| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import typing |
| 4 | from collections.abc import Callable, Iterable |
| 5 | from inspect import signature |
| 6 | from itertools import islice |
| 7 | |
| 8 | import regex as re |
| 9 | import torch |
| 10 | import torch.nn as nn |
| 11 | |
| 12 | import vllm.envs as envs |
| 13 | from vllm.config import VllmConfig |
| 14 | from vllm.config.kernel import MEGA_MOE_BACKENDS |
| 15 | from vllm.distributed import ( |
| 16 | get_ep_group, |
| 17 | get_pp_group, |
| 18 | get_tensor_model_parallel_rank, |
| 19 | get_tensor_model_parallel_world_size, |
| 20 | ) |
| 21 | from vllm.distributed.eplb.eplb_state import EplbLayerState |
| 22 | from vllm.forward_context import get_forward_context, is_forward_context_available |
| 23 | from vllm.logger import init_logger |
| 24 | from vllm.model_executor.kernels.mhc.tilelang import ( |
| 25 | hc_head_fused_kernel_tilelang, |
| 26 | mhc_fused_post_pre_tilelang, |
| 27 | mhc_post_tilelang, |
| 28 | mhc_pre_broadcast_tilelang, |
| 29 | mhc_pre_tilelang, |
| 30 | ) |
| 31 | from vllm.model_executor.layers.activation import SiluAndMul, SiluAndMulWithClamp |
| 32 | from vllm.model_executor.layers.fused_moe import ( |
| 33 | FusedMoEFactory, |
| 34 | fused_moe_make_expert_params_mapping, |
| 35 | ) |
| 36 | from vllm.model_executor.layers.fused_moe.router.base_router import ( |
| 37 | eplb_map_to_physical_and_record, |
| 38 | ) |
| 39 | from vllm.model_executor.layers.fused_moe.router.fused_topk_bias_router import ( |
| 40 | fused_topk_bias, |
| 41 | ) |
| 42 | from vllm.model_executor.layers.fused_moe.router.gate_linear import GateLinear |
| 43 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 44 | from vllm.model_executor.layers.linear import ( |
| 45 | MergedColumnParallelLinear, |
| 46 | RowParallelLinear, |
| 47 | ) |
| 48 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 49 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 50 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 51 | ParallelLMHead, |
| 52 | VocabParallelEmbedding, |
| 53 | ) |
| 54 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 55 | from vllm.model_executor.models.interfaces import ( |
| 56 | EagleModelMixin, |
| 57 | MixtureOfExperts, |
| 58 | SupportsEagle3, |
| 59 | SupportsLoRA, |
| 60 | SupportsPP, |
| 61 | ) |
| 62 | from vllm.model_executor.models.utils import ( |
| 63 | AutoWeightsLoader, |
| 64 | PPMissingLayer, |
| 65 | WeightsMapper, |
| 66 | extract_layer_index, |
| 67 | is_pp_missing_parameter, |
| 68 | make_layers, |
| 69 | maybe_prefix, |
| 70 | spec_decode_needs_target_embed, |
| 71 | ) |
| 72 | from vllm.model_executor.utils import set_weight_attrs |
| 73 | from vllm.models.common.ops.sequence_parallel import ( |
| 74 | sp_all_gather, |
| 75 | sp_padding_mask, |
| 76 | sp_reduce_scatter, |
| 77 | sp_shard, |
| 78 | ) |
| 79 | from vllm.models.deepseek_v4.attention import DeepseekV4Attention |
| 80 | from vllm.models.deepseek_v4.nvidia.flashinfer_sparse import ( |
| 81 | DeepseekV4FlashInferMLAAttention, |
| 82 | DeepseekV4FlashInferSM120Attention, |
| 83 | ) |
| 84 | from vllm.models.deepseek_v4.nvidia.flashmla import DeepseekV4FlashMLAAttention |
| 85 | from vllm.models.deepseek_v4.nvidia.ops.prepare_megamoe import prepare_megamoe_inputs |
| 86 | from vllm.platforms import current_platform |
| 87 | from vllm.sequence import IntermediateTensors |
| 88 | from vllm.utils.flashinfer_moe_ep import ( |
| 89 | is_fi_moe_ep_backend, |
| 90 | validate_fi_moe_ep_config, |
| 91 | ) |
| 92 | from vllm.utils.math_utils import cdiv |
| 93 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 94 | from vllm.v1.worker.ubatching import dbo_current_ubatch_id |
| 95 | |
| 96 | from ..common.mm_preprocess import IMAGE_SENTINEL_BASE_ID |
| 97 | |
| 98 | logger = init_logger(__name__) |
| 99 | |
| 100 | |
| 101 | class MegaGateRoutingMetadata(typing.NamedTuple): |
| 102 | safe_hash_input_ids: torch.Tensor | None |
| 103 | image_token_mask: torch.Tensor | None |
| 104 | hash_token_mask: torch.Tensor | None |
| 105 | |
| 106 | |
| 107 | def prepare_mega_gate_routing_metadata( |
| 108 | input_ids: torch.Tensor, |
| 109 | *, |
| 110 | has_hash_routing: bool, |
| 111 | image_sentinel_base_id: int | None, |
| 112 | ) -> MegaGateRoutingMetadata: |
| 113 | if image_sentinel_base_id is None: |
| 114 | return MegaGateRoutingMetadata( |
| 115 | input_ids if has_hash_routing else None, |
| 116 | None, |
| 117 | torch.ones_like(input_ids, dtype=torch.bool) if has_hash_routing else None, |
| 118 | ) |
| 119 | |
| 120 | image_token_mask = (input_ids >= image_sentinel_base_id) & ( |
| 121 | input_ids < image_sentinel_base_id + 5 |
| 122 | ) |
| 123 | if not has_hash_routing: |
| 124 | return MegaGateRoutingMetadata(None, image_token_mask, None) |
| 125 | return MegaGateRoutingMetadata( |
| 126 | torch.where(image_token_mask, 0, input_ids), |
| 127 | image_token_mask, |
| 128 | ~image_token_mask, |
| 129 | ) |
| 130 | |
| 131 | |
| 132 | class DeepseekV4MLP(nn.Module): |
| 133 | def __init__( |
| 134 | self, |
| 135 | hidden_size: int, |
| 136 | intermediate_size: int, |
| 137 | hidden_act: str, |
| 138 | swiglu_limit: float | None = None, |
| 139 | quant_config: QuantizationConfig | None = None, |
| 140 | reduce_results: bool = True, |
| 141 | is_sequence_parallel: bool = False, |
| 142 | prefix: str = "", |
| 143 | ) -> None: |
| 144 | super().__init__() |
| 145 | |
| 146 | # If is_sequence_parallel, the input and output tensors are sharded |
| 147 | # across the ranks within the tp_group. In this case the weights are |
| 148 | # replicated and no collective ops are needed. |
| 149 | # Otherwise we use standard TP with an allreduce at the end. |
| 150 | # |
| 151 | # Block-FP8 shards in whole 128-blocks; cdiv rounds the per-rank block |
| 152 | # count up so the linear's even TP split stays block-aligned, with the |
| 153 | # trailing ranks zero-filled by load_weights. |
| 154 | block_size = getattr(quant_config, "weight_block_size", None) |
| 155 | if block_size is not None and not is_sequence_parallel: |
| 156 | tp_size = get_tensor_model_parallel_world_size() |
| 157 | n_local = cdiv(intermediate_size // block_size[0], tp_size) |
| 158 | intermediate_size = n_local * block_size[0] * tp_size |
| 159 | self.gate_up_proj = MergedColumnParallelLinear( |
| 160 | hidden_size, |
| 161 | [intermediate_size] * 2, |
| 162 | bias=False, |
| 163 | quant_config=quant_config, |
| 164 | disable_tp=is_sequence_parallel, |
| 165 | prefix=f"{prefix}.gate_up_proj", |
| 166 | ) |
| 167 | self.down_proj = RowParallelLinear( |
| 168 | intermediate_size, |
| 169 | hidden_size, |
| 170 | bias=False, |
| 171 | quant_config=quant_config, |
| 172 | reduce_results=reduce_results, |
| 173 | disable_tp=is_sequence_parallel, |
| 174 | prefix=f"{prefix}.down_proj", |
| 175 | ) |
| 176 | if hidden_act != "silu": |
| 177 | raise ValueError( |
| 178 | f"Unsupported activation: {hidden_act}. Only silu is supported for now." |
| 179 | ) |
| 180 | if swiglu_limit is not None: |
| 181 | self.act_fn = SiluAndMulWithClamp(swiglu_limit) |
| 182 | else: |
| 183 | self.act_fn = SiluAndMul() |
| 184 | |
| 185 | def forward(self, x): |
| 186 | gate_up, _ = self.gate_up_proj(x) |
| 187 | x = self.act_fn(gate_up) |
| 188 | x, _ = self.down_proj(x) |
| 189 | return x |
| 190 | |
| 191 | |
| 192 | def make_deepseek_v4_expert_params_mapping( |
| 193 | num_experts: int, |
| 194 | ) -> list[tuple[str, str, int, str]]: |
| 195 | return [ |
| 196 | ( |
| 197 | "experts.w13_" if shard_id in ("w1", "w3") else "experts.w2_", |
| 198 | f"experts.{expert_id}.{weight_name}.", |
| 199 | expert_id, |
| 200 | shard_id, |
| 201 | ) |
| 202 | for expert_id in range(num_experts) |
| 203 | for shard_id, weight_name in [ |
| 204 | ("w1", "w1"), |
| 205 | ("w2", "w2"), |
| 206 | ("w3", "w3"), |
| 207 | ] |
| 208 | ] |
| 209 | |
| 210 | |
| 211 | class DeepseekV4MegaMoEExperts(nn.Module): |
| 212 | _symm_buffer_cache: dict[tuple[int, int, int, int, int, int, int, int], object] = {} |
| 213 | |
| 214 | def __init__( |
| 215 | self, |
| 216 | vllm_config: VllmConfig, |
| 217 | *, |
| 218 | num_experts: int, |
| 219 | num_local_experts: int, |
| 220 | experts_start_idx: int, |
| 221 | top_k: int, |
| 222 | hidden_size: int, |
| 223 | intermediate_size: int, |
| 224 | num_shared_experts: int = 0, |
| 225 | prefix: str = "", |
| 226 | num_logical_experts: int | None = None, |
| 227 | ): |
| 228 | super().__init__() |
| 229 | self.prefix = prefix |
| 230 | self.capture_fn: Callable[[torch.Tensor], None] | None = None |
| 231 | self.num_experts = num_experts |
| 232 | self.num_local_experts = num_local_experts |
| 233 | self.experts_start_idx = experts_start_idx |
| 234 | self.experts_end_idx = experts_start_idx + num_local_experts |
| 235 | self.top_k = top_k |
| 236 | self.hidden_size = hidden_size |
| 237 | self.intermediate_size = intermediate_size |
| 238 | self.num_shared_experts = num_shared_experts |
| 239 | self.max_num_tokens = vllm_config.scheduler_config.max_num_batched_tokens |
| 240 | |
| 241 | self.num_logical_experts = ( |
| 242 | num_logical_experts if num_logical_experts is not None else num_experts |
| 243 | ) |
| 244 | |
| 245 | self.eplb_state = EplbLayerState() |
| 246 | |
| 247 | weight_attrs = {"weight_loader": self.weight_loader} |
| 248 | self.w13_weight = nn.Parameter( |
| 249 | torch.zeros( |
| 250 | num_local_experts, |
| 251 | 2 * intermediate_size, |
| 252 | hidden_size // 2, |
| 253 | dtype=torch.uint8, |
| 254 | ), |
| 255 | requires_grad=False, |
| 256 | ) |
| 257 | set_weight_attrs(self.w13_weight, weight_attrs) |
| 258 | |
| 259 | self.w13_weight_scale = nn.Parameter( |
| 260 | torch.zeros( |
| 261 | num_local_experts, |
| 262 | 2 * intermediate_size, |
| 263 | hidden_size // 32, |
| 264 | dtype=torch.uint8, |
| 265 | ), |
| 266 | requires_grad=False, |
| 267 | ) |
| 268 | set_weight_attrs(self.w13_weight_scale, weight_attrs) |
| 269 | self.w13_weight_scale.quant_method = "block" |
| 270 | |
| 271 | self.w2_weight = nn.Parameter( |
| 272 | torch.zeros( |
| 273 | num_local_experts, |
| 274 | hidden_size, |
| 275 | intermediate_size // 2, |
| 276 | dtype=torch.uint8, |
| 277 | ), |
| 278 | requires_grad=False, |
| 279 | ) |
| 280 | set_weight_attrs(self.w2_weight, weight_attrs) |
| 281 | |
| 282 | self.w2_weight_scale = nn.Parameter( |
| 283 | torch.zeros( |
| 284 | num_local_experts, |
| 285 | hidden_size, |
| 286 | intermediate_size // 32, |
| 287 | dtype=torch.uint8, |
| 288 | ), |
| 289 | requires_grad=False, |
| 290 | ) |
| 291 | set_weight_attrs(self.w2_weight_scale, weight_attrs) |
| 292 | self.w2_weight_scale.quant_method = "block" |
| 293 | |
| 294 | self._transformed_l1_weights: tuple[torch.Tensor, torch.Tensor] | None = None |
| 295 | self._transformed_l2_weights: tuple[torch.Tensor, torch.Tensor] | None = None |
| 296 | self._transformed_shared_l1_weights: ( |
| 297 | tuple[torch.Tensor, torch.Tensor] | None |
| 298 | ) = None |
| 299 | self._transformed_shared_l2_weights: ( |
| 300 | tuple[torch.Tensor, torch.Tensor] | None |
| 301 | ) = None |
| 302 | |
| 303 | # Register in the static forward context so the custom-op wrapper |
| 304 | # can look up this module by name from within a torch.compile graph. |
| 305 | compilation_config = vllm_config.compilation_config |
| 306 | if prefix in compilation_config.static_forward_context: |
| 307 | raise ValueError(f"Duplicate layer name: {prefix}") |
| 308 | compilation_config.static_forward_context[prefix] = self |
| 309 | |
| 310 | def _map_global_expert_id(self, expert_id: int) -> list[int]: |
| 311 | """Return local (per-rank) slot offsets where logical expert |
| 312 | `expert_id` should land on this rank. |
| 313 | """ |
| 314 | physical_ids: list[int] = [] |
| 315 | for p in range(self.experts_start_idx, self.experts_end_idx): |
| 316 | if p % self.num_logical_experts == expert_id: |
| 317 | physical_ids.append(p - self.experts_start_idx) |
| 318 | return physical_ids |
| 319 | |
| 320 | def weight_loader( |
| 321 | self, |
| 322 | param: nn.Parameter, |
| 323 | loaded_weight: torch.Tensor, |
| 324 | weight_name: str, |
| 325 | shard_id: str, |
| 326 | expert_id: int, |
| 327 | return_success: bool = False, |
| 328 | ) -> bool | None: |
| 329 | local_expert_ids = self._map_global_expert_id(expert_id) |
| 330 | if not local_expert_ids: |
| 331 | return False if return_success else None |
| 332 | |
| 333 | loaded_any = False |
| 334 | for local_expert_id in local_expert_ids: |
| 335 | expert_data = param.data[local_expert_id] |
| 336 | if shard_id in ("w1", "w3"): |
| 337 | if "w13_" not in weight_name: |
| 338 | continue |
| 339 | shard_offset = 0 if shard_id == "w1" else self.intermediate_size |
| 340 | expert_data = expert_data.narrow( |
| 341 | 0, shard_offset, self.intermediate_size |
| 342 | ) |
| 343 | elif shard_id == "w2": |
| 344 | if "w2_" not in weight_name: |
| 345 | continue |
| 346 | else: |
| 347 | raise ValueError(f"Unsupported expert shard id: {shard_id}") |
| 348 | |
| 349 | if expert_data.shape != loaded_weight.shape: |
| 350 | raise ValueError( |
| 351 | f"DeepSeek V4 MegaMoE expert weight shape mismatch for " |
| 352 | f"{weight_name}: parameter shard {tuple(expert_data.shape)} " |
| 353 | f"vs checkpoint {tuple(loaded_weight.shape)}" |
| 354 | ) |
| 355 | expert_data.copy_(loaded_weight) |
| 356 | loaded_any = True |
| 357 | |
| 358 | if return_success: |
| 359 | return loaded_any |
| 360 | return None |
| 361 | |
| 362 | @staticmethod |
| 363 | def _ue8m0_uint8_to_float(sf: torch.Tensor) -> torch.Tensor: |
| 364 | return (sf.to(torch.int32) << 23).view(torch.float32) |
| 365 | |
| 366 | def _check_runtime_supported(self) -> None: |
| 367 | device = self.w13_weight.device |
| 368 | if torch.cuda.get_device_capability(device)[0] != 10: |
| 369 | raise NotImplementedError("DeepGEMM MegaMoE requires SM100 GPUs.") |
| 370 | if self.hidden_size % 128 != 0 or self.intermediate_size % 128 != 0: |
| 371 | raise ValueError( |
| 372 | "DeepGEMM MegaMoE requires hidden and intermediate sizes " |
| 373 | "to be multiples of 128." |
| 374 | ) |
| 375 | |
| 376 | @staticmethod |
| 377 | def _deep_gemm_supports_shared_experts(deep_gemm) -> bool: |
| 378 | """Check the Python API before touching a symmetric-memory group. |
| 379 | |
| 380 | This also gives users of an older precompiled vLLM wheel a safe serial |
| 381 | fallback instead of failing halfway through multi-rank buffer setup. |
| 382 | """ |
| 383 | try: |
| 384 | buffer_params = signature(deep_gemm.get_symm_buffer_for_mega_moe).parameters |
| 385 | kernel_params = signature(deep_gemm.fp8_fp4_mega_moe).parameters |
| 386 | except (TypeError, ValueError): |
| 387 | return False |
| 388 | return ( |
| 389 | hasattr(deep_gemm, "get_block_m_for_mega_moe") |
| 390 | and hasattr(deep_gemm, "transform_weights_for_mega_moe") |
| 391 | and "num_shared_experts" in buffer_params |
| 392 | and "shared_l1_weights" in kernel_params |
| 393 | and "shared_l2_weights" in kernel_params |
| 394 | ) |
| 395 | |
| 396 | def _finalize_shared_expert_weights( |
| 397 | self, deep_gemm, shared_experts: DeepseekV4MLP |
| 398 | ) -> None: |
| 399 | gate_up = shared_experts.gate_up_proj |
| 400 | down = shared_experts.down_proj |
| 401 | gate_up_weight = gate_up.weight.data |
| 402 | gate_up_scale = ( |
| 403 | gate_up.weight_scale |
| 404 | if hasattr(gate_up, "weight_scale") |
| 405 | else gate_up.weight_scale_inv |
| 406 | ).data |
| 407 | down_weight = down.weight.data |
| 408 | down_scale = ( |
| 409 | down.weight_scale |
| 410 | if hasattr(down, "weight_scale") |
| 411 | else down.weight_scale_inv |
| 412 | ).data |
| 413 | |
| 414 | # MegaMoE's shared FP8 MMA consumes a 1x32 scale for every weight row, |
| 415 | # while the checkpoint uses coarser block-FP8 scales (usually |
| 416 | # 128x128). Build a dedicated, numerically equivalent scale view before |
| 417 | # the generic linear post-load hook replaces the raw checkpoint scales |
| 418 | # with its 128x128 DeepGEMM layout. |
| 419 | checkpoint_scale_dtypes = (torch.float8_e8m0fnu, torch.uint8) |
| 420 | if ( |
| 421 | gate_up_scale.dtype in checkpoint_scale_dtypes |
| 422 | and down_scale.dtype in checkpoint_scale_dtypes |
| 423 | ): |
| 424 | gate_up_scale = self._prepare_shared_expert_scale( |
| 425 | deep_gemm, |
| 426 | gate_up, |
| 427 | gate_up_scale, |
| 428 | gate_up_weight.shape[0], |
| 429 | gate_up_weight.shape[1], |
| 430 | ) |
| 431 | down_scale = self._prepare_shared_expert_scale( |
| 432 | deep_gemm, |
| 433 | down, |
| 434 | down_scale, |
| 435 | down_weight.shape[0], |
| 436 | down_weight.shape[1], |
| 437 | ) |
| 438 | |
| 439 | if gate_up_scale is None or down_scale is None: |
| 440 | self.num_shared_experts = 0 |
| 441 | return |
| 442 | |
| 443 | shared_intermediate_size = self.intermediate_size * self.num_shared_experts |
| 444 | expected_gate_up_shape = ( |
| 445 | 2 * shared_intermediate_size, |
| 446 | self.hidden_size, |
| 447 | ) |
| 448 | expected_down_shape = (self.hidden_size, shared_intermediate_size) |
| 449 | if ( |
| 450 | gate_up_weight.dtype != torch.float8_e4m3fn |
| 451 | or down_weight.dtype != torch.float8_e4m3fn |
| 452 | or gate_up_scale.dtype != torch.int32 |
| 453 | or down_scale.dtype != torch.int32 |
| 454 | or tuple(gate_up_weight.shape) != expected_gate_up_shape |
| 455 | or tuple(down_weight.shape) != expected_down_shape |
| 456 | ): |
| 457 | logger.warning( |
| 458 | "Disabling native MegaMoE shared-expert fusion for %s: expected " |
| 459 | "replicated block-FP8 weights with gate_up=%s, down=%s, and " |
| 460 | "DeepGEMM int32 scales; got gate_up=%s/%s/%s and down=%s/%s/%s.", |
| 461 | self.prefix, |
| 462 | expected_gate_up_shape, |
| 463 | expected_down_shape, |
| 464 | tuple(gate_up_weight.shape), |
| 465 | gate_up_weight.dtype, |
| 466 | gate_up_scale.dtype, |
| 467 | tuple(down_weight.shape), |
| 468 | down_weight.dtype, |
| 469 | down_scale.dtype, |
| 470 | ) |
| 471 | self.num_shared_experts = 0 |
| 472 | return |
| 473 | |
| 474 | transformed_l1, transformed_l2 = deep_gemm.transform_weights_for_mega_moe( |
| 475 | (gate_up_weight, gate_up_scale), |
| 476 | (down_weight, down_scale), |
| 477 | ) |
| 478 | # L1 interleaving allocates a full copy. Re-home the loader Parameter on |
| 479 | # that storage so the original 2*intermediate*hidden FP8 tensor can be |
| 480 | # released instead of adding roughly 0.7 GiB per rank on DSV4-Flash. |
| 481 | # The generic linear post-load hook may still repack the serial scales, |
| 482 | # but this shared MLP is never called after native fusion is enabled. |
| 483 | gate_up.weight.data = transformed_l1[0] |
| 484 | self._transformed_shared_l1_weights = ( |
| 485 | gate_up.weight.data, |
| 486 | transformed_l1[1], |
| 487 | ) |
| 488 | self._transformed_shared_l2_weights = transformed_l2 |
| 489 | |
| 490 | def _prepare_shared_expert_scale( |
| 491 | self, |
| 492 | deep_gemm, |
| 493 | linear: nn.Module, |
| 494 | scale: torch.Tensor, |
| 495 | mn: int, |
| 496 | k: int, |
| 497 | ) -> torch.Tensor | None: |
| 498 | block_size = getattr(linear, "weight_block_size", None) |
| 499 | if block_size is None or len(block_size) != 2: |
| 500 | logger.warning( |
| 501 | "Disabling native MegaMoE shared-expert fusion for %s: " |
| 502 | "shared FP8 weight block size is unavailable.", |
| 503 | self.prefix, |
| 504 | ) |
| 505 | return None |
| 506 | |
| 507 | block_m, block_k = block_size |
| 508 | expected_shape = ( |
| 509 | (mn + block_m - 1) // block_m, |
| 510 | (k + block_k - 1) // block_k, |
| 511 | ) |
| 512 | if block_k % 32 != 0 or tuple(scale.shape) != expected_shape: |
| 513 | logger.warning( |
| 514 | "Disabling native MegaMoE shared-expert fusion for %s: " |
| 515 | "cannot convert shared scale shape %s with block size %s " |
| 516 | "to MegaMoE's 1x32 layout for weight (%d, %d).", |
| 517 | self.prefix, |
| 518 | tuple(scale.shape), |
| 519 | tuple(block_size), |
| 520 | mn, |
| 521 | k, |
| 522 | ) |
| 523 | return None |
| 524 | |
| 525 | scale_fp32 = self._ue8m0_uint8_to_float(scale.view(torch.uint8)) |
| 526 | scale_1x32 = ( |
| 527 | scale_fp32.repeat_interleave(block_m, dim=0) |
| 528 | .repeat_interleave(block_k // 32, dim=1)[:mn, : k // 32] |
| 529 | .contiguous() |
| 530 | ) |
| 531 | # The grouped API is used with a singleton dimension to request the |
| 532 | # MN-major, TMA-aligned packed-UE8M0 strides, then squeezed back to the |
| 533 | # 2D layout required for a shared expert. |
| 534 | return deep_gemm.transform_sf_into_required_layout( |
| 535 | scale_1x32.unsqueeze(0), |
| 536 | mn, |
| 537 | k, |
| 538 | (1, 32), |
| 539 | 1, |
| 540 | ).squeeze(0) |
| 541 | |
| 542 | def finalize_weights(self, shared_experts: DeepseekV4MLP | None = None) -> None: |
| 543 | from vllm.utils.deep_gemm import _import_deep_gemm |
| 544 | |
| 545 | deep_gemm = _import_deep_gemm() |
| 546 | |
| 547 | if self._transformed_l1_weights is None: |
| 548 | self._check_runtime_supported() |
| 549 | w13_scale = deep_gemm.transform_sf_into_required_layout( |
| 550 | self._ue8m0_uint8_to_float(self.w13_weight_scale.data).contiguous(), |
| 551 | 2 * self.intermediate_size, |
| 552 | self.hidden_size, |
| 553 | (1, 32), |
| 554 | self.num_local_experts, |
| 555 | ) |
| 556 | w2_scale = deep_gemm.transform_sf_into_required_layout( |
| 557 | self._ue8m0_uint8_to_float(self.w2_weight_scale.data).contiguous(), |
| 558 | self.hidden_size, |
| 559 | self.intermediate_size, |
| 560 | (1, 32), |
| 561 | self.num_local_experts, |
| 562 | ) |
| 563 | self._transformed_l1_weights, self._transformed_l2_weights = ( |
| 564 | deep_gemm.transform_weights_for_mega_moe( |
| 565 | (self.w13_weight.data.view(torch.int8).contiguous(), w13_scale), |
| 566 | (self.w2_weight.data.view(torch.int8).contiguous(), w2_scale), |
| 567 | ) |
| 568 | ) |
| 569 | # Drop the original loader-side parameters: the MegaMoE kernels only |
| 570 | # consume the transformed views above. transform_weights_for_mega_moe |
| 571 | # allocates a fresh tensor for the L1 weight (see |
| 572 | # _interleave_l1_weights) and fresh SF tensors for L1/L2; the L2 |
| 573 | # weight is the only tensor that aliases the original storage, and |
| 574 | # _transformed_l2_weights still holds it, so the storage stays live |
| 575 | # after we drop the Parameter. |
| 576 | self.w13_weight = None |
| 577 | self.w13_weight_scale = None |
| 578 | self.w2_weight = None |
| 579 | self.w2_weight_scale = None |
| 580 | |
| 581 | if shared_experts is None or self.num_shared_experts == 0: |
| 582 | return |
| 583 | if self._transformed_shared_l1_weights is not None: |
| 584 | return |
| 585 | if not self._deep_gemm_supports_shared_experts(deep_gemm): |
| 586 | logger.warning_once( |
| 587 | "Disabling native MegaMoE shared-expert fusion because the " |
| 588 | "installed DeepGEMM Python API is older than the vLLM " |
| 589 | "source. Rebuild the vendored _deep_gemm_C extension to enable it.", |
| 590 | ) |
| 591 | self.num_shared_experts = 0 |
| 592 | return |
| 593 | self._finalize_shared_expert_weights(deep_gemm, shared_experts) |
| 594 | |
| 595 | @property |
| 596 | def has_fused_shared_experts(self) -> bool: |
| 597 | return self._transformed_shared_l1_weights is not None |
| 598 | |
| 599 | def get_symm_buffer(self): |
| 600 | from vllm.utils.deep_gemm import _import_deep_gemm |
| 601 | |
| 602 | deep_gemm = _import_deep_gemm() |
| 603 | |
| 604 | group = get_ep_group().device_group |
| 605 | device = torch.accelerator.current_device_index() |
| 606 | key = ( |
| 607 | id(group), |
| 608 | device, |
| 609 | self.num_experts, |
| 610 | self.max_num_tokens, |
| 611 | self.top_k, |
| 612 | self.hidden_size, |
| 613 | self.intermediate_size, |
| 614 | self.num_shared_experts if self.has_fused_shared_experts else 0, |
| 615 | ) |
| 616 | symm_buffer = self._symm_buffer_cache.get(key) |
| 617 | if symm_buffer is None: |
| 618 | symm_buffer = deep_gemm.get_symm_buffer_for_mega_moe( |
| 619 | group, |
| 620 | self.num_experts, |
| 621 | self.max_num_tokens, |
| 622 | self.top_k, |
| 623 | self.hidden_size, |
| 624 | self.intermediate_size, |
| 625 | num_shared_experts=( |
| 626 | self.num_shared_experts if self.has_fused_shared_experts else 0 |
| 627 | ), |
| 628 | ) |
| 629 | self._symm_buffer_cache[key] = symm_buffer |
| 630 | return symm_buffer |
| 631 | |
| 632 | def set_eplb_state( |
| 633 | self, |
| 634 | moe_layer_idx: int, |
| 635 | expert_load_view: torch.Tensor, |
| 636 | logical_to_physical_map: torch.Tensor, |
| 637 | logical_replica_count: torch.Tensor, |
| 638 | ) -> None: |
| 639 | self.eplb_state.set_layer_state( |
| 640 | moe_layer_idx, |
| 641 | expert_load_view, |
| 642 | logical_to_physical_map, |
| 643 | logical_replica_count, |
| 644 | ) |
| 645 | |
| 646 | def get_expert_weights(self) -> list[torch.Tensor]: |
| 647 | self.finalize_weights() |
| 648 | assert self._transformed_l1_weights is not None |
| 649 | assert self._transformed_l2_weights is not None |
| 650 | |
| 651 | def _to_eplb_view(name: str, t: torch.Tensor) -> torch.Tensor: |
| 652 | """Return a (num_local_experts, -1) view with contiguous memory layout.""" |
| 653 | assert t.shape[0] == self.num_local_experts |
| 654 | if t.is_contiguous(): |
| 655 | return t.view(self.num_local_experts, -1) |
| 656 | elif t.dim() == 3 and t.stride(1) == 1 and t.stride(2) == t.shape[1]: |
| 657 | # scales have shape (E, M, N) with memory layout (E, N, M) |
| 658 | back = torch.transpose(t, 1, 2) |
| 659 | assert back.is_contiguous() |
| 660 | return back.view(self.num_local_experts, -1) |
| 661 | |
| 662 | raise AssertionError( |
| 663 | f"DSv4 EPLB {name}: non-contiguous expert tensor with " |
| 664 | f"unexpected layout shape={tuple(t.shape)} " |
| 665 | f"stride={tuple(t.stride())} dtype={t.dtype}" |
| 666 | ) |
| 667 | |
| 668 | return [ |
| 669 | _to_eplb_view("l1_packed", self._transformed_l1_weights[0]), |
| 670 | _to_eplb_view("l1_scale", self._transformed_l1_weights[1]), |
| 671 | _to_eplb_view("l2_weight", self._transformed_l2_weights[0]), |
| 672 | _to_eplb_view("l2_scale", self._transformed_l2_weights[1]), |
| 673 | ] |
| 674 | |
| 675 | def update_expert_map(self) -> None: |
| 676 | pass |
| 677 | |
| 678 | @property |
| 679 | def layer_id(self) -> int: |
| 680 | return extract_layer_index(self.prefix) |
| 681 | |
| 682 | def forward( |
| 683 | self, |
| 684 | hidden_states: torch.Tensor, |
| 685 | topk_weights: torch.Tensor, |
| 686 | topk_ids: torch.Tensor, |
| 687 | *, |
| 688 | activation_clamp: float | None, |
| 689 | fast_math: bool = True, |
| 690 | ) -> torch.Tensor: |
| 691 | if hidden_states.shape[0] > self.max_num_tokens: |
| 692 | raise ValueError( |
| 693 | f"DeepSeek V4 MegaMoE got {hidden_states.shape[0]} tokens, " |
| 694 | f"but the symmetric buffer was sized for {self.max_num_tokens}." |
| 695 | ) |
| 696 | y = torch.empty_like(hidden_states, dtype=torch.bfloat16) |
| 697 | |
| 698 | from vllm.utils.deep_gemm import _import_deep_gemm |
| 699 | |
| 700 | deep_gemm = _import_deep_gemm() |
| 701 | |
| 702 | symm_buffer = self.get_symm_buffer() |
| 703 | num_tokens = hidden_states.shape[0] |
| 704 | is_padding = None |
| 705 | if envs.VLLM_MOE_SKIP_PADDING and is_forward_context_available(): |
| 706 | is_padding = get_forward_context().is_padding |
| 707 | if is_padding is not None: |
| 708 | is_padding = is_padding[:num_tokens] |
| 709 | |
| 710 | if self.capture_fn is not None: |
| 711 | self.capture_fn(topk_ids) |
| 712 | |
| 713 | # EPLB: map logical expert IDs to physical replicas and record load. |
| 714 | eplb_state = self.eplb_state |
| 715 | if eplb_state.logical_to_physical_map is not None: |
| 716 | assert eplb_state.expert_load_view is not None |
| 717 | assert eplb_state.logical_replica_count is not None |
| 718 | assert eplb_state.should_record_tensor is not None |
| 719 | if is_padding is not None: |
| 720 | topk_ids = torch.where(is_padding.unsqueeze(1), -1, topk_ids) |
| 721 | topk_ids = eplb_map_to_physical_and_record( |
| 722 | topk_ids=topk_ids, |
| 723 | expert_load_view=eplb_state.expert_load_view, |
| 724 | logical_to_physical_map=eplb_state.logical_to_physical_map, |
| 725 | logical_replica_count=eplb_state.logical_replica_count, |
| 726 | record_enabled=eplb_state.should_record_tensor, |
| 727 | num_unpadded_tokens=eplb_state.num_unpadded_tokens_tensors[ |
| 728 | dbo_current_ubatch_id() |
| 729 | ] |
| 730 | if eplb_state.num_unpadded_tokens_tensors is not None |
| 731 | else None, |
| 732 | ) |
| 733 | |
| 734 | shared_x_sf = None |
| 735 | shared_block_m = None |
| 736 | if self.has_fused_shared_experts: |
| 737 | shared_x_sf = symm_buffer.shared_l1_acts_sf |
| 738 | shared_block_m = deep_gemm.get_block_m_for_mega_moe( |
| 739 | get_ep_group().world_size, |
| 740 | self.num_experts, |
| 741 | symm_buffer.num_max_tokens_per_rank, |
| 742 | num_tokens, |
| 743 | self.top_k, |
| 744 | "fp8xfp4", |
| 745 | ) |
| 746 | |
| 747 | prepare_megamoe_inputs( |
| 748 | hidden_states, |
| 749 | topk_weights, |
| 750 | topk_ids, |
| 751 | symm_buffer.x[:num_tokens], |
| 752 | symm_buffer.x_sf[:num_tokens], |
| 753 | symm_buffer.topk_idx[:num_tokens], |
| 754 | symm_buffer.topk_weights[:num_tokens], |
| 755 | is_padding=is_padding, |
| 756 | shared_x_sf=shared_x_sf, |
| 757 | shared_block_m=shared_block_m, |
| 758 | ) |
| 759 | |
| 760 | assert self._transformed_l1_weights is not None |
| 761 | assert self._transformed_l2_weights is not None |
| 762 | if self.has_fused_shared_experts: |
| 763 | deep_gemm.fp8_fp4_mega_moe( |
| 764 | y, |
| 765 | self._transformed_l1_weights, |
| 766 | self._transformed_l2_weights, |
| 767 | symm_buffer, |
| 768 | shared_l1_weights=self._transformed_shared_l1_weights, |
| 769 | shared_l2_weights=self._transformed_shared_l2_weights, |
| 770 | activation_clamp=activation_clamp, |
| 771 | fast_math=fast_math, |
| 772 | ) |
| 773 | else: |
| 774 | deep_gemm.fp8_fp4_mega_moe( |
| 775 | y, |
| 776 | self._transformed_l1_weights, |
| 777 | self._transformed_l2_weights, |
| 778 | symm_buffer, |
| 779 | activation_clamp=activation_clamp, |
| 780 | fast_math=fast_math, |
| 781 | ) |
| 782 | return y |
| 783 | |
| 784 | |
| 785 | DeepseekV4MegaMoEExperts.weight_loader.supports_moe_loading = True # type: ignore[attr-defined] |
| 786 | |
| 787 | |
| 788 | class DeepseekV4MoE(nn.Module): |
| 789 | def __init__( |
| 790 | self, |
| 791 | vllm_config: VllmConfig, |
| 792 | prefix: str = "", |
| 793 | use_sequence_parallel: bool = False, |
| 794 | *, |
| 795 | num_hash_layers: int, |
| 796 | reduce_results: bool = True, |
| 797 | n_routed_experts: int | None = None, |
| 798 | n_activated_experts: int | None = None, |
| 799 | image_sentinel_lo: int = IMAGE_SENTINEL_BASE_ID, |
| 800 | ): |
| 801 | super().__init__() |
| 802 | |
| 803 | self.tp_size = get_tensor_model_parallel_world_size() |
| 804 | config = vllm_config.model_config.hf_config |
| 805 | quant_config = vllm_config.quant_config |
| 806 | self.prefix = prefix |
| 807 | self.use_sequence_parallel = use_sequence_parallel |
| 808 | self.reduce_results = reduce_results |
| 809 | moe_backend = vllm_config.kernel_config.moe_backend |
| 810 | validate_fi_moe_ep_config(vllm_config) |
| 811 | self.use_mega_moe = moe_backend in MEGA_MOE_BACKENDS |
| 812 | self.use_fi_mega_moe = is_fi_moe_ep_backend(moe_backend) |
| 813 | if self.use_mega_moe and not vllm_config.parallel_config.enable_expert_parallel: |
| 814 | raise NotImplementedError( |
| 815 | "DeepSeek V4 MegaMoE currently requires expert parallel. " |
| 816 | "Enable it with --enable-expert-parallel, or pick a different " |
| 817 | "moe backend." |
| 818 | ) |
| 819 | |
| 820 | self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0) |
| 821 | self.hidden_size = config.hidden_size |
| 822 | |
| 823 | self.n_routed_experts = ( |
| 824 | config.n_routed_experts if n_routed_experts is None else n_routed_experts |
| 825 | ) |
| 826 | self.n_activated_experts = ( |
| 827 | config.num_experts_per_tok |
| 828 | if n_activated_experts is None |
| 829 | else n_activated_experts |
| 830 | ) |
| 831 | self.moe_intermediate_size = config.moe_intermediate_size |
| 832 | self.swiglu_limit = config.swiglu_limit |
| 833 | self.renormalize = config.norm_topk_prob |
| 834 | self.scoring_func = getattr(config, "scoring_func", "sqrtsoftplus") |
| 835 | if self.use_mega_moe and self.scoring_func != "sqrtsoftplus": |
| 836 | raise NotImplementedError( |
| 837 | "DeepSeek V4 MegaMoE currently supports sqrtsoftplus routing only." |
| 838 | ) |
| 839 | if self.use_mega_moe and not self.renormalize: |
| 840 | raise NotImplementedError( |
| 841 | "DeepSeek V4 MegaMoE requires normalized top-k probabilities." |
| 842 | ) |
| 843 | if self.use_mega_moe and getattr(config, "expert_dtype", "fp4") != "fp4": |
| 844 | raise NotImplementedError( |
| 845 | "DeepSeek V4 MegaMoE only supports fp4 experts; got expert_dtype=" |
| 846 | f"{config.expert_dtype!r}. Drop --kernel-config moe_backend=" |
| 847 | f"{moe_backend} for this checkpoint." |
| 848 | ) |
| 849 | |
| 850 | self.gate = GateLinear( |
| 851 | input_size=config.hidden_size, |
| 852 | output_size=self.n_routed_experts, |
| 853 | bias=False, |
| 854 | out_dtype=torch.float32, |
| 855 | prefix=f"{prefix}.gate", |
| 856 | ) |
| 857 | |
| 858 | self.gate.e_score_correction_bias = None |
| 859 | self.gate.tid2eid = None |
| 860 | self.gate.bias_vl = None |
| 861 | # Image tokens borrow five consecutive reserved in-vocab ids starting |
| 862 | # at IMAGE_SENTINEL_BASE_ID; 0 disables vision routing (text model). |
| 863 | self.image_sentinel_lo = ( |
| 864 | image_sentinel_lo if getattr(config, "vision_n_layers", 0) > 0 else 0 |
| 865 | ) |
| 866 | is_hash_moe = extract_layer_index(prefix) < num_hash_layers |
| 867 | self.hash_indices_dtype = torch.int64 if self.use_mega_moe else torch.int32 |
| 868 | if is_hash_moe: |
| 869 | # hash MoE doesn't use e_score_correction_bias |
| 870 | # Use randint instead of empty to avoid garbage values causing |
| 871 | # invalid memory access in dummy mode (--load-format="dummy") |
| 872 | self.gate.tid2eid = nn.Parameter( |
| 873 | torch.randint( |
| 874 | 0, |
| 875 | self.n_routed_experts, |
| 876 | (config.vocab_size, self.n_activated_experts), |
| 877 | dtype=self.hash_indices_dtype, |
| 878 | ), |
| 879 | requires_grad=False, |
| 880 | ) |
| 881 | if getattr(config, "topk_method", None) == "noaux_tc" and ( |
| 882 | not is_hash_moe or getattr(config, "vision_n_layers", 0) > 0 |
| 883 | ): |
| 884 | # Vision checkpoints ship a gate bias on hash layers too (it is |
| 885 | # unused for routing there; image tokens use bias_vl instead). |
| 886 | self.gate.e_score_correction_bias = nn.Parameter( |
| 887 | torch.empty(self.n_routed_experts, dtype=torch.float32), |
| 888 | requires_grad=False, |
| 889 | ) |
| 890 | |
| 891 | if getattr(config, "vision_n_layers", 0) > 0: |
| 892 | # Vision checkpoints route image sentinel tokens with bias_vl |
| 893 | # instead of e_score_correction_bias / the hash table. Created on |
| 894 | # every MoE layer, hash layers included. |
| 895 | self.gate.bias_vl = nn.Parameter( |
| 896 | torch.empty(self.n_routed_experts, dtype=torch.float32), |
| 897 | requires_grad=False, |
| 898 | ) |
| 899 | |
| 900 | if config.n_shared_experts is None: |
| 901 | self.shared_experts = None |
| 902 | else: |
| 903 | intermediate_size = config.moe_intermediate_size * config.n_shared_experts |
| 904 | |
| 905 | self.shared_experts = DeepseekV4MLP( |
| 906 | hidden_size=config.hidden_size, |
| 907 | intermediate_size=intermediate_size, |
| 908 | hidden_act=config.hidden_act, |
| 909 | swiglu_limit=self.swiglu_limit, |
| 910 | quant_config=quant_config, |
| 911 | reduce_results=self.use_mega_moe, |
| 912 | is_sequence_parallel=use_sequence_parallel, |
| 913 | prefix=f"{prefix}.shared_experts", |
| 914 | ) |
| 915 | |
| 916 | if self.use_mega_moe: |
| 917 | self._init_mega_moe_experts(vllm_config, config, prefix) |
| 918 | else: |
| 919 | self._init_fused_moe_experts(vllm_config, config, quant_config, prefix) |
| 920 | |
| 921 | def _init_mega_moe_experts( |
| 922 | self, |
| 923 | vllm_config: VllmConfig, |
| 924 | config, |
| 925 | prefix: str, |
| 926 | ) -> None: |
| 927 | self.ep_group = get_ep_group() |
| 928 | self.ep_size = self.ep_group.world_size |
| 929 | self.ep_rank = self.ep_group.rank_in_group |
| 930 | |
| 931 | eplb_config = vllm_config.parallel_config.eplb_config |
| 932 | self.n_redundant_experts = eplb_config.num_redundant_experts |
| 933 | self.n_shared_experts = config.n_shared_experts or 0 |
| 934 | self.n_logical_experts = self.n_routed_experts |
| 935 | self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts |
| 936 | assert self.n_physical_experts % self.ep_size == 0, ( |
| 937 | f"n_physical_experts={self.n_physical_experts} must be divisible by " |
| 938 | f"ep_size={self.ep_size}. Adjust num_redundant_experts." |
| 939 | ) |
| 940 | self.n_local_physical_experts = self.n_physical_experts // self.ep_size |
| 941 | self.physical_expert_start = self.ep_rank * self.n_local_physical_experts |
| 942 | self.physical_expert_end = ( |
| 943 | self.physical_expert_start + self.n_local_physical_experts |
| 944 | ) |
| 945 | |
| 946 | self.n_local_experts = self.n_local_physical_experts |
| 947 | self.experts_start_idx = self.physical_expert_start |
| 948 | self.experts_end_idx = self.physical_expert_end |
| 949 | |
| 950 | # Native DeepGEMM fusion requires each EP rank to own the complete |
| 951 | # shared MLP. Sequence parallel replicates those weights while sharding |
| 952 | # tokens. TP=1 is also naturally replicated. With PP+TP the shared MLP |
| 953 | # remains tensor-sharded, so retain the serial path. The FlashInfer |
| 954 | # megakernel has no shared-expert fusion, so it keeps the serial path. |
| 955 | fuse_shared_experts = bool( |
| 956 | self.shared_experts is not None |
| 957 | and not envs.VLLM_DISABLE_DSV4_MEGAMOE_SHARED_EXPERT_FUSION |
| 958 | and (self.use_sequence_parallel or self.tp_size == 1) |
| 959 | and not self.use_fi_mega_moe |
| 960 | ) |
| 961 | |
| 962 | activation_clamp = ( |
| 963 | float(self.swiglu_limit) if self.swiglu_limit is not None else None |
| 964 | ) |
| 965 | if self.use_fi_mega_moe: |
| 966 | # Deferred: fi_moe subclasses DeepseekV4MegaMoEExperts, so a |
| 967 | # module-level import here would be circular. |
| 968 | from vllm.models.deepseek_v4.nvidia.fi_moe import ( |
| 969 | DeepseekV4MegaMoEExpertsFI, |
| 970 | ) |
| 971 | |
| 972 | experts_cls: type[DeepseekV4MegaMoEExperts] = DeepseekV4MegaMoEExpertsFI |
| 973 | else: |
| 974 | experts_cls = DeepseekV4MegaMoEExperts |
| 975 | expert_kwargs: dict[str, typing.Any] = dict( |
| 976 | num_experts=self.n_physical_experts, |
| 977 | num_local_experts=self.n_local_physical_experts, |
| 978 | experts_start_idx=self.physical_expert_start, |
| 979 | num_logical_experts=self.n_logical_experts, |
| 980 | top_k=self.n_activated_experts, |
| 981 | hidden_size=config.hidden_size, |
| 982 | intermediate_size=config.moe_intermediate_size, |
| 983 | num_shared_experts=(self.n_shared_experts if fuse_shared_experts else 0), |
| 984 | prefix=f"{prefix}.experts", |
| 985 | ) |
| 986 | if self.use_fi_mega_moe: |
| 987 | expert_kwargs["activation_clamp"] = activation_clamp |
| 988 | self.experts = experts_cls(vllm_config, **expert_kwargs) |
| 989 | |
| 990 | def _init_fused_moe_experts( |
| 991 | self, |
| 992 | vllm_config: VllmConfig, |
| 993 | config, |
| 994 | quant_config, |
| 995 | prefix: str, |
| 996 | ) -> None: |
| 997 | parallel_config = vllm_config.parallel_config |
| 998 | ep_group = get_ep_group() |
| 999 | ep_size = ep_group.world_size |
| 1000 | ep_rank = ep_group.rank_in_group |
| 1001 | |
| 1002 | eplb_config = parallel_config.eplb_config |
| 1003 | self.n_redundant_experts = eplb_config.num_redundant_experts |
| 1004 | self.n_shared_experts = config.n_shared_experts or 0 |
| 1005 | self.n_logical_experts = self.n_routed_experts |
| 1006 | self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts |
| 1007 | self.n_local_physical_experts = self.n_physical_experts // ep_size |
| 1008 | self.n_local_experts = self.n_local_physical_experts |
| 1009 | self.experts_start_idx = ep_rank * self.n_local_experts |
| 1010 | self.experts_end_idx = self.experts_start_idx + self.n_local_experts |
| 1011 | self.physical_expert_start = self.experts_start_idx |
| 1012 | self.physical_expert_end = self.experts_end_idx |
| 1013 | |
| 1014 | self.experts = FusedMoEFactory( |
| 1015 | reduce_results=self.reduce_results, |
| 1016 | shared_experts=self.shared_experts, |
| 1017 | gate=self.gate, |
| 1018 | num_experts=self.n_routed_experts, |
| 1019 | top_k=self.n_activated_experts, |
| 1020 | hidden_size=config.hidden_size, |
| 1021 | intermediate_size=config.moe_intermediate_size, |
| 1022 | renormalize=config.norm_topk_prob, |
| 1023 | quant_config=quant_config, |
| 1024 | prefix=f"{prefix}.experts", |
| 1025 | scoring_func=self.scoring_func, |
| 1026 | routed_scaling_factor=self.routed_scaling_factor, |
| 1027 | e_score_correction_bias=self.gate.e_score_correction_bias, |
| 1028 | hash_indices_table=self.gate.tid2eid, |
| 1029 | bias_vl=getattr(self.gate, "bias_vl", None), |
| 1030 | image_sentinel_lo=self.image_sentinel_lo, |
| 1031 | swiglu_limit=self.swiglu_limit, |
| 1032 | router_logits_dtype=torch.float32, |
| 1033 | enable_eplb=parallel_config.enable_eplb, |
| 1034 | num_redundant_experts=eplb_config.num_redundant_experts, |
| 1035 | is_sequence_parallel=self.use_sequence_parallel, |
| 1036 | ) |
| 1037 | |
| 1038 | def forward( |
| 1039 | self, |
| 1040 | hidden_states: torch.Tensor, |
| 1041 | input_ids: torch.Tensor | None = None, |
| 1042 | mega_gate_metadata: MegaGateRoutingMetadata | None = None, |
| 1043 | ) -> torch.Tensor: |
| 1044 | if self.gate.tid2eid is not None and input_ids is None: |
| 1045 | raise ValueError("DeepSeek V4 hash MoE routing requires input_ids.") |
| 1046 | # getattr for test doubles that fake the gate module. |
| 1047 | bias_vl = getattr(self.gate, "bias_vl", None) |
| 1048 | if bias_vl is not None and input_ids is None: |
| 1049 | raise ValueError("DeepSeek V4 vision MoE routing requires input_ids.") |
| 1050 | |
| 1051 | if not self.use_mega_moe: |
| 1052 | return self._forward_fused_moe(hidden_states, input_ids) |
| 1053 | |
| 1054 | org_shape = hidden_states.shape |
| 1055 | # Small local padded batches favor GateLinear; 128-expert gates cross earlier. |
| 1056 | gate_threshold = 1 if self.gate.weight.shape[0] == 128 else 16 |
| 1057 | if hidden_states.shape[0] <= gate_threshold: |
| 1058 | router_logits, _ = self.gate(hidden_states) |
| 1059 | topk_weights, topk_ids = fused_topk_bias( |
| 1060 | hidden_states=hidden_states, |
| 1061 | gating_output=router_logits, |
| 1062 | scoring_func=self.scoring_func, |
| 1063 | e_score_correction_bias=self.gate.e_score_correction_bias.data |
| 1064 | if self.gate.e_score_correction_bias is not None |
| 1065 | else None, |
| 1066 | topk=self.n_activated_experts, |
| 1067 | renormalize=self.renormalize, |
| 1068 | indices_type=self.hash_indices_dtype, |
| 1069 | input_tokens=input_ids, |
| 1070 | hash_indices_table=self.gate.tid2eid, |
| 1071 | routed_scaling_factor=self.routed_scaling_factor, |
| 1072 | bias_vl=bias_vl.data if bias_vl is not None else None, |
| 1073 | image_sentinel_lo=self.image_sentinel_lo if bias_vl is not None else 0, |
| 1074 | ) |
| 1075 | else: |
| 1076 | from vllm.utils.deep_gemm import bf16_mega_gate |
| 1077 | |
| 1078 | if mega_gate_metadata is None: |
| 1079 | raise RuntimeError( |
| 1080 | "MegaMoE routing metadata must be prepared once by the model." |
| 1081 | ) |
| 1082 | |
| 1083 | unmapped_topk_idx = None |
| 1084 | fix_routing_mask = None |
| 1085 | if self.gate.tid2eid is not None: |
| 1086 | if mega_gate_metadata.safe_hash_input_ids is None: |
| 1087 | raise RuntimeError("Hash routing requires prepared input IDs.") |
| 1088 | if mega_gate_metadata.hash_token_mask is None: |
| 1089 | raise RuntimeError("Hash routing requires a prepared routing mask.") |
| 1090 | unmapped_topk_idx = self.gate.tid2eid[ |
| 1091 | mega_gate_metadata.safe_hash_input_ids |
| 1092 | ] |
| 1093 | fix_routing_mask = mega_gate_metadata.hash_token_mask |
| 1094 | |
| 1095 | topk_weights, topk_ids = bf16_mega_gate( |
| 1096 | hidden_states, |
| 1097 | self.gate.weight, |
| 1098 | self.n_activated_experts, |
| 1099 | scoring_func=self.scoring_func, |
| 1100 | routed_scaling_factor=self.routed_scaling_factor, |
| 1101 | ep_rank=self.ep_rank, |
| 1102 | bias=self.gate.e_score_correction_bias.data |
| 1103 | if self.gate.e_score_correction_bias is not None |
| 1104 | else None, |
| 1105 | image_bias=bias_vl.data if bias_vl is not None else None, |
| 1106 | image_token_mask=mega_gate_metadata.image_token_mask, |
| 1107 | fix_routing_mask=fix_routing_mask, |
| 1108 | unmapped_topk_idx=unmapped_topk_idx, |
| 1109 | ) |
| 1110 | activation_clamp = ( |
| 1111 | float(self.swiglu_limit) if self.swiglu_limit is not None else None |
| 1112 | ) |
| 1113 | final_hidden_states = self.experts( |
| 1114 | hidden_states, |
| 1115 | topk_weights, |
| 1116 | topk_ids, |
| 1117 | activation_clamp=activation_clamp, |
| 1118 | ) |
| 1119 | |
| 1120 | if ( |
| 1121 | self.shared_experts is not None |
| 1122 | and not self.experts.has_fused_shared_experts |
| 1123 | ): |
| 1124 | shared_output = self.shared_experts(hidden_states) |
| 1125 | final_hidden_states += shared_output |
| 1126 | |
| 1127 | return final_hidden_states.view(org_shape) |
| 1128 | |
| 1129 | def _forward_fused_moe( |
| 1130 | self, hidden_states: torch.Tensor, input_ids: torch.Tensor | None = None |
| 1131 | ) -> torch.Tensor: |
| 1132 | org_shape = hidden_states.shape |
| 1133 | final_hidden_states = self.experts( |
| 1134 | hidden_states=hidden_states, |
| 1135 | router_logits=hidden_states, |
| 1136 | input_ids=input_ids, |
| 1137 | ) |
| 1138 | |
| 1139 | return final_hidden_states.view(org_shape) |
| 1140 | |
| 1141 | def finalize_mega_moe_weights(self) -> None: |
| 1142 | if self.use_mega_moe: |
| 1143 | self.experts.finalize_weights(self.shared_experts) |
| 1144 | |
| 1145 | |
| 1146 | def _select_dsv4_attn_cls(vllm_config: VllmConfig) -> type[DeepseekV4Attention]: |
| 1147 | """Pick the CUDA sparse-MLA attention class for the configured backend. |
| 1148 | |
| 1149 | The generic CUDA backend selector does not instantiate DSv4 layers directly, |
| 1150 | so map generic sparse-MLA choices to the DSv4-specialized attention class. |
| 1151 | Without an explicit backend, SM12 defaults to FlashInfer while the other |
| 1152 | CUDA arches keep the FlashMLA path. |
| 1153 | """ |
| 1154 | backend = vllm_config.attention_config.backend |
| 1155 | device_capability = current_platform.get_device_capability() |
| 1156 | if backend in ( |
| 1157 | AttentionBackendEnum.FLASHINFER_MLA_SPARSE, |
| 1158 | AttentionBackendEnum.FLASHINFER_MLA_SPARSE_SM120, |
| 1159 | ): |
| 1160 | raise ValueError( |
| 1161 | f"{backend.name} is not a DeepSeek V4 attention backend. " |
| 1162 | "Use FLASHINFER_MLA_SPARSE_DSV4 for DeepSeek V4 FlashInfer " |
| 1163 | "sparse MLA." |
| 1164 | ) |
| 1165 | if backend == AttentionBackendEnum.FLASHINFER_MLA_SPARSE_DSV4: |
| 1166 | if device_capability is not None and device_capability.major == 12: |
| 1167 | return DeepseekV4FlashInferSM120Attention |
| 1168 | return DeepseekV4FlashInferMLAAttention |
| 1169 | if backend in ( |
| 1170 | AttentionBackendEnum.FLASHMLA_SPARSE, |
| 1171 | AttentionBackendEnum.FLASHMLA_SPARSE_DSV4, |
| 1172 | ): |
| 1173 | return DeepseekV4FlashMLAAttention |
| 1174 | |
| 1175 | if device_capability is not None and device_capability.major == 12: |
| 1176 | return DeepseekV4FlashInferSM120Attention |
| 1177 | return DeepseekV4FlashMLAAttention |
| 1178 | |
| 1179 | |
| 1180 | def _use_sequence_parallel(vllm_config: VllmConfig) -> bool: |
| 1181 | parallel_config = vllm_config.parallel_config |
| 1182 | use_mega_moe = vllm_config.kernel_config.moe_backend in MEGA_MOE_BACKENDS |
| 1183 | return ( |
| 1184 | parallel_config.pipeline_parallel_size == 1 |
| 1185 | and parallel_config.enable_expert_parallel |
| 1186 | and parallel_config.tensor_parallel_size > 1 |
| 1187 | and (use_mega_moe or parallel_config.data_parallel_size > 1) |
| 1188 | ) |
| 1189 | |
| 1190 | |
| 1191 | class DeepseekV4DecoderLayer(nn.Module): |
| 1192 | def __init__( |
| 1193 | self, |
| 1194 | vllm_config, |
| 1195 | prefix, |
| 1196 | topk_indices_buffer: torch.Tensor | None = None, |
| 1197 | aux_stream_list: list[torch.cuda.Stream] | None = None, |
| 1198 | ): |
| 1199 | super().__init__() |
| 1200 | |
| 1201 | config = vllm_config.model_config.hf_config |
| 1202 | self.hidden_size = config.hidden_size |
| 1203 | self.use_sequence_parallel = _use_sequence_parallel(vllm_config) |
| 1204 | |
| 1205 | self.rms_norm_eps = config.rms_norm_eps |
| 1206 | self.attn = _select_dsv4_attn_cls(vllm_config)( |
| 1207 | vllm_config, |
| 1208 | prefix=f"{prefix}.attn", |
| 1209 | topk_indices_buffer=topk_indices_buffer, |
| 1210 | aux_stream_list=aux_stream_list, |
| 1211 | ) |
| 1212 | if self.use_sequence_parallel: |
| 1213 | self.attn.wo_b.reduce_results = False |
| 1214 | self.ffn = DeepseekV4MoE( |
| 1215 | vllm_config, |
| 1216 | prefix=f"{prefix}.ffn", |
| 1217 | use_sequence_parallel=self.use_sequence_parallel, |
| 1218 | num_hash_layers=config.num_hash_layers, |
| 1219 | ) |
| 1220 | |
| 1221 | self.attn_norm = RMSNorm(self.hidden_size, self.rms_norm_eps) |
| 1222 | self.ffn_norm = RMSNorm(self.hidden_size, self.rms_norm_eps) |
| 1223 | self.hc_mult = config.hc_mult |
| 1224 | self.hc_sinkhorn_iters = config.hc_sinkhorn_iters |
| 1225 | self.hc_eps = config.hc_eps |
| 1226 | self.hc_post_alpha = 2.0 |
| 1227 | mix_hc = (2 + self.hc_mult) * self.hc_mult |
| 1228 | hc_dim = self.hc_mult * self.hidden_size |
| 1229 | self.hc_attn_fn = nn.Parameter( |
| 1230 | torch.empty( |
| 1231 | (mix_hc, hc_dim), |
| 1232 | dtype=torch.float32, |
| 1233 | ), |
| 1234 | requires_grad=False, |
| 1235 | ) |
| 1236 | self.hc_attn_fn_broadcast: torch.Tensor | None = None |
| 1237 | self.hc_ffn_fn = nn.Parameter( |
| 1238 | torch.empty( |
| 1239 | (mix_hc, hc_dim), |
| 1240 | dtype=torch.float32, |
| 1241 | ), |
| 1242 | requires_grad=False, |
| 1243 | ) |
| 1244 | self.hc_attn_base = nn.Parameter( |
| 1245 | torch.empty( |
| 1246 | mix_hc, |
| 1247 | dtype=torch.float32, |
| 1248 | ), |
| 1249 | requires_grad=False, |
| 1250 | ) |
| 1251 | self.hc_ffn_base = nn.Parameter( |
| 1252 | torch.empty( |
| 1253 | mix_hc, |
| 1254 | dtype=torch.float32, |
| 1255 | ), |
| 1256 | requires_grad=False, |
| 1257 | ) |
| 1258 | self.hc_attn_scale = nn.Parameter( |
| 1259 | torch.empty( |
| 1260 | 3, |
| 1261 | dtype=torch.float32, |
| 1262 | ), |
| 1263 | requires_grad=False, |
| 1264 | ) |
| 1265 | self.hc_ffn_scale = nn.Parameter( |
| 1266 | torch.empty( |
| 1267 | 3, |
| 1268 | dtype=torch.float32, |
| 1269 | ), |
| 1270 | requires_grad=False, |
| 1271 | ) |
| 1272 | |
| 1273 | if vllm_config.kernel_config.enable_jit_warmup: |
| 1274 | from vllm.model_executor.kernels.mhc.tilelang_kernels import ( |
| 1275 | _HC_PRENORM_GEMM_TILELANG_KERNEL, |
| 1276 | _MHC_FUSED_TILELANG_KERNEL, |
| 1277 | _MHC_POST_TILELANG_KERNEL, |
| 1278 | _MHC_PRE_BIG_FUSE_TILELANG_KERNEL, |
| 1279 | ) |
| 1280 | from vllm.utils.deep_gemm import is_deep_gemm_supported |
| 1281 | |
| 1282 | include_pre_gemm_splits = is_deep_gemm_supported() |
| 1283 | _MHC_PRE_BIG_FUSE_TILELANG_KERNEL.register_warmup( |
| 1284 | vllm_config, |
| 1285 | hidden_size=self.hidden_size, |
| 1286 | hc_mult=self.hc_mult, |
| 1287 | use_norm_weight=True, |
| 1288 | include_pre_gemm_splits=include_pre_gemm_splits, |
| 1289 | include_broadcast_splits=( |
| 1290 | get_pp_group().is_first_rank and extract_layer_index(prefix) == 0 |
| 1291 | ), |
| 1292 | rms_eps=self.rms_norm_eps, |
| 1293 | hc_pre_eps=self.hc_eps, |
| 1294 | hc_sinkhorn_eps=self.hc_eps, |
| 1295 | hc_post_mult_value=self.hc_post_alpha, |
| 1296 | sinkhorn_repeat=self.hc_sinkhorn_iters, |
| 1297 | norm_eps=( |
| 1298 | self.attn_norm.variance_epsilon, |
| 1299 | self.ffn_norm.variance_epsilon, |
| 1300 | ), |
| 1301 | broadcast_norm_eps=self.attn_norm.variance_epsilon, |
| 1302 | ) |
| 1303 | if not include_pre_gemm_splits: |
| 1304 | _HC_PRENORM_GEMM_TILELANG_KERNEL.register_warmup( |
| 1305 | vllm_config, |
| 1306 | hidden_size=self.hidden_size, |
| 1307 | hc_mult=self.hc_mult, |
| 1308 | n_out=self.hc_mult * (2 + self.hc_mult), |
| 1309 | ) |
| 1310 | _MHC_POST_TILELANG_KERNEL.register_warmup( |
| 1311 | hidden_size=self.hidden_size, |
| 1312 | hc_mult=self.hc_mult, |
| 1313 | ) |
| 1314 | _MHC_FUSED_TILELANG_KERNEL.register_warmup( |
| 1315 | vllm_config, |
| 1316 | hidden_size=self.hidden_size, |
| 1317 | hc_mult=self.hc_mult, |
| 1318 | ) |
| 1319 | |
| 1320 | def forward( |
| 1321 | self, |
| 1322 | x: torch.Tensor, |
| 1323 | positions: torch.Tensor, |
| 1324 | input_ids: torch.Tensor | None, |
| 1325 | post_mix: torch.Tensor | None = None, |
| 1326 | res_mix: torch.Tensor | None = None, |
| 1327 | residual: torch.Tensor | None = None, |
| 1328 | mega_gate_metadata: MegaGateRoutingMetadata | None = None, |
| 1329 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: |
| 1330 | attn_norm_weight = self.attn_norm.weight.data |
| 1331 | attn_norm_eps = self.attn_norm.variance_epsilon |
| 1332 | if residual is None: |
| 1333 | # Run standalone mhc_pre on first layer |
| 1334 | if x.dim() == 2: |
| 1335 | assert self.hc_attn_fn_broadcast is not None |
| 1336 | residual, post_mix, res_mix, x = mhc_pre_broadcast_tilelang( |
| 1337 | x, |
| 1338 | self.hc_attn_fn, |
| 1339 | self.hc_attn_scale, |
| 1340 | self.hc_attn_base, |
| 1341 | self.rms_norm_eps, |
| 1342 | self.hc_eps, |
| 1343 | self.hc_eps, |
| 1344 | self.hc_post_alpha, |
| 1345 | self.hc_sinkhorn_iters, |
| 1346 | norm_weight=attn_norm_weight, |
| 1347 | norm_eps=attn_norm_eps, |
| 1348 | fn_broadcast=self.hc_attn_fn_broadcast, |
| 1349 | ) |
| 1350 | else: |
| 1351 | residual = x |
| 1352 | post_mix, res_mix, x = mhc_pre_tilelang( |
| 1353 | x, |
| 1354 | self.hc_attn_fn, |
| 1355 | self.hc_attn_scale, |
| 1356 | self.hc_attn_base, |
| 1357 | self.rms_norm_eps, |
| 1358 | self.hc_eps, |
| 1359 | self.hc_eps, |
| 1360 | self.hc_post_alpha, |
| 1361 | self.hc_sinkhorn_iters, |
| 1362 | norm_weight=attn_norm_weight, |
| 1363 | norm_eps=attn_norm_eps, |
| 1364 | ) |
| 1365 | else: |
| 1366 | residual, post_mix, res_mix, x = mhc_fused_post_pre_tilelang( |
| 1367 | x, |
| 1368 | residual, |
| 1369 | post_mix, |
| 1370 | res_mix, |
| 1371 | self.hc_attn_fn, |
| 1372 | self.hc_attn_scale, |
| 1373 | self.hc_attn_base, |
| 1374 | self.rms_norm_eps, |
| 1375 | self.hc_eps, |
| 1376 | self.hc_eps, |
| 1377 | self.hc_post_alpha, |
| 1378 | self.hc_sinkhorn_iters, |
| 1379 | n_splits=1, |
| 1380 | tile_n=1, |
| 1381 | norm_weight=attn_norm_weight, |
| 1382 | norm_eps=attn_norm_eps, |
| 1383 | ) |
| 1384 | |
| 1385 | if self.use_sequence_parallel: |
| 1386 | x = sp_all_gather(x)[: positions.shape[0]] |
| 1387 | |
| 1388 | x = self.attn(positions, x, None) |
| 1389 | if self.use_sequence_parallel: |
| 1390 | x = sp_reduce_scatter(x) |
| 1391 | |
| 1392 | ffn_norm_weight = self.ffn_norm.weight.data |
| 1393 | ffn_norm_eps = self.ffn_norm.variance_epsilon |
| 1394 | residual, post_mix, res_mix, x = mhc_fused_post_pre_tilelang( |
| 1395 | x, |
| 1396 | residual, |
| 1397 | post_mix, |
| 1398 | res_mix, |
| 1399 | self.hc_ffn_fn, |
| 1400 | self.hc_ffn_scale, |
| 1401 | self.hc_ffn_base, |
| 1402 | self.rms_norm_eps, |
| 1403 | self.hc_eps, |
| 1404 | self.hc_eps, |
| 1405 | self.hc_post_alpha, |
| 1406 | self.hc_sinkhorn_iters, |
| 1407 | n_splits=1, |
| 1408 | tile_n=1, |
| 1409 | norm_weight=ffn_norm_weight, |
| 1410 | norm_eps=ffn_norm_eps, |
| 1411 | ) |
| 1412 | |
| 1413 | x = self.ffn(x, input_ids, mega_gate_metadata) |
| 1414 | return x, residual, post_mix, res_mix |
| 1415 | |
| 1416 | |
| 1417 | class DeepseekV4Model(nn.Module, EagleModelMixin): |
| 1418 | supports_aux_hidden_states_over_pp = True |
| 1419 | |
| 1420 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1421 | super().__init__() |
| 1422 | |
| 1423 | config = vllm_config.model_config.hf_config |
| 1424 | quant_config = vllm_config.quant_config |
| 1425 | self.config = config |
| 1426 | self.quant_config = quant_config |
| 1427 | self.parallel_config = vllm_config.parallel_config |
| 1428 | self.use_mega_moe = vllm_config.kernel_config.moe_backend in MEGA_MOE_BACKENDS |
| 1429 | self.use_sequence_parallel = _use_sequence_parallel(vllm_config) |
| 1430 | if self.use_mega_moe and not vllm_config.parallel_config.enable_expert_parallel: |
| 1431 | raise NotImplementedError( |
| 1432 | "DeepSeek V4 MegaMoE currently requires expert parallel. " |
| 1433 | "Enable it with --enable-expert-parallel, or pick a different " |
| 1434 | "moe backend." |
| 1435 | ) |
| 1436 | self.vocab_size = config.vocab_size |
| 1437 | self.hc_eps = config.hc_eps |
| 1438 | self.hc_mult = config.hc_mult |
| 1439 | self.hc_dim = self.hc_mult * config.hidden_size |
| 1440 | self.rms_norm_eps = config.rms_norm_eps |
| 1441 | |
| 1442 | # Three aux streams: one per non-default input GEMM in |
| 1443 | # DeepseekV4Attention._run_parallel_input_projections |
| 1444 | # (compressor kv_score, indexer.weights_proj, indexer.compressor |
| 1445 | # kv_score). fused_wqa_wkv stays on the default stream. |
| 1446 | aux_stream_list = [torch.cuda.Stream() for _ in range(3)] |
| 1447 | |
| 1448 | # Reserved topk indices buffer for all Indexer layers to reuse. |
| 1449 | self.topk_indices_buffer = torch.empty( |
| 1450 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 1451 | config.index_topk, |
| 1452 | dtype=torch.int32, |
| 1453 | ) |
| 1454 | |
| 1455 | if get_pp_group().is_first_rank or spec_decode_needs_target_embed(vllm_config): |
| 1456 | self.embed_tokens = VocabParallelEmbedding( |
| 1457 | config.vocab_size, |
| 1458 | config.hidden_size, |
| 1459 | quant_config=quant_config, |
| 1460 | prefix=f"{prefix}.embed_tokens", |
| 1461 | ) |
| 1462 | else: |
| 1463 | self.embed_tokens = PPMissingLayer() |
| 1464 | |
| 1465 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 1466 | config.num_hidden_layers, |
| 1467 | lambda prefix: DeepseekV4DecoderLayer( |
| 1468 | vllm_config, |
| 1469 | prefix=prefix, |
| 1470 | topk_indices_buffer=self.topk_indices_buffer, |
| 1471 | aux_stream_list=aux_stream_list, |
| 1472 | ), |
| 1473 | prefix=f"{prefix}.layers", |
| 1474 | ) |
| 1475 | self.has_local_hash_moe = self.start_layer < min( |
| 1476 | self.end_layer, config.num_hash_layers |
| 1477 | ) |
| 1478 | |
| 1479 | if get_pp_group().is_last_rank: |
| 1480 | self.norm = RMSNorm(config.hidden_size, self.rms_norm_eps) |
| 1481 | else: |
| 1482 | self.norm = PPMissingLayer() |
| 1483 | |
| 1484 | self.hc_head_fn = nn.Parameter( |
| 1485 | torch.empty( |
| 1486 | self.hc_mult, |
| 1487 | self.hc_dim, |
| 1488 | dtype=torch.float32, |
| 1489 | ), |
| 1490 | requires_grad=False, |
| 1491 | ) |
| 1492 | self.hc_head_base = nn.Parameter( |
| 1493 | torch.empty( |
| 1494 | self.hc_mult, |
| 1495 | dtype=torch.float32, |
| 1496 | ), |
| 1497 | requires_grad=False, |
| 1498 | ) |
| 1499 | self.hc_head_scale = nn.Parameter( |
| 1500 | torch.empty(1, dtype=torch.float32), |
| 1501 | requires_grad=False, |
| 1502 | ) |
| 1503 | spec_config = vllm_config.speculative_config |
| 1504 | needs_mtp_hidden_states = spec_config is not None and ( |
| 1505 | spec_config.use_eagle() or spec_config.uses_draft_model() |
| 1506 | ) |
| 1507 | if get_pp_group().is_last_rank and needs_mtp_hidden_states: |
| 1508 | self._mtp_hidden_buffer = torch.empty( |
| 1509 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 1510 | self.hc_dim, |
| 1511 | dtype=vllm_config.model_config.dtype, |
| 1512 | ) |
| 1513 | else: |
| 1514 | self._mtp_hidden_buffer = None |
| 1515 | |
| 1516 | if vllm_config.kernel_config.enable_jit_warmup and get_pp_group().is_last_rank: |
| 1517 | from vllm.model_executor.kernels.mhc.tilelang_kernels import ( |
| 1518 | _HC_HEAD_FUSED_TILELANG_KERNEL, |
| 1519 | ) |
| 1520 | |
| 1521 | _HC_HEAD_FUSED_TILELANG_KERNEL.register_warmup( |
| 1522 | hidden_size=config.hidden_size, |
| 1523 | hc_mult=self.hc_mult, |
| 1524 | rms_eps=self.rms_norm_eps, |
| 1525 | hc_eps=self.hc_eps, |
| 1526 | ) |
| 1527 | |
| 1528 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1529 | return self.embed_tokens(input_ids) |
| 1530 | |
| 1531 | def make_empty_intermediate_tensors( |
| 1532 | self, |
| 1533 | batch_size: int, |
| 1534 | dtype: torch.dtype, |
| 1535 | device: torch.device, |
| 1536 | ) -> IntermediateTensors: |
| 1537 | # PP intermediate tensors carry the multi-stream hidden_states |
| 1538 | # of shape (num_tokens, hc_mult, hidden_size) — V4 expands the |
| 1539 | # token embedding to hc_mult streams before the first decoder |
| 1540 | # layer and keeps that shape until hc_head() collapses it. |
| 1541 | return IntermediateTensors( |
| 1542 | { |
| 1543 | "hidden_states": torch.zeros( |
| 1544 | (batch_size, self.hc_mult, self.config.hidden_size), |
| 1545 | dtype=dtype, |
| 1546 | device=device, |
| 1547 | ), |
| 1548 | } |
| 1549 | ) |
| 1550 | |
| 1551 | def forward( |
| 1552 | self, |
| 1553 | input_ids: torch.Tensor, |
| 1554 | positions: torch.Tensor, |
| 1555 | intermediate_tensors: IntermediateTensors | None, |
| 1556 | inputs_embeds: torch.Tensor | None = None, |
| 1557 | ) -> torch.Tensor | IntermediateTensors: |
| 1558 | if get_pp_group().is_first_rank: |
| 1559 | if inputs_embeds is not None: |
| 1560 | hidden_states = inputs_embeds |
| 1561 | else: |
| 1562 | hidden_states = self.embed_input_ids(input_ids) |
| 1563 | else: |
| 1564 | assert intermediate_tensors is not None |
| 1565 | hidden_states = intermediate_tensors["hidden_states"] |
| 1566 | |
| 1567 | if self.use_mega_moe: |
| 1568 | input_ids = input_ids.to(torch.int64) |
| 1569 | |
| 1570 | full_num_tokens = positions.shape[0] |
| 1571 | if self.use_sequence_parallel: |
| 1572 | if envs.VLLM_MOE_SKIP_PADDING and is_forward_context_available(): |
| 1573 | forward_context = get_forward_context() |
| 1574 | forward_context.is_padding = sp_padding_mask( |
| 1575 | forward_context.is_padding, hidden_states |
| 1576 | ) |
| 1577 | hidden_states = sp_shard(hidden_states) |
| 1578 | input_ids = sp_shard(input_ids) |
| 1579 | |
| 1580 | mega_gate_metadata = None |
| 1581 | if self.use_mega_moe: |
| 1582 | mega_gate_metadata = prepare_mega_gate_routing_metadata( |
| 1583 | input_ids, |
| 1584 | has_hash_routing=self.has_local_hash_moe, |
| 1585 | image_sentinel_base_id=IMAGE_SENTINEL_BASE_ID |
| 1586 | if getattr(self.config, "vision_n_layers", 0) > 0 |
| 1587 | else None, |
| 1588 | ) |
| 1589 | |
| 1590 | residual, post_mix, res_mix = None, None, None |
| 1591 | remote_aux = self.collect_remote_aux_hidden_states(intermediate_tensors) |
| 1592 | aux_hidden_states: list[torch.Tensor] = [] |
| 1593 | final_aux_recon: torch.Tensor | None = None # avoid duplicate mhc_post call |
| 1594 | for idx, layer in enumerate( |
| 1595 | islice(self.layers, self.start_layer, self.end_layer), |
| 1596 | start=self.start_layer, |
| 1597 | ): |
| 1598 | hidden_states, residual, post_mix, res_mix = layer( |
| 1599 | hidden_states, |
| 1600 | positions, |
| 1601 | input_ids, |
| 1602 | post_mix, |
| 1603 | res_mix, |
| 1604 | residual, |
| 1605 | mega_gate_metadata=mega_gate_metadata, |
| 1606 | ) |
| 1607 | if idx + 1 in self.aux_hidden_state_layers: |
| 1608 | # Reconstruct the aux hidden state for draft models |
| 1609 | aux_recon = mhc_post_tilelang( |
| 1610 | hidden_states, residual, post_mix, res_mix |
| 1611 | ) |
| 1612 | aux_hidden_state = aux_recon.mean(dim=1) |
| 1613 | if self.use_sequence_parallel: |
| 1614 | aux_hidden_state = sp_all_gather(aux_hidden_state)[:full_num_tokens] |
| 1615 | aux_hidden_states.append(aux_hidden_state) |
| 1616 | final_aux_recon = aux_recon |
| 1617 | if layer is not None: |
| 1618 | # Reuse if the last layer was captured as an aux hidden state |
| 1619 | if self.end_layer in self.aux_hidden_state_layers: |
| 1620 | hidden_states = final_aux_recon |
| 1621 | else: |
| 1622 | hidden_states = mhc_post_tilelang( |
| 1623 | hidden_states, residual, post_mix, res_mix |
| 1624 | ) |
| 1625 | |
| 1626 | if not get_pp_group().is_last_rank: |
| 1627 | return IntermediateTensors( |
| 1628 | { |
| 1629 | "hidden_states": hidden_states, |
| 1630 | **self.pack_local_aux_hidden_states(aux_hidden_states), |
| 1631 | } |
| 1632 | ) |
| 1633 | |
| 1634 | if self.use_sequence_parallel: |
| 1635 | hidden_states = sp_all_gather(hidden_states)[:full_num_tokens] |
| 1636 | |
| 1637 | if self._mtp_hidden_buffer is not None: |
| 1638 | num_tokens = hidden_states.shape[0] |
| 1639 | self._mtp_hidden_buffer[:num_tokens].copy_(hidden_states.flatten(1)) |
| 1640 | |
| 1641 | hidden_states = hc_head_fused_kernel_tilelang( |
| 1642 | hidden_states, |
| 1643 | self.hc_head_fn, |
| 1644 | self.hc_head_scale, |
| 1645 | self.hc_head_base, |
| 1646 | self.rms_norm_eps, |
| 1647 | self.hc_eps, |
| 1648 | ) |
| 1649 | hidden_states = self.norm(hidden_states) |
| 1650 | aux_hidden_states = remote_aux + aux_hidden_states |
| 1651 | if len(aux_hidden_states) > 0: |
| 1652 | return hidden_states, aux_hidden_states |
| 1653 | return hidden_states |
| 1654 | |
| 1655 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1656 | stacked_params_mapping = [ |
| 1657 | # (param_name, shard_name, shard_id) |
| 1658 | ("gate_up_proj", "w1", 0), |
| 1659 | ("gate_up_proj", "w3", 1), |
| 1660 | ("attn.fused_wqa_wkv", "attn.wq_a", 0), |
| 1661 | ("attn.fused_wqa_wkv", "attn.wkv", 1), |
| 1662 | ("compressor.fused_wkv_wgate", "compressor.wkv", 0), |
| 1663 | ("compressor.fused_wkv_wgate", "compressor.wgate", 1), |
| 1664 | ] |
| 1665 | params_dict = dict(self.named_parameters()) |
| 1666 | loaded_params: set[str] = set() |
| 1667 | |
| 1668 | # TP for attention |
| 1669 | tp_size = get_tensor_model_parallel_world_size() |
| 1670 | tp_rank = get_tensor_model_parallel_rank() |
| 1671 | n_head = self.config.num_attention_heads |
| 1672 | n_local_head = n_head // tp_size |
| 1673 | head_rank_start = n_local_head * tp_rank |
| 1674 | head_rank_end = n_local_head * (tp_rank + 1) |
| 1675 | |
| 1676 | # Pre-compute expert mapping ONCE. |
| 1677 | expert_mapping = self.get_expert_mapping() |
| 1678 | |
| 1679 | # Block-FP8 shared experts: pad the intermediate up to the TP-uniform |
| 1680 | # block count so the standard loaders below slice it evenly (trailing |
| 1681 | # ranks land on the zero pad). SP / unquantized ones need no padding. |
| 1682 | pad_shared_expert = ( |
| 1683 | getattr(self.quant_config, "weight_block_size", None) is not None |
| 1684 | and not self.use_sequence_parallel |
| 1685 | ) |
| 1686 | |
| 1687 | for name, loaded_weight in weights: |
| 1688 | if pad_shared_expert and ".shared_experts." in name: |
| 1689 | loaded_weight = self._pad_shared_expert_weight( |
| 1690 | self.quant_config, name, loaded_weight |
| 1691 | ) |
| 1692 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 1693 | # Skip non-stacked layers and experts (experts handled below). |
| 1694 | if ".experts." in name: |
| 1695 | continue |
| 1696 | if weight_name not in name: |
| 1697 | continue |
| 1698 | name = name.replace(weight_name, param_name) |
| 1699 | |
| 1700 | if is_pp_missing_parameter(name, self): |
| 1701 | break |
| 1702 | if name not in params_dict: |
| 1703 | head, _, leaf = name.rpartition(".") |
| 1704 | suffixed = f"{head}.base_layer.{leaf}" |
| 1705 | if suffixed in params_dict: |
| 1706 | name = suffixed |
| 1707 | param = params_dict[name] |
| 1708 | weight_loader = param.weight_loader |
| 1709 | weight_loader(param, loaded_weight, shard_id) |
| 1710 | loaded_params.add(name) |
| 1711 | break |
| 1712 | else: |
| 1713 | if ".experts." in name: |
| 1714 | # E8M0 scales are stored as float8_e8m0fnu in |
| 1715 | # checkpoints but the MoE param is uint8. copy_() |
| 1716 | # would do a numeric conversion (e.g. 2^-7 → 0), |
| 1717 | # destroying the raw exponent bytes. |
| 1718 | if ( |
| 1719 | "weight_scale" in name |
| 1720 | and loaded_weight.dtype == torch.float8_e8m0fnu |
| 1721 | ): |
| 1722 | loaded_weight = loaded_weight.view(torch.uint8) |
| 1723 | for mapping in expert_mapping: |
| 1724 | param_name, weight_name, expert_id, expert_shard_id = mapping |
| 1725 | if weight_name not in name: |
| 1726 | continue |
| 1727 | name_mapped = name.replace(weight_name, param_name) |
| 1728 | if is_pp_missing_parameter(name_mapped, self): |
| 1729 | continue |
| 1730 | param = params_dict[name_mapped] |
| 1731 | # We should ask the weight loader to return success or not |
| 1732 | # here since otherwise we may skip experts with other |
| 1733 | # available replicas. |
| 1734 | weight_loader = typing.cast( |
| 1735 | Callable[..., bool], param.weight_loader |
| 1736 | ) |
| 1737 | success = weight_loader( |
| 1738 | param, |
| 1739 | loaded_weight, |
| 1740 | name_mapped, |
| 1741 | shard_id=expert_shard_id, |
| 1742 | expert_id=expert_id, |
| 1743 | return_success=True, |
| 1744 | ) |
| 1745 | if success: |
| 1746 | name = name_mapped |
| 1747 | break |
| 1748 | loaded_params.add(name_mapped) |
| 1749 | continue |
| 1750 | elif "attn_sink" in name: |
| 1751 | if is_pp_missing_parameter(name, self): |
| 1752 | continue |
| 1753 | narrow_weight = loaded_weight[head_rank_start:head_rank_end] |
| 1754 | n = narrow_weight.shape[0] |
| 1755 | params_dict[name][:n].copy_(narrow_weight) |
| 1756 | loaded_params.add(name) |
| 1757 | continue |
| 1758 | else: |
| 1759 | if is_pp_missing_parameter(name, self): |
| 1760 | continue |
| 1761 | # Non-LoRA params on a LoRA-wrapped module live at |
| 1762 | # ``<head>.base_layer.<leaf>``; the checkpoint is plain. |
| 1763 | if name not in params_dict: |
| 1764 | head, _, leaf = name.rpartition(".") |
| 1765 | suffixed = f"{head}.base_layer.{leaf}" |
| 1766 | if suffixed in params_dict: |
| 1767 | name = suffixed |
| 1768 | param = params_dict[name] |
| 1769 | weight_loader = getattr( |
| 1770 | param, "weight_loader", default_weight_loader |
| 1771 | ) |
| 1772 | weight_loader(param, loaded_weight) |
| 1773 | loaded_params.add(name) |
| 1774 | continue |
| 1775 | |
| 1776 | return loaded_params |
| 1777 | |
| 1778 | @staticmethod |
| 1779 | def _pad_shared_expert_weight( |
| 1780 | quant_config: QuantizationConfig | None, |
| 1781 | name: str, |
| 1782 | loaded_weight: torch.Tensor, |
| 1783 | ) -> torch.Tensor: |
| 1784 | """Zero-pad a block-FP8 shared-expert weight/scale on its intermediate |
| 1785 | axis so the standard TP loaders split it into even, block-aligned shards |
| 1786 | (trailing ranks get the zero pad). gate (w1)/up (w3) [I, H] pad dim 0; |
| 1787 | down (w2 -> down_proj) [H, I] pads dim 1. |
| 1788 | """ |
| 1789 | block_size = getattr(quant_config, "weight_block_size", None) |
| 1790 | assert block_size is not None |
| 1791 | # Round the intermediate axis up to a whole number of TP shards. The axis |
| 1792 | # is in elements for weights (step = block) and in blocks for scales. |
| 1793 | step = 1 if name.endswith("weight_scale_inv") else block_size[0] |
| 1794 | dim = 1 if ".down_proj." in name else 0 |
| 1795 | mult = get_tensor_model_parallel_world_size() * step |
| 1796 | pad = cdiv(loaded_weight.shape[dim], mult) * mult - loaded_weight.shape[dim] |
| 1797 | if pad == 0: |
| 1798 | return loaded_weight |
| 1799 | pad_shape = list(loaded_weight.shape) |
| 1800 | pad_shape[dim] = pad |
| 1801 | return torch.cat([loaded_weight, loaded_weight.new_zeros(pad_shape)], dim=dim) |
| 1802 | |
| 1803 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 1804 | first_layer = next(iter(islice(self.layers, self.start_layer, self.end_layer))) |
| 1805 | if first_layer.ffn.use_mega_moe: |
| 1806 | return make_deepseek_v4_expert_params_mapping(self.config.n_routed_experts) |
| 1807 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 1808 | # (param_name, weight_name, expert_id, shard_id) |
| 1809 | return fused_moe_make_expert_params_mapping( |
| 1810 | self, |
| 1811 | ckpt_gate_proj_name="w1", |
| 1812 | ckpt_down_proj_name="w2", |
| 1813 | ckpt_up_proj_name="w3", |
| 1814 | num_experts=self.config.n_routed_experts, |
| 1815 | ) |
| 1816 | |
| 1817 | def finalize_mega_moe_weights(self) -> None: |
| 1818 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 1819 | layer.ffn.finalize_mega_moe_weights() |
| 1820 | |
| 1821 | def finalize_mhc_broadcast_weights(self) -> None: |
| 1822 | if not get_pp_group().is_first_rank or self.start_layer >= self.end_layer: |
| 1823 | return |
| 1824 | layer = self.layers[self.start_layer] |
| 1825 | if isinstance(layer, DeepseekV4DecoderLayer): |
| 1826 | broadcast = ( |
| 1827 | layer.hc_attn_fn.detach() |
| 1828 | .view(-1, layer.hc_mult, layer.hidden_size) |
| 1829 | .sum(dim=1) |
| 1830 | ) |
| 1831 | if layer.hc_attn_fn_broadcast is None: |
| 1832 | layer.hc_attn_fn_broadcast = broadcast |
| 1833 | else: |
| 1834 | layer.hc_attn_fn_broadcast.copy_(broadcast) |
| 1835 | |
| 1836 | |
| 1837 | def _make_deepseek_v4_weights_mapper(expert_dtype: str) -> WeightsMapper: |
| 1838 | if expert_dtype == "fp4": |
| 1839 | # MXFP4 experts use Mxfp4MoEMethod, which registers scales as |
| 1840 | # ``w{1,2,3}_weight_scale`` (no _inv suffix). FP8 linear and |
| 1841 | # shared experts use Fp8LinearMethod's block scales, which |
| 1842 | # register as ``weight_scale_inv``. |
| 1843 | scale_regex = { |
| 1844 | # ``.base_layer.``-namespace variant (LoRA-wrapped experts). |
| 1845 | re.compile( |
| 1846 | r"(\.experts\.\d+\.w[123]\.base_layer)\.scale$" |
| 1847 | ): r"\1.weight_scale", |
| 1848 | re.compile(r"(\.experts\.\d+\.w[123])\.scale$"): r"\1.weight_scale", |
| 1849 | re.compile(r"\.scale$"): ".weight_scale_inv", |
| 1850 | } |
| 1851 | else: |
| 1852 | # FP8 experts use Fp8MoEMethod (block_quant=True), which registers |
| 1853 | # scales as ``w{13,2}_weight_scale_inv``. Map all ``.scale`` keys |
| 1854 | # there. |
| 1855 | scale_regex = { |
| 1856 | # ``.base_layer.``-namespace variant of the above. |
| 1857 | re.compile( |
| 1858 | r"(\.experts\.\d+\.w[123]\.base_layer)\.scale$" |
| 1859 | ): r"\1.weight_scale_inv", |
| 1860 | re.compile(r"\.scale$"): ".weight_scale_inv", |
| 1861 | } |
| 1862 | return WeightsMapper( |
| 1863 | orig_to_new_prefix={ |
| 1864 | "layers.": "model.layers.", |
| 1865 | "embed.": "model.embed.", |
| 1866 | "norm.": "model.norm.", |
| 1867 | "hc_head": "model.hc_head", |
| 1868 | "mtp.": "model.mtp.", |
| 1869 | }, |
| 1870 | orig_to_new_regex=scale_regex, |
| 1871 | orig_to_new_suffix={ |
| 1872 | "head.weight": "lm_head.weight", |
| 1873 | "embed.weight": "embed_tokens.weight", |
| 1874 | ".ffn.gate.bias": ".ffn.gate.e_score_correction_bias", |
| 1875 | }, |
| 1876 | orig_to_new_substr={ |
| 1877 | ".shared_experts.w2": ".shared_experts.down_proj", |
| 1878 | "mtp.": None, |
| 1879 | }, |
| 1880 | ) |
| 1881 | |
| 1882 | |
| 1883 | class DeepseekV4MixtureOfExperts(MixtureOfExperts): |
| 1884 | moe_mlp_layers: list["DeepseekV4MoE"] |
| 1885 | |
| 1886 | def extract_moe_parameters(self, example_moe: "DeepseekV4MoE | None") -> None: |
| 1887 | if example_moe is None: |
| 1888 | self.num_moe_layers = 0 |
| 1889 | self.num_expert_groups = 0 |
| 1890 | self.num_logical_experts = 0 |
| 1891 | self.num_physical_experts = 0 |
| 1892 | self.num_local_physical_experts = 0 |
| 1893 | self.num_routed_experts = 0 |
| 1894 | self.num_shared_experts = 0 |
| 1895 | self.num_redundant_experts = 0 |
| 1896 | return |
| 1897 | self.num_logical_experts = example_moe.n_logical_experts |
| 1898 | self.num_physical_experts = example_moe.n_physical_experts |
| 1899 | self.num_local_physical_experts = example_moe.n_local_physical_experts |
| 1900 | self.num_routed_experts = example_moe.n_routed_experts |
| 1901 | self.num_shared_experts = example_moe.n_shared_experts |
| 1902 | self.num_redundant_experts = example_moe.n_redundant_experts |
| 1903 | |
| 1904 | def update_physical_experts_metadata( |
| 1905 | self, |
| 1906 | num_physical_experts: int, |
| 1907 | num_local_physical_experts: int, |
| 1908 | ) -> None: |
| 1909 | assert self.num_local_physical_experts == num_local_physical_experts |
| 1910 | self.num_physical_experts = num_physical_experts |
| 1911 | self.num_local_physical_experts = num_local_physical_experts |
| 1912 | self.num_redundant_experts = num_physical_experts - self.num_logical_experts |
| 1913 | for moe in self.moe_mlp_layers: |
| 1914 | moe.n_local_physical_experts = num_local_physical_experts |
| 1915 | moe.n_physical_experts = num_physical_experts |
| 1916 | moe.n_redundant_experts = self.num_redundant_experts |
| 1917 | moe.experts.update_expert_map() |
| 1918 | |
| 1919 | |
| 1920 | class DeepseekV4ForCausalLM( |
| 1921 | nn.Module, |
| 1922 | SupportsPP, |
| 1923 | SupportsEagle3, |
| 1924 | SupportsLoRA, |
| 1925 | DeepseekV4MixtureOfExperts, |
| 1926 | ): |
| 1927 | finalizes_weights_during_load = True |
| 1928 | model_cls = DeepseekV4Model |
| 1929 | |
| 1930 | # Default mapper assumes the original FP4-expert checkpoint layout. |
| 1931 | # Overridden per-instance in __init__ when expert_dtype != "fp4". |
| 1932 | hf_to_vllm_mapper = _make_deepseek_v4_weights_mapper("fp4") |
| 1933 | |
| 1934 | packed_modules_mapping = { |
| 1935 | "gate_up_proj": ["w1", "w3"], |
| 1936 | "fused_wqa_wkv": ["wq_a", "wkv"], |
| 1937 | "fused_wkv_wgate": ["wkv", "wgate"], |
| 1938 | } |
| 1939 | |
| 1940 | # The MTP draft head is not LoRA-adapted. |
| 1941 | lora_skip_prefixes = ["mtp."] |
| 1942 | |
| 1943 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1944 | super().__init__() |
| 1945 | |
| 1946 | config = vllm_config.model_config.hf_config |
| 1947 | self.config = config |
| 1948 | expert_dtype = getattr(config, "expert_dtype", "fp4") |
| 1949 | if expert_dtype != "fp4": |
| 1950 | self.hf_to_vllm_mapper = _make_deepseek_v4_weights_mapper(expert_dtype) |
| 1951 | |
| 1952 | self.model = self.model_cls( |
| 1953 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 1954 | ) |
| 1955 | if get_pp_group().is_last_rank: |
| 1956 | self.lm_head = ParallelLMHead( |
| 1957 | config.vocab_size, |
| 1958 | config.hidden_size, |
| 1959 | prefix=maybe_prefix(prefix, "lm_head"), |
| 1960 | ) |
| 1961 | else: |
| 1962 | self.lm_head = PPMissingLayer() |
| 1963 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 1964 | self.make_empty_intermediate_tensors = ( # type: ignore[method-assign] |
| 1965 | self.model.make_empty_intermediate_tensors |
| 1966 | ) |
| 1967 | |
| 1968 | self.set_moe_parameters() |
| 1969 | |
| 1970 | def set_moe_parameters(self) -> None: |
| 1971 | self.num_expert_groups = getattr(self.config, "n_group", 1) |
| 1972 | self.num_moe_layers = self.config.num_hidden_layers |
| 1973 | self.moe_layers: list[nn.Module] = [] |
| 1974 | self.moe_mlp_layers: list[DeepseekV4MoE] = [] |
| 1975 | example_moe: DeepseekV4MoE | None = None |
| 1976 | for layer in self.model.layers: |
| 1977 | if isinstance(layer, PPMissingLayer): |
| 1978 | continue |
| 1979 | if not isinstance(layer, DeepseekV4DecoderLayer): |
| 1980 | continue |
| 1981 | if isinstance(layer.ffn, DeepseekV4MoE): |
| 1982 | example_moe = layer.ffn |
| 1983 | self.moe_mlp_layers.append(layer.ffn) |
| 1984 | self.moe_layers.append(layer.ffn.experts) |
| 1985 | |
| 1986 | self.num_moe_layers = len(self.moe_layers) |
| 1987 | self.extract_moe_parameters(example_moe) |
| 1988 | |
| 1989 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1990 | return self.model.embed_input_ids(input_ids) |
| 1991 | |
| 1992 | def compute_logits( |
| 1993 | self, |
| 1994 | hidden_states: torch.Tensor, |
| 1995 | ) -> torch.Tensor | None: |
| 1996 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 1997 | return logits |
| 1998 | |
| 1999 | def compute_logits_local( |
| 2000 | self, |
| 2001 | hidden_states: torch.Tensor, |
| 2002 | ) -> torch.Tensor: |
| 2003 | return self.logits_processor(self.lm_head, hidden_states, skip_gather=True) |
| 2004 | |
| 2005 | def forward( |
| 2006 | self, |
| 2007 | input_ids: torch.Tensor, |
| 2008 | positions: torch.Tensor, |
| 2009 | intermediate_tensors: IntermediateTensors | None = None, |
| 2010 | inputs_embeds: torch.Tensor | None = None, |
| 2011 | ) -> torch.Tensor | IntermediateTensors: |
| 2012 | hidden_states = self.model( |
| 2013 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 2014 | ) |
| 2015 | return hidden_states |
| 2016 | |
| 2017 | def get_mtp_target_hidden_states(self) -> torch.Tensor | None: |
| 2018 | """Pre-hc_head residual stream buffer (max_num_batched_tokens, |
| 2019 | hc_mult * hidden_size) for the MTP draft model. Populated by |
| 2020 | forward(); valid after each target step.""" |
| 2021 | return getattr(self.model, "_mtp_hidden_buffer", None) |
| 2022 | |
| 2023 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 2024 | loader = AutoWeightsLoader(self) |
| 2025 | loaded_params = loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 2026 | self.process_weights_after_loading() |
| 2027 | return loaded_params |
| 2028 | |
| 2029 | def process_weights_after_loading(self) -> None: |
| 2030 | self.model.finalize_mega_moe_weights() |
| 2031 | self.model.finalize_mhc_broadcast_weights() |
| 2032 | |
| 2033 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 2034 | return self.model.get_expert_mapping() |