| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Adapted from |
| 5 | # https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py |
| 6 | # Copyright 2023 The vLLM team. |
| 7 | # Copyright 2023 DeepSeek-AI and the HuggingFace Inc. team. All rights reserved. |
| 8 | # |
| 9 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 10 | # and OPT implementations in this library. It has been modified from its |
| 11 | # original forms to accommodate minor architectural differences compared |
| 12 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 13 | # |
| 14 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 15 | # you may not use this file except in compliance with the License. |
| 16 | # You may obtain a copy of the License at |
| 17 | # |
| 18 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 19 | # |
| 20 | # Unless required by applicable law or agreed to in writing, software |
| 21 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 22 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 23 | # See the License for the specific language governing permissions and |
| 24 | # limitations under the License. |
| 25 | """Inference-only DeepseekV2/DeepseekV3 model.""" |
| 26 | |
| 27 | import typing |
| 28 | from collections.abc import Callable, Iterable |
| 29 | from itertools import islice |
| 30 | |
| 31 | import torch |
| 32 | from torch import nn |
| 33 | from transformers import DeepseekV2Config, DeepseekV3Config |
| 34 | |
| 35 | import vllm._custom_ops as ops |
| 36 | import vllm.envs as envs |
| 37 | from vllm._aiter_ops import rocm_aiter_ops |
| 38 | from vllm.compilation.decorators import support_torch_compile |
| 39 | from vllm.config import CacheConfig, ParallelConfig, VllmConfig, get_current_vllm_config |
| 40 | from vllm.distributed import ( |
| 41 | get_ep_group, |
| 42 | get_pp_group, |
| 43 | get_tensor_model_parallel_rank, |
| 44 | get_tensor_model_parallel_world_size, |
| 45 | tensor_model_parallel_all_gather, |
| 46 | tensor_model_parallel_reduce_scatter, |
| 47 | ) |
| 48 | from vllm.logger import init_logger |
| 49 | from vllm.model_executor.layers.activation import SiluAndMul |
| 50 | from vllm.model_executor.layers.attention import ( |
| 51 | Attention, |
| 52 | EncoderOnlyAttention, |
| 53 | RSWAAttention, |
| 54 | ) |
| 55 | from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase |
| 56 | from vllm.model_executor.layers.fused_moe import ( |
| 57 | FusedMoEFactory, |
| 58 | GateLinear, |
| 59 | fused_moe_make_expert_params_mapping, |
| 60 | ) |
| 61 | from vllm.model_executor.layers.fused_moe.utils import ( |
| 62 | is_model_fused_shared_expert_compatible, |
| 63 | resolve_layer_fused_shared_expert, |
| 64 | ) |
| 65 | from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm |
| 66 | from vllm.model_executor.layers.linear import ( |
| 67 | ColumnParallelLinear, |
| 68 | DCPGroupColumnParallelLinear, |
| 69 | MergedColumnParallelLinear, |
| 70 | QKVParallelLinear, |
| 71 | ReplicatedLinear, |
| 72 | RowParallelLinear, |
| 73 | ) |
| 74 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 75 | from vllm.model_executor.layers.mla import MLAModules, MultiHeadLatentAttentionWrapper |
| 76 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 77 | from vllm.model_executor.layers.quantization.utils.fp8_utils import ( |
| 78 | per_token_group_quant_fp8, |
| 79 | ) |
| 80 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 81 | GroupShape, |
| 82 | scaled_dequantize, |
| 83 | ) |
| 84 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 85 | from vllm.model_executor.layers.sparse_attn_indexer import ( |
| 86 | SparseAttnIndexer, |
| 87 | fused_indexer_q_rope_quant, |
| 88 | ) |
| 89 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 90 | ParallelLMHead, |
| 91 | VocabParallelEmbedding, |
| 92 | ) |
| 93 | from vllm.model_executor.model_loader.weight_utils import ( |
| 94 | default_weight_loader, |
| 95 | maybe_remap_kv_scale_name, |
| 96 | ) |
| 97 | from vllm.model_executor.models.utils import ( |
| 98 | AutoWeightsLoader, |
| 99 | extract_layer_index, |
| 100 | sequence_parallel_chunk, |
| 101 | ) |
| 102 | from vllm.platforms import current_platform |
| 103 | from vllm.sequence import IntermediateTensors |
| 104 | from vllm.utils.torch_utils import direct_register_custom_op |
| 105 | from vllm.v1.attention.backend import AttentionBackend, AttentionType |
| 106 | from vllm.v1.attention.backends.mla.index_group import ( |
| 107 | SparseMLAIndexGroupBuilder, |
| 108 | get_sparse_mla_index_group_max_rows, |
| 109 | ) |
| 110 | from vllm.v1.attention.backends.mla.indexer import ( |
| 111 | DeepseekV32IndexerBackend, |
| 112 | ) |
| 113 | from vllm.v1.kv_cache_interface import KVCacheSpec, MLAAttentionSpec, SparseCacheRole |
| 114 | |
| 115 | from .interfaces import ( |
| 116 | MixtureOfExperts, |
| 117 | SupportsEagle, |
| 118 | SupportsEagle3, |
| 119 | SupportsLoRA, |
| 120 | SupportsPP, |
| 121 | ) |
| 122 | from .utils import ( |
| 123 | PPMissingLayer, |
| 124 | get_pp_missing_layer_names, |
| 125 | get_spec_layer_idx_from_weight_name, |
| 126 | is_pp_missing_parameter, |
| 127 | make_empty_intermediate_tensors_factory, |
| 128 | make_layers, |
| 129 | maybe_prefix, |
| 130 | ) |
| 131 | |
| 132 | logger = init_logger(__name__) |
| 133 | |
| 134 | |
| 135 | def _get_moe_router_dtype( |
| 136 | config: DeepseekV2Config | DeepseekV3Config, |
| 137 | ) -> torch.dtype | None: |
| 138 | router_dtype = getattr(config, "moe_router_dtype", None) |
| 139 | if getattr(config, "model_type", None) == "glm_moe_dsa": |
| 140 | # Older GLM-5/5.2 configs require fp32 routing but do not expose |
| 141 | # moe_router_dtype yet. |
| 142 | return torch.float32 |
