| 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 functools import partial |
| 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_engram_dp_size, |
| 17 | get_pp_group, |
| 18 | get_tensor_model_parallel_rank, |
| 19 | get_tensor_model_parallel_world_size, |
| 20 | tensor_model_parallel_all_reduce, |
| 21 | ) |
| 22 | from vllm.forward_context import ( |
| 23 | get_forward_context, |
| 24 | in_piecewise_cudagraph, |
| 25 | is_forward_context_available, |
| 26 | ) |
| 27 | from vllm.logger import init_logger |
| 28 | from vllm.model_executor.kernels.mhc.tilelang import ( |
| 29 | mhc_post_tilelang, |
| 30 | mhc_pre_delayed_tilelang, |
| 31 | ) |
| 32 | from vllm.model_executor.kernels.mhc.triton import hc_collapse_triton |
| 33 | from vllm.model_executor.layers.fused_moe import ( |
| 34 | fused_moe_make_expert_params_mapping, |
| 35 | ) |
| 36 | from vllm.model_executor.layers.fused_moe.experts.trtllm_mxfp4_moe import ( |
| 37 | TrtLlmMxfp4ExpertsModular, |
| 38 | ) |
| 39 | from vllm.model_executor.layers.fused_moe.moe_output import ( |
| 40 | MoEOutput, |
| 41 | UnfinalizedMoEOutput, |
| 42 | ) |
| 43 | from vllm.model_executor.layers.fused_moe.runner.moe_runner import _unpack |
| 44 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 45 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 46 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 47 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 48 | ParallelLMHead, |
| 49 | VocabParallelEmbedding, |
| 50 | ) |
| 51 | from vllm.model_executor.model_loader.weight_utils import ( |
| 52 | default_weight_loader, |
| 53 | drop_checkpoint_cache, |
| 54 | ) |
| 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 | ) |
| 71 | from vllm.models.common.ops.sequence_parallel import ( |
| 72 | sp_all_gather, |
| 73 | sp_padding_mask, |
| 74 | sp_reduce_scatter, |
| 75 | sp_shard, |
| 76 | ) |
| 77 | from vllm.models.deepseek_v4.nvidia.model import ( |
| 78 | DeepseekV4MoE as DeepseekV4MoEBase, |
| 79 | ) |
| 80 | from vllm.models.deepseek_v4.nvidia.model import ( |
| 81 | MegaGateRoutingMetadata, |
| 82 | make_deepseek_v4_expert_params_mapping, |
| 83 | prepare_mega_gate_routing_metadata, |
| 84 | ) |
| 85 | from vllm.models.deepseek_v41.attention import DeepseekV4Attention |
| 86 | from vllm.models.deepseek_v41.decoder_replay_layers import DecoderReplayLayers |
| 87 | from vllm.models.deepseek_v41.nvidia.flash_mla_mega_attn import ( |
| 88 | DeepseekV4MegaAttnAttention, |
| 89 | ) |
| 90 | from vllm.models.deepseek_v41.nvidia.flashinfer_sparse import ( |
| 91 | DeepseekV4FlashInferMLAAttention, |
| 92 | DeepseekV4FlashInferSM120Attention, |
| 93 | ) |
| 94 | from vllm.models.deepseek_v41.nvidia.flashmla import DeepseekV4FlashMLAAttention |
| 95 | from vllm.platforms import current_platform |
| 96 | from vllm.sequence import IntermediateTensors |
| 97 | from vllm.utils.math_utils import cdiv |
| 98 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 99 | from vllm.v1.worker.ubatching import dbo_current_ubatch_id |
| 100 | |
| 101 | from ..common.engram import ( |
| 102 | Engram, |
| 103 | EngramLayout, |
| 104 | NgramHashState, |
| 105 | can_share_engram_tables, |
| 106 | gather_engram_hashes, |
| 107 | ) |
| 108 | from ..common.mm_preprocess import IMAGE_SENTINEL_BASE_ID, image_sentinel_mask |
| 109 | from .ops.mhc import ( |
| 110 | MHC_OVERLAP_MAX_TOKENS, |
| 111 | init_mhc_all_reduce, |
| 112 | mhc_pre_delayed_overlap, |
| 113 | mhc_shifted_post_pre, |
| 114 | supports_mhc_all_reduce, |
| 115 | supports_mhc_overlap, |
| 116 | ) |
| 117 | |
| 118 | if typing.TYPE_CHECKING: |
| 119 | from vllm.v1.attention.backends.mla.sparse_swa import DeepseekSparseSWAMetadata |
| 120 | |
| 121 | logger = init_logger(__name__) |
| 122 | |
| 123 | |
| 124 | class DeepseekV4MoE(DeepseekV4MoEBase): |
| 125 | def __init__( |
| 126 | self, |
| 127 | vllm_config: VllmConfig, |
| 128 | prefix: str = "", |
| 129 | use_sequence_parallel: bool = False, |
| 130 | reduce_results: bool = True, |
| 131 | ): |
| 132 | config = vllm_config.model_config.hf_config |
| 133 | n_routed_experts = config.n_routed_experts |
| 134 | n_activated_experts = config.num_experts_per_tok |
| 135 | if extract_layer_index(prefix) >= config.num_hidden_layers: |
| 136 | n_routed_experts = ( |
| 137 | getattr(config, "dspark_n_routed_experts", 0) or n_routed_experts |
| 138 | ) |
| 139 | n_activated_experts = ( |
| 140 | getattr(config, "dspark_num_experts_per_tok", 0) or n_activated_experts |
| 141 | ) |
| 142 | super().__init__( |
| 143 | vllm_config, |
| 144 | prefix=prefix, |
| 145 | use_sequence_parallel=use_sequence_parallel, |
| 146 | n_routed_experts=n_routed_experts, |
| 147 | n_activated_experts=n_activated_experts, |
| 148 | num_hash_layers=0, |
| 149 | reduce_results=reduce_results, |
| 150 | image_sentinel_lo=IMAGE_SENTINEL_BASE_ID, |
| 151 | ) |
| 152 | |
| 153 | def defer_finalize(self) -> None: |
| 154 | """Leave the routed top-k reduction to the next fused all-reduce + mHC. |
| 155 | |
| 156 | Only the modular TRTLLM MXFP4 experts can stop after GEMM2; the fused |
| 157 | kernel takes as many tokens as the mHC overlap path. |
| 158 | """ |
| 159 | moe_config = self.experts.moe_config |
| 160 | experts_cls = getattr( |
| 161 | self.experts.routed_experts.quant_method, "experts_cls", None |
| 162 | ) |
| 163 | if ( |
| 164 | experts_cls is TrtLlmMxfp4ExpertsModular |
| 165 | and self.experts.routed_scaling_factor == 1.0 |
| 166 | ): |
| 167 | moe_config.defer_moe_finalize(MHC_OVERLAP_MAX_TOKENS) |
| 168 | if moe_config.use_deferred_moe_finalize: |
| 169 | logger.info_once( |
| 170 | "DSV4.1 mHC: MoE top-k finalize fused into the all-reduce for " |
| 171 | "up to %d tokens.", |
| 172 | MHC_OVERLAP_MAX_TOKENS, |
| 173 | ) |
| 174 | |
| 175 | def defers_finalize(self, num_tokens: int) -> bool: |
| 176 | return ( |
| 177 | not self.use_mega_moe |
| 178 | and self.experts.moe_config.should_defer_moe_finalize(num_tokens) |
| 179 | ) |
| 180 | |
| 181 | def forward_unfinalized( |
| 182 | self, hidden_states: torch.Tensor, input_ids: torch.Tensor | None |
| 183 | ) -> MoEOutput: |
| 184 | """``forward`` with the routed top-k reduction and all-reduce left open.""" |
| 185 | # The runner's custom op returns tensors only, so run its body directly. |
| 186 | shared_output, routed = _unpack( |
| 187 | self.experts._forward_impl( |
| 188 | hidden_states, hidden_states, hidden_states, input_ids |
| 189 | ) |
| 190 | ) |
| 191 | assert shared_output is not None |
| 192 | assert isinstance(routed, UnfinalizedMoEOutput) |
| 193 | return MoEOutput(routed=routed, shared_output=shared_output) |
| 194 | |
| 195 | |
| 196 | def _select_dsv4_attn_cls(vllm_config: VllmConfig) -> type[DeepseekV4Attention]: |
| 197 | """Pick the CUDA sparse-MLA attention class for the configured backend. |
| 198 | |
| 199 | The generic CUDA backend selector does not instantiate DSv4 layers directly, |
| 200 | so map generic sparse-MLA choices to the DSv4-specialized attention class. |
| 201 | Without an explicit backend: SM12 takes FlashInfer, SM100 takes mega |
| 202 | attention where the topology allows it, and everything else keeps the |
| 203 | FlashMLA path. |
| 204 | """ |
| 205 | backend = vllm_config.attention_config.backend |
| 206 | device_capability = current_platform.get_device_capability() |
| 207 | if backend in ( |
| 208 | AttentionBackendEnum.FLASHINFER_MLA_SPARSE, |
| 209 | AttentionBackendEnum.FLASHINFER_MLA_SPARSE_SM120, |
| 210 | ): |
| 211 | raise ValueError( |
| 212 | f"{backend.name} is not a DeepSeek V4.1 attention backend. " |
| 213 | "Use FLASHINFER_MLA_SPARSE_DSV41 for DeepSeek V4.1 FlashInfer " |
| 214 | "sparse MLA." |
| 215 | ) |
| 216 | if backend in ( |
| 217 | AttentionBackendEnum.FLASHINFER_MLA_SPARSE_DSV4, |
| 218 | AttentionBackendEnum.FLASHINFER_MLA_SPARSE_DSV41, |
| 219 | ): |
| 220 | if device_capability is not None and device_capability.major == 12: |
| 221 | return DeepseekV4FlashInferSM120Attention |
| 222 | return DeepseekV4FlashInferMLAAttention |
| 223 | if backend is AttentionBackendEnum.FLASHMLA_MEGA_ATTN_DSV41: |
| 224 | return DeepseekV4MegaAttnAttention |
| 225 | if backend in ( |
| 226 | AttentionBackendEnum.FLASHMLA_SPARSE, |
| 227 | AttentionBackendEnum.FLASHMLA_SPARSE_DSV4, |
| 228 | AttentionBackendEnum.FLASHMLA_SPARSE_DSV41, |
| 229 | ): |
| 230 | return DeepseekV4FlashMLAAttention |
| 231 | |
