITON_BACKENDS or ( |
| 822 | self.mxfp4_backend == Mxfp4MoeBackend.AITER_MXFP4_BF16 and is_gfx1250 |
| 823 | ): |
| 824 | # TRITON backends free w13/w2_weight_scale after swizzling; the |
| 825 | # swizzled scales live inside the precision configs instead. |
| 826 | assert self.w13_precision_config is not None |
| 827 | assert self.w2_precision_config is not None |
| 828 | w1_scale = self.w13_precision_config |
| 829 | w2_scale = self.w2_precision_config |
| 830 | else: |
| 831 | w1_scale = layer.w13_weight_scale |
| 832 | w2_scale = layer.w2_weight_scale |
| 833 | |
| 834 | if self.mxfp4_backend == Mxfp4MoeBackend.EMULATION: |
| 835 | # Canonical ``mxfp4`` checkpoints are weight-only W4A16. The |
| 836 | # generic EMULATION config is W4A4, so preserve BF16 activations |
| 837 | # while the fallback dequantizes only the weights. |
| 838 | return mxfp4_w4a16_moe_quant_config( |
| 839 | w1_scale=w1_scale, |
| 840 | w2_scale=w2_scale, |
| 841 | w1_bias=w1_bias, |
| 842 | w2_bias=w2_bias, |
| 843 | gemm1_clamp_limit=swiglu_limit, |
| 844 | ) |
| 845 | |
| 846 | return make_mxfp4_moe_quant_config( |
| 847 | mxfp4_backend=self.mxfp4_backend, |
| 848 | w1_scale=w1_scale, |
| 849 | w2_scale=w2_scale, |
| 850 | w1_bias=w1_bias, |
| 851 | w2_bias=w2_bias, |
| 852 | swiglu_limit=swiglu_limit, |
| 853 | layer=layer, |
| 854 | ) |
| 855 | |
| 856 | def apply( |
| 857 | self, |
| 858 | layer: RoutedExperts, |
| 859 | x: torch.Tensor, |
| 860 | topk_weights: torch.Tensor, |
| 861 | topk_ids: torch.Tensor, |
| 862 | shared_experts: SharedExperts | None, |
| 863 | shared_experts_input: torch.Tensor | None, |
| 864 | ) -> torch.Tensor | UnfinalizedMoEOutput: |
| 865 | assert not self.is_monolithic |
| 866 | assert self.moe_kernel is not None |
| 867 | return self.moe_kernel.apply( |
| 868 | hidden_states=x, |
| 869 | w1=layer.w13_weight, |
| 870 | w2=layer.w2_weight, |
| 871 | topk_weights=topk_weights, |
| 872 | topk_ids=topk_ids, |
| 873 | activation=layer.activation, |
| 874 | global_num_experts=layer.global_num_experts, |
| 875 | apply_router_weight_on_input=layer.apply_router_weight_on_input, |
| 876 | expert_map=layer.expert_map, |
| 877 | shared_experts=shared_experts, |
| 878 | shared_experts_input=shared_experts_input, |
| 879 | ) |
| 880 | |
| 881 | def apply_monolithic( |
| 882 | self, |
| 883 | layer: RoutedExperts, |
| 884 | x: torch.Tensor, |
| 885 | router_logits: torch.Tensor, |
| 886 | input_ids: torch.Tensor | None = None, |
| 887 | ) -> torch.Tensor | UnfinalizedMoEOutput: |
| 888 | assert self.is_monolithic |
| 889 | assert self.moe_kernel is not None |
| 890 | return self.moe_kernel.apply_monolithic( |
| 891 | hidden_states=x, |
| 892 | w1=layer.w13_weight, |
| 893 | w2=layer.w2_weight, |
| 894 | router_logits=router_logits, |
| 895 | activation=layer.activation, |
| 896 | global_num_experts=layer.global_num_experts, |
| 897 | expert_map=layer.expert_map, |
| 898 | apply_router_weight_on_input=layer.apply_router_weight_on_input, |
| 899 | num_expert_group=layer.num_expert_group, |
| 900 | topk_group=layer.topk_group, |
| 901 | e_score_correction_bias=layer.e_score_correction_bias, |
| 902 | routed_scaling_factor=layer.routed_scaling_factor, |
| 903 | ) |