| 143 | if router_dtype == "float32": |
| 144 | return torch.float32 |
| 145 | return None |
| 146 | |
| 147 | |
| 148 | class DeepseekAttention(nn.Module): |
| 149 | """Normal MHA implementation used by Deepseek v1.""" |
| 150 | |
| 151 | def __init__( |
| 152 | self, |
| 153 | vllm_config: VllmConfig, |
| 154 | config: DeepseekV2Config | DeepseekV3Config, |
| 155 | hidden_size: int, |
| 156 | num_heads: int, |
| 157 | max_position_embeddings: int = 8192, |
| 158 | cache_config: CacheConfig | None = None, |
| 159 | quant_config: QuantizationConfig | None = None, |
| 160 | reduce_results: bool = True, |
| 161 | prefix: str = "", |
| 162 | **kwargs, |
| 163 | ) -> None: |
| 164 | super().__init__() |
| 165 | self.hidden_size = hidden_size |
| 166 | tp_size = get_tensor_model_parallel_world_size() |
| 167 | self.total_num_heads = num_heads |
| 168 | assert self.total_num_heads % tp_size == 0 |
| 169 | self.num_heads = self.total_num_heads // tp_size |
| 170 | self.total_num_kv_heads = config.num_key_value_heads |
| 171 | if self.total_num_kv_heads >= tp_size: |
| 172 | # Number of KV heads is greater than TP size, so we partition |
| 173 | # the KV heads across multiple tensor parallel GPUs. |
| 174 | assert self.total_num_kv_heads % tp_size == 0 |
| 175 | else: |
| 176 | # Number of KV heads is less than TP size, so we replicate |
| 177 | # the KV heads across multiple tensor parallel GPUs. |
| 178 | assert tp_size % self.total_num_kv_heads == 0 |
| 179 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 180 | self.head_dim = hidden_size // self.total_num_heads |
| 181 | self.q_size = self.num_heads * self.head_dim |
| 182 | self.kv_size = self.num_kv_heads * self.head_dim |
| 183 | self.scaling = self.head_dim**-0.5 |
| 184 | self.max_position_embeddings = max_position_embeddings |
| 185 | |
| 186 | self.qkv_proj = QKVParallelLinear( |
| 187 | hidden_size, |
| 188 | self.head_dim, |
| 189 | self.total_num_heads, |
| 190 | self.total_num_kv_heads, |
| 191 | bias=False, |
| 192 | quant_config=quant_config, |
| 193 | ) |
| 194 | |
| 195 | self.o_proj = RowParallelLinear( |
| 196 | self.total_num_heads * self.head_dim, |
| 197 | hidden_size, |
| 198 | bias=False, |
| 199 | reduce_results=reduce_results, |
| 200 | quant_config=quant_config, |
| 201 | ) |
| 202 | |
| 203 | self.rotary_emb = get_rope( |
| 204 | self.head_dim, |
| 205 | max_position=max_position_embeddings, |
| 206 | rope_parameters=config.rope_parameters, |
| 207 | ) |
| 208 | rswa_window = getattr(vllm_config.model_config.hf_config, "rswa_window", None) |
| 209 | if rswa_window is not None: |
| 210 | self.attn = RSWAAttention( |
| 211 | self.num_heads, |
| 212 | self.head_dim, |
| 213 | self.scaling, |
| 214 | num_kv_heads=self.num_kv_heads, |
| 215 | cache_config=cache_config, |
| 216 | quant_config=quant_config, |
| 217 | prefix=f"{prefix}.attn", |
| 218 | rswa_window=rswa_window, |
| 219 | ) |
| 220 | else: |
| 221 | self.attn = Attention( |
| 222 | self.num_heads, |
| 223 | self.head_dim, |
| 224 | self.scaling, |
| 225 | num_kv_heads=self.num_kv_heads, |
| 226 | cache_config=cache_config, |
| 227 | quant_config=quant_config, |
| 228 | prefix=f"{prefix}.attn", |
| 229 | ) |
| 230 | |
| 231 | def forward( |
| 232 | self, |
| 233 | positions: torch.Tensor, |
| 234 | hidden_states: torch.Tensor, |
| 235 | ) -> torch.Tensor: |
| 236 | qkv, _ = self.qkv_proj(hidden_states) |
| 237 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 238 | q, k = self.rotary_emb(positions, q, k) |
| 239 | attn_output = self.attn(q, k, v) |
| 240 | output, _ = self.o_proj(attn_output) |
| 241 | return output |
| 242 | |
| 243 | |
| 244 | class DeepseekV2MLP(nn.Module): |
| 245 | def __init__( |
| 246 | self, |
| 247 | hidden_size: int, |
| 248 | intermediate_size: int, |
| 249 | hidden_act: str, |
| 250 | quant_config: QuantizationConfig | None = None, |
| 251 | reduce_results: bool = True, |
| 252 | is_sequence_parallel=False, |
| 253 | prefix: str = "", |
| 254 | ) -> None: |
| 255 | super().__init__() |
| 256 | |
| 257 | # If is_sequence_parallel, the input and output tensors are sharded |
| 258 | # across the ranks within the tp_group. In this case the weights are |
| 259 | # replicated and no collective ops are needed. |
| 260 | # Otherwise we use standard TP with an allreduce at the end. |
| 261 | self.gate_up_proj = MergedColumnParallelLinear( |
| 262 | hidden_size, |
| 263 | [intermediate_size] * 2, |
| 264 | bias=False, |
| 265 | quant_config=quant_config, |
| 266 | disable_tp=is_sequence_parallel, |
| 267 | prefix=f"{prefix}.gate_up_proj", |
| 268 | ) |
| 269 | self.down_proj = RowParallelLinear( |
| 270 | intermediate_size, |
| 271 | hidden_size, |
| 272 | bias=False, |
| 273 | quant_config=quant_config, |
| 274 | reduce_results=reduce_results, |
| 275 | disable_tp=is_sequence_parallel, |
| 276 | prefix=f"{prefix}.down_proj", |
| 277 | ) |
| 278 | if hidden_act != "silu": |
| 279 | raise ValueError( |
| 280 | f"Unsupported activation: {hidden_act}. Only silu is supported for now." |
| 281 | ) |
| 282 | self.act_fn = SiluAndMul() |
| 283 | |
| 284 | def forward(self, x): |
| 285 | gate_up, _ = self.gate_up_proj(x) |
| 286 | x = self.act_fn(gate_up) |
| 287 | x, _ = self.down_proj(x) |
| 288 | return x |
| 289 | |
| 290 | |
| 291 | class DeepseekV2MoE(nn.Module): |
| 292 | def __init__( |
| 293 | self, |
| 294 | config: DeepseekV2Config | DeepseekV3Config, |
| 295 | parallel_config: ParallelConfig, |
| 296 | quant_config: QuantizationConfig | None = None, |
| 297 | reduce_results: bool = True, |
| 298 | prefix: str = "", |
| 299 | apply_routed_scale_to_output: bool = False, |
| 300 | ): |
| 301 | super().__init__() |
| 302 | self.tp_size = get_tensor_model_parallel_world_size() |