| 232 | if device_capability is not None and device_capability.major == 12: |
| 233 | return DeepseekV4FlashInferSM120Attention |
| 234 | # Mega attention is the SM100 default: it fuses Q RoPE, sparse attention, |
| 235 | # the output's inverse RoPE and its FP8 cast into one launch, and brings |
| 236 | # the 288 B NVFP4 compressed record -- the format the reference |
| 237 | # implementation itself stores. It declines topologies it cannot serve |
| 238 | # (non-SM100, TP that leaves fewer than WV_GROUP_SIZE heads per wo_a |
| 239 | # group, a build without the kernel), which then fall through to FlashMLA. |
| 240 | if DeepseekV4MegaAttnAttention.is_available_for(vllm_config): |
| 241 | return DeepseekV4MegaAttnAttention |
| 242 | return DeepseekV4FlashMLAAttention |
| 243 | |
| 244 | |
| 245 | def _use_sequence_parallel(vllm_config: VllmConfig) -> bool: |
| 246 | parallel_config = vllm_config.parallel_config |
| 247 | use_mega_moe = vllm_config.kernel_config.moe_backend in MEGA_MOE_BACKENDS |
| 248 | return ( |
| 249 | parallel_config.pipeline_parallel_size == 1 |
| 250 | and parallel_config.enable_expert_parallel |
| 251 | and parallel_config.tensor_parallel_size > 1 |
| 252 | and (use_mega_moe or parallel_config.data_parallel_size > 1) |
| 253 | ) |
| 254 | |
| 255 | |
| 256 | def maybe_init_gemm_rs(vllm_config: VllmConfig, use_sequence_parallel: bool) -> bool: |
| 257 | """Set up the fused ``wo_b`` GEMM + reduce-scatter when opted in. |
| 258 | |
| 259 | Sequence parallel is the only topology where ``wo_b`` ends in a |
| 260 | reduce-scatter, so the kernel is bound to RS mode. The workspace is a |
| 261 | process-wide singleton shared with any other model code in this worker |
| 262 | (the DSpark drafter reuses it), and its NVLink multicast rendezvous is |
| 263 | collective, so every TP rank must take the same decision here. |
| 264 | """ |
| 265 | if not (use_sequence_parallel and envs.VLLM_ENABLE_GEMM_RS): |
| 266 | return False |
| 267 | |
| 268 | # The kernel module pulls in cute_dsl, so import it only once opted in. |
| 269 | from vllm.model_executor.kernels.linear.cute_dsl.gemm_rs_ar import ( |
| 270 | maybe_init_gemm_rs_ar, |
| 271 | ) |
| 272 | |
| 273 | hidden_size = vllm_config.model_config.hf_config.hidden_size |
| 274 | if not maybe_init_gemm_rs_ar(vllm_config, N=hidden_size, all_reduce=False): |
| 275 | return False |
| 276 | logger.info_once("To disable DeepSeek-V4.1 GEMM-RS, set VLLM_ENABLE_GEMM_RS=0.") |
| 277 | return True |
| 278 | |
| 279 | |
| 280 | class DeepseekV4DecoderLayer(nn.Module): |
| 281 | def __init__( |
| 282 | self, |
| 283 | vllm_config, |
| 284 | prefix, |
| 285 | topk_indices_buffer: torch.Tensor | None = None, |
| 286 | aux_stream_list: list[torch.cuda.Stream] | None = None, |
| 287 | candidate_block_buffer: torch.Tensor | None = None, |
| 288 | engram_layout: EngramLayout | None = None, |
| 289 | engram_prefetch_stream: torch.cuda.Stream | None = None, |
| 290 | run_gemm_rs: bool = False, |
| 291 | mhc_stream: torch.cuda.Stream | None = None, |
| 292 | fuse_mhc_all_reduce: bool = False, |
| 293 | ): |
| 294 | super().__init__() |
| 295 | |
| 296 | config = vllm_config.model_config.hf_config |
| 297 | self.hidden_size = config.hidden_size |
| 298 | self.mhc_stream = mhc_stream |
| 299 | self.fuse_mhc_all_reduce = fuse_mhc_all_reduce |
| 300 | self.use_sequence_parallel = _use_sequence_parallel(vllm_config) |
| 301 | |
| 302 | self.engram: Engram | None = None |
| 303 | if engram_layout is not None: |
| 304 | layer_id = extract_layer_index(prefix) |
| 305 | if layer_id in engram_layout.layer_ids: |
| 306 | self.engram = Engram( |
| 307 | config, |
| 308 | vllm_config.quant_config, |
| 309 | engram_layout, |
| 310 | engram_layout.layer_ids.index(layer_id), |
| 311 | use_sequence_parallel=self.use_sequence_parallel, |
| 312 | prefix=f"{prefix}.engram", |
| 313 | prefetch_stream=engram_prefetch_stream, |
| 314 | ) |
| 315 | |
| 316 | self.rms_norm_eps = config.rms_norm_eps |
| 317 | self.attn = _select_dsv4_attn_cls(vllm_config)( |
| 318 | vllm_config, |
| 319 | prefix=f"{prefix}.attn", |
| 320 | topk_indices_buffer=topk_indices_buffer, |
| 321 | aux_stream_list=aux_stream_list, |
| 322 | candidate_block_buffer=candidate_block_buffer, |
| 323 | ) |
| 324 | if self.use_sequence_parallel or fuse_mhc_all_reduce: |
| 325 | self.attn.wo_b.reduce_results = False |
| 326 | if run_gemm_rs: |
| 327 | # Binds only when wo_b's kernel and shape qualify; otherwise |
| 328 | # forward keeps the separate reduce-scatter below. |
| 329 | self.attn.bind_gemm_rs() |
| 330 | self.ffn = DeepseekV4MoE( |
| 331 | vllm_config, |
| 332 | prefix=f"{prefix}.ffn", |
| 333 | use_sequence_parallel=self.use_sequence_parallel, |
| 334 | reduce_results=not fuse_mhc_all_reduce, |
| 335 | ) |
| 336 | |
| 337 | self.attn_norm = RMSNorm(self.hidden_size, self.rms_norm_eps) |
| 338 | self.ffn_norm = RMSNorm(self.hidden_size, self.rms_norm_eps) |
| 339 | self.hc_mult = config.hc_mult |
| 340 | self.hc_sinkhorn_iters = config.hc_sinkhorn_iters |
| 341 | self.hc_eps = config.hc_eps |
| 342 | self.hc_post_alpha = 2.0 |
| 343 | if vllm_config.kernel_config.enable_jit_warmup and current_platform.is_cuda(): |
| 344 | from vllm.model_executor.kernels.mhc.tilelang_kernels import ( |
| 345 | mhc_fused_post_pre_splits, |
| 346 | ) |
| 347 | from vllm.model_executor.kernels.mhc.warmup import MHC_PRE_NORM_KERNEL |
| 348 | |
| 349 | max_tokens = vllm_config.scheduler_config.max_num_batched_tokens |
| 350 | if self.use_sequence_parallel: |
| 351 | tp_size = vllm_config.parallel_config.tensor_parallel_size |
| 352 | max_tokens = (max_tokens + tp_size - 1) // tp_size |
| 353 | # The epilogue compiles per projection width and per pre-mix mode. |
| 354 | # The first layer projects the broadcast embedding, so it reads one |
| 355 | # hidden_size-wide row and selects stream zero; every later sublayer |
| 356 | # projects the full hc stream, and collapses it with the pre-mix the |
| 357 | # previous sublayer carried in. Those later shapes also arrive |
| 358 | # through the fused post + pre-norm GEMM, which picks split-k |
| 359 | # factors the token sweep below never produces. |
| 360 | broadcast_embedding = { |
| 361 | "rms_numel": self.hidden_size, |
| 362 | "use_pre_mix_in": False, |
| 363 | "extra_splits": (), |
| 364 | } |
| 365 | hc_stream = { |
| 366 | "rms_numel": self.hc_mult * self.hidden_size, |
| 367 | "extra_splits": mhc_fused_post_pre_splits( |
| 368 | self.hidden_size, self.hc_mult |
| 369 | ), |
| 370 | } |
| 371 | variants = [ |
| 372 | broadcast_embedding, |
| 373 | {**hc_stream, "use_pre_mix_in": False}, |
| 374 | {**hc_stream, "use_pre_mix_in": True}, |
| 375 | ] |
| 376 | if vllm_config.speculative_config is not None: |
| 377 | # Draft setups read aux hidden states out of the same collapse, |
| 378 | # which is its own specialization of the kernel. |
| 379 | variants.append( |
| 380 | {**hc_stream, "use_pre_mix_in": True, "write_aux": True} |
| 381 | ) |
| 382 | # Shifted mHC splits the epilogue in two: the input collapse on the |
| 383 | # caller stream and coefficient generation on a side stream. |
| 384 | split_modes = ( |
| 385 | ("fused", "stats", "input") if mhc_stream is not None else ("fused",) |
| 386 | ) |
| 387 | for variant in variants: |
| 388 | for split_mode in split_modes: |
| 389 | MHC_PRE_NORM_KERNEL.register_warmup( |
| 390 | split_mode=split_mode, |
| 391 | max_tokens=( |
| 392 | max_tokens |
| 393 | if split_mode == "fused" |
| 394 | else min(max_tokens, MHC_OVERLAP_MAX_TOKENS) |
| 395 | ), |
| 396 | hidden_size=self.hidden_size, |
| 397 | rms_eps=self.rms_norm_eps, |
| 398 | hc_pre_eps=self.hc_eps, |
| 399 | hc_sinkhorn_eps=self.hc_eps, |
| 400 | hc_post_mult_value=self.hc_post_alpha, |
| 401 | sinkhorn_repeat=self.hc_sinkhorn_iters, |
| 402 | norm_eps=self.rms_norm_eps, |
| 403 | hc_mult=self.hc_mult, |
| 404 | **variant, |
| 405 | ) |
| 406 | mix_hc = (2 + self.hc_mult) * self.hc_mult |
| 407 | hc_dim = self.hc_mult * self.hidden_size |
| 408 | self.hc_attn_fn = nn.Parameter( |
| 409 | torch.empty( |
| 410 | (mix_hc, hc_dim), |
| 411 | dtype=torch.float32, |
| 412 | ), |
| 413 | requires_grad=False, |
| 414 | ) |
| 415 | self.hc_attn_fn_broadcast: torch.Tensor | None = None |
| 416 | self.hc_ffn_fn = nn.Parameter( |
| 417 | torch.empty( |
| 418 | (mix_hc, hc_dim), |
| 419 | dtype=torch.float32, |
| 420 | ), |
| 421 | requires_grad=False, |
| 422 | ) |
| 423 | self.hc_attn_base = nn.Parameter( |
| 424 | torch.empty( |
| 425 | mix_hc, |
| 426 | dtype=torch.float32, |
| 427 | ), |
| 428 | requires_grad=False, |
| 429 | ) |
| 430 | self.hc_ffn_base = nn.Parameter( |
| 431 | torch.empty( |
| 432 | mix_hc, |
| 433 | dtype=torch.float32, |
| 434 | ), |
| 435 | requires_grad=False, |
| 436 | ) |
| 437 | self.hc_attn_scale = nn.Parameter( |
| 438 | torch.empty( |
| 439 | 3, |
| 440 | dtype=torch.float32, |
| 441 | ), |
| 442 | requires_grad=False, |
| 443 | ) |
| 444 | self.hc_ffn_scale = nn.Parameter( |
| 445 | torch.empty( |
| 446 | 3, |
| 447 | dtype=torch.float32, |
| 448 | ), |
| 449 | requires_grad=False, |
| 450 | ) |
| 451 | |
| 452 | def forward( |
| 453 | self, |
| 454 | x: torch.Tensor | MoEOutput, |
| 455 | positions: torch.Tensor, |
| 456 | input_ids: torch.Tensor | None, |
| 457 | pre_mix: torch.Tensor | None = None, |
| 458 | post_mix: torch.Tensor | None = None, |
| 459 | res_mix: torch.Tensor | None = None, |
| 460 | residual: torch.Tensor | None = None, |
| 461 | engram_hashes: torch.Tensor | None = None, |
| 462 | engram_mask: torch.Tensor | None = None, |
| 463 | *, |
| 464 | capture_previous_aux: bool = False, |
| 465 | mega_gate_metadata: MegaGateRoutingMetadata | None = None, |
| 466 | ) -> tuple[ |
| 467 | torch.Tensor | MoEOutput, |
| 468 | torch.Tensor, |
| 469 | torch.Tensor, |
| 470 | torch.Tensor, |
| 471 | torch.Tensor, |
| 472 | torch.Tensor | None, |
| 473 | ]: |
| 474 | previous_aux: torch.Tensor | None = None |
| 475 | mhc_stream = self.mhc_stream |
| 476 | if mhc_stream is not None and ( |
| 477 | in_piecewise_cudagraph() |
| 478 | or not 0 < positions.shape[0] <= MHC_OVERLAP_MAX_TOKENS |
| 479 | or not torch.cuda.is_current_stream_capturing() |
| 480 | ): |
| 481 | # Use overlap only in FULL graphs; eager warmup initializes Mega mHC. |
| 482 | mhc_stream = None |
| 483 | mhc_pre = ( |
| 484 | partial(mhc_pre_delayed_overlap, stream=mhc_stream) |
| 485 | if mhc_stream is not None |
| 486 | else mhc_pre_delayed_tilelang |
| 487 | ) |
| 488 | # The reference collapses each sublayer's input with the *previous* |
| 489 | # sublayer's pre-mix: attention uses the pre-mix carried in (identity |
| 490 | # for the first layer), the FFN uses this layer's attention pre-mix. |
| 491 | if residual is None: |
| 492 | assert isinstance(x, torch.Tensor) |
| 493 | if x.dim() == 2: |
| 494 | # First layer: the stream is the embedding broadcast to hc |
| 495 | # copies and the identity pre-mix selects copy 0. |
| 496 | assert self.hc_attn_fn_broadcast is not None |
| 497 | residual = x.unsqueeze(1).expand(-1, self.hc_mult, -1).contiguous() |
| 498 | post_mix, res_mix, x, attn_pre = mhc_pre( |
| 499 | residual, |
| 500 | self.hc_attn_fn_broadcast, |
| 501 | self.hc_attn_scale, |
| 502 | self.hc_attn_base, |
| 503 | self.rms_norm_eps, |
| 504 | self.hc_eps, |
| 505 | self.hc_eps, |
| 506 | self.hc_post_alpha, |
| 507 | self.hc_sinkhorn_iters, |
| 508 | x=x, |
| 509 | norm_weight=self.attn_norm.weight, |
| 510 | norm_eps=self.attn_norm.variance_epsilon, |
| 511 | ) |
| 512 | else: |
| 513 | residual = x |
| 514 | post_mix, res_mix, x, attn_pre = mhc_pre( |
| 515 | residual, |
| 516 | self.hc_attn_fn, |
| 517 | self.hc_attn_scale, |
| 518 | self.hc_attn_base, |
| 519 | self.rms_norm_eps, |
| 520 | self.hc_eps, |
| 521 | self.hc_eps, |
| 522 | self.hc_post_alpha, |
| 523 | self.hc_sinkhorn_iters, |
| 524 | pre_mix=pre_mix, |
| 525 | norm_weight=self.attn_norm.weight, |
| 526 | norm_eps=self.attn_norm.variance_epsilon, |
| 527 | ) |
| 528 | elif self.engram is not None and engram_hashes is not None: |
| 529 | # Engram injection happens between the previous sublayer's post |
| 530 | # and this block's pre, on the full hc stream, so the mix |
| 531 | # coefficients see the injected stream. The injection also keeps |
| 532 | # the post out of the pre-norm GEMM's fused prologue. The layer |
| 533 | # before an engram layer finalizes its own MoE. |
| 534 | assert isinstance(x, torch.Tensor) |
| 535 | if self.fuse_mhc_all_reduce: |
| 536 | x = tensor_model_parallel_all_reduce(x) |
| 537 | previous_post = mhc_post_tilelang(x, residual, post_mix, res_mix) |
| 538 | if capture_previous_aux: |
| 539 | previous_aux = previous_post.mean(dim=1) |
| 540 | residual = self.engram( |
| 541 | previous_post, |
| 542 | engram_hashes[:, self.engram.layer_hash_index], |
| 543 | engram_mask, |
| 544 | ) |
| 545 | post_mix, res_mix, x, attn_pre = mhc_pre( |
| 546 | residual, |
| 547 | self.hc_attn_fn, |
| 548 | self.hc_attn_scale, |
| 549 | self.hc_attn_base, |
| 550 | self.rms_norm_eps, |
| 551 | self.hc_eps, |
| 552 | self.hc_eps, |
| 553 | self.hc_post_alpha, |
| 554 | self.hc_sinkhorn_iters, |
| 555 | pre_mix=pre_mix, |
| 556 | norm_weight=self.attn_norm.weight, |
| 557 | norm_eps=self.attn_norm.variance_epsilon, |
| 558 | ) |
| 559 | else: |
| 560 | # The collapse already reads the post-mapped streams, so the mean |
| 561 | # aux consumers want comes out of the same kernel. |
| 562 | residual, post_mix, res_mix, x, attn_pre, aux = mhc_shifted_post_pre( |
| 563 | x, |
| 564 | residual, |
| 565 | post_mix, |
| 566 | res_mix, |
| 567 | self.hc_attn_fn, |
| 568 | self.hc_attn_scale, |
| 569 | self.hc_attn_base, |
| 570 | self.rms_norm_eps, |
| 571 | self.hc_eps, |
| 572 | self.hc_eps, |
| 573 | self.hc_post_alpha, |
| 574 | self.hc_sinkhorn_iters, |
| 575 | pre_mix=pre_mix, |
| 576 | norm_weight=self.attn_norm.weight, |
| 577 | norm_eps=self.attn_norm.variance_epsilon, |
| 578 | capture_aux=capture_previous_aux, |
| 579 | stream=mhc_stream, |
| 580 | reduce_results=self.fuse_mhc_all_reduce, |
| 581 | ) |
| 582 | if capture_previous_aux: |
| 583 | previous_aux = aux |
| 584 | |
| 585 | if self.use_sequence_parallel: |
| 586 | x = sp_all_gather(x)[: positions.shape[0]] |
| 587 | |
| 588 | x = self.attn(positions, x, None) |
| 589 | # With GEMM-RS bound, the attention output is already this rank's |
| 590 | # sequence-parallel shard (see DeepseekV4Attention._wo_b_proj). |
| 591 | if self.use_sequence_parallel and self.attn.gemm_rs is None: |
| 592 | x = sp_reduce_scatter(x) |
| 593 | |
| 594 | if mhc_stream is not None: |
| 595 | torch.cuda.current_stream().wait_stream(mhc_stream) |
| 596 | residual, post_mix, res_mix, x, ffn_pre, _ = mhc_shifted_post_pre( |
| 597 | x, |
| 598 | residual, |
| 599 | post_mix, |
| 600 | res_mix, |
| 601 | self.hc_ffn_fn, |
| 602 | self.hc_ffn_scale, |
| 603 | self.hc_ffn_base, |
| 604 | self.rms_norm_eps, |
| 605 | self.hc_eps, |
| 606 | self.hc_eps, |
| 607 | self.hc_post_alpha, |
| 608 | self.hc_sinkhorn_iters, |
| 609 | pre_mix=attn_pre, |
| 610 | norm_weight=self.ffn_norm.weight, |
| 611 | norm_eps=self.ffn_norm.variance_epsilon, |
| 612 | stream=mhc_stream, |
| 613 | reduce_results=self.fuse_mhc_all_reduce, |
| 614 | ) |
| 615 | if self.fuse_mhc_all_reduce and self.ffn.defers_finalize(x.shape[0]): |
| 616 | # The next layer's first boundary finalizes it with the all-reduce. |
| 617 | x = self.ffn.forward_unfinalized(x, input_ids) |
| 618 | else: |
| 619 | x = self.ffn(x, input_ids, mega_gate_metadata) |
| 620 | if mhc_stream is not None: |
| 621 | torch.cuda.current_stream().wait_stream(mhc_stream) |
| 622 | return x, residual, post_mix, res_mix, ffn_pre, previous_aux |
| 623 | |
| 624 | |
| 625 | class DeepseekV4Model(nn.Module, EagleModelMixin): |
| 626 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 627 | super().__init__() |
| 628 | |
| 629 | config = vllm_config.model_config.hf_config |
| 630 | quant_config = vllm_config.quant_config |
| 631 | self.config = config |
| 632 | self.quant_config = quant_config |
| 633 | self.parallel_config = vllm_config.parallel_config |
| 634 | self.use_mega_moe = vllm_config.kernel_config.moe_backend in MEGA_MOE_BACKENDS |
| 635 | self.use_sequence_parallel = _use_sequence_parallel(vllm_config) |
| 636 | if self.use_mega_moe and not vllm_config.parallel_config.enable_expert_parallel: |
| 637 | raise NotImplementedError( |
| 638 | "DeepSeek V4 MegaMoE currently requires expert parallel. " |
| 639 | "Enable it with --enable-expert-parallel, or pick a different " |
| 640 | "moe backend." |
| 641 | ) |
| 642 | self.vocab_size = config.vocab_size |
| 643 | self.hc_eps = config.hc_eps |
| 644 | self.hc_mult = config.hc_mult |
| 645 | self.hc_dim = self.hc_mult * config.hidden_size |
| 646 | self.rms_norm_eps = config.rms_norm_eps |
| 647 | |
| 648 | # Three aux streams: one per non-default input GEMM in |
| 649 | # DeepseekV4Attention._run_parallel_input_projections |
| 650 | # (compressor kv_score, indexer.weights_proj). fused_wqa_wkv stays on |
| 651 | # the default stream. |
| 652 | aux_stream_list = [torch.cuda.Stream() for _ in range(3)] |
| 653 | # Keep mHC independent of the streams used inside attention. |
| 654 | mhc_stream = torch.cuda.Stream() if supports_mhc_overlap(vllm_config) else None |
| 655 | self.fuse_mhc_all_reduce = mhc_stream is not None and supports_mhc_all_reduce( |
| 656 | vllm_config |
| 657 | ) |
| 658 | if self.fuse_mhc_all_reduce: |
| 659 | init_mhc_all_reduce(vllm_config) |
| 660 | |
| 661 | # Reserved topk indices buffer for all Indexer layers to reuse. |
| 662 | self.topk_indices_buffer = torch.empty( |
| 663 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 664 | config.index_topk, |
| 665 | dtype=torch.int32, |
| 666 | ) |
| 667 | |
| 668 | # Two-level candidate filtering: the indexer at |
| 669 | # candidate_source_layer_id publishes the top candidate blocks of |
| 670 | # compressed positions here; later ratio-1 indexers (24/28/32/36) |
| 671 | # mask their scores with it. |
| 672 | candidate_source_layer = getattr(config, "candidate_source_layer_id", -1) |
| 673 | candidate_topk_blocks = getattr(config, "candidate_topk_blocks", 0) |
| 674 | if candidate_source_layer >= 0 and candidate_topk_blocks > 0: |
| 675 | self.candidate_block_buffer = torch.empty( |
| 676 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 677 | candidate_topk_blocks, |
| 678 | dtype=torch.int32, |
| 679 | ) |
| 680 | else: |
| 681 | self.candidate_block_buffer = None |
| 682 | |
| 683 | if get_pp_group().is_first_rank: |
| 684 | self.embed_tokens = VocabParallelEmbedding( |
| 685 | config.vocab_size, |
| 686 | config.hidden_size, |
| 687 | quant_config=quant_config, |
| 688 | prefix=f"{prefix}.embed_tokens", |
| 689 | ) |
| 690 | else: |
| 691 | self.embed_tokens = PPMissingLayer() |
| 692 | |
| 693 | self.engram_layout = EngramLayout.from_config(config) |
| 694 | engram_config = vllm_config.engram_config |
| 695 | self.vllm_config = vllm_config |
| 696 | engram_prefetch_stream = ( |
| 697 | torch.cuda.Stream() |
| 698 | if self.engram_layout is not None |
| 699 | and engram_config is not None |
| 700 | and engram_config.cpu_offload |
| 701 | else None |
| 702 | ) |
| 703 | if ( |
| 704 | self.engram_layout is not None |
| 705 | and engram_config is not None |
| 706 | and engram_config.dp_shared_memory |
| 707 | ): |
| 708 | engram_config.dp_shared_memory = can_share_engram_tables(self.engram_layout) |
| 709 | |
| 710 | if self.engram_layout is not None and engram_config and engram_config.use_thp: |
| 711 | # Release old checkpoint cache before allocating the Engram host tables. |
| 712 | model_config = vllm_config.model_config |
| 713 | drop_checkpoint_cache( |
| 714 | model_config.model_weights or model_config.model, |
| 715 | revision=model_config.revision, |
| 716 | cache_dir=vllm_config.load_config.download_dir, |
| 717 | ) |
| 718 | |
| 719 | # GEMM-RS uses NCCL symmetric-memory multicast, which requires all TP |
| 720 | # ranks to belong to one NVLink domain. Collective: run before layers. |
| 721 | self.run_gemm_rs = maybe_init_gemm_rs(vllm_config, self.use_sequence_parallel) |
| 722 | |
| 723 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 724 | config.num_hidden_layers, |
| 725 | lambda prefix: DeepseekV4DecoderLayer( |
| 726 | vllm_config, |
| 727 | prefix=prefix, |
| 728 | topk_indices_buffer=self.topk_indices_buffer, |
| 729 | aux_stream_list=aux_stream_list, |
| 730 | candidate_block_buffer=self.candidate_block_buffer, |
| 731 | engram_layout=self.engram_layout, |
| 732 | engram_prefetch_stream=engram_prefetch_stream, |
| 733 | run_gemm_rs=self.run_gemm_rs, |
| 734 | mhc_stream=mhc_stream, |
| 735 | fuse_mhc_all_reduce=self.fuse_mhc_all_reduce, |
| 736 | ), |
| 737 | prefix=f"{prefix}.layers", |
| 738 | ) |
| 739 | if self.fuse_mhc_all_reduce: |
| 740 | # A MoE's top-k finalize folds into the next layer's first mHC |
| 741 | # boundary, which the last local layer lacks and an engram layer |
| 742 | # replaces with its own all-reduce. |
| 743 | local_layers = list(islice(self.layers, self.start_layer, self.end_layer)) |
| 744 | for layer, successor in zip(local_layers, local_layers[1:]): |
| 745 | assert isinstance(layer, DeepseekV4DecoderLayer) |
| 746 | assert isinstance(successor, DeepseekV4DecoderLayer) |
| 747 | if successor.engram is None: |
| 748 | layer.ffn.defer_finalize() |
| 749 | |
| 750 | # Decoder-side SWA bounded replay: in eager prefill steps the layers past |
| 751 | # the last KV source run on each request's trailing window only |
| 752 | # (decoder_replay_layers.py). |
| 753 | self.decoder_replay_layers: DecoderReplayLayers | None = None |
| 754 | self.decoder_replay_start = self.end_layer |
| 755 | cut = max(config.kv_source_layer_ids) |
| 756 | if ( |
| 757 | cut < self.end_layer - 1 |
| 758 | and self._decoder_replay_supported(vllm_config, cut) |
| 759 | and self.layers[cut].attn.swa_cache_layer.bounded_replay |
| 760 | ): |
| 761 | self.decoder_replay_start = cut + 1 |
| 762 | self.decoder_replay_layers = DecoderReplayLayers( |
| 763 | config.sliding_window, |
| 764 | self._run_replay_layers, |
| 765 | [ |
| 766 | buf |
| 767 | for buf in (self.topk_indices_buffer, self.candidate_block_buffer) |
| 768 | if buf is not None |
| 769 | ], |
| 770 | ) |
| 771 | logger.info_once( |
| 772 | "Decoder SWA bounded replay: in eager prefill steps, layers " |
| 773 | "%d-%d run on each request's last %d tokens only.", |
| 774 | cut + 1, |
| 775 | self.end_layer - 1, |
| 776 | config.sliding_window, |
| 777 | ) |
| 778 | |
| 779 | # The n-gram hash needs a slot-keyed rolling store of compressed ids |
| 780 | # (chunked prefill / decode lookback); key it off the first local |
| 781 | # layer's sliding-window KV cache. Only PP ranks owning an engram |
| 782 | # layer need it. |
| 783 | self.engram_hash: NgramHashState | None = None |
| 784 | self.engram_dp_shared_memory = bool( |
| 785 | vllm_config.engram_config and vllm_config.engram_config.dp_shared_memory |
| 786 | ) |
| 787 | self.engram_swa_prefix: str | None = None |
| 788 | if self.engram_layout is not None: |
| 789 | local_engram = any( |
| 790 | isinstance(layer, DeepseekV4DecoderLayer) and layer.engram is not None |
| 791 | for layer in islice(self.layers, self.start_layer, self.end_layer) |
| 792 | ) |
| 793 | if local_engram: |
| 794 | first_layer = next( |
| 795 | iter(islice(self.layers, self.start_layer, self.end_layer)) |
| 796 | ) |
| 797 | swa_cache_module = first_layer.attn.swa_cache_layer |
| 798 | self.engram_hash = NgramHashState( |
| 799 | vllm_config, self.engram_layout, swa_cache_module |
| 800 | ) |
| 801 | self.engram_swa_prefix = swa_cache_module.prefix |
| 802 | |
| 803 | if get_pp_group().is_last_rank: |
| 804 | self.norm = RMSNorm(config.hidden_size, self.rms_norm_eps) |
| 805 | else: |
| 806 | self.norm = PPMissingLayer() |
| 807 | |
| 808 | spec_config = vllm_config.speculative_config |
| 809 | needs_mtp_hidden_states = spec_config is not None and ( |
| 810 | spec_config.use_eagle() or spec_config.uses_draft_model() |
| 811 | ) |
| 812 | if get_pp_group().is_last_rank and needs_mtp_hidden_states: |
| 813 | self._mtp_hidden_buffer = torch.empty( |
| 814 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 815 | self.hc_dim, |
| 816 | dtype=vllm_config.model_config.dtype, |
| 817 | ) |
| 818 | else: |
| 819 | self._mtp_hidden_buffer = None |
| 820 | |
| 821 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 822 | return self.embed_tokens(input_ids) |
| 823 | |
| 824 | def make_empty_intermediate_tensors( |
| 825 | self, |
| 826 | batch_size: int, |
| 827 | dtype: torch.dtype, |
| 828 | device: torch.device, |
| 829 | ) -> IntermediateTensors: |
| 830 | # PP intermediate tensors carry the multi-stream hidden_states |
| 831 | # of shape (num_tokens, hc_mult, hidden_size) — V4 expands the |
| 832 | # token embedding to hc_mult streams before the first decoder |
| 833 | # layer and keeps that shape until the final hc collapse — plus the |
| 834 | # (num_tokens, hc_mult) pre-mix the next rank's first layer needs |
| 835 | # for its attention collapse. |
| 836 | return IntermediateTensors( |
| 837 | { |
| 838 | "hidden_states": torch.zeros( |
| 839 | (batch_size, self.hc_mult, self.config.hidden_size), |
| 840 | dtype=dtype, |
| 841 | device=device, |
| 842 | ), |
| 843 | "pre_mix": torch.zeros( |
| 844 | (batch_size, self.hc_mult), |
| 845 | dtype=torch.float32, |
| 846 | device=device, |
| 847 | ), |
| 848 | } |
| 849 | ) |
| 850 | |
| 851 | def forward( |
| 852 | self, |
| 853 | input_ids: torch.Tensor, |
| 854 | positions: torch.Tensor, |
| 855 | intermediate_tensors: IntermediateTensors | None, |
| 856 | inputs_embeds: torch.Tensor | None = None, |
| 857 | lookback_token_ids: torch.Tensor | None = None, |
| 858 | ) -> torch.Tensor | IntermediateTensors: |
| 859 | if get_pp_group().is_first_rank: |
| 860 | if inputs_embeds is not None: |
| 861 | hidden_states = inputs_embeds |
| 862 | else: |
| 863 | hidden_states = self.embed_input_ids(input_ids) |
| 864 | else: |
| 865 | assert intermediate_tensors is not None |
| 866 | hidden_states = intermediate_tensors["hidden_states"] |
| 867 | |
| 868 | if self.use_mega_moe: |
| 869 | input_ids = input_ids.to(torch.int64) |
| 870 | |
| 871 | # Engram n-gram hashes for the whole (flattened) batch, computed once |
| 872 | # on the full token stream — before any sequence-parallel sharding — |
| 873 | # and consumed by the engram layers (1 and 14) below. Skipped on |
| 874 | # profile runs (KV cache unbound). |
| 875 | engram_hashes: torch.Tensor | None = None |
| 876 | engram_mask: torch.Tensor | None = None |
| 877 | if ( |
| 878 | self.engram_hash is not None |
| 879 | and input_ids is not None |
| 880 | and is_forward_context_available() |
| 881 | ): |
| 882 | attn_metadata = get_forward_context().attn_metadata |
| 883 | if isinstance(attn_metadata, list): |
| 884 | attn_metadata = attn_metadata[dbo_current_ubatch_id()] |
| 885 | if isinstance(attn_metadata, dict) and self.engram_hash.ensure_cache(): |
| 886 | assert self.engram_swa_prefix is not None |
| 887 | swa_metadata = typing.cast( |
| 888 | "DeepseekSparseSWAMetadata", attn_metadata[self.engram_swa_prefix] |
| 889 | ) |
| 890 | # Image-span tokens are dead: they break n-grams (hash op |
| 891 | # takes True=dead) and their gate is zeroed (Engram.forward |
| 892 | # takes True=keep). |
| 893 | image_mask = image_sentinel_mask(input_ids) |
| 894 | engram_mask = ~image_mask |
| 895 | if lookback_token_ids is None: |
| 896 | if not self.engram_hash.use_slot_cache: |
| 897 | raise NotImplementedError( |
| 898 | "engram needs `lookback_token_ids` from the model " |
| 899 | "runner (the DBO/ubatch wrapper drops model kwargs)" |
| 900 | ) |
| 901 | num_reqs = swa_metadata.num_decodes + swa_metadata.num_prefills |
| 902 | lookback_token_ids = input_ids.new_full( |
| 903 | (num_reqs, self.engram_hash.lookback_depth), -1 |
| 904 | ) |
| 905 | engram_hashes = self.engram_hash( |
| 906 | input_ids, |
| 907 | positions, |
| 908 | swa_metadata.query_start_loc, |
| 909 | image_mask, |
| 910 | lookback_token_ids, |
| 911 | image_sentinel_mask(lookback_token_ids), |
| 912 | swa_metadata.slot_mapping, |
| 913 | swa_metadata.block_table, |
| 914 | ) |
| 915 | elif not self.engram_dp_shared_memory and get_engram_dp_size() > 1: |
| 916 | # DP-sharded lookups are collective, so a replica skipping the |
| 917 | # hash still has to reach them. |
| 918 | engram_hashes, engram_mask = self.engram_hash.dummy_hashes(input_ids) |
| 919 | if engram_hashes is not None: |
| 920 | # Gather all Engram rows before entering the decoder layers. |
| 921 | # One gather feeds every layer sharing the DP-split table. |
| 922 | gathered_hashes = gather_engram_hashes( |
| 923 | engram_hashes, dp_shared_memory=self.engram_dp_shared_memory |
| 924 | ) |
| 925 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 926 | engram = getattr(layer, "engram", None) |
| 927 | if engram is not None: |
| 928 | engram.prepare_embeddings( |
| 929 | gathered_hashes[:, engram.layer_hash_index] |
| 930 | ) |
| 931 | |
| 932 | full_num_tokens = positions.shape[0] |
| 933 | if self.use_sequence_parallel: |
| 934 | if envs.VLLM_MOE_SKIP_PADDING and is_forward_context_available(): |
| 935 | forward_context = get_forward_context() |
| 936 | forward_context.is_padding = sp_padding_mask( |
| 937 | forward_context.is_padding, hidden_states |
| 938 | ) |
| 939 | hidden_states = sp_shard(hidden_states) |
| 940 | input_ids = sp_shard(input_ids) |
| 941 | |
| 942 | residual, post_mix, res_mix = None, None, None |
| 943 | pre_mix: torch.Tensor | None = None |
| 944 | if not get_pp_group().is_first_rank: |
| 945 | assert intermediate_tensors is not None |
| 946 | pre_mix = intermediate_tensors["pre_mix"] |
| 947 | aux_hidden_by_layer: dict[int, torch.Tensor] = {} |
| 948 | hidden_states, residual, post_mix, res_mix, pre_mix = self._run_layers( |
| 949 | range(self.start_layer, self.decoder_replay_start), |
| 950 | hidden_states, |
| 951 | positions, |
| 952 | input_ids, |
| 953 | pre_mix, |
| 954 | post_mix, |
| 955 | res_mix, |
| 956 | residual, |
| 957 | aux_hidden_by_layer, |
| 958 | engram_hashes, |
| 959 | engram_mask, |
| 960 | ) |
| 961 | late_aux: list[torch.Tensor] = [] |
| 962 | if self.decoder_replay_layers is not None: |
| 963 | hidden_states, pre_mix, *late_aux = self.decoder_replay_layers( |
| 964 | hidden_states, |
| 965 | positions, |
| 966 | input_ids, |
| 967 | pre_mix, |
| 968 | post_mix, |
| 969 | res_mix, |
| 970 | residual, |
| 971 | ) |
| 972 | else: |
| 973 | hidden_states = self._collapse( |
| 974 | hidden_states, |
| 975 | residual, |
| 976 | post_mix, |
| 977 | res_mix, |
| 978 | aux_hidden_by_layer, |
| 979 | full_num_tokens, |
| 980 | ) |
| 981 | |
| 982 | aux_hidden_states = [ |
| 983 | aux_hidden_by_layer[layer_id] |
| 984 | for layer_id in self.aux_hidden_state_layers |
| 985 | if layer_id in aux_hidden_by_layer |
| 986 | ] + late_aux |
| 987 | |
| 988 | if not get_pp_group().is_last_rank: |
| 989 | return IntermediateTensors( |
| 990 | {"hidden_states": hidden_states, "pre_mix": pre_mix} |
| 991 | ) |
| 992 | |
| 993 | # MTP needs full HC states; otherwise collapse and normalize locally |
| 994 | # before gathering to reduce communication. |
| 995 | if self._mtp_hidden_buffer is not None: |
| 996 | if self.use_sequence_parallel: |
| 997 | hidden_states = sp_all_gather(hidden_states)[:full_num_tokens] |
| 998 | pre_mix = sp_all_gather(pre_mix)[:full_num_tokens] |
| 999 | num_tokens = hidden_states.shape[0] |
| 1000 | self._mtp_hidden_buffer[:num_tokens].copy_(hidden_states.flatten(1)) |
| 1001 | |
| 1002 | # Collapse the hc copies with the pre-mix from the last layer's FFN |
| 1003 | # mixes — the mix the reference applies via |
| 1004 | # ``last_layer.hc_pre(h, pre_mix)`` (v4.1 has no learned hc_head). |
| 1005 | assert pre_mix is not None |
| 1006 | hidden_states = hc_collapse_triton(hidden_states, pre_mix) |
| 1007 | hidden_states = self.norm(hidden_states) |
| 1008 | if self.use_sequence_parallel and self._mtp_hidden_buffer is None: |
| 1009 | # Without MTP, gather only the collapsed and normalized hidden states. |
| 1010 | hidden_states = sp_all_gather(hidden_states)[:full_num_tokens] |
| 1011 | if len(aux_hidden_states) > 0: |
| 1012 | return hidden_states, aux_hidden_states |
| 1013 | return hidden_states |
| 1014 | |
| 1015 | def _mega_gate_metadata( |
| 1016 | self, input_ids: torch.Tensor | None |
| 1017 | ) -> MegaGateRoutingMetadata | None: |
| 1018 | if not self.use_mega_moe: |
| 1019 | return None |
| 1020 | assert input_ids is not None |
| 1021 | return prepare_mega_gate_routing_metadata( |
| 1022 | input_ids, |
| 1023 | has_hash_routing=False, |
| 1024 | image_sentinel_base_id=IMAGE_SENTINEL_BASE_ID |
| 1025 | if getattr(self.config, "vision_n_layers", 0) > 0 |
| 1026 | else None, |
| 1027 | ) |
| 1028 | |
| 1029 | def _run_layers( |
| 1030 | self, |
| 1031 | layer_ids: range, |
| 1032 | hidden_states: torch.Tensor | MoEOutput, |
| 1033 | positions: torch.Tensor, |
| 1034 | input_ids: torch.Tensor | None, |
| 1035 | pre_mix: torch.Tensor | None, |
| 1036 | post_mix: torch.Tensor | None, |
| 1037 | res_mix: torch.Tensor | None, |
| 1038 | residual: torch.Tensor | None, |
| 1039 | aux_hidden_by_layer: dict[int, torch.Tensor], |
| 1040 | engram_hashes: torch.Tensor | None = None, |
| 1041 | engram_mask: torch.Tensor | None = None, |
| 1042 | ) -> tuple[ |
| 1043 | torch.Tensor | MoEOutput, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor |
| 1044 | ]: |
| 1045 | # Every layer's post runs inside the next layer's fused pre, so aux |
| 1046 | # hidden states are read back from there instead of recomputed. |
| 1047 | full_num_tokens = positions.shape[0] |
| 1048 | mega_gate_metadata = self._mega_gate_metadata(input_ids) |
| 1049 | for idx in layer_ids: |
| 1050 | hidden_states, residual, post_mix, res_mix, pre_mix, previous_aux = ( |
| 1051 | self.layers[idx]( |
| 1052 | hidden_states, |
| 1053 | positions, |
| 1054 | input_ids, |
| 1055 | pre_mix, |
| 1056 | post_mix, |
| 1057 | res_mix, |
| 1058 | residual, |
| 1059 | engram_hashes, |
| 1060 | engram_mask, |
| 1061 | capture_previous_aux=idx in self.aux_hidden_state_layers, |
| 1062 | mega_gate_metadata=mega_gate_metadata, |
| 1063 | ) |
| 1064 | ) |
| 1065 | if previous_aux is not None: |
| 1066 | # idx is the one-based id of the layer whose post this is. |
| 1067 | if self.use_sequence_parallel: |
| 1068 | previous_aux = sp_all_gather(previous_aux)[:full_num_tokens] |
| 1069 | aux_hidden_by_layer[idx] = previous_aux |
| 1070 | assert residual is not None and post_mix is not None |
| 1071 | assert res_mix is not None and pre_mix is not None |
| 1072 | return hidden_states, residual, post_mix, res_mix, pre_mix |
| 1073 | |
| 1074 | def _collapse( |
| 1075 | self, |
| 1076 | hidden_states: torch.Tensor | MoEOutput, |
| 1077 | residual: torch.Tensor, |
| 1078 | post_mix: torch.Tensor, |
| 1079 | res_mix: torch.Tensor, |
| 1080 | aux_hidden_by_layer: dict[int, torch.Tensor], |
| 1081 | full_num_tokens: int, |
| 1082 | ) -> torch.Tensor: |
| 1083 | # Without a successor boundary, the last layer finalized its own MoE. |
| 1084 | assert isinstance(hidden_states, torch.Tensor) |
| 1085 | # The last layer has no successor to fold its post into. |
| 1086 | if self.fuse_mhc_all_reduce: |
| 1087 | hidden_states = tensor_model_parallel_all_reduce(hidden_states) |
| 1088 | hidden_states = mhc_post_tilelang(hidden_states, residual, post_mix, res_mix) |
| 1089 | if self.end_layer in self.aux_hidden_state_layers: |
| 1090 | final_aux = hidden_states.mean(dim=1) |
| 1091 | if self.use_sequence_parallel: |
| 1092 | final_aux = sp_all_gather(final_aux)[:full_num_tokens] |
| 1093 | aux_hidden_by_layer[self.end_layer] = final_aux |
| 1094 | return hidden_states |
| 1095 | |
| 1096 | def _run_replay_layers( |
| 1097 | self, |
| 1098 | hidden_states: torch.Tensor | MoEOutput, |
| 1099 | positions: torch.Tensor, |
| 1100 | input_ids: torch.Tensor | None, |
| 1101 | pre_mix: torch.Tensor, |
| 1102 | post_mix: torch.Tensor, |
| 1103 | res_mix: torch.Tensor, |
| 1104 | residual: torch.Tensor, |
| 1105 | ) -> tuple[torch.Tensor, ...]: |
| 1106 | """The layers past the last KV source, on whatever rows they are given; |
| 1107 | returns their output, the last FFN's pre-mix and the aux hidden states |
| 1108 | they capture.""" |
| 1109 | aux_hidden_by_layer: dict[int, torch.Tensor] = {} |
| 1110 | hidden_states, residual, post_mix, res_mix, pre_mix = self._run_layers( |
| 1111 | range(self.decoder_replay_start, self.end_layer), |
| 1112 | hidden_states, |
| 1113 | positions, |
| 1114 | input_ids, |
| 1115 | pre_mix, |
| 1116 | post_mix, |
| 1117 | res_mix, |
| 1118 | residual, |
| 1119 | aux_hidden_by_layer, |
| 1120 | ) |
| 1121 | hidden_states = self._collapse( |
| 1122 | hidden_states, |
| 1123 | residual, |
| 1124 | post_mix, |
| 1125 | res_mix, |
| 1126 | aux_hidden_by_layer, |
| 1127 | positions.shape[0], |
| 1128 | ) |
| 1129 | return ( |
| 1130 | hidden_states, |
| 1131 | pre_mix, |
| 1132 | *( |
| 1133 | aux_hidden_by_layer[layer_id] |
| 1134 | for layer_id in self.aux_hidden_state_layers |
| 1135 | if layer_id in aux_hidden_by_layer |
| 1136 | ), |
| 1137 | ) |
| 1138 | |
| 1139 | def _decoder_replay_supported(self, vllm_config: VllmConfig, cut: int) -> bool: |
| 1140 | """Whether this rank may trim the layers after ``cut``; warns when not.""" |
| 1141 | parallel_config = vllm_config.parallel_config |
| 1142 | spec_config = vllm_config.speculative_config |
| 1143 | draft_config = spec_config.draft_model_config if spec_config else None |
| 1144 | draft_hf_config = getattr(draft_config, "hf_config", None) |
| 1145 | draft_window = getattr(draft_hf_config, "sliding_window", None) |
| 1146 | draft_layer_types = getattr(draft_hf_config, "layer_types", None) or () |
| 1147 | window = self.config.sliding_window |
| 1148 | if self.start_layer > cut or self.end_layer < self.config.num_hidden_layers: |
| 1149 | reason = ( |
| 1150 | "the pipeline stage holding the last KV source layer must also " |
| 1151 | "hold every layer after it" |
| 1152 | ) |
| 1153 | elif ( |
| 1154 | self.use_sequence_parallel |
| 1155 | or parallel_config.prefill_context_parallel_size > 1 |
| 1156 | or parallel_config.use_ubatching |
| 1157 | ): |
| 1158 | reason = ( |
| 1159 | "the replay-layer batch shrinks per rank, which sequence and " |
| 1160 | "prefill-context parallelism and microbatching cannot follow" |
| 1161 | ) |
| 1162 | elif any(i > cut for i in getattr(self.config, "engram_layer_ids", ())): |
| 1163 | reason = "an Engram layer sits after the last KV source layer" |
| 1164 | elif draft_config is not None and ( |
| 1165 | draft_window is None |
| 1166 | or draft_window > window |
| 1167 | or any(t != "sliding_attention" for t in draft_layer_types) |
| 1168 | ): |
| 1169 | reason = ( |
| 1170 | f"the drafter (sliding window {draft_window}) reads hidden states " |
| 1171 | f"outside the target's {window}-token window" |
| 1172 | ) |
| 1173 | else: |
| 1174 | return True |
| 1175 | logger.warning_once("Decoder SWA bounded replay is off: %s.", reason) |
| 1176 | return False |
| 1177 | |
| 1178 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1179 | stacked_params_mapping = [ |
| 1180 | # (param_name, shard_name, shard_id) |
| 1181 | ("gate_up_proj", "w1", 0), |
| 1182 | ("gate_up_proj", "w3", 1), |
| 1183 | ("attn.fused_wqa_wkv", "attn.wq_a", 0), |
| 1184 | ("attn.fused_wqa_wkv", "attn.wkv", 1), |
| 1185 | ("compressor.fused_wkv_wgate", "compressor.wkv", 0), |
| 1186 | ("compressor.fused_wkv_wgate", "compressor.wgate", 1), |
| 1187 | ] |
| 1188 | params_dict = dict(self.named_parameters()) |
| 1189 | loaded_params: set[str] = set() |
| 1190 | |
| 1191 | # TP for attention |
| 1192 | tp_size = get_tensor_model_parallel_world_size() |
| 1193 | tp_rank = get_tensor_model_parallel_rank() |
| 1194 | n_head = self.config.num_attention_heads |
| 1195 | n_local_head = n_head // tp_size |
| 1196 | head_rank_start = n_local_head * tp_rank |
| 1197 | head_rank_end = n_local_head * (tp_rank + 1) |
| 1198 | |
| 1199 | # Pre-compute expert mapping ONCE. |
| 1200 | expert_mapping = self.get_expert_mapping() |
| 1201 | |
| 1202 | # Block-FP8 shared experts: pad the intermediate up to the TP-uniform |
| 1203 | # block count so the standard loaders below slice it evenly (trailing |
| 1204 | # ranks land on the zero pad). SP / unquantized ones need no padding. |
| 1205 | pad_shared_expert = ( |
| 1206 | getattr(self.quant_config, "weight_block_size", None) is not None |
| 1207 | and not self.use_sequence_parallel |
| 1208 | ) |
| 1209 | |
| 1210 | for name, loaded_weight in weights: |
| 1211 | if name.startswith(("vision.", "aligner.", "image_")): |
| 1212 | # Vision weights are loaded by the outer multimodal wrapper. |
| 1213 | logger.warning_once("Skipping non-text weight: %s", name) |
| 1214 | continue |
| 1215 | if pad_shared_expert and ".shared_experts." in name: |
| 1216 | loaded_weight = self._pad_shared_expert_weight( |
| 1217 | self.quant_config, name, loaded_weight |
| 1218 | ) |
| 1219 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 1220 | # Skip non-stacked layers and experts (experts handled below). |
| 1221 | if ".experts." in name: |
| 1222 | continue |
| 1223 | if weight_name not in name: |
| 1224 | continue |
| 1225 | name = name.replace(weight_name, param_name) |
| 1226 | |
| 1227 | if is_pp_missing_parameter(name, self): |
| 1228 | break |
| 1229 | if name not in params_dict: |
| 1230 | head, _, leaf = name.rpartition(".") |
| 1231 | suffixed = f"{head}.base_layer.{leaf}" |
| 1232 | if suffixed in params_dict: |
| 1233 | name = suffixed |
| 1234 | param = params_dict[name] |
| 1235 | weight_loader = param.weight_loader |
| 1236 | weight_loader(param, loaded_weight, shard_id) |
| 1237 | loaded_params.add(name) |
| 1238 | break |
| 1239 | else: |
| 1240 | if ".experts." in name: |
| 1241 | # E8M0 scales are stored as float8_e8m0fnu in |
| 1242 | # checkpoints but the MoE param is uint8. copy_() |
| 1243 | # would do a numeric conversion (e.g. 2^-7 → 0), |
| 1244 | # destroying the raw exponent bytes. |
| 1245 | if ( |
| 1246 | "weight_scale" in name |
| 1247 | and loaded_weight.dtype == torch.float8_e8m0fnu |
| 1248 | ): |
| 1249 | loaded_weight = loaded_weight.view(torch.uint8) |
| 1250 | for mapping in expert_mapping: |
| 1251 | param_name, weight_name, expert_id, expert_shard_id = mapping |
| 1252 | if weight_name not in name: |
| 1253 | continue |
| 1254 | name_mapped = name.replace(weight_name, param_name) |
| 1255 | if is_pp_missing_parameter(name_mapped, self): |
| 1256 | continue |
| 1257 | param = params_dict[name_mapped] |
| 1258 | # We should ask the weight loader to return success or not |
| 1259 | # here since otherwise we may skip experts with other |
| 1260 | # available replicas. |
| 1261 | weight_loader = typing.cast( |
| 1262 | Callable[..., bool], param.weight_loader |
| 1263 | ) |
| 1264 | success = weight_loader( |
| 1265 | param, |
| 1266 | loaded_weight, |
| 1267 | name_mapped, |
| 1268 | shard_id=expert_shard_id, |
| 1269 | expert_id=expert_id, |
| 1270 | return_success=True, |
| 1271 | ) |
| 1272 | if success: |
| 1273 | name = name_mapped |
| 1274 | break |
| 1275 | loaded_params.add(name_mapped) |
| 1276 | continue |
| 1277 | elif "attn_sink" in name: |
| 1278 | if is_pp_missing_parameter(name, self): |
| 1279 | continue |
| 1280 | narrow_weight = loaded_weight[head_rank_start:head_rank_end] |
| 1281 | n = narrow_weight.shape[0] |
| 1282 | params_dict[name][:n].copy_(narrow_weight) |
| 1283 | loaded_params.add(name) |
| 1284 | continue |
| 1285 | else: |
| 1286 | if is_pp_missing_parameter(name, self): |
| 1287 | continue |
| 1288 | # Non-LoRA params on a LoRA-wrapped module live at |
| 1289 | # ``<head>.base_layer.<leaf>``; the checkpoint is plain. |
| 1290 | if name not in params_dict: |
| 1291 | head, _, leaf = name.rpartition(".") |
| 1292 | suffixed = f"{head}.base_layer.{leaf}" |
| 1293 | if suffixed in params_dict: |
| 1294 | name = suffixed |
| 1295 | param = params_dict[name] |
| 1296 | weight_loader = getattr( |
| 1297 | param, "weight_loader", default_weight_loader |
| 1298 | ) |
| 1299 | weight_loader(param, loaded_weight) |
| 1300 | loaded_params.add(name) |
| 1301 | continue |
| 1302 | |
| 1303 | return loaded_params |
| 1304 | |
| 1305 | @staticmethod |
| 1306 | def _pad_shared_expert_weight( |
| 1307 | quant_config: QuantizationConfig | None, |
| 1308 | name: str, |
| 1309 | loaded_weight: torch.Tensor, |
| 1310 | ) -> torch.Tensor: |
| 1311 | """Zero-pad a block-FP8 shared-expert weight/scale on its intermediate |
| 1312 | axis so the standard TP loaders split it into even, block-aligned shards |
| 1313 | (trailing ranks get the zero pad). gate (w1)/up (w3) [I, H] pad dim 0; |
| 1314 | down (w2 -> down_proj) [H, I] pads dim 1. |
| 1315 | """ |
| 1316 | block_size = getattr(quant_config, "weight_block_size", None) |
| 1317 | assert block_size is not None |
| 1318 | # Round the intermediate axis up to a whole number of TP shards. The axis |
| 1319 | # is in elements for weights (step = block) and in blocks for scales. |
| 1320 | step = ( |
| 1321 | 1 if name.endswith(("weight_scale_inv", "weight_scale")) else block_size[0] |
| 1322 | ) |
| 1323 | dim = 1 if ".down_proj." in name else 0 |
| 1324 | mult = get_tensor_model_parallel_world_size() * step |
| 1325 | pad = cdiv(loaded_weight.shape[dim], mult) * mult - loaded_weight.shape[dim] |
| 1326 | if pad == 0: |
| 1327 | return loaded_weight |
| 1328 | pad_shape = list(loaded_weight.shape) |
| 1329 | pad_shape[dim] = pad |
| 1330 | return torch.cat([loaded_weight, loaded_weight.new_zeros(pad_shape)], dim=dim) |
| 1331 | |
| 1332 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 1333 | first_layer = next(iter(islice(self.layers, self.start_layer, self.end_layer))) |
| 1334 | if first_layer.ffn.use_mega_moe: |
| 1335 | return make_deepseek_v4_expert_params_mapping(self.config.n_routed_experts) |
| 1336 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 1337 | # (param_name, weight_name, expert_id, shard_id) |
| 1338 | return fused_moe_make_expert_params_mapping( |
| 1339 | self, |
| 1340 | ckpt_gate_proj_name="w1", |
| 1341 | ckpt_down_proj_name="w2", |
| 1342 | ckpt_up_proj_name="w3", |
| 1343 | num_experts=self.config.n_routed_experts, |
| 1344 | ) |
| 1345 | |
| 1346 | def finalize_mega_moe_weights(self) -> None: |
| 1347 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 1348 | layer.ffn.finalize_mega_moe_weights() |
| 1349 | |
| 1350 | def finalize_mega_attn_weights(self) -> None: |
| 1351 | """Permute wq_b / wo_a into FlashMLA's mega-attention layouts. |
| 1352 | |
| 1353 | A no-op for every other attention layer, and idempotent, so a second |
| 1354 | post-load pass cannot permute twice. |
| 1355 | """ |
| 1356 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 1357 | finalize = getattr(layer.attn, "finalize_loaded_weights", None) |
| 1358 | if finalize is not None: |
| 1359 | finalize() |
| 1360 | |
| 1361 | def finalize_mhc_broadcast_weights(self) -> None: |
| 1362 | if not get_pp_group().is_first_rank or self.start_layer >= self.end_layer: |
| 1363 | return |
| 1364 | layer = self.layers[self.start_layer] |
| 1365 | if isinstance(layer, DeepseekV4DecoderLayer): |
| 1366 | broadcast = ( |
| 1367 | layer.hc_attn_fn.detach() |
| 1368 | .view(-1, layer.hc_mult, layer.hidden_size) |
| 1369 | .sum(dim=1) |
| 1370 | ) |
| 1371 | if layer.hc_attn_fn_broadcast is None: |
| 1372 | layer.hc_attn_fn_broadcast = broadcast |
| 1373 | else: |
| 1374 | layer.hc_attn_fn_broadcast.copy_(broadcast) |
| 1375 | |
| 1376 | |
| 1377 | def _linear_scale_param_name(vllm_config: VllmConfig, expert_dtype: str) -> str: |
| 1378 | """Parameter name the linear quant method registers for weight scales. |
| 1379 | |
| 1380 | Native MXFP8 checkpoints ([32, 32] blocks with MXFP4 experts) route linear |
| 1381 | layers through ModelOptLinearMethod, which registers ``weight_scale``; |
| 1382 | block-FP8 linear layers register ``weight_scale_inv``. |
| 1383 | """ |
| 1384 | use_mxfp8 = ( |
| 1385 | getattr(vllm_config.quant_config, "weight_block_size", None) == [32, 32] |
| 1386 | and expert_dtype == "fp4" |
| 1387 | ) |
| 1388 | return "weight_scale" if use_mxfp8 else "weight_scale_inv" |
| 1389 | |
| 1390 | |
| 1391 | def _make_deepseek_v4_weights_mapper( |
| 1392 | expert_dtype: str, linear_scale_name: str = "weight_scale_inv" |
| 1393 | ) -> WeightsMapper: |
| 1394 | if expert_dtype == "fp4": |
| 1395 | # MXFP4 experts use Mxfp4MoEMethod, which registers scales as |
| 1396 | # ``w{1,2,3}_weight_scale`` (no _inv suffix). Linear scales use the |
| 1397 | # parameter name registered by their quantization method. |
| 1398 | scale_regex = { |
| 1399 | # ``.base_layer.``-namespace variant (LoRA-wrapped experts). |
| 1400 | re.compile( |
| 1401 | r"(\.experts\.\d+\.w[123]\.base_layer)\.scale$" |
| 1402 | ): r"\1.weight_scale", |
| 1403 | re.compile(r"(\.experts\.\d+\.w[123])\.scale$"): r"\1.weight_scale", |
| 1404 | # The ``embed.weight`` -> ``embed_tokens.weight`` suffix rule |
| 1405 | # renames the engram fp8 table but not its scale; route the |
| 1406 | # scale explicitly to the same module. |
| 1407 | re.compile(r"(engram\.embed)\.scale$"): r"\1_tokens.weight_scale_inv", |
| 1408 | re.compile(r"\.scale$"): f".{linear_scale_name}", |
| 1409 | } |
| 1410 | else: |
| 1411 | # FP8 experts use Fp8MoEMethod (block_quant=True), which registers |
| 1412 | # scales as ``w{13,2}_weight_scale_inv``. Map all ``.scale`` keys |
| 1413 | # there. |
| 1414 | scale_regex = { |
| 1415 | # ``.base_layer.``-namespace variant of the above. |
| 1416 | re.compile( |
| 1417 | r"(\.experts\.\d+\.w[123]\.base_layer)\.scale$" |
| 1418 | ): r"\1.weight_scale_inv", |
| 1419 | # Same engram reroute as the fp4 branch above. |
| 1420 | re.compile(r"(engram\.embed)\.scale$"): r"\1_tokens.weight_scale_inv", |
| 1421 | re.compile(r"\.scale$"): f".{linear_scale_name}", |
| 1422 | } |
| 1423 | return WeightsMapper( |
| 1424 | orig_to_new_prefix={ |
| 1425 | "layers.": "model.layers.", |
| 1426 | "embed.": "model.embed.", |
| 1427 | "norm.": "model.norm.", |
| 1428 | "mtp.": "model.mtp.", |
| 1429 | }, |
| 1430 | orig_to_new_regex=scale_regex, |
| 1431 | orig_to_new_suffix={ |
| 1432 | "head.weight": "lm_head.weight", |
| 1433 | "embed.weight": "embed_tokens.weight", |
| 1434 | ".ffn.gate.bias": ".ffn.gate.e_score_correction_bias", |
| 1435 | }, |
| 1436 | orig_to_new_substr={ |
| 1437 | ".shared_experts.w2": ".shared_experts.down_proj", |
| 1438 | "mtp.": None, |
| 1439 | # The v4.1 checkpoint declares the VL arch even for text-only |
| 1440 | # use; the text backbone drops the vision-tower weights. |
| 1441 | "vision.": None, |
| 1442 | "aligner.": None, |
| 1443 | "image_": None, |
| 1444 | }, |
| 1445 | ) |
| 1446 | |
| 1447 | |
| 1448 | class DeepseekV4MixtureOfExperts(MixtureOfExperts): |
| 1449 | moe_mlp_layers: list["DeepseekV4MoE"] |
| 1450 | |
| 1451 | def extract_moe_parameters(self, example_moe: "DeepseekV4MoE | None") -> None: |
| 1452 | if example_moe is None: |
| 1453 | self.num_moe_layers = 0 |
| 1454 | self.num_expert_groups = 0 |
| 1455 | self.num_logical_experts = 0 |
| 1456 | self.num_physical_experts = 0 |
| 1457 | self.num_local_physical_experts = 0 |
| 1458 | self.num_routed_experts = 0 |
| 1459 | self.num_shared_experts = 0 |
| 1460 | self.num_redundant_experts = 0 |
| 1461 | return |
| 1462 | self.num_logical_experts = example_moe.n_logical_experts |
| 1463 | self.num_physical_experts = example_moe.n_physical_experts |
| 1464 | self.num_local_physical_experts = example_moe.n_local_physical_experts |
| 1465 | self.num_routed_experts = example_moe.n_routed_experts |
| 1466 | self.num_shared_experts = example_moe.n_shared_experts |
| 1467 | self.num_redundant_experts = example_moe.n_redundant_experts |
| 1468 | |
| 1469 | def update_physical_experts_metadata( |
| 1470 | self, |
| 1471 | num_physical_experts: int, |
| 1472 | num_local_physical_experts: int, |
| 1473 | ) -> None: |
| 1474 | assert self.num_local_physical_experts == num_local_physical_experts |
| 1475 | self.num_physical_experts = num_physical_experts |
| 1476 | self.num_local_physical_experts = num_local_physical_experts |
| 1477 | self.num_redundant_experts = num_physical_experts - self.num_logical_experts |
| 1478 | for moe in self.moe_mlp_layers: |
| 1479 | moe.n_local_physical_experts = num_local_physical_experts |
| 1480 | moe.n_physical_experts = num_physical_experts |
| 1481 | moe.n_redundant_experts = self.num_redundant_experts |
| 1482 | moe.experts.update_expert_map() |
| 1483 | |
| 1484 | |
| 1485 | class DeepseekV41LLMForCausalLM( |
| 1486 | nn.Module, |
| 1487 | SupportsPP, |
| 1488 | SupportsEagle3, |
| 1489 | SupportsLoRA, |
| 1490 | DeepseekV4MixtureOfExperts, |
| 1491 | ): |
| 1492 | model_cls = DeepseekV4Model |
| 1493 | |
| 1494 | # Default mapper assumes the original FP4-expert checkpoint layout. |
| 1495 | # Overridden per-instance in __init__ when expert_dtype != "fp4". |
| 1496 | hf_to_vllm_mapper = _make_deepseek_v4_weights_mapper("fp4") |
| 1497 | |
| 1498 | packed_modules_mapping = { |
| 1499 | "gate_up_proj": ["w1", "w3"], |
| 1500 | "fused_wqa_wkv": ["wq_a", "wkv"], |
| 1501 | "fused_wkv_wgate": ["wkv", "wgate"], |
| 1502 | } |
| 1503 | |
| 1504 | # The MTP draft head is not LoRA-adapted. |
| 1505 | lora_skip_prefixes = ["mtp."] |
| 1506 | |
| 1507 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1508 | super().__init__() |
| 1509 | |
| 1510 | config = vllm_config.model_config.hf_config |
| 1511 | self.config = config |
| 1512 | expert_dtype = getattr(config, "expert_dtype", "fp4") |
| 1513 | self.hf_to_vllm_mapper = _make_deepseek_v4_weights_mapper( |
| 1514 | expert_dtype, _linear_scale_param_name(vllm_config, expert_dtype) |
| 1515 | ) |
| 1516 | |
| 1517 | self.model = self.model_cls( |
| 1518 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 1519 | ) |
| 1520 | if get_pp_group().is_last_rank: |
| 1521 | self.lm_head = ParallelLMHead( |
| 1522 | config.vocab_size, |
| 1523 | config.hidden_size, |
| 1524 | prefix=maybe_prefix(prefix, "lm_head"), |
| 1525 | ) |
| 1526 | else: |
| 1527 | self.lm_head = PPMissingLayer() |
| 1528 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 1529 | self.make_empty_intermediate_tensors = ( # type: ignore[method-assign] |
| 1530 | self.model.make_empty_intermediate_tensors |
| 1531 | ) |
| 1532 | |
| 1533 | self.set_moe_parameters() |
| 1534 | |
| 1535 | def set_moe_parameters(self) -> None: |
| 1536 | self.num_expert_groups = getattr(self.config, "n_group", 1) |
| 1537 | self.num_moe_layers = self.config.num_hidden_layers |
| 1538 | self.moe_layers: list[nn.Module] = [] |
| 1539 | self.moe_mlp_layers: list[DeepseekV4MoE] = [] |
| 1540 | example_moe: DeepseekV4MoE | None = None |
| 1541 | for layer in self.model.layers: |
| 1542 | if isinstance(layer, PPMissingLayer): |
| 1543 | continue |
| 1544 | if not isinstance(layer, DeepseekV4DecoderLayer): |
| 1545 | continue |
| 1546 | if isinstance(layer.ffn, DeepseekV4MoE): |
| 1547 | example_moe = layer.ffn |
| 1548 | self.moe_mlp_layers.append(layer.ffn) |
| 1549 | self.moe_layers.append(layer.ffn.experts) |
| 1550 | |
| 1551 | self.num_moe_layers = len(self.moe_layers) |
| 1552 | self.extract_moe_parameters(example_moe) |
| 1553 | |
| 1554 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1555 | return self.model.embed_input_ids(input_ids) |
| 1556 | |
| 1557 | def compute_logits( |
| 1558 | self, |
| 1559 | hidden_states: torch.Tensor, |
| 1560 | ) -> torch.Tensor | None: |
| 1561 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 1562 | return logits |
| 1563 | |
| 1564 | def compute_logits_local( |
| 1565 | self, |
| 1566 | hidden_states: torch.Tensor, |
| 1567 | ) -> torch.Tensor: |
| 1568 | return self.logits_processor(self.lm_head, hidden_states, skip_gather=True) |
| 1569 | |
| 1570 | @staticmethod |
| 1571 | def get_model_state_cls(): |
| 1572 | from .model_state import DeepseekV41ModelState |
| 1573 | |
| 1574 | return DeepseekV41ModelState |
| 1575 | |
| 1576 | @property |
| 1577 | def token_lookback_depth(self) -> int: |
| 1578 | """Tokens before a chunk start the engram hash needs; the model runner |
| 1579 | passes them as `lookback_token_ids`.""" |
| 1580 | engram_hash = self.model.engram_hash |
| 1581 | return engram_hash.lookback_depth if engram_hash is not None else 0 |
| 1582 | |
| 1583 | @property |
| 1584 | def decoder_replay_layers(self) -> DecoderReplayLayers | None: |
| 1585 | return self.model.decoder_replay_layers |
| 1586 | |
| 1587 | def forward( |
| 1588 | self, |
| 1589 | input_ids: torch.Tensor, |
| 1590 | positions: torch.Tensor, |
| 1591 | intermediate_tensors: IntermediateTensors | None = None, |
| 1592 | inputs_embeds: torch.Tensor | None = None, |
| 1593 | lookback_token_ids: torch.Tensor | None = None, |
| 1594 | ) -> torch.Tensor | IntermediateTensors: |
| 1595 | hidden_states = self.model( |
| 1596 | input_ids, |
| 1597 | positions, |
| 1598 | intermediate_tensors, |
| 1599 | inputs_embeds, |
| 1600 | lookback_token_ids=lookback_token_ids, |
| 1601 | ) |
| 1602 | return hidden_states |
| 1603 | |
| 1604 | def get_mtp_target_hidden_states(self) -> torch.Tensor | None: |
| 1605 | """Pre-collapse residual stream buffer (max_num_batched_tokens, |
| 1606 | hc_mult * hidden_size) for the MTP draft model. Populated by |
| 1607 | forward(); valid after each target step.""" |
| 1608 | return getattr(self.model, "_mtp_hidden_buffer", None) |
| 1609 | |
| 1610 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1611 | loader = AutoWeightsLoader(self) |
| 1612 | loaded_params = loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 1613 | self.process_weights_after_loading() |
| 1614 | config = self.model.vllm_config |
| 1615 | if config.engram_config and config.engram_config.use_thp: |
| 1616 | # Loading weights refills the file cache; release it before MADV_COLLAPSE. |
| 1617 | model_config = config.model_config |
| 1618 | drop_checkpoint_cache( |
| 1619 | model_config.model_weights or model_config.model, |
| 1620 | revision=model_config.revision, |
| 1621 | cache_dir=config.load_config.download_dir, |
| 1622 | ) |
| 1623 | for layer in islice( |
| 1624 | self.model.layers, self.model.start_layer, self.model.end_layer |
| 1625 | ): |
| 1626 | if layer.engram is not None: |
| 1627 | layer.engram.embed_tokens.collapse_huge_pages() |
| 1628 | return loaded_params |
| 1629 | |
| 1630 | def process_weights_after_loading(self) -> None: |
| 1631 | self.model.finalize_mega_moe_weights() |
| 1632 | self.model.finalize_mhc_broadcast_weights() |
| 1633 | self.model.finalize_mega_attn_weights() |
| 1634 | |
| 1635 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 1636 | return self.model.get_expert_mapping() |