| 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 torch |
| 8 | |
| 9 | import vllm.envs as envs |
| 10 | from vllm.config import VllmConfig |
| 11 | from vllm.distributed import get_pp_group |
| 12 | from vllm.forward_context import get_forward_context, is_forward_context_available |
| 13 | from vllm.model_executor.layers.fused_embed_norm import ( |
| 14 | fused_embed_norm, |
| 15 | has_full_vocab_on_rank, |
| 16 | make_input_embedding, |
| 17 | ) |
| 18 | from vllm.model_executor.layers.fused_moe import ( |
| 19 | fused_moe_make_expert_params_mapping, |
| 20 | ) |
| 21 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 22 | from vllm.model_executor.model_loader.weight_utils import ( |
| 23 | default_weight_loader, |
| 24 | maybe_remap_kv_scale_name, |
| 25 | ) |
| 26 | from vllm.model_executor.models.deepseek_v2 import ( |
| 27 | DeepseekV2ForCausalLM, |
| 28 | DeepseekV2MLP, |
| 29 | DeepseekV2MoE, |
| 30 | _try_load_fp8_indexer_wk, |
| 31 | get_spec_layer_idx_from_weight_name, |
| 32 | ) |
| 33 | from vllm.model_executor.models.utils import ( |
| 34 | PPMissingLayer, |
| 35 | get_pp_missing_layer_names, |
| 36 | is_pp_missing_parameter, |
| 37 | make_empty_intermediate_tensors_factory, |
| 38 | make_layers, |
| 39 | ) |
| 40 | from vllm.models.common.ops.fused_allreduce_rms_norm import fused_allreduce_rms_norm |
| 41 | from vllm.models.common.ops.sequence_parallel import ( |
| 42 | sp_all_gather, |
| 43 | sp_padding_mask, |
| 44 | sp_reduce_scatter, |
| 45 | sp_shard, |
| 46 | ) |
| 47 | from vllm.models.deepseek_v32.attention import DeepseekV32Attention |
| 48 | from vllm.sequence import IntermediateTensors |
| 49 | from vllm.v1.attention.backends.mla.index_group import ( |
| 50 | SparseMLAIndexGroupBuilder, |
| 51 | get_sparse_mla_index_group_max_rows, |
| 52 | ) |
| 53 | |
| 54 | from .glm52_low_latency_gemm import enable_glm52_low_latency_gemm |
| 55 | |
| 56 | |
| 57 | class DeepseekV32DecoderLayer(torch.nn.Module): |
| 58 | def __init__( |
| 59 | self, |
| 60 | vllm_config: VllmConfig, |
| 61 | prefix: str, |
| 62 | config=None, |
| 63 | topk_indices_buffer: torch.Tensor | None = None, |
| 64 | index_group_builder: SparseMLAIndexGroupBuilder | None = None, |
| 65 | ) -> None: |
| 66 | super().__init__() |
| 67 | |
| 68 | if config is None: |
| 69 | config = vllm_config.model_config.hf_config |
| 70 | quant_config = vllm_config.quant_config |
| 71 | parallel_config = vllm_config.parallel_config |
| 72 | |
| 73 | self.hidden_size = config.hidden_size |
| 74 | moe_layer_freq = getattr(config, "moe_layer_freq", 1) |
| 75 | layer_idx = int(prefix.split(sep=".")[-1]) |
| 76 | self.layer_idx = layer_idx |
| 77 | self.use_mha = False |
| 78 | self.use_sequence_parallel = ( |
| 79 | parallel_config.use_sequence_parallel_moe |
| 80 | and parallel_config.pipeline_parallel_size == 1 |
| 81 | ) |
| 82 | |
| 83 | self.self_attn = DeepseekV32Attention( |
| 84 | vllm_config=vllm_config, |
| 85 | config=config, |
| 86 | prefix=f"{prefix}.self_attn", |
| 87 | topk_indices_buffer=topk_indices_buffer, |
| 88 | index_group_builder=index_group_builder, |
| 89 | ) |
| 90 | |
| 91 | if ( |
| 92 | config.n_routed_experts is not None |
| 93 | and layer_idx >= config.first_k_dense_replace |
| 94 | and layer_idx % moe_layer_freq == 0 |
| 95 | ): |
| 96 | # Keep the MoE output un-reduced. Non-SP fuses its all-reduce into |
| 97 | # the next layer norm; SP keeps the token shard local. |
| 98 | self.mlp = DeepseekV2MoE( |
| 99 | config=config, |
| 100 | parallel_config=parallel_config, |
| 101 | quant_config=quant_config, |
| 102 | reduce_results=False, |
| 103 | prefix=f"{prefix}.mlp", |
| 104 | apply_routed_scale_to_output=False, |
| 105 | ) |
| 106 | else: |
| 107 | self.mlp = DeepseekV2MLP( |
| 108 | hidden_size=config.hidden_size, |
| 109 | intermediate_size=config.intermediate_size, |
| 110 | hidden_act=config.hidden_act, |
| 111 | quant_config=quant_config, |
| 112 | prefix=f"{prefix}.mlp", |
| 113 | reduce_results=False, |
| 114 | is_sequence_parallel=self.use_sequence_parallel, |
| 115 | ) |
| 116 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 117 | self.post_attention_layernorm = RMSNorm( |
| 118 | config.hidden_size, eps=config.rms_norm_eps |
| 119 | ) |
| 120 | self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0) |
| 121 | |
| 122 | def forward( |
| 123 | self, |
| 124 | positions: torch.Tensor, |
| 125 | hidden_states: torch.Tensor, |
| 126 | residual: torch.Tensor | None, |
| 127 | attn_in: torch.Tensor | None = None, |
| 128 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 129 | full_num_tokens = positions.shape[0] |
| 130 | |
| 131 | if residual is None: |
| 132 | # First layer: hidden_states is the embedding (the residual). |
| 133 | residual = hidden_states |
| 134 | # ``attn_in`` is input_layernorm(embedding) already computed fused |
| 135 | # with the embedding gather; otherwise apply it here. |
| 136 | hidden_states = ( |
| 137 | attn_in if attn_in is not None else self.input_layernorm(hidden_states) |
| 138 | ) |
| 139 | elif self.use_sequence_parallel: |
| 140 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 141 | else: |
| 142 | # The previous layer's MLP/MoE output is left un-reduced; fuse its |
| 143 | # all-reduce into this input_layernorm. |
| 144 | hidden_states, residual = fused_allreduce_rms_norm( |
| 145 | hidden_states, residual, self.input_layernorm |
| 146 | ) |
| 147 | if self.use_sequence_parallel: |
| 148 | hidden_states = sp_all_gather(hidden_states)[:full_num_tokens] |
| 149 | |
| 150 | # self_attn's o_proj runs reduce_results=False; reduce before RMSNorm. |
| 151 | hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states) |
| 152 | if self.use_sequence_parallel: |
| 153 | hidden_states = sp_reduce_scatter(hidden_states) |
| 154 | hidden_states, residual = self.post_attention_layernorm( |
| 155 | hidden_states, residual |
| 156 | ) |
| 157 | else: |
| 158 | hidden_states, residual = fused_allreduce_rms_norm( |
| 159 | hidden_states, residual, self.post_attention_layernorm |
| 160 | ) |
| 161 | |
| 162 | if self.use_sequence_parallel and isinstance(self.mlp, DeepseekV2MoE): |
| 163 | hidden_states = self.mlp(hidden_states, already_sequence_parallel=True) |
| 164 | else: |
| 165 | hidden_states = self.mlp(hidden_states) |
| 166 | return hidden_states, residual |
| 167 | |
| 168 | |
| 169 | class DeepseekV32Model(torch.nn.Module): |
| 170 | fall_back_to_pt_during_load = False |
| 171 | |
| 172 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 173 | super().__init__() |
| 174 | |
| 175 | config = vllm_config.model_config.hf_config |
| 176 | quant_config = vllm_config.quant_config |
| 177 | self.config = config |
| 178 | from vllm.platforms import current_platform |
| 179 | |
| 180 | self.device = current_platform.device_type |
| 181 | |
| 182 | self.vocab_size = config.vocab_size |
| 183 | parallel_config = vllm_config.parallel_config |
| 184 | self.use_sequence_parallel = ( |
| 185 | parallel_config.use_sequence_parallel_moe |
| 186 | and parallel_config.pipeline_parallel_size == 1 |
| 187 | ) |
| 188 | # DSA is always sparse (has index_topk); allocate the shared top-k |
| 189 | # buffer the indexer writes and the sparse MLA backend reads. |
| 190 | self.is_v32 = True |
| 191 | # On the model, not a local: the MTP proposer shares it with the draft. |
| 192 | self.topk_indices_buffer = torch.empty( |
| 193 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 194 | config.index_topk, |
| 195 | dtype=torch.int32, |
| 196 | device=self.device, |
| 197 | ) |
| 198 | index_group_builder = SparseMLAIndexGroupBuilder( |
| 199 | self.topk_indices_buffer, |
| 200 | get_sparse_mla_index_group_max_rows(vllm_config), |
| 201 | ) |
| 202 | |
| 203 | if get_pp_group().is_first_rank: |
| 204 | self.embed_tokens = make_input_embedding( |
| 205 | config.vocab_size, |
| 206 | config.hidden_size, |
| 207 | quant_config=quant_config, |
| 208 | prefix=f"{prefix}.embed_tokens", |
| 209 | tie_word_embeddings=getattr(config, "tie_word_embeddings", False), |
| 210 | ) |
| 211 | else: |
| 212 | self.embed_tokens = PPMissingLayer() |
| 213 | # The fused embed+norm gather needs the full table on-rank. |
| 214 | self.replicated_embed = has_full_vocab_on_rank(self.embed_tokens) |
| 215 | |
| 216 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 217 | config.num_hidden_layers, |
| 218 | lambda prefix: DeepseekV32DecoderLayer( |
| 219 | vllm_config=vllm_config, |
| 220 | prefix=prefix, |
| 221 | topk_indices_buffer=self.topk_indices_buffer, |
| 222 | index_group_builder=index_group_builder, |
| 223 | ), |
| 224 | prefix=f"{prefix}.layers", |
| 225 | ) |
| 226 | |
| 227 | if get_pp_group().is_last_rank: |
| 228 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 229 | else: |
| 230 | self.norm = PPMissingLayer() |
| 231 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 232 | ["hidden_states", "residual"], config.hidden_size |
| 233 | ) |
| 234 | |
| 235 | self.aux_hidden_state_layers = tuple[int, ...]() |
| 236 | self.num_redundant_experts = parallel_config.eplb_config.num_redundant_experts |
| 237 | |
| 238 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 239 | return self.embed_tokens(input_ids) |
| 240 | |
| 241 | def forward( |
| 242 | self, |
| 243 | input_ids: torch.Tensor | None, |
| 244 | positions: torch.Tensor, |
| 245 | intermediate_tensors: IntermediateTensors | None = None, |
| 246 | inputs_embeds: torch.Tensor | None = None, |
| 247 | ) -> torch.Tensor | IntermediateTensors: |
| 248 | attn_in = None |
| 249 | if get_pp_group().is_first_rank: |
| 250 | if inputs_embeds is not None: |
| 251 | hidden_states = inputs_embeds |
| 252 | elif self.replicated_embed: |
| 253 | assert input_ids is not None |
| 254 | # Full table on-rank: gather the embedding and the first layer's |
| 255 | # input_layernorm in one launch. ``attn_in`` is the pre-normed |
| 256 | # attention input; ``hidden_states`` is the (residual) embedding. |
| 257 | hidden_states, attn_in = fused_embed_norm( |
| 258 | input_ids, |
| 259 | self.embed_tokens.weight, |
| 260 | chain_weight=self.layers[self.start_layer].input_layernorm.weight, |
| 261 | eps=self.config.rms_norm_eps, |
| 262 | ) |
| 263 | else: |
| 264 | assert input_ids is not None |
| 265 | hidden_states = self.embed_input_ids(input_ids) |
| 266 | residual = None |
| 267 | else: |
| 268 | assert intermediate_tensors is not None |
| 269 | hidden_states = intermediate_tensors["hidden_states"] |
| 270 | residual = intermediate_tensors["residual"] |
| 271 | |
| 272 | full_num_tokens = positions.shape[0] |
| 273 | if self.use_sequence_parallel: |
| 274 | if envs.VLLM_MOE_SKIP_PADDING and is_forward_context_available(): |
| 275 | forward_context = get_forward_context() |
| 276 | forward_context.is_padding = sp_padding_mask( |
| 277 | forward_context.is_padding, hidden_states |
| 278 | ) |
| 279 | hidden_states = sp_shard(hidden_states) |
| 280 | if attn_in is not None: |
| 281 | attn_in = sp_shard(attn_in) |
| 282 | assert residual is None, "Currently, SP is not supported with PP" |
| 283 | |
| 284 | aux_hidden_states = [] |
| 285 | for idx, layer in enumerate( |
| 286 | islice(self.layers, self.start_layer, self.end_layer), |
| 287 | start=self.start_layer, |
| 288 | ): |
| 289 | if idx in self.aux_hidden_state_layers: |
| 290 | aux_hidden_states.append( |
| 291 | hidden_states if residual is None else hidden_states + residual |
| 292 | ) |
| 293 | hidden_states, residual = layer(positions, hidden_states, residual, attn_in) |
| 294 | attn_in = None |
| 295 | |
| 296 | if not get_pp_group().is_last_rank: |
| 297 | assert not self.use_sequence_parallel, ( |
| 298 | "Currently, SP is not supported with PP" |
| 299 | ) |
| 300 | return IntermediateTensors( |
| 301 | {"hidden_states": hidden_states, "residual": residual} |
| 302 | ) |
| 303 | |
| 304 | if self.use_sequence_parallel: |
| 305 | hidden_states, _ = self.norm(hidden_states, residual) |
| 306 | if aux_hidden_states: |
| 307 | hidden_size = hidden_states.shape[-1] |
| 308 | packed_hidden_states = torch.cat( |
| 309 | [hidden_states, *aux_hidden_states], dim=-1 |
| 310 | ) |
| 311 | packed_hidden_states = sp_all_gather(packed_hidden_states) |
| 312 | packed_hidden_states = packed_hidden_states[:full_num_tokens] |
| 313 | hidden_states, *aux_hidden_states = packed_hidden_states.split( |
| 314 | hidden_size, dim=-1 |
| 315 | ) |
| 316 | else: |
| 317 | hidden_states = sp_all_gather(hidden_states)[:full_num_tokens] |
| 318 | else: |
| 319 | hidden_states, _ = fused_allreduce_rms_norm( |
| 320 | hidden_states, residual, self.norm |
| 321 | ) |
| 322 | if len(aux_hidden_states) > 0: |
| 323 | return hidden_states, aux_hidden_states |
| 324 | return hidden_states |
| 325 | |
| 326 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 327 | # DSA-only: MLA (fused_qkv_a_proj) + the fused indexer wk/weights_proj + |
| 328 | # routed experts. No MHA (qkv_proj) or ROCm shared-expert-fusion paths. |
| 329 | stacked_params_mapping = [ |
| 330 | ("gate_up_proj", "gate_proj", 0), |
| 331 | ("gate_up_proj", "up_proj", 1), |
| 332 | ("fused_qkv_a_proj", "q_a_proj", 0), |
| 333 | ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1), |
| 334 | ("wk_weights_proj", "wk", 0), |
| 335 | ("wk_weights_proj", "weights_proj", 1), |
| 336 | ] |
| 337 | expert_params_mapping = fused_moe_make_expert_params_mapping( |
| 338 | self, |
| 339 | ckpt_gate_proj_name="gate_proj", |
| 340 | ckpt_down_proj_name="down_proj", |
| 341 | ckpt_up_proj_name="up_proj", |
| 342 | num_experts=self.config.n_routed_experts, |
| 343 | num_redundant_experts=self.num_redundant_experts, |
| 344 | ) |
| 345 | |
| 346 | pp_missing_layer_names = get_pp_missing_layer_names(self) |
| 347 | params_dict = dict(self.named_parameters()) |
| 348 | loaded_params: set[str] = set() |
| 349 | _pending_wk_fp8: dict = {} |
| 350 | for name, loaded_weight in weights: |
| 351 | if "rotary_emb.inv_freq" in name: |
| 352 | continue |
| 353 | # MTP / nextn layers are loaded by the MTP model, not here. |
| 354 | if get_spec_layer_idx_from_weight_name(self.config, name) is not None: |
| 355 | continue |
| 356 | if _try_load_fp8_indexer_wk( |
| 357 | name, |
| 358 | loaded_weight, |
| 359 | _pending_wk_fp8, |
| 360 | params_dict, |
| 361 | loaded_params, |
| 362 | pp_missing_layer_names, |
| 363 | ): |
| 364 | continue |
| 365 | |
| 366 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 367 | if weight_name not in name: |
| 368 | continue |
| 369 | # Experts are handled below; skip here before the name rewrite. |
| 370 | if ("mlp.experts." in name) and name not in params_dict: |
| 371 | continue |
| 372 | name_mapped = name.replace(weight_name, param_name) |
| 373 | if ( |
| 374 | param_name == "fused_qkv_a_proj" |
| 375 | ) and name_mapped not in params_dict: |
| 376 | continue |
| 377 | name = name_mapped |
| 378 | if name.endswith(".bias") and name not in params_dict: |
| 379 | continue |
| 380 | if is_pp_missing_parameter(name, self): |
| 381 | continue |
| 382 | param = params_dict[name] |
| 383 | param.weight_loader(param, loaded_weight, shard_id) |
| 384 | break |
| 385 | else: |
| 386 | is_expert_weight = False |
| 387 | for mapping in expert_params_mapping: |
| 388 | param_name, weight_name, expert_id, shard_id = mapping # type: ignore[assignment] |
| 389 | if weight_name not in name: |
| 390 | continue |
| 391 | is_expert_weight = True |
| 392 | name_mapped = name.replace(weight_name, param_name) |
| 393 | if is_pp_missing_parameter(name_mapped, self): |
| 394 | continue |
| 395 | param = params_dict[name_mapped] |
| 396 | weight_loader = typing.cast( |
| 397 | Callable[..., bool], param.weight_loader |
| 398 | ) |
| 399 | success = weight_loader( |
| 400 | param, |
| 401 | loaded_weight, |
| 402 | name_mapped, |
| 403 | shard_id=shard_id, |
| 404 | expert_id=expert_id, |
| 405 | return_success=True, |
| 406 | ) |
| 407 | if success: |
| 408 | name = name_mapped |
| 409 | break |
| 410 | else: |
| 411 | if is_expert_weight: |
| 412 | continue |
| 413 | if name.endswith(".bias") and name not in params_dict: |
| 414 | continue |
| 415 | name = maybe_remap_kv_scale_name(name, params_dict) # type: ignore[assignment] |
| 416 | if name is None: |
| 417 | continue |
| 418 | if is_pp_missing_parameter(name, self): |
| 419 | continue |
| 420 | param = params_dict[name] |
| 421 | loader = getattr(param, "weight_loader", default_weight_loader) |
| 422 | loader(param, loaded_weight) |
| 423 | loaded_params.add(name) |
| 424 | return loaded_params |
| 425 | |
| 426 | |
| 427 | class DeepseekV32ForCausalLM(DeepseekV2ForCausalLM): |
| 428 | """DSA causal LM — DeepSeek V2/V3 orchestration with the DSA backbone. |
| 429 | |
| 430 | Serves DeepSeek V3.2 and any architecture reusing DSA (e.g. GLM-5.2). |
| 431 | """ |
| 432 | |
| 433 | model_cls = DeepseekV32Model |
| 434 | |
| 435 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 436 | super().__init__(vllm_config=vllm_config, prefix=prefix) |
| 437 | if self.config.model_type == "glm_moe_dsa": |
| 438 | enable_glm52_low_latency_gemm(self, vllm_config.model_config.dtype) |
| 439 | |
| 440 | def set_moe_parameters(self): |
| 441 | # Same as the base, but keyed on the MoE block type rather than the |
| 442 | # decoder-layer type (DeepseekV32DecoderLayer is a plain nn.Module). |
| 443 | self.num_expert_groups = getattr(self.config, "n_group", 1) |
| 444 | self.moe_layers = [] |
| 445 | self.moe_mlp_layers = [] |
| 446 | example_moe = None |
| 447 | for layer in self.model.layers: |
| 448 | if isinstance(layer, PPMissingLayer): |
| 449 | continue |
| 450 | if isinstance(layer.mlp, DeepseekV2MoE): |
| 451 | example_moe = layer.mlp |
| 452 | self.moe_mlp_layers.append(layer.mlp) |
| 453 | self.moe_layers.append(layer.mlp.experts) |
| 454 | self.extract_moe_parameters(example_moe) |