| 303 | self.tp_rank = get_tensor_model_parallel_rank() |
| 304 | |
| 305 | self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0) |
| 306 | |
| 307 | self.ep_group = get_ep_group().device_group |
| 308 | self.ep_size = self.ep_group.size() |
| 309 | self.n_routed_experts: int = config.n_routed_experts |
| 310 | self.n_shared_experts: int = config.n_shared_experts |
| 311 | |
| 312 | self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe |
| 313 | |
| 314 | if config.hidden_act != "silu": |
| 315 | raise ValueError( |
| 316 | f"Unsupported activation: {config.hidden_act}. " |
| 317 | "Only silu is supported for now." |
| 318 | ) |
| 319 | |
| 320 | self.router_dtype = _get_moe_router_dtype(config) |
| 321 | self.gate = GateLinear( |
| 322 | config.hidden_size, |
| 323 | config.n_routed_experts, |
| 324 | out_dtype=self.router_dtype, |
| 325 | prefix=f"{prefix}.gate", |
| 326 | ) |
| 327 | if getattr(config, "topk_method", None) == "noaux_tc": |
| 328 | self.gate.e_score_correction_bias = nn.Parameter( |
| 329 | torch.empty(config.n_routed_experts, dtype=torch.float32) |
| 330 | ) |
| 331 | else: |
| 332 | self.gate.e_score_correction_bias = None |
| 333 | |
| 334 | # Load balancing settings. |
| 335 | eplb_config = parallel_config.eplb_config |
| 336 | self.enable_eplb = parallel_config.enable_eplb |
| 337 | |
| 338 | self.n_redundant_experts = eplb_config.num_redundant_experts |
| 339 | self.n_logical_experts = self.n_routed_experts |
| 340 | self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts |
| 341 | self.n_local_physical_experts = self.n_physical_experts // self.ep_size |
| 342 | |
| 343 | self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled() |
| 344 | |
| 345 | self.is_fused_shared_expert_enabled = False |
| 346 | if config.n_shared_experts is not None: |
| 347 | self.is_fused_shared_expert_enabled = resolve_layer_fused_shared_expert( |
| 348 | quant_config, prefix |
| 349 | ) |
| 350 | |
| 351 | if ( |
| 352 | self.is_rocm_aiter_moe_enabled |
| 353 | and self.gate.e_score_correction_bias is not None |
| 354 | and self.gate.out_dtype is None |
| 355 | ): |
| 356 | # Accumulates in fp32; avoids bf16->fp32 cast. |
| 357 | self.gate.set_out_dtype(self.gate.weight.dtype) |
| 358 | |
| 359 | if config.n_shared_experts is None or self.is_fused_shared_expert_enabled: |
| 360 | self.shared_experts = None |
| 361 | else: |
| 362 | intermediate_size = config.moe_intermediate_size * config.n_shared_experts |
| 363 | |
| 364 | self.shared_experts = DeepseekV2MLP( |
| 365 | hidden_size=config.hidden_size, |
| 366 | intermediate_size=intermediate_size, |
| 367 | hidden_act=config.hidden_act, |
| 368 | quant_config=quant_config, |
| 369 | is_sequence_parallel=self.is_sequence_parallel, |
| 370 | reduce_results=False, |
| 371 | prefix=f"{prefix}.shared_experts", |
| 372 | ) |
| 373 | |
| 374 | self.experts = FusedMoEFactory( |
| 375 | shared_experts=self.shared_experts, |
| 376 | gate=self.gate, |
| 377 | num_experts=config.n_routed_experts, |
| 378 | top_k=config.num_experts_per_tok, |
| 379 | hidden_size=config.hidden_size, |
| 380 | intermediate_size=config.moe_intermediate_size, |
| 381 | renormalize=config.norm_topk_prob, |
| 382 | quant_config=quant_config, |
| 383 | use_grouped_topk=True, |
| 384 | num_expert_group=getattr(config, "n_group", 1), |
| 385 | topk_group=getattr(config, "topk_group", 1), |
| 386 | prefix=f"{prefix}.experts", |
| 387 | scoring_func=getattr(config, "scoring_func", "softmax"), |
| 388 | routed_scaling_factor=self.routed_scaling_factor, |
| 389 | apply_routed_scale_to_output=apply_routed_scale_to_output, |
| 390 | e_score_correction_bias=self.gate.e_score_correction_bias, |
| 391 | enable_eplb=self.enable_eplb, |
| 392 | num_redundant_experts=self.n_redundant_experts, |
| 393 | is_sequence_parallel=self.is_sequence_parallel, |
| 394 | reduce_results=reduce_results, |
| 395 | n_shared_experts=config.n_shared_experts |
| 396 | if self.is_fused_shared_expert_enabled |
| 397 | else None, |
| 398 | fuse_shared_experts=self.is_fused_shared_expert_enabled, |
| 399 | router_logits_dtype=self.gate.out_dtype, |
| 400 | ) |
| 401 | |
| 402 | if ( |
| 403 | self.is_rocm_aiter_moe_enabled |
| 404 | and self.gate.e_score_correction_bias is not None |
| 405 | ): |
| 406 | self.gate.e_score_correction_bias.data = ( |
| 407 | self.gate.e_score_correction_bias.data.to(self.gate.out_dtype) |
| 408 | ) |
| 409 | |
| 410 | def forward( |
| 411 | self, |
| 412 | hidden_states: torch.Tensor, |
| 413 | already_sequence_parallel: bool = False, |
| 414 | ) -> torch.Tensor: |
| 415 | num_tokens, hidden_dim = hidden_states.shape |
| 416 | hidden_states = hidden_states.view(-1, hidden_dim) |
| 417 | |
| 418 | # Chunk the hidden states so they aren't replicated across TP ranks. |
| 419 | # This avoids duplicate computation in self.experts. |
| 420 | if self.is_sequence_parallel and not already_sequence_parallel: |
| 421 | hidden_states = sequence_parallel_chunk(hidden_states) |
| 422 | |
| 423 | final_hidden_states = self.experts( |
| 424 | hidden_states=hidden_states, router_logits=hidden_states |
| 425 | ) |
| 426 | |
| 427 | if self.is_sequence_parallel and not already_sequence_parallel: |
| 428 | final_hidden_states = tensor_model_parallel_all_gather( |
| 429 | final_hidden_states, 0 |
| 430 | ) |
| 431 | final_hidden_states = final_hidden_states[:num_tokens] |
| 432 | |
| 433 | return final_hidden_states.view(num_tokens, hidden_dim) |
| 434 | |
| 435 | |
| 436 | def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float: |
| 437 | import math |
| 438 | |
| 439 | if scale <= 1: |
| 440 | return 1.0 |
| 441 | return 0.1 * mscale * math.log(scale) + 1.0 |
| 442 | |
| 443 | |
| 444 | def _get_llama_4_scaling( |
| 445 | original_max_position_embeddings: int, scaling_beta: float, positions: torch.Tensor |
| 446 | ) -> torch.Tensor: |
| 447 | scaling = 1 + scaling_beta * torch.log( |
| 448 | 1 + torch.floor(positions / original_max_position_embeddings) |
| 449 | ) |
| 450 | # Broadcast over num_heads and head_dim |
| 451 | return scaling[..., None, None] |
| 452 | |
| 453 | |
| 454 | class DeepseekV2Attention(nn.Module): |
| 455 | def __init__( |
| 456 | self, |
| 457 | vllm_config: VllmConfig, |
| 458 | config: DeepseekV2Config | DeepseekV3Config, |
| 459 | hidden_size: int, |
| 460 | num_heads: int, |
| 461 | qk_nope_head_dim: int, |
| 462 | qk_rope_head_dim: int, |
| 463 | v_head_dim: int, |
| 464 | q_lora_rank: int | None, |
| 465 | kv_lora_rank: int, |
| 466 | max_position_embeddings: int = 8192, |
| 467 | cache_config: CacheConfig | None = None, |
| 468 | quant_config: QuantizationConfig | None = None, |
| 469 | topk_indices_buffer: torch.Tensor | None = None, |
| 470 | reduce_results: bool = True, |
| 471 | prefix: str = "", |
| 472 | ) -> None: |
| 473 | super().__init__() |
| 474 | self.hidden_size = hidden_size |
| 475 | self.qk_nope_head_dim = qk_nope_head_dim |
| 476 | self.qk_rope_head_dim = qk_rope_head_dim |
| 477 | self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim |
| 478 | self.v_head_dim = v_head_dim |
| 479 | self.q_lora_rank = q_lora_rank |
| 480 | self.kv_lora_rank = kv_lora_rank |
| 481 | self.num_heads = num_heads |
| 482 | tp_size = get_tensor_model_parallel_world_size() |
| 483 | assert num_heads % tp_size == 0 |
| 484 | self.num_local_heads = num_heads // tp_size |
| 485 | self.scaling = self.qk_head_dim**-0.5 |
| 486 | self.max_position_embeddings = max_position_embeddings |
| 487 | assert topk_indices_buffer is None, ( |
| 488 | "topk_indices_buffer is not \ |
| 489 | supported for DeepseekV2Attention" |
| 490 | ) |
| 491 | |
| 492 | if self.q_lora_rank is not None: |
| 493 | self.q_a_proj = ReplicatedLinear( |
| 494 | self.hidden_size, |
| 495 | self.q_lora_rank, |
| 496 | bias=False, |
| 497 | quant_config=quant_config, |
| 498 | prefix=f"{prefix}.q_a_proj", |
| 499 | ) |
| 500 | self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) |
| 501 | self.q_b_proj = ColumnParallelLinear( |
| 502 | self.q_lora_rank, |
| 503 | self.num_heads * self.qk_head_dim, |
| 504 | bias=False, |
| 505 | quant_config=quant_config, |
| 506 | prefix=f"{prefix}.q_b_proj", |
| 507 | ) |
| 508 | else: |
| 509 | self.q_proj = ColumnParallelLinear( |
| 510 | self.hidden_size, |
| 511 | self.num_heads * self.qk_head_dim, |
| 512 | bias=False, |
| 513 | quant_config=quant_config, |
| 514 | prefix=f"{prefix}.q_proj", |
| 515 | ) |
| 516 | |
| 517 | self.kv_a_proj_with_mqa = ReplicatedLinear( |
| 518 | self.hidden_size, |
| 519 | self.kv_lora_rank + self.qk_rope_head_dim, |
| 520 | bias=False, |
| 521 | quant_config=quant_config, |
| 522 | prefix=f"{prefix}.kv_a_proj_with_mqa", |
| 523 | ) |
| 524 | self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) |
| 525 | self.kv_b_proj = ColumnParallelLinear( |
| 526 | self.kv_lora_rank, |
| 527 | self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), |
| 528 | bias=False, |
| 529 | quant_config=quant_config, |
| 530 | prefix=f"{prefix}.kv_b_proj", |
| 531 | ) |
| 532 | # O projection. |
| 533 | self.o_proj = RowParallelLinear( |
| 534 | self.num_heads * self.v_head_dim, |
| 535 | self.hidden_size, |
| 536 | bias=False, |
| 537 | reduce_results=reduce_results, |
| 538 | quant_config=quant_config, |
| 539 | prefix=f"{prefix}.o_proj", |
| 540 | ) |
| 541 | if config.rope_parameters["rope_type"] != "default": |
| 542 | config.rope_parameters["rope_type"] = ( |
| 543 | "deepseek_llama_scaling" |
| 544 | if config.rope_parameters.get("attention_factor") == 1.0 |
| 545 | else "deepseek_yarn" |
| 546 | ) |
| 547 | |
| 548 | self.rotary_emb = get_rope( |
| 549 | qk_rope_head_dim, |
| 550 | max_position=max_position_embeddings, |
| 551 | rope_parameters=config.rope_parameters, |
| 552 | is_neox_style=False, |
| 553 | ) |
| 554 | |
| 555 | if ( |
| 556 | config.rope_parameters["rope_type"] != "default" |
| 557 | and config.rope_parameters["rope_type"] == "deepseek_yarn" |
| 558 | ): |
| 559 | mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False) |
| 560 | scaling_factor = config.rope_parameters["factor"] |
| 561 | mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim)) |
| 562 | self.scaling = self.scaling * mscale * mscale |
| 563 | |
| 564 | # DeepSeek is causal by default (decoder-only). Bidirectional |
| 565 | # (encoder-only) attention is used by embedding models derived from |
| 566 | # this architecture, which set `is_causal=False` on the HF config. |
| 567 | # This only applies on the non-MLA path, since the MLA kernels are |
| 568 | # causal-only; `ModelConfig.use_mla` already disables MLA for these |
| 569 | # models. |
| 570 | if getattr(config, "is_causal", True): |
| 571 | attn_type = AttentionType.DECODER |
| 572 | else: |
| 573 | attn_type = AttentionType.ENCODER_ONLY |
| 574 | attn_cls = ( |
| 575 | EncoderOnlyAttention |
| 576 | if attn_type == AttentionType.ENCODER_ONLY |
| 577 | else Attention |
| 578 | ) |
| 579 | self.attn = attn_cls( |
| 580 | self.num_local_heads, |
| 581 | self.qk_head_dim, |
| 582 | self.scaling, |
| 583 | num_kv_heads=self.num_local_heads, |
| 584 | cache_config=cache_config, |
| 585 | quant_config=quant_config, |
| 586 | prefix=f"{prefix}.attn", |
| 587 | attn_type=attn_type, |
| 588 | ) |
| 589 | |
| 590 | def forward( |
| 591 | self, |
| 592 | positions: torch.Tensor, |
| 593 | hidden_states: torch.Tensor, |
| 594 | llama_4_scaling: torch.Tensor | None, |
| 595 | ) -> torch.Tensor: |
| 596 | if self.q_lora_rank is not None: |
| 597 | q = self.q_a_proj(hidden_states)[0] |
| 598 | q = self.q_a_layernorm(q) |
| 599 | q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim) |
| 600 | else: |
| 601 | q = self.q_proj(hidden_states)[0].view( |
| 602 | -1, self.num_local_heads, self.qk_head_dim |
| 603 | ) |
| 604 | q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1) |
| 605 | latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0] |
| 606 | kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1) |
| 607 | latent_cache = latent_cache.unsqueeze(1) |
| 608 | kv_a = self.kv_a_layernorm(kv_a) |
| 609 | kv = self.kv_b_proj(kv_a)[0] |
| 610 | kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim) |
| 611 | k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1) |
| 612 | k_pe = latent_cache[:, :, self.kv_lora_rank :] |
| 613 | |
| 614 | q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe) |
| 615 | |
| 616 | q[..., self.qk_nope_head_dim :] = q_pe |
| 617 | k = torch.empty_like(q) |
| 618 | k[..., : self.qk_nope_head_dim] = k_nope |
| 619 | k[..., self.qk_nope_head_dim :] = k_pe |
| 620 | |
| 621 | # Apply llama 4 scaling if provided |
| 622 | if llama_4_scaling is not None: |
| 623 | q *= llama_4_scaling |
| 624 | |
| 625 | # padding value to qk_head_dim for alignment |
| 626 | v = torch.nn.functional.pad( |
| 627 | v, [0, self.qk_head_dim - self.v_head_dim], value=0 |
| 628 | ).view(-1, self.num_local_heads * self.qk_head_dim) |
| 629 | attn_output = self.attn(q, k, v) |
| 630 | attn_output = attn_output.view(-1, self.num_local_heads, self.qk_head_dim)[ |
| 631 | ..., : self.v_head_dim |
| 632 | ].reshape(-1, self.num_local_heads * self.v_head_dim) |
| 633 | output, _ = self.o_proj(attn_output) |
| 634 | return output |
| 635 | |
| 636 | |
| 637 | class DeepseekV32IndexerCache(torch.nn.Module, AttentionLayerBase): |
| 638 | def __init__( |
| 639 | self, head_dim: int, dtype: torch.dtype, prefix: str, cache_config: CacheConfig |
| 640 | ): |
| 641 | super().__init__() |
| 642 | self.kv_cache = torch.tensor([]) |
| 643 | self.head_dim = head_dim |
| 644 | self.prefix = prefix |
| 645 | self.cache_config = cache_config |
| 646 | self.dtype = dtype |
| 647 | compilation_config = get_current_vllm_config().compilation_config |
| 648 | if prefix in compilation_config.static_forward_context: |
| 649 | raise ValueError(f"Duplicate layer name: {prefix}") |
| 650 | compilation_config.static_forward_context[prefix] = self |
| 651 | |
| 652 | def bind_kv_cache(self, kv_cache: torch.Tensor) -> None: |
| 653 | # [B, H=1, N, C] -> [B, N, C]: the indexer kernels and |
| 654 | # kv_cache_as_quant_view index a 3-D block-major cache. |
| 655 | self.kv_cache = kv_cache.squeeze(1) |
| 656 | |
| 657 | def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec: |
| 658 | return MLAAttentionSpec( |
| 659 | block_size=self.cache_config.block_size, |
| 660 | num_kv_heads=1, |
| 661 | max_tp_shards=1, |
| 662 | head_size=self.head_dim, |
| 663 | dtype=self.dtype, |
| 664 | cache_role=SparseCacheRole.INDEXER, |
| 665 | ) # Only has one vector instead of K + V |
| 666 | |
| 667 | def forward(self): ... |
| 668 | |
| 669 | @property |
| 670 | def uses_shuffled_layout(self) -> bool: |
| 671 | """Whether this cache's reader expects the shuffled value layout.""" |
| 672 | return False |
| 673 | |
| 674 | def get_attn_backend(self) -> type[AttentionBackend]: |
| 675 | return DeepseekV32IndexerBackend |
| 676 | |
| 677 | |
| 678 | class Indexer(nn.Module): |
| 679 | def __init__( |
| 680 | self, |
| 681 | vllm_config: VllmConfig, |
| 682 | config: DeepseekV2Config | DeepseekV3Config, |
| 683 | hidden_size: int, |
| 684 | q_lora_rank: int, |
| 685 | quant_config: QuantizationConfig | None, |
| 686 | cache_config: CacheConfig | None, |
| 687 | topk_indices_buffer: torch.Tensor | None, |
| 688 | prefix: str = "", |
| 689 | is_inplace_rope: bool = False, |
| 690 | ): |
| 691 | super().__init__() |
| 692 | self.vllm_config = vllm_config |
| 693 | self.config = config |
| 694 | self.quant_config = quant_config |
| 695 | # self.indexer_cfg = config.attn_module_list_cfg[0]["attn_index"] |
| 696 | self.topk_tokens = config.index_topk |
| 697 | self.n_head = config.index_n_heads # 64 |
| 698 | self.head_dim = config.index_head_dim # 128 |
| 699 | self.rope_dim = config.qk_rope_head_dim # 64 |
| 700 | self.q_lora_rank = q_lora_rank # 1536 |
| 701 | # no tensor parallel, just replicated |
| 702 | self.wq_b = ReplicatedLinear( |
| 703 | self.q_lora_rank, |
| 704 | self.head_dim * self.n_head, |
| 705 | bias=False, |
| 706 | quant_config=quant_config, |
| 707 | prefix=f"{prefix}.wq_b", |
| 708 | ) |
| 709 | # Fused wk + weights_proj: single GEMM producing [head_dim + n_head]. |
| 710 | # FP8 wk weights are upcasted to BF16 during loading to maintain fusion. |
| 711 | self.wk_weights_proj = MergedColumnParallelLinear( |
| 712 | hidden_size, |
| 713 | [self.head_dim, self.n_head], |
| 714 | bias=False, |
| 715 | quant_config=None, |
| 716 | disable_tp=True, |
| 717 | prefix=f"{prefix}.wk_weights_proj", |
| 718 | ) |
| 719 | self.k_norm = LayerNorm(self.head_dim, eps=1e-6, dtype=torch.float32) |
| 720 | self.softmax_scale = self.head_dim**-0.5 |
| 721 | |
| 722 | self.scale_fmt = "ue8m0" |
| 723 | self.quant_block_size = 128 # TODO: get from config |
| 724 | self.topk_indices_buffer = topk_indices_buffer |
| 725 | |
| 726 | # NOTE: (zyongye) we use fp8 naive cache, |
| 727 | # where we store value in fp8 and scale in fp32 |
| 728 | # per self.quant_block_size element |
| 729 | assert cache_config is not None |
| 730 | self.k_cache = DeepseekV32IndexerCache( |
| 731 | head_dim=self.head_dim + self.head_dim // self.quant_block_size * 4, |
| 732 | dtype=torch.uint8, |
| 733 | prefix=f"{prefix}.k_cache", |
| 734 | cache_config=cache_config, |
| 735 | ) |
| 736 | self.max_model_len = vllm_config.model_config.max_model_len |
| 737 | self.prefix = prefix |
| 738 | from vllm.v1.attention.backends.mla.indexer import get_max_prefill_buffer_size |
| 739 | |
| 740 | self.max_total_seq_len = get_max_prefill_buffer_size(vllm_config) |
| 741 | self.indexer_op = SparseAttnIndexer( |
| 742 | self.k_cache, |
| 743 | self.quant_block_size, |
| 744 | self.scale_fmt, |
| 745 | self.topk_tokens, |
| 746 | self.head_dim, |
| 747 | self.max_model_len, |
| 748 | self.max_total_seq_len, |
| 749 | self.topk_indices_buffer, |
| 750 | ) |
| 751 | |
| 752 | self.is_inplace_rope = is_inplace_rope |
| 753 | self.n_head_scale = self.n_head**-0.5 |
| 754 | self.use_fused_indexer_q = ( |
| 755 | current_platform.is_cuda() |
| 756 | and self.quant_block_size == self.head_dim |
| 757 | and self.head_dim == 128 |
| 758 | and self.rope_dim == 64 |
| 759 | and self.scale_fmt is not None |
| 760 | ) |
| 761 | |
| 762 | def forward( |
| 763 | self, hidden_states: torch.Tensor, qr: torch.Tensor, positions, rotary_emb |
| 764 | ) -> torch.Tensor: |
| 765 | q, _ = self.wq_b(qr) |
| 766 | q = q.view(-1, self.n_head, self.head_dim) |
| 767 | |
| 768 | if current_platform.is_rocm() and self.is_inplace_rope: |
| 769 | # This path should works on all platform, will remove extra |
| 770 | # branches in the future |
| 771 | # This fast path relies on rotary_emb mutating q and k inplace. |
| 772 | # On ROCm, this is only valid for kernels used as custom ops. |
| 773 | # In pytorch-native rope for inductor fusion, rotated q/k tensors |
| 774 | # are not mutated inplace but returned as new tensors. |
| 775 | # Fused wk + weights_proj: one GEMM, then split |
| 776 | kw, _ = self.wk_weights_proj(hidden_states) |
| 777 | k = kw[:, : self.head_dim] |
| 778 | weights = kw[:, self.head_dim :] |
| 779 | |
| 780 | k = self.k_norm(k) |
| 781 | |
| 782 | rotary_emb( |
| 783 | positions, q[..., : self.rope_dim], k[..., : self.rope_dim].unsqueeze(1) |
| 784 | ) |
| 785 | elif self.use_fused_indexer_q and q.dtype == torch.bfloat16: |
| 786 | # fused wk + weights_proj: one GEMM, then split |
| 787 | kw, _ = self.wk_weights_proj(hidden_states) |
| 788 | k = kw[:, : self.head_dim] |
| 789 | weights = kw[:, self.head_dim :] |
| 790 | |
| 791 | k = self.k_norm(k) |
| 792 | k_pe, k_nope = torch.split( |
| 793 | k, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1 |
| 794 | ) |
| 795 | |
| 796 | q_fp8, weights = fused_indexer_q_rope_quant( |
| 797 | positions, |
| 798 | q, |
| 799 | rotary_emb.cos_sin_cache, |
| 800 | weights, |
| 801 | self.softmax_scale, |
| 802 | self.n_head_scale, |
| 803 | rotary_emb.is_neox_style, |
| 804 | ) |
| 805 | |
| 806 | # rotate only the MQA K |
| 807 | k_pe = k_pe.unsqueeze(1) |
| 808 | q_dummy = torch.empty_like(k_pe) |
| 809 | _, k_pe = rotary_emb(positions, q_dummy, k_pe) |
| 810 | k_pe = k_pe.reshape(-1, self.rope_dim) |
| 811 | k = torch.cat([k_pe, k_nope], dim=-1) |
| 812 | |
| 813 | return self.indexer_op(hidden_states, q_fp8, k, weights) |
| 814 | else: |
| 815 | q_pe, q_nope = torch.split( |
| 816 | q, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1 |
| 817 | ) |
| 818 | # Fused wk + weights_proj: one GEMM, then split |
| 819 | kw, _ = self.wk_weights_proj(hidden_states) |
| 820 | k = kw[:, : self.head_dim] |
| 821 | weights = kw[:, self.head_dim :] |
| 822 | |
| 823 | k = self.k_norm(k) |
| 824 | k_pe, k_nope = torch.split( |
| 825 | k, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1 |
| 826 | ) |
| 827 | |
| 828 | q_pe, k_pe = rotary_emb(positions, q_pe, k_pe.unsqueeze(1)) |
| 829 | # Note: RoPE (NeoX) can introduce extra leading dimensions during |
| 830 | # compilation so we need to reshape back to token-flattened shapes |
| 831 | q_pe = q_pe.reshape(-1, self.n_head, self.rope_dim) |
| 832 | k_pe = k_pe.reshape(-1, self.rope_dim) |
| 833 | |
| 834 | # `rotary_emb` is shape-preserving; `q_pe` is already |
| 835 | # [num_tokens, n_head, rope_dim]. |
| 836 | q = torch.cat([q_pe, q_nope], dim=-1) |
| 837 | # `k_pe` is [num_tokens, rope_dim] (MQA). |
| 838 | k = torch.cat([k_pe, k_nope], dim=-1) |
| 839 | |
| 840 | # we only quant q here since k quant is fused with cache insertion |
| 841 | q = q.view(-1, self.head_dim) |
| 842 | q_fp8, q_scale = per_token_group_quant_fp8( |
| 843 | q, |
| 844 | self.quant_block_size, |
| 845 | column_major_scales=False, |
| 846 | use_ue8m0=self.scale_fmt is not None, |
| 847 | ) |
| 848 | q_fp8 = q_fp8.view(-1, self.n_head, self.head_dim) |
| 849 | q_scale = q_scale.view(-1, self.n_head) |
| 850 | |
| 851 | weights = weights * q_scale * self.softmax_scale * self.n_head_scale |
| 852 | |
| 853 | return self.indexer_op(hidden_states, q_fp8, k, weights) |
| 854 | |
| 855 | |
| 856 | def _try_load_fp8_indexer_wk( |
| 857 | name, tensor, buf, params_dict, loaded_params, pp_missing_layer_names |
| 858 | ): |
| 859 | """We fuse the WK and weights_proj projections, but in some checkpoints WK is stored |
| 860 | in FP8 with a separate weight_scale_inv, while weights_proj is stored in BF16. |
| 861 | Upcasting to BF16 during loading enables the fusion. This function loads the FP8 WK |
| 862 | weights and scale, and when both are available, dequantizes to BF16 and stores into |
| 863 | the fused wk_weights_proj.weight parameter. |
| 864 | """ |
| 865 | if "indexer.wk." not in name or "wk_weights" in name: |
| 866 | return False # Weight is not an isolated WK weight for the indexer, ignore. |
| 867 | is_weight = name.endswith(".weight") and tensor.dtype == torch.float8_e4m3fn |
| 868 | is_scale = "weight_scale" in name |
| 869 | if not is_weight and not is_scale: |
| 870 | return False # WK is not in FP8 format, ignore. |
| 871 | # Buffer this tensor (weight or scale) until both have arrived. |
| 872 | layer_prefix = name.rsplit(".wk.", 1)[0] # e.g. "model.layers.0.self_attn.indexer" |
| 873 | fused_name = f"{layer_prefix}.wk_weights_proj.weight" |
| 874 | if any( |
| 875 | name.startswith(missing_layer_name) |
| 876 | for missing_layer_name in pp_missing_layer_names |
| 877 | ): |
| 878 | return True |
| 879 | entry = buf.setdefault(layer_prefix, {}) |
| 880 | entry["weight" if is_weight else "scale"] = tensor |
| 881 | if "weight" not in entry or "scale" not in entry: |
| 882 | return True # still waiting for the other param |
| 883 | |
| 884 | # We have both weight and scale: dequantize FP8 to BF16. |
| 885 | weight_fp8, scale_inv = entry["weight"], entry["scale"] |
| 886 | del buf[layer_prefix] |
| 887 | if scale_inv.dtype in (torch.uint8, torch.float8_e8m0fnu): |
| 888 | # MXFP8 checkpoints store power-of-two E8M0 scales, decode them to float. |
| 889 | scale_inv = scale_inv.view(torch.float8_e8m0fnu).to(torch.float32) |
| 890 | if scale_inv.ndim == 1: |
| 891 | # Per-channel scale: one scale per row of [out, in] |
| 892 | group_shape = GroupShape(1, weight_fp8.shape[1]) |
| 893 | else: |
| 894 | # Derive the block size independently per dim so that per-channel |
| 895 | # ([out, 1]), MX ([out, in / 32]) and 2D block ([out / 128, in / 128]) |
| 896 | # scale layouts are all handled. |
| 897 | group_shape = GroupShape( |
| 898 | weight_fp8.shape[0] // scale_inv.shape[0], |
| 899 | weight_fp8.shape[1] // scale_inv.shape[1], |
| 900 | ) |
| 901 | weight_bf16 = scaled_dequantize( |
| 902 | weight_fp8, |
| 903 | scale_inv, |
| 904 | group_shape=group_shape, |
| 905 | out_dtype=torch.bfloat16, |
| 906 | ) |
| 907 | |
| 908 | # Load the dequantized weight into shard 0 of the fused buffer. |
| 909 | param = params_dict[fused_name] |
| 910 | param.weight_loader(param, weight_bf16, 0) |
| 911 | loaded_params.add(fused_name) |
| 912 | return True |
| 913 | |
| 914 | |
| 915 | def _min_latency_fused_qkv_a_proj_impl( |
| 916 | input_: torch.Tensor, |
| 917 | weight: torch.Tensor, |
| 918 | ) -> torch.Tensor: |
| 919 | """Dynamically run min-latency gemm if num_tokens <= 16. |
| 920 | This must be wrapped in a custom op because our torch.compile integration |
| 921 | does not support runtime dispatching on num_tokens. |
| 922 | """ |
| 923 | num_tokens = input_.shape[0] |
| 924 | if 0 < num_tokens <= 16: |
| 925 | output = torch.empty( |
| 926 | num_tokens, |
| 927 | weight.shape[0], |
| 928 | dtype=torch.bfloat16, |
| 929 | device=input_.device, |
| 930 | ) |
| 931 | ops.dsv3_fused_a_gemm(output, input_, weight.T) |
| 932 | return output |
| 933 | else: |
| 934 | return torch.nn.functional.linear(input_, weight) |
| 935 | |
| 936 | |
| 937 | def _min_latency_fused_qkv_a_proj_fake( |
| 938 | input_: torch.Tensor, |
| 939 | weight: torch.Tensor, |
| 940 | ) -> torch.Tensor: |
| 941 | return input_.new_empty(input_.shape[0], weight.shape[0]) |
| 942 | |
| 943 | |
| 944 | direct_register_custom_op( |
| 945 | op_name="min_latency_fused_qkv_a_proj", |
| 946 | op_func=_min_latency_fused_qkv_a_proj_impl, |
| 947 | mutates_args=[], |
| 948 | fake_impl=_min_latency_fused_qkv_a_proj_fake, |
| 949 | ) |
| 950 | |
| 951 | |
| 952 | class DeepSeekV2FusedQkvAProjLinear(MergedColumnParallelLinear): |
| 953 | def __init__( |
| 954 | self, |
| 955 | input_size: int, |
| 956 | output_size: list[int], |
| 957 | quant_config: QuantizationConfig | None = None, |
| 958 | prefix: str = "", |
| 959 | ): |
| 960 | super().__init__( |
| 961 | input_size, |
| 962 | output_size, |
| 963 | bias=False, |
| 964 | quant_config=quant_config, |
| 965 | disable_tp=True, |
| 966 | prefix=prefix, |
| 967 | ) |
| 968 | |
| 969 | # Check if the DeepSeek V3 fused A GEMM kernel can be used. |
| 970 | # This kernel supports PDL and is optimized for low batch size. |
| 971 | self._use_min_latency_gemm = ( |
| 972 | hasattr(self, "weight") |
| 973 | and self.weight.dtype == torch.bfloat16 |
| 974 | and self.weight.shape[0] == 2112 |
| 975 | and self.weight.shape[1] == 7168 |
| 976 | and current_platform.is_cuda() |
| 977 | and ( |
| 978 | current_platform.is_device_capability(90) |
| 979 | or current_platform.is_device_capability_family(100) |
| 980 | ) |
| 981 | ) |
| 982 | |
| 983 | def forward( |
| 984 | self, |
| 985 | input_, |
| 986 | ) -> torch.Tensor | tuple[torch.Tensor, torch.nn.Parameter | None]: |
| 987 | if self._use_min_latency_gemm: |
| 988 | output = torch.ops.vllm.min_latency_fused_qkv_a_proj(input_, self.weight) |
| 989 | if not self.return_bias: |
| 990 | return output |
| 991 | output_bias = self.bias if self.skip_bias_add else None |
| 992 | return output, output_bias |
| 993 | else: |
| 994 | # Fallback to the standard forward method when |
| 995 | # the fused A GEMM kernel cannot be used. |
| 996 | return super().forward(input_) |
| 997 | |
| 998 | |
| 999 | class DeepseekV2MLAAttention(nn.Module): |
| 1000 | """Main reference: DeepseekV2 paper, and FlashInfer Implementation |
| 1001 | (https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551). |
| 1002 | |
| 1003 | For more info see MLACommonImpl in: |
| 1004 | vllm/v1/attention/backends/mla/utils.py |
| 1005 | """ |
| 1006 | |
| 1007 | def __init__( |
| 1008 | self, |
| 1009 | vllm_config: VllmConfig, |
| 1010 | config: DeepseekV2Config | DeepseekV3Config, |
| 1011 | hidden_size: int, |
| 1012 | num_heads: int, |
| 1013 | qk_nope_head_dim: int, |
| 1014 | qk_rope_head_dim: int, |
| 1015 | v_head_dim: int, |
| 1016 | q_lora_rank: int | None, |
| 1017 | kv_lora_rank: int, |
| 1018 | max_position_embeddings: int = 8192, |
| 1019 | cache_config: CacheConfig | None = None, |
| 1020 | quant_config: QuantizationConfig | None = None, |
| 1021 | prefix: str = "", |
| 1022 | topk_indices_buffer: torch.Tensor | None = None, |
| 1023 | index_group_builder: SparseMLAIndexGroupBuilder | None = None, |
| 1024 | input_size: int | None = None, |
| 1025 | reduce_results: bool = True, |
| 1026 | non_causal_multi_token_decode: bool = False, |
| 1027 | ) -> None: |
| 1028 | super().__init__() |
| 1029 | self.hidden_size = hidden_size |
| 1030 | self.qk_nope_head_dim = qk_nope_head_dim |
| 1031 | self.qk_rope_head_dim = qk_rope_head_dim |
| 1032 | self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim |
| 1033 | self.v_head_dim = v_head_dim |
| 1034 | |
| 1035 | self.q_lora_rank = q_lora_rank |
| 1036 | self.kv_lora_rank = kv_lora_rank |
| 1037 | |
| 1038 | self.num_heads = num_heads |
| 1039 | tp_size = get_tensor_model_parallel_world_size() |
| 1040 | assert num_heads % tp_size == 0 |
| 1041 | self.num_local_heads = num_heads // tp_size |
| 1042 | |
| 1043 | self.scaling = self.qk_head_dim**-0.5 |
| 1044 | self.max_position_embeddings = max_position_embeddings |
| 1045 | |
| 1046 | # Use input_size for projection input dimensions if provided, |
| 1047 | # otherwise default to hidden_size (used in Eagle3 Deepseek with MLA) |
| 1048 | proj_input_size = input_size if input_size is not None else self.hidden_size |
| 1049 | |
| 1050 | if self.q_lora_rank is not None: |
| 1051 | self.fused_qkv_a_proj = DeepSeekV2FusedQkvAProjLinear( |
| 1052 | proj_input_size, |
| 1053 | [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], |
| 1054 | quant_config=quant_config, |
| 1055 | prefix=f"{prefix}.fused_qkv_a_proj", |
| 1056 | ) |
| 1057 | else: |
| 1058 | self.kv_a_proj_with_mqa = ReplicatedLinear( |
| 1059 | proj_input_size, |
| 1060 | self.kv_lora_rank + self.qk_rope_head_dim, |
| 1061 | bias=False, |
| 1062 | quant_config=quant_config, |
| 1063 | prefix=f"{prefix}.kv_a_proj_with_mqa", |
| 1064 | ) |
| 1065 | |
| 1066 | # The env var predates the config field and still wins if set explicitly. |
| 1067 | qrep_requested = ( |
| 1068 | envs.VLLM_DCP_Q_REPLICATE |
| 1069 | if envs.is_set("VLLM_DCP_Q_REPLICATE") |
| 1070 | else bool(vllm_config.parallel_config.dcp_q_replicate) |
| 1071 | ) |
| 1072 | qrep_enabled = ( |
| 1073 | qrep_requested |
| 1074 | and vllm_config.parallel_config.decode_context_parallel_size > 1 |
| 1075 | and vllm_config.parallel_config.prefill_context_parallel_size <= 1 |
| 1076 | ) |
| 1077 | q_proj_cls = ( |
| 1078 | DCPGroupColumnParallelLinear if qrep_enabled else ColumnParallelLinear |
| 1079 | ) |
| 1080 | if self.q_lora_rank is not None: |
| 1081 | self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) |
| 1082 | self.q_b_proj = q_proj_cls( |
| 1083 | self.q_lora_rank, |
| 1084 | self.num_heads * self.qk_head_dim, |
| 1085 | bias=False, |
| 1086 | quant_config=quant_config, |
| 1087 | prefix=f"{prefix}.q_b_proj", |
| 1088 | ) |
| 1089 | else: |
| 1090 | self.q_proj = q_proj_cls( |
| 1091 | proj_input_size, |
| 1092 | self.num_heads * self.qk_head_dim, |
| 1093 | bias=False, |
| 1094 | quant_config=quant_config, |
| 1095 | prefix=f"{prefix}.q_proj", |
| 1096 | ) |
| 1097 | self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) |
| 1098 | self.kv_b_proj = ColumnParallelLinear( |
| 1099 | self.kv_lora_rank, |
| 1100 | self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), |
| 1101 | bias=False, |
| 1102 | quant_config=quant_config, |
| 1103 | prefix=f"{prefix}.kv_b_proj", |
| 1104 | ) |
| 1105 | self.o_proj = RowParallelLinear( |
| 1106 | self.num_heads * self.v_head_dim, |
| 1107 | self.hidden_size, |
| 1108 | bias=False, |
| 1109 | reduce_results=reduce_results, |
| 1110 | quant_config=quant_config, |
| 1111 | prefix=f"{prefix}.o_proj", |
| 1112 | ) |
| 1113 | |
| 1114 | if config.rope_parameters["rope_type"] != "default": |
| 1115 | config.rope_parameters["rope_type"] = ( |
| 1116 | "deepseek_llama_scaling" |
| 1117 | if config.rope_parameters.get("attention_factor") == 1.0 |
| 1118 | else "deepseek_yarn" |
| 1119 | ) |
| 1120 | |
| 1121 | self.rotary_emb = get_rope( |
| 1122 | qk_rope_head_dim, |
| 1123 | max_position=max_position_embeddings, |
| 1124 | rope_parameters=config.rope_parameters, |
| 1125 | is_neox_style=False, |
| 1126 | ) |
| 1127 | |
| 1128 | if ( |
| 1129 | config.rope_parameters["rope_type"] != "default" |
| 1130 | and config.rope_parameters["rope_type"] == "deepseek_yarn" |
| 1131 | ): |
| 1132 | mscale_all_dim = config.rope_parameters.get("mscale_all_dim", False) |
| 1133 | scaling_factor = config.rope_parameters["factor"] |
| 1134 | mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim)) |
| 1135 | self.scaling = self.scaling * mscale * mscale |
| 1136 | |
| 1137 | self.is_v32 = hasattr(config, "index_topk") |
| 1138 | |
| 1139 | # IndexCache config |
| 1140 | # Refer: https://arxiv.org/abs/2603.12201 for more details. |
| 1141 | _skip_topk = False |
| 1142 | is_mtp_layer = False |
| 1143 | if self.is_v32: |
| 1144 | _index_topk_freq = getattr(config, "index_topk_freq", 1) |
| 1145 | _index_topk_pattern = getattr(config, "index_topk_pattern", None) |
| 1146 | _index_skip_topk_offset = getattr(config, "index_skip_topk_offset", 2) |
| 1147 | layer_id = extract_layer_index(prefix) |
| 1148 | |
| 1149 | if _index_topk_pattern is None: |
| 1150 | _skip_topk = ( |
| 1151 | max(layer_id - _index_skip_topk_offset + 1, 0) % _index_topk_freq |
| 1152 | != 0 |
| 1153 | ) |
| 1154 | elif 0 <= layer_id < len(_index_topk_pattern): |
| 1155 | _skip_topk = _index_topk_pattern[layer_id] == "S" |
| 1156 | |
| 1157 | # The skip pattern only governs backbone layers. MTP/nextