| 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 itertools import islice |
| 6 | |
| 7 | import regex as re |
| 8 | import torch |
| 9 | import torch.nn as nn |
| 10 | |
| 11 | import vllm.envs as envs |
| 12 | from vllm.config import VllmConfig |
| 13 | from vllm.distributed import ( |
| 14 | get_pp_group, |
| 15 | get_tensor_model_parallel_rank, |
| 16 | get_tensor_model_parallel_world_size, |
| 17 | ) |
| 18 | from vllm.forward_context import get_forward_context, is_forward_context_available |
| 19 | from vllm.logger import init_logger |
| 20 | from vllm.model_executor.layers.fused_moe import ( |
| 21 | fused_moe_make_expert_params_mapping, |
| 22 | ) |
| 23 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 24 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 25 | from vllm.model_executor.layers.mhc import ( |
| 26 | HAS_AITER_MHC_FUSED_POST_PRE_DELAYED_RMS_NORM, |
| 27 | MHCPostOp, |
| 28 | MHCPreDelayedOp, |
| 29 | ) |
| 30 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 31 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 32 | ParallelLMHead, |
| 33 | VocabParallelEmbedding, |
| 34 | ) |
| 35 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 36 | from vllm.model_executor.models.interfaces import ( |
| 37 | EagleModelMixin, |
| 38 | MixtureOfExperts, |
| 39 | SupportsEagle3, |
| 40 | SupportsLoRA, |
| 41 | SupportsPP, |
| 42 | ) |
| 43 | from vllm.model_executor.models.utils import ( |
| 44 | AutoWeightsLoader, |
| 45 | PPMissingLayer, |
| 46 | WeightsMapper, |
| 47 | extract_layer_index, |
| 48 | is_pp_missing_parameter, |
| 49 | make_layers, |
| 50 | maybe_prefix, |
| 51 | ) |
| 52 | from vllm.models.common.ops.sequence_parallel import ( |
| 53 | sp_all_gather, |
| 54 | sp_padding_mask, |
| 55 | sp_reduce_scatter, |
| 56 | sp_shard, |
| 57 | ) |
| 58 | from vllm.models.deepseek_v4.amd.model import ( |
| 59 | DeepseekV4MoE as DeepseekV4MoEBase, |
| 60 | ) |
| 61 | from vllm.models.deepseek_v41.amd.rocm import DeepseekV41ROCMAiterMLAAttention |
| 62 | from vllm.models.deepseek_v41.attention import DeepseekV4Attention |
| 63 | from vllm.sequence import IntermediateTensors |
| 64 | from vllm.utils.math_utils import cdiv |
| 65 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 66 | from vllm.v1.worker.ubatching import dbo_current_ubatch_id |
| 67 | |
| 68 | from ..common.engram import Engram, EngramLayout, NgramHashState |
| 69 | from ..common.mm_preprocess import IMAGE_SENTINEL_BASE_ID, image_sentinel_mask |
| 70 | |
| 71 | if typing.TYPE_CHECKING: |
| 72 | from vllm.v1.attention.backends.mla.sparse_swa import DeepseekSparseSWAMetadata |
| 73 | |
| 74 | logger = init_logger(__name__) |
| 75 | |
| 76 | |
| 77 | class DeepseekV4MoE(DeepseekV4MoEBase): |
| 78 | def __init__( |
| 79 | self, |
| 80 | vllm_config: VllmConfig, |
| 81 | prefix: str = "", |
| 82 | use_sequence_parallel: bool = False, |
| 83 | ): |
| 84 | config = vllm_config.model_config.hf_config |
| 85 | n_routed_experts = config.n_routed_experts |
| 86 | n_activated_experts = config.num_experts_per_tok |
| 87 | if extract_layer_index(prefix) >= config.num_hidden_layers: |
| 88 | n_routed_experts = ( |
| 89 | getattr(config, "dspark_n_routed_experts", 0) or n_routed_experts |
| 90 | ) |
| 91 | n_activated_experts = ( |
| 92 | getattr(config, "dspark_num_experts_per_tok", 0) or n_activated_experts |
| 93 | ) |
| 94 | # The reused ROCm V4 MoE reads expert counts from the HF config. Draft |
| 95 | # layers use V4.1's smaller DSpark expert set, so temporarily expose |
| 96 | # the per-layer counts while constructing the module, then restore the |
| 97 | # shared config for the remaining backbone layers. |
| 98 | old_routed = config.n_routed_experts |
| 99 | old_activated = config.num_experts_per_tok |
| 100 | old_hash_layers = getattr(config, "num_hash_layers", None) |
| 101 | config.n_routed_experts = n_routed_experts |
| 102 | config.num_experts_per_tok = n_activated_experts |
| 103 | config.num_hash_layers = 0 |
| 104 | try: |
| 105 | super().__init__(vllm_config, prefix=prefix) |
| 106 | finally: |
| 107 | config.n_routed_experts = old_routed |
| 108 | config.num_experts_per_tok = old_activated |
| 109 | if old_hash_layers is None: |
| 110 | delattr(config, "num_hash_layers") |
| 111 | else: |
| 112 | config.num_hash_layers = old_hash_layers |
| 113 | |
| 114 | # The V4.1 checkpoint contains a second per-expert routing bias for |
| 115 | # image sentinel tokens. The reused ROCm V4 MoE predates this field, |
| 116 | # so register it here and attach the same Parameter to the router, |
| 117 | # matching the CUDA V4.1 implementation. Text tokens continue to use |
| 118 | # e_score_correction_bias. |
| 119 | self.gate.bias_vl = None |
| 120 | self.image_sentinel_lo = 0 |
| 121 | if getattr(config, "vision_n_layers", 0) > 0: |
| 122 | self.gate.bias_vl = nn.Parameter( |
| 123 | torch.empty(self.n_routed_experts, dtype=torch.float32), |
| 124 | requires_grad=False, |
| 125 | ) |
| 126 | self.image_sentinel_lo = IMAGE_SENTINEL_BASE_ID |
| 127 | self.experts.router.bias_vl = self.gate.bias_vl |
| 128 | self.experts.router.image_sentinel_lo = self.image_sentinel_lo |
| 129 | self.use_mega_moe = False |
| 130 | # Expose the MixtureOfExperts protocol fields expected by the V4.1 |
| 131 | # outer model. The current ROCm runner has no redundant EPLB copies, |
| 132 | # so logical and physical expert counts are identical. |
| 133 | self.n_logical_experts = self.n_routed_experts |
| 134 | self.n_physical_experts = self.n_routed_experts |
| 135 | self.n_local_physical_experts = self.n_local_experts |
| 136 | self.n_redundant_experts = 0 |
| 137 | |
| 138 | |
| 139 | def _select_dsv4_attn_cls(vllm_config: VllmConfig) -> type[DeepseekV4Attention]: |
| 140 | """Pick the ROCm sparse-MLA implementation for DeepSeek V4.1.""" |
| 141 | backend = vllm_config.attention_config.backend |
| 142 | if backend is not None and ( |
| 143 | backend is not AttentionBackendEnum.ROCM_FLASHMLA_SPARSE_DSV4 |
| 144 | ): |
| 145 | raise ValueError( |
| 146 | f"{backend.name} is not supported for DeepSeek V4.1 on ROCm; " |
| 147 | "use ROCM_FLASHMLA_SPARSE_DSV4." |
| 148 | ) |
| 149 | return DeepseekV41ROCMAiterMLAAttention |
| 150 | |
| 151 | |
| 152 | def _use_sequence_parallel(vllm_config: VllmConfig) -> bool: |
| 153 | # The ROCm fused-MoE path currently uses ordinary tensor/expert parallel. |
| 154 | return False |
| 155 | |
| 156 | |
| 157 | class DeepseekV4DecoderLayer(nn.Module): |
| 158 | def __init__( |
| 159 | self, |
| 160 | vllm_config, |
| 161 | prefix, |
| 162 | topk_indices_buffer: torch.Tensor | None = None, |
| 163 | aux_stream_list: list[torch.cuda.Stream] | None = None, |
| 164 | candidate_block_buffer: torch.Tensor | None = None, |
| 165 | engram_layout: EngramLayout | None = None, |
| 166 | ): |
| 167 | super().__init__() |
| 168 | config = vllm_config.model_config.hf_config |
| 169 | self.hidden_size = config.hidden_size |
| 170 | self.use_sequence_parallel = _use_sequence_parallel(vllm_config) |
| 171 | |
| 172 | self.engram: Engram | None = None |
| 173 | if engram_layout is not None: |
| 174 | layer_id = extract_layer_index(prefix) |
| 175 | if layer_id in engram_layout.layer_ids: |
| 176 | self.engram = Engram( |
| 177 | config, |
| 178 | vllm_config.quant_config, |
| 179 | engram_layout, |
| 180 | engram_layout.layer_ids.index(layer_id), |
| 181 | use_sequence_parallel=self.use_sequence_parallel, |
| 182 | prefix=f"{prefix}.engram", |
| 183 | ) |
| 184 | |
| 185 | self.rms_norm_eps = config.rms_norm_eps |
| 186 | self.attn = _select_dsv4_attn_cls(vllm_config)( |
| 187 | vllm_config, |
| 188 | prefix=f"{prefix}.attn", |
| 189 | topk_indices_buffer=topk_indices_buffer, |
| 190 | aux_stream_list=aux_stream_list, |
| 191 | candidate_block_buffer=candidate_block_buffer, |
| 192 | ) |
| 193 | if self.use_sequence_parallel: |
| 194 | self.attn.wo_b.reduce_results = False |
| 195 | self.ffn = DeepseekV4MoE( |
| 196 | vllm_config, |
| 197 | prefix=f"{prefix}.ffn", |
| 198 | use_sequence_parallel=self.use_sequence_parallel, |
| 199 | ) |
| 200 | |
| 201 | self.attn_norm = RMSNorm(self.hidden_size, self.rms_norm_eps) |
| 202 | self.ffn_norm = RMSNorm(self.hidden_size, self.rms_norm_eps) |
| 203 | self.hc_mult = config.hc_mult |
| 204 | self.hc_sinkhorn_iters = config.hc_sinkhorn_iters |
| 205 | self.hc_eps = config.hc_eps |
| 206 | self.hc_post_alpha = 2.0 |
| 207 | mix_hc = (2 + self.hc_mult) * self.hc_mult |
| 208 | hc_dim = self.hc_mult * self.hidden_size |
| 209 | self.hc_attn_fn = nn.Parameter( |
| 210 | torch.empty( |
| 211 | (mix_hc, hc_dim), |
| 212 | dtype=torch.float32, |
| 213 | ), |
| 214 | requires_grad=False, |
| 215 | ) |
| 216 | self.hc_attn_fn_broadcast: torch.Tensor | None = None |
| 217 | self.hc_ffn_fn = nn.Parameter( |
| 218 | torch.empty( |
| 219 | (mix_hc, hc_dim), |
| 220 | dtype=torch.float32, |
| 221 | ), |
| 222 | requires_grad=False, |
| 223 | ) |
| 224 | self.hc_attn_base = nn.Parameter( |
| 225 | torch.empty( |
| 226 | mix_hc, |
| 227 | dtype=torch.float32, |
| 228 | ), |
| 229 | requires_grad=False, |
| 230 | ) |
| 231 | self.hc_ffn_base = nn.Parameter( |
| 232 | torch.empty( |
| 233 | mix_hc, |
| 234 | dtype=torch.float32, |
| 235 | ), |
| 236 | requires_grad=False, |
| 237 | ) |
| 238 | self.hc_attn_scale = nn.Parameter( |
| 239 | torch.empty( |
| 240 | 3, |
| 241 | dtype=torch.float32, |
| 242 | ), |
| 243 | requires_grad=False, |
| 244 | ) |
| 245 | self.hc_ffn_scale = nn.Parameter( |
| 246 | torch.empty( |
| 247 | 3, |
| 248 | dtype=torch.float32, |
| 249 | ), |
| 250 | requires_grad=False, |
| 251 | ) |
| 252 | self.mhc_pre_delayed = MHCPreDelayedOp() |
| 253 | self.mhc_post = MHCPostOp() |
| 254 | # Where aiter's fused seam kernel runs (gfx950), it folds the following |
| 255 | # attn_norm / ffn_norm into its collapse, so the separate norms are |
| 256 | # skipped for the seams it takes. |
| 257 | self.fuse_seam_norm = HAS_AITER_MHC_FUSED_POST_PRE_DELAYED_RMS_NORM |
| 258 | |
| 259 | @staticmethod |
| 260 | def _hc_collapse(x: torch.Tensor, pre_mix: torch.Tensor) -> torch.Tensor: |
| 261 | """Hyper-connection collapse used by DSpark on ROCm.""" |
| 262 | return torch.ops.vllm.hc_collapse_triton(x, pre_mix) |
| 263 | |
| 264 | def forward( |
| 265 | self, |
| 266 | x: torch.Tensor, |
| 267 | positions: torch.Tensor, |
| 268 | input_ids: torch.Tensor | None, |
| 269 | pre_mix: torch.Tensor | None = None, |
| 270 | post_mix: torch.Tensor | None = None, |
| 271 | res_mix: torch.Tensor | None = None, |
| 272 | residual: torch.Tensor | None = None, |
| 273 | engram_hashes: torch.Tensor | None = None, |
| 274 | engram_mask: torch.Tensor | None = None, |
| 275 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: |
| 276 | # Layer 0's attention seam projects the 2-D embedding with the folded |
| 277 | # hc_attn_fn_broadcast instead of the 4-stream residual with hc_attn_fn. |
| 278 | # The fused kernel only takes the latter, so that seam keeps the |
| 279 | # separate attn_norm. |
| 280 | fuse_attn_norm = self.fuse_seam_norm and not (residual is None and x.dim() == 2) |
| 281 | # The reference collapses each sublayer's input with the *previous* |
| 282 | # sublayer's pre-mix: attention uses the pre-mix carried in (identity |
| 283 | # for the first layer), the FFN uses this layer's attention pre-mix. |
| 284 | if residual is None: |
| 285 | if x.dim() == 2: |
| 286 | # First layer: the stream is the embedding broadcast to hc |
| 287 | # copies and the identity pre-mix selects copy 0. |
| 288 | assert self.hc_attn_fn_broadcast is not None |
| 289 | residual = x.unsqueeze(1).expand(-1, self.hc_mult, -1).contiguous() |
| 290 | residual, post_mix, res_mix, x, attn_pre = self.mhc_pre_delayed( |
| 291 | residual, |
| 292 | self.hc_attn_fn_broadcast, |
| 293 | self.hc_attn_scale, |
| 294 | self.hc_attn_base, |
| 295 | self.rms_norm_eps, |
| 296 | self.hc_eps, |
| 297 | self.hc_eps, |
| 298 | self.hc_post_alpha, |
| 299 | self.hc_sinkhorn_iters, |
| 300 | x=x, |
| 301 | ) |
| 302 | else: |
| 303 | residual = x |
| 304 | residual, post_mix, res_mix, x, attn_pre = self.mhc_pre_delayed( |
| 305 | residual, |
| 306 | self.hc_attn_fn, |
| 307 | self.hc_attn_scale, |
| 308 | self.hc_attn_base, |
| 309 | self.rms_norm_eps, |
| 310 | self.hc_eps, |
| 311 | self.hc_eps, |
| 312 | self.hc_post_alpha, |
| 313 | self.hc_sinkhorn_iters, |
| 314 | pre_mix=pre_mix, |
| 315 | norm_weight=self.attn_norm.weight if self.fuse_seam_norm else None, |
| 316 | norm_eps=self.attn_norm.variance_epsilon, |
| 317 | ) |
| 318 | else: |
| 319 | pre_args = ( |
| 320 | self.hc_attn_fn, |
| 321 | self.hc_attn_scale, |
| 322 | self.hc_attn_base, |
| 323 | self.rms_norm_eps, |
| 324 | self.hc_eps, |
| 325 | self.hc_eps, |
| 326 | self.hc_post_alpha, |
| 327 | self.hc_sinkhorn_iters, |
| 328 | ) |
| 329 | if self.engram is not None and engram_hashes is not None: |
| 330 | # Engram injection happens between the previous sublayer's |
| 331 | # post and this block's pre, on the full hc stream, so the |
| 332 | # mix coefficients see the injected stream. That read of the |
| 333 | # residual is what stops the post from fusing into the pre |
| 334 | # here, unlike the attention seam below. |
| 335 | residual = self.mhc_post(x, residual, post_mix, res_mix) |
| 336 | residual = self.engram( |
| 337 | residual, |
| 338 | engram_hashes[:, self.engram.layer_hash_index], |
| 339 | engram_mask, |
| 340 | ) |
| 341 | residual, post_mix, res_mix, x, attn_pre = self.mhc_pre_delayed( |
| 342 | residual, |
| 343 | *pre_args, |
| 344 | pre_mix=pre_mix, |
| 345 | norm_weight=self.attn_norm.weight if self.fuse_seam_norm else None, |
| 346 | norm_eps=self.attn_norm.variance_epsilon, |
| 347 | ) |
| 348 | else: |
| 349 | ( |
| 350 | residual, |
| 351 | post_mix, |
| 352 | res_mix, |
| 353 | x, |
| 354 | attn_pre, |
| 355 | ) = self.mhc_pre_delayed( |
| 356 | residual, |
| 357 | *pre_args, |
| 358 | pre_mix=pre_mix, |
| 359 | sublayer_out=x, |
| 360 | post_layer_mix=post_mix, |
| 361 | comb_res_mix=res_mix, |
| 362 | norm_weight=self.attn_norm.weight if self.fuse_seam_norm else None, |
| 363 | norm_eps=self.attn_norm.variance_epsilon, |
| 364 | ) |
| 365 | if not fuse_attn_norm: |
| 366 | x = self.attn_norm(x) |
| 367 | |
| 368 | if self.use_sequence_parallel: |
| 369 | x = sp_all_gather(x)[: positions.shape[0]] |
| 370 | |
| 371 | x = self.attn(positions, x, None) |
| 372 | if self.use_sequence_parallel: |
| 373 | x = sp_reduce_scatter(x) |
| 374 | |
| 375 | residual, post_mix, res_mix, x, ffn_pre = self.mhc_pre_delayed( |
| 376 | residual, |
| 377 | self.hc_ffn_fn, |
| 378 | self.hc_ffn_scale, |
| 379 | self.hc_ffn_base, |
| 380 | self.rms_norm_eps, |
| 381 | self.hc_eps, |
| 382 | self.hc_eps, |
| 383 | self.hc_post_alpha, |
| 384 | self.hc_sinkhorn_iters, |
| 385 | pre_mix=attn_pre, |
| 386 | sublayer_out=x, |
| 387 | post_layer_mix=post_mix, |
| 388 | comb_res_mix=res_mix, |
| 389 | norm_weight=self.ffn_norm.weight if self.fuse_seam_norm else None, |
| 390 | norm_eps=self.ffn_norm.variance_epsilon, |
| 391 | ) |
| 392 | if not self.fuse_seam_norm: |
| 393 | x = self.ffn_norm(x) |
| 394 | x = self.ffn(x, input_ids) |
| 395 | return x, residual, post_mix, res_mix, ffn_pre |
| 396 | |
| 397 | |
| 398 | class DeepseekV4Model(nn.Module, EagleModelMixin): |
| 399 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 400 | super().__init__() |
| 401 | |
| 402 | config = vllm_config.model_config.hf_config |
| 403 | quant_config = vllm_config.quant_config |
| 404 | self.config = config |
| 405 | self.quant_config = quant_config |
| 406 | self.parallel_config = vllm_config.parallel_config |
| 407 | self.use_mega_moe = False |
| 408 | self.use_sequence_parallel = _use_sequence_parallel(vllm_config) |
| 409 | if self.use_mega_moe and not vllm_config.parallel_config.enable_expert_parallel: |
| 410 | raise NotImplementedError( |
| 411 | "DeepSeek V4 MegaMoE currently requires expert parallel. " |
| 412 | "Enable it with --enable-expert-parallel, or pick a different " |
| 413 | "moe backend." |
| 414 | ) |
| 415 | self.vocab_size = config.vocab_size |
| 416 | self.hc_eps = config.hc_eps |
| 417 | self.hc_mult = config.hc_mult |
| 418 | self.hc_dim = self.hc_mult * config.hidden_size |
| 419 | self.rms_norm_eps = config.rms_norm_eps |
| 420 | |
| 421 | # Three aux streams: one per non-default input GEMM in |
| 422 | # DeepseekV4Attention._run_parallel_input_projections |
| 423 | # (compressor kv_score, indexer.weights_proj). fused_wqa_wkv stays on |
| 424 | # the default stream. |
| 425 | aux_stream_list = [torch.cuda.Stream() for _ in range(3)] |
| 426 | |
| 427 | # Reserved topk indices buffer for all Indexer layers to reuse. |
| 428 | self.topk_indices_buffer = torch.empty( |
| 429 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 430 | config.index_topk, |
| 431 | dtype=torch.int32, |
| 432 | ) |
| 433 | |
| 434 | # Two-level candidate filtering: the indexer at |
| 435 | # candidate_source_layer_id publishes the top candidate blocks of |
| 436 | # compressed positions here; later ratio-1 indexers (24/28/32/36) |
| 437 | # mask their scores with it. |
| 438 | candidate_source_layer = getattr(config, "candidate_source_layer_id", -1) |
| 439 | candidate_topk_blocks = getattr(config, "candidate_topk_blocks", 0) |
| 440 | if candidate_source_layer >= 0 and candidate_topk_blocks > 0: |
| 441 | self.candidate_block_buffer = torch.empty( |
| 442 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 443 | candidate_topk_blocks, |
| 444 | dtype=torch.int32, |
| 445 | ) |
| 446 | else: |
| 447 | self.candidate_block_buffer = None |
| 448 | |
| 449 | if get_pp_group().is_first_rank: |
| 450 | self.embed_tokens = VocabParallelEmbedding( |
| 451 | config.vocab_size, |
| 452 | config.hidden_size, |
| 453 | quant_config=quant_config, |
| 454 | prefix=f"{prefix}.embed_tokens", |
| 455 | ) |
| 456 | else: |
| 457 | self.embed_tokens = PPMissingLayer() |
| 458 | |
| 459 | self.engram_layout = EngramLayout.from_config(config) |
| 460 | |
| 461 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 462 | config.num_hidden_layers, |
| 463 | lambda prefix: DeepseekV4DecoderLayer( |
| 464 | vllm_config, |
| 465 | prefix=prefix, |
| 466 | topk_indices_buffer=self.topk_indices_buffer, |
| 467 | aux_stream_list=aux_stream_list, |
| 468 | candidate_block_buffer=self.candidate_block_buffer, |
| 469 | engram_layout=self.engram_layout, |
| 470 | ), |
| 471 | prefix=f"{prefix}.layers", |
| 472 | ) |
| 473 | |
| 474 | # The n-gram hash needs a slot-keyed rolling store of compressed ids |
| 475 | # (chunked prefill / decode lookback); key it off the first local |
| 476 | # layer's sliding-window KV cache. Only PP ranks owning an engram |
| 477 | # layer need it. |
| 478 | self.engram_hash: NgramHashState | None = None |
| 479 | self.engram_swa_prefix: str | None = None |
| 480 | if self.engram_layout is not None: |
| 481 | local_engram = any( |
| 482 | isinstance(layer, DeepseekV4DecoderLayer) and layer.engram is not None |
| 483 | for layer in islice(self.layers, self.start_layer, self.end_layer) |
| 484 | ) |
| 485 | if local_engram: |
| 486 | first_layer = next( |
| 487 | iter(islice(self.layers, self.start_layer, self.end_layer)) |
| 488 | ) |
| 489 | swa_cache_module = first_layer.attn.swa_cache_layer |
| 490 | self.engram_hash = NgramHashState( |
| 491 | vllm_config, self.engram_layout, swa_cache_module |
| 492 | ) |
| 493 | self.engram_swa_prefix = swa_cache_module.prefix |
| 494 | |
| 495 | if get_pp_group().is_last_rank: |
| 496 | self.norm = RMSNorm(config.hidden_size, self.rms_norm_eps) |
| 497 | else: |
| 498 | self.norm = PPMissingLayer() |
| 499 | |
| 500 | self.mhc_post = MHCPostOp() |
| 501 | |
| 502 | spec_config = vllm_config.speculative_config |
| 503 | needs_mtp_hidden_states = spec_config is not None and ( |
| 504 | spec_config.use_eagle() or spec_config.uses_draft_model() |
| 505 | ) |
| 506 | if get_pp_group().is_last_rank and needs_mtp_hidden_states: |
| 507 | self._mtp_hidden_buffer = torch.empty( |
| 508 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 509 | self.hc_dim, |
| 510 | dtype=vllm_config.model_config.dtype, |
| 511 | ) |
| 512 | else: |
| 513 | self._mtp_hidden_buffer = None |
| 514 | |
| 515 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 516 | return self.embed_tokens(input_ids) |
| 517 | |
| 518 | def make_empty_intermediate_tensors( |
| 519 | self, |
| 520 | batch_size: int, |
| 521 | dtype: torch.dtype, |
| 522 | device: torch.device, |
| 523 | ) -> IntermediateTensors: |
| 524 | # PP intermediate tensors carry the multi-stream hidden_states |
| 525 | # of shape (num_tokens, hc_mult, hidden_size) — V4 expands the |
| 526 | # token embedding to hc_mult streams before the first decoder |
| 527 | # layer and keeps that shape until the final hc collapse — plus the |
| 528 | # (num_tokens, hc_mult) pre-mix the next rank's first layer needs |
| 529 | # for its attention collapse. |
| 530 | return IntermediateTensors( |
| 531 | { |
| 532 | "hidden_states": torch.zeros( |
| 533 | (batch_size, self.hc_mult, self.config.hidden_size), |
| 534 | dtype=dtype, |
| 535 | device=device, |
| 536 | ), |
| 537 | "pre_mix": torch.zeros( |
| 538 | (batch_size, self.hc_mult), |
| 539 | dtype=torch.float32, |
| 540 | device=device, |
| 541 | ), |
| 542 | } |
| 543 | ) |
| 544 | |
| 545 | def forward( |
| 546 | self, |
| 547 | input_ids: torch.Tensor, |
| 548 | positions: torch.Tensor, |
| 549 | intermediate_tensors: IntermediateTensors | None, |
| 550 | inputs_embeds: torch.Tensor | None = None, |
| 551 | lookback_token_ids: torch.Tensor | None = None, |
| 552 | ) -> torch.Tensor | IntermediateTensors: |
| 553 | if get_pp_group().is_first_rank: |
| 554 | if inputs_embeds is not None: |
| 555 | hidden_states = inputs_embeds |
| 556 | else: |
| 557 | hidden_states = self.embed_input_ids(input_ids) |
| 558 | else: |
| 559 | assert intermediate_tensors is not None |
| 560 | hidden_states = intermediate_tensors["hidden_states"] |
| 561 | |
| 562 | if self.use_mega_moe: |
| 563 | input_ids = input_ids.to(torch.int64) |
| 564 | |
| 565 | # Engram n-gram hashes for the whole (flattened) batch, computed once |
| 566 | # on the full token stream — before any sequence-parallel sharding — |
| 567 | # and consumed by the engram layers (1 and 14) below. Skipped on |
| 568 | # profile runs (KV cache unbound -> no slot space to key the rolling |
| 569 | # compressed-id cache into). |
| 570 | engram_hashes: torch.Tensor | None = None |
| 571 | engram_mask: torch.Tensor | None = None |
| 572 | if ( |
| 573 | self.engram_hash is not None |
| 574 | and input_ids is not None |
| 575 | and is_forward_context_available() |
| 576 | ): |
| 577 | attn_metadata = get_forward_context().attn_metadata |
| 578 | if isinstance(attn_metadata, list): |
| 579 | attn_metadata = attn_metadata[dbo_current_ubatch_id()] |
| 580 | if isinstance(attn_metadata, dict) and self.engram_hash.ensure_cache(): |
| 581 | assert self.engram_swa_prefix is not None |
| 582 | swa_metadata = typing.cast( |
| 583 | "DeepseekSparseSWAMetadata", attn_metadata[self.engram_swa_prefix] |
| 584 | ) |
| 585 | # Image-span tokens are dead: they break n-grams (hash op |
| 586 | # takes True=dead) and their gate is zeroed (Engram.forward |
| 587 | # takes True=keep). |
| 588 | image_mask = image_sentinel_mask(input_ids) |
| 589 | engram_mask = ~image_mask |
| 590 | if lookback_token_ids is None: |
| 591 | if not self.engram_hash.use_slot_cache: |
| 592 | raise NotImplementedError( |
| 593 | "engram needs `lookback_token_ids` from the model " |
| 594 | "runner (the DBO/ubatch wrapper drops model kwargs)" |
| 595 | ) |
| 596 | num_reqs = swa_metadata.num_decodes + swa_metadata.num_prefills |
| 597 | lookback_token_ids = input_ids.new_full( |
| 598 | (num_reqs, self.engram_hash.lookback_depth), -1 |
| 599 | ) |
| 600 | engram_hashes = self.engram_hash( |
| 601 | input_ids, |
| 602 | positions, |
| 603 | swa_metadata.query_start_loc, |
| 604 | image_mask, |
| 605 | lookback_token_ids, |
| 606 | image_sentinel_mask(lookback_token_ids), |
| 607 | swa_metadata.slot_mapping, |
| 608 | swa_metadata.block_table, |
| 609 | ) |
| 610 | # Gather all Engram rows before entering the decoder layers. |
| 611 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 612 | engram = getattr(layer, "engram", None) |
| 613 | if engram is not None: |
| 614 | engram.prepare_embeddings( |
| 615 | engram_hashes[:, engram.layer_hash_index] |
| 616 | ) |
| 617 | |
| 618 | full_num_tokens = positions.shape[0] |
| 619 | if self.use_sequence_parallel: |
| 620 | if envs.VLLM_MOE_SKIP_PADDING and is_forward_context_available(): |
| 621 | forward_context = get_forward_context() |
| 622 | forward_context.is_padding = sp_padding_mask( |
| 623 | forward_context.is_padding, hidden_states |
| 624 | ) |
| 625 | hidden_states = sp_shard(hidden_states) |
| 626 | input_ids = sp_shard(input_ids) |
| 627 | |
| 628 | residual, post_mix, res_mix = None, None, None |
| 629 | pre_mix: torch.Tensor | None = None |
| 630 | if not get_pp_group().is_first_rank: |
| 631 | assert intermediate_tensors is not None |
| 632 | pre_mix = intermediate_tensors["pre_mix"] |
| 633 | aux_hidden_states: list[torch.Tensor] = [] |
| 634 | final_aux_recon: torch.Tensor | None = None # avoid duplicate mhc_post call |
| 635 | for idx, layer in enumerate( |
| 636 | islice(self.layers, self.start_layer, self.end_layer), |
| 637 | start=self.start_layer, |
| 638 | ): |
| 639 | hidden_states, residual, post_mix, res_mix, pre_mix = layer( |
| 640 | hidden_states, |
| 641 | positions, |
| 642 | input_ids, |
| 643 | pre_mix, |
| 644 | post_mix, |
| 645 | res_mix, |
| 646 | residual, |
| 647 | engram_hashes, |
| 648 | engram_mask, |
| 649 | ) |
| 650 | if idx + 1 in self.aux_hidden_state_layers: |
| 651 | # Reconstruct the aux hidden state for draft models |
| 652 | aux_recon = self.mhc_post(hidden_states, residual, post_mix, res_mix) |
| 653 | aux_hidden_state = aux_recon.mean(dim=1) |
| 654 | if self.use_sequence_parallel: |
| 655 | aux_hidden_state = sp_all_gather(aux_hidden_state)[:full_num_tokens] |
| 656 | aux_hidden_states.append(aux_hidden_state) |
| 657 | final_aux_recon = aux_recon |
| 658 | if layer is not None: |
| 659 | # Reuse if the last layer was captured as an aux hidden state |
| 660 | if self.end_layer in self.aux_hidden_state_layers: |
| 661 | hidden_states = final_aux_recon |
| 662 | else: |
| 663 | hidden_states = self.mhc_post( |
| 664 | hidden_states, residual, post_mix, res_mix |
| 665 | ) |
| 666 | |
| 667 | if not get_pp_group().is_last_rank: |
| 668 | return IntermediateTensors( |
| 669 | {"hidden_states": hidden_states, "pre_mix": pre_mix} |
| 670 | ) |
| 671 | |
| 672 | if self.use_sequence_parallel: |
| 673 | hidden_states = sp_all_gather(hidden_states)[:full_num_tokens] |
| 674 | pre_mix = sp_all_gather(pre_mix)[:full_num_tokens] |
| 675 | |
| 676 | if self._mtp_hidden_buffer is not None: |
| 677 | num_tokens = hidden_states.shape[0] |
| 678 | self._mtp_hidden_buffer[:num_tokens].copy_(hidden_states.flatten(1)) |
| 679 | |
| 680 | # Collapse the hc copies with the pre-mix from the last layer's FFN |
| 681 | # mixes — the mix the reference applies via |
| 682 | # ``last_layer.hc_pre(h, pre_mix)`` (v4.1 has no learned hc_head). |
| 683 | assert pre_mix is not None |
| 684 | hidden_states = torch.ops.vllm.hc_collapse_triton(hidden_states, pre_mix) |
| 685 | hidden_states = self.norm(hidden_states) |
| 686 | if len(aux_hidden_states) > 0: |
| 687 | return hidden_states, aux_hidden_states |
| 688 | return hidden_states |
| 689 | |
| 690 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 691 | stacked_params_mapping = [ |
| 692 | # (param_name, shard_name, shard_id) |
| 693 | ("gate_up_proj", "w1", 0), |
| 694 | ("gate_up_proj", "w3", 1), |
| 695 | ("attn.fused_wqa_wkv", "attn.wq_a", 0), |
| 696 | ("attn.fused_wqa_wkv", "attn.wkv", 1), |
| 697 | ("compressor.fused_wkv_wgate", "compressor.wkv", 0), |
| 698 | ("compressor.fused_wkv_wgate", "compressor.wgate", 1), |
| 699 | ] |
| 700 | params_dict = dict(self.named_parameters()) |
| 701 | loaded_params: set[str] = set() |
| 702 | |
| 703 | def _resolve_param_name(name: str) -> str: |
| 704 | """Resolve scale naming across ROCm FP8 and MXFP8 methods.""" |
| 705 | if name.endswith("_inv"): |
| 706 | direct_name = name.removesuffix("_inv") |
| 707 | if direct_name in params_dict: |
| 708 | return direct_name |
| 709 | inv_name = f"{name}_inv" |
| 710 | if name not in params_dict and inv_name in params_dict: |
| 711 | return inv_name |
| 712 | return name |
| 713 | |
| 714 | # TP for attention |
| 715 | tp_size = get_tensor_model_parallel_world_size() |
| 716 | tp_rank = get_tensor_model_parallel_rank() |
| 717 | n_head = self.config.num_attention_heads |
| 718 | n_local_head = n_head // tp_size |
| 719 | head_rank_start = n_local_head * tp_rank |
| 720 | head_rank_end = n_local_head * (tp_rank + 1) |
| 721 | |
| 722 | # Pre-compute expert mapping ONCE. |
| 723 | expert_mapping = self.get_expert_mapping() |
| 724 | |
| 725 | # Block-FP8 shared experts: pad the intermediate up to the TP-uniform |
| 726 | # block count so the standard loaders below slice it evenly (trailing |
| 727 | # ranks land on the zero pad). SP / unquantized ones need no padding. |
| 728 | pad_shared_expert = ( |
| 729 | getattr(self.quant_config, "weight_block_size", None) is not None |
| 730 | and not self.use_sequence_parallel |
| 731 | ) |
| 732 | |
| 733 | for name, loaded_weight in weights: |
| 734 | if name.startswith(("vision.", "aligner.", "image_")): |
| 735 | # Vision weights are loaded by the outer multimodal wrapper. |
| 736 | logger.warning_once("Skipping non-text weight: %s", name) |
| 737 | continue |
| 738 | if pad_shared_expert and ".shared_experts." in name: |
| 739 | loaded_weight = self._pad_shared_expert_weight( |
| 740 | self.quant_config, name, loaded_weight |
| 741 | ) |
| 742 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 743 | # Skip non-stacked layers and experts (experts handled below). |
| 744 | if ".experts." in name: |
| 745 | continue |
| 746 | if weight_name not in name: |
| 747 | continue |
| 748 | name = name.replace(weight_name, param_name) |
| 749 | |
| 750 | if is_pp_missing_parameter(name, self): |
| 751 | break |
| 752 | if name not in params_dict: |
| 753 | head, _, leaf = name.rpartition(".") |
| 754 | suffixed = f"{head}.base_layer.{leaf}" |
| 755 | if suffixed in params_dict: |
| 756 | name = suffixed |
| 757 | name = _resolve_param_name(name) |
| 758 | param = params_dict[name] |
| 759 | weight_loader = param.weight_loader |
| 760 | weight_loader(param, loaded_weight, shard_id) |
| 761 | loaded_params.add(name) |
| 762 | break |
| 763 | else: |
| 764 | if ".experts." in name: |
| 765 | # E8M0 scales are stored as float8_e8m0fnu in |
| 766 | # checkpoints but the MoE param is uint8. copy_() |
| 767 | # would do a numeric conversion (e.g. 2^-7 → 0), |
| 768 | # destroying the raw exponent bytes. |
| 769 | if ( |
| 770 | "weight_scale" in name |
| 771 | and loaded_weight.dtype == torch.float8_e8m0fnu |
| 772 | ): |
| 773 | loaded_weight = loaded_weight.view(torch.uint8) |
| 774 | for mapping in expert_mapping: |
| 775 | param_name, weight_name, expert_id, expert_shard_id = mapping |
| 776 | if weight_name not in name: |
| 777 | continue |
| 778 | name_mapped = name.replace(weight_name, param_name) |
| 779 | if is_pp_missing_parameter(name_mapped, self): |
| 780 | continue |
| 781 | name_mapped = _resolve_param_name(name_mapped) |
| 782 | param = params_dict[name_mapped] |
| 783 | # We should ask the weight loader to return success or not |
| 784 | # here since otherwise we may skip experts with other |
| 785 | # available replicas. |
| 786 | weight_loader = typing.cast( |
| 787 | Callable[..., bool], param.weight_loader |
| 788 | ) |
| 789 | success = weight_loader( |
| 790 | param, |
| 791 | loaded_weight, |
| 792 | name_mapped, |
| 793 | shard_id=expert_shard_id, |
| 794 | expert_id=expert_id, |
| 795 | return_success=True, |
| 796 | ) |
| 797 | if success: |
| 798 | name = name_mapped |
| 799 | break |
| 800 | loaded_params.add(name_mapped) |
| 801 | continue |
| 802 | elif "attn_sink" in name: |
| 803 | if is_pp_missing_parameter(name, self): |
| 804 | continue |
| 805 | narrow_weight = loaded_weight[head_rank_start:head_rank_end] |
| 806 | n = narrow_weight.shape[0] |
| 807 | params_dict[name][:n].copy_(narrow_weight) |
| 808 | loaded_params.add(name) |
| 809 | continue |
| 810 | else: |
| 811 | if is_pp_missing_parameter(name, self): |
| 812 | continue |
| 813 | # Non-LoRA params on a LoRA-wrapped module live at |
| 814 | # ``<head>.base_layer.<leaf>``; the checkpoint is plain. |
| 815 | if name not in params_dict: |
| 816 | head, _, leaf = name.rpartition(".") |
| 817 | suffixed = f"{head}.base_layer.{leaf}" |
| 818 | if suffixed in params_dict: |
| 819 | name = suffixed |
| 820 | name = _resolve_param_name(name) |
| 821 | param = params_dict[name] |
| 822 | weight_loader = getattr( |
| 823 | param, "weight_loader", default_weight_loader |
| 824 | ) |
| 825 | weight_loader(param, loaded_weight) |
| 826 | loaded_params.add(name) |
| 827 | continue |
| 828 | |
| 829 | return loaded_params |
| 830 | |
| 831 | @staticmethod |
| 832 | def _pad_shared_expert_weight( |
| 833 | quant_config: QuantizationConfig | None, |
| 834 | name: str, |
| 835 | loaded_weight: torch.Tensor, |
| 836 | ) -> torch.Tensor: |
| 837 | """Zero-pad a block-FP8 shared-expert weight/scale on its intermediate |
| 838 | axis so the standard TP loaders split it into even, block-aligned shards |
| 839 | (trailing ranks get the zero pad). gate (w1)/up (w3) [I, H] pad dim 0; |
| 840 | down (w2 -> down_proj) [H, I] pads dim 1. |
| 841 | """ |
| 842 | block_size = getattr(quant_config, "weight_block_size", None) |
| 843 | assert block_size is not None |
| 844 | # Round the intermediate axis up to a whole number of TP shards. The axis |
| 845 | # is in elements for weights (step = block) and in blocks for scales. |
| 846 | step = ( |
| 847 | 1 if name.endswith(("weight_scale_inv", "weight_scale")) else block_size[0] |
| 848 | ) |
| 849 | dim = 1 if ".down_proj." in name else 0 |
| 850 | mult = get_tensor_model_parallel_world_size() * step |
| 851 | pad = cdiv(loaded_weight.shape[dim], mult) * mult - loaded_weight.shape[dim] |
| 852 | if pad == 0: |
| 853 | return loaded_weight |
| 854 | pad_shape = list(loaded_weight.shape) |
| 855 | pad_shape[dim] = pad |
| 856 | return torch.cat([loaded_weight, loaded_weight.new_zeros(pad_shape)], dim=dim) |
| 857 | |
| 858 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 859 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 860 | # (param_name, weight_name, expert_id, shard_id) |
| 861 | return fused_moe_make_expert_params_mapping( |
| 862 | self, |
| 863 | ckpt_gate_proj_name="w1", |
| 864 | ckpt_down_proj_name="w2", |
| 865 | ckpt_up_proj_name="w3", |
| 866 | num_experts=self.config.n_routed_experts, |
| 867 | ) |
| 868 | |
| 869 | def finalize_mega_moe_weights(self) -> None: |
| 870 | return |
| 871 | |
| 872 | def finalize_mhc_broadcast_weights(self) -> None: |
| 873 | if not get_pp_group().is_first_rank or self.start_layer >= self.end_layer: |
| 874 | return |
| 875 | layer = self.layers[self.start_layer] |
| 876 | if isinstance(layer, DeepseekV4DecoderLayer): |
| 877 | broadcast = ( |
| 878 | layer.hc_attn_fn.detach() |
| 879 | .view(-1, layer.hc_mult, layer.hidden_size) |
| 880 | .sum(dim=1) |
| 881 | ) |
| 882 | if layer.hc_attn_fn_broadcast is None: |
| 883 | layer.hc_attn_fn_broadcast = broadcast |
| 884 | else: |
| 885 | layer.hc_attn_fn_broadcast.copy_(broadcast) |
| 886 | |
| 887 | |
| 888 | def _linear_scale_param_name(vllm_config: VllmConfig, expert_dtype: str) -> str: |
| 889 | """Parameter name the linear quant method registers for weight scales. |
| 890 | |
| 891 | Native MXFP8 checkpoints ([32, 32] blocks with MXFP4 experts) route linear |
| 892 | layers through ModelOptLinearMethod, which registers ``weight_scale``; |
| 893 | block-FP8 linear layers register ``weight_scale_inv``. |
| 894 | """ |
| 895 | use_mxfp8 = ( |
| 896 | getattr(vllm_config.quant_config, "weight_block_size", None) == [32, 32] |
| 897 | and expert_dtype == "fp4" |
| 898 | ) |
| 899 | return "weight_scale" if use_mxfp8 else "weight_scale_inv" |
| 900 | |
| 901 | |
| 902 | def _make_deepseek_v4_weights_mapper( |
| 903 | expert_dtype: str, linear_scale_name: str = "weight_scale_inv" |
| 904 | ) -> WeightsMapper: |
| 905 | if expert_dtype == "fp4": |
| 906 | # MXFP4 experts use Mxfp4MoEMethod, which registers scales as |
| 907 | # ``w{1,2,3}_weight_scale`` (no _inv suffix). Linear scales use the |
| 908 | # parameter name registered by their quantization method. |
| 909 | scale_regex = { |
| 910 | # ``.base_layer.``-namespace variant (LoRA-wrapped experts). |
| 911 | re.compile( |
| 912 | r"(\.experts\.\d+\.w[123]\.base_layer)\.scale$" |
| 913 | ): r"\1.weight_scale", |
| 914 | re.compile(r"(\.experts\.\d+\.w[123])\.scale$"): r"\1.weight_scale", |
| 915 | # The ``embed.weight`` -> ``embed_tokens.weight`` suffix rule |
| 916 | # renames the engram fp8 table but not its scale; route the |
| 917 | # scale explicitly to the same module. |
| 918 | re.compile(r"(engram\.embed)\.scale$"): r"\1_tokens.weight_scale_inv", |
| 919 | # Quark exports spell the same tensor ``embed.weight_scale``, |
| 920 | # which the ``\.scale$`` rules never match. |
| 921 | re.compile( |
| 922 | r"(engram\.embed)\.weight_scale$" |
| 923 | ): r"\1_tokens.weight_scale_inv", |
| 924 | re.compile(r"\.scale$"): f".{linear_scale_name}", |
| 925 | } |
| 926 | else: |
| 927 | # FP8 experts use Fp8MoEMethod (block_quant=True), which registers |
| 928 | # scales as ``w{13,2}_weight_scale_inv``. Map all ``.scale`` keys |
| 929 | # there. |
| 930 | scale_regex = { |
| 931 | # ``.base_layer.``-namespace variant of the above. |
| 932 | re.compile( |
| 933 | r"(\.experts\.\d+\.w[123]\.base_layer)\.scale$" |
| 934 | ): r"\1.weight_scale_inv", |
| 935 | # Same engram reroute as the fp4 branch above. |
| 936 | re.compile(r"(engram\.embed)\.scale$"): r"\1_tokens.weight_scale_inv", |
| 937 | # Quark exports spell the same tensor ``embed.weight_scale``, |
| 938 | # which the ``\.scale$`` rules never match. |
| 939 | re.compile( |
| 940 | r"(engram\.embed)\.weight_scale$" |
| 941 | ): r"\1_tokens.weight_scale_inv", |
| 942 | re.compile(r"\.scale$"): f".{linear_scale_name}", |
| 943 | } |
| 944 | return WeightsMapper( |
| 945 | orig_to_new_prefix={ |
| 946 | "layers.": "model.layers.", |
| 947 | "embed.": "model.embed.", |
| 948 | "norm.": "model.norm.", |
| 949 | "mtp.": "model.mtp.", |
| 950 | }, |
| 951 | orig_to_new_regex=scale_regex, |
| 952 | orig_to_new_suffix={ |
| 953 | "head.weight": "lm_head.weight", |
| 954 | "embed.weight": "embed_tokens.weight", |
| 955 | ".ffn.gate.bias": ".ffn.gate.e_score_correction_bias", |
| 956 | }, |
| 957 | orig_to_new_substr={ |
| 958 | ".shared_experts.w2": ".shared_experts.down_proj", |
| 959 | "mtp.": None, |
| 960 | # The v4.1 checkpoint declares the VL arch even for text-only |
| 961 | # use; the text backbone drops the vision-tower weights. |
| 962 | "vision.": None, |
| 963 | "aligner.": None, |
| 964 | "image_": None, |
| 965 | }, |
| 966 | ) |
| 967 | |
| 968 | |
| 969 | class DeepseekV4MixtureOfExperts(MixtureOfExperts): |
| 970 | moe_mlp_layers: list["DeepseekV4MoE"] |
| 971 | |
| 972 | def extract_moe_parameters(self, example_moe: "DeepseekV4MoE | None") -> None: |
| 973 | if example_moe is None: |
| 974 | self.num_moe_layers = 0 |
| 975 | self.num_expert_groups = 0 |
| 976 | self.num_logical_experts = 0 |
| 977 | self.num_physical_experts = 0 |
| 978 | self.num_local_physical_experts = 0 |
| 979 | self.num_routed_experts = 0 |
| 980 | self.num_shared_experts = 0 |
| 981 | self.num_redundant_experts = 0 |
| 982 | return |
| 983 | self.num_logical_experts = example_moe.n_logical_experts |
| 984 | self.num_physical_experts = example_moe.n_physical_experts |
| 985 | self.num_local_physical_experts = example_moe.n_local_physical_experts |
| 986 | self.num_routed_experts = example_moe.n_routed_experts |
| 987 | self.num_shared_experts = example_moe.n_shared_experts |
| 988 | self.num_redundant_experts = example_moe.n_redundant_experts |
| 989 | |
| 990 | def update_physical_experts_metadata( |
| 991 | self, |
| 992 | num_physical_experts: int, |
| 993 | num_local_physical_experts: int, |
| 994 | ) -> None: |
| 995 | assert self.num_local_physical_experts == num_local_physical_experts |
| 996 | self.num_physical_experts = num_physical_experts |
| 997 | self.num_local_physical_experts = num_local_physical_experts |
| 998 | self.num_redundant_experts = num_physical_experts - self.num_logical_experts |
| 999 | for moe in self.moe_mlp_layers: |
| 1000 | moe.n_local_physical_experts = num_local_physical_experts |
| 1001 | moe.n_physical_experts = num_physical_experts |
| 1002 | moe.n_redundant_experts = self.num_redundant_experts |
| 1003 | moe.experts.update_expert_map() |
| 1004 | |
| 1005 | |
| 1006 | class DeepseekV41LLMForCausalLM( |
| 1007 | nn.Module, |
| 1008 | SupportsPP, |
| 1009 | SupportsEagle3, |
| 1010 | SupportsLoRA, |
| 1011 | DeepseekV4MixtureOfExperts, |
| 1012 | ): |
| 1013 | model_cls = DeepseekV4Model |
| 1014 | |
| 1015 | # Default mapper assumes the original FP4-expert checkpoint layout. |
| 1016 | # Overridden per-instance in __init__ when expert_dtype != "fp4". |
| 1017 | hf_to_vllm_mapper = _make_deepseek_v4_weights_mapper("fp4") |
| 1018 | |
| 1019 | packed_modules_mapping = { |
| 1020 | "gate_up_proj": ["w1", "w3"], |
| 1021 | "fused_wqa_wkv": ["wq_a", "wkv"], |
| 1022 | "fused_wkv_wgate": ["wkv", "wgate"], |
| 1023 | } |
| 1024 | |
| 1025 | # The MTP draft head is not LoRA-adapted. |
| 1026 | lora_skip_prefixes = ["mtp."] |
| 1027 | |
| 1028 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1029 | super().__init__() |
| 1030 | |
| 1031 | config = vllm_config.model_config.hf_config |
| 1032 | self.config = config |
| 1033 | expert_dtype = getattr(config, "expert_dtype", "fp4") |
| 1034 | self.hf_to_vllm_mapper = _make_deepseek_v4_weights_mapper( |
| 1035 | expert_dtype, _linear_scale_param_name(vllm_config, expert_dtype) |
| 1036 | ) |
| 1037 | |
| 1038 | self.model = self.model_cls( |
| 1039 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 1040 | ) |
| 1041 | if get_pp_group().is_last_rank: |
| 1042 | self.lm_head = ParallelLMHead( |
| 1043 | config.vocab_size, |
| 1044 | config.hidden_size, |
| 1045 | prefix=maybe_prefix(prefix, "lm_head"), |
| 1046 | ) |
| 1047 | else: |
| 1048 | self.lm_head = PPMissingLayer() |
| 1049 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 1050 | self.make_empty_intermediate_tensors = ( # type: ignore[method-assign] |
| 1051 | self.model.make_empty_intermediate_tensors |
| 1052 | ) |
| 1053 | |
| 1054 | self.set_moe_parameters() |
| 1055 | |
| 1056 | def set_moe_parameters(self) -> None: |
| 1057 | self.num_expert_groups = getattr(self.config, "n_group", 1) |
| 1058 | self.num_moe_layers = self.config.num_hidden_layers |
| 1059 | self.moe_layers: list[nn.Module] = [] |
| 1060 | self.moe_mlp_layers: list[DeepseekV4MoE] = [] |
| 1061 | example_moe: DeepseekV4MoE | None = None |
| 1062 | for layer in self.model.layers: |
| 1063 | if isinstance(layer, PPMissingLayer): |
| 1064 | continue |
| 1065 | if not isinstance(layer, DeepseekV4DecoderLayer): |
| 1066 | continue |
| 1067 | if isinstance(layer.ffn, DeepseekV4MoE): |
| 1068 | example_moe = layer.ffn |
| 1069 | self.moe_mlp_layers.append(layer.ffn) |
| 1070 | self.moe_layers.append(layer.ffn.experts) |
| 1071 | |
| 1072 | self.num_moe_layers = len(self.moe_layers) |
| 1073 | self.extract_moe_parameters(example_moe) |
| 1074 | |
| 1075 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1076 | return self.model.embed_input_ids(input_ids) |
| 1077 | |
| 1078 | def compute_logits( |
| 1079 | self, |
| 1080 | hidden_states: torch.Tensor, |
| 1081 | ) -> torch.Tensor | None: |
| 1082 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 1083 | return logits |
| 1084 | |
| 1085 | def compute_logits_local( |
| 1086 | self, |
| 1087 | hidden_states: torch.Tensor, |
| 1088 | ) -> torch.Tensor: |
| 1089 | return self.logits_processor(self.lm_head, hidden_states, skip_gather=True) |
| 1090 | |
| 1091 | @staticmethod |
| 1092 | def get_model_state_cls(): |
| 1093 | from ..nvidia.model_state import DeepseekV41ModelState |
| 1094 | |
| 1095 | return DeepseekV41ModelState |
| 1096 | |
| 1097 | @property |
| 1098 | def token_lookback_depth(self) -> int: |
| 1099 | """Tokens before a chunk start the engram hash needs; the model runner |
| 1100 | passes them as `lookback_token_ids`.""" |
| 1101 | engram_hash = self.model.engram_hash |
| 1102 | return engram_hash.lookback_depth if engram_hash is not None else 0 |
| 1103 | |
| 1104 | def forward( |
| 1105 | self, |
| 1106 | input_ids: torch.Tensor, |
| 1107 | positions: torch.Tensor, |
| 1108 | intermediate_tensors: IntermediateTensors | None = None, |
| 1109 | inputs_embeds: torch.Tensor | None = None, |
| 1110 | lookback_token_ids: torch.Tensor | None = None, |
| 1111 | ) -> torch.Tensor | IntermediateTensors: |
| 1112 | hidden_states = self.model( |
| 1113 | input_ids, |
| 1114 | positions, |
| 1115 | intermediate_tensors, |
| 1116 | inputs_embeds, |
| 1117 | lookback_token_ids=lookback_token_ids, |
| 1118 | ) |
| 1119 | return hidden_states |
| 1120 | |
| 1121 | def get_mtp_target_hidden_states(self) -> torch.Tensor | None: |
| 1122 | """Pre-collapse residual stream buffer (max_num_batched_tokens, |
| 1123 | hc_mult * hidden_size) for the MTP draft model. Populated by |
| 1124 | forward(); valid after each target step.""" |
| 1125 | return getattr(self.model, "_mtp_hidden_buffer", None) |
| 1126 | |
| 1127 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1128 | loader = AutoWeightsLoader(self) |
| 1129 | loaded_params = loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 1130 | self.process_weights_after_loading() |
| 1131 | return loaded_params |
| 1132 | |
| 1133 | def process_weights_after_loading(self) -> None: |
| 1134 | self.model.finalize_mega_moe_weights() |
| 1135 | self.model.finalize_mhc_broadcast_weights() |
| 1136 | |
| 1137 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 1138 | return self.model.get_expert_mapping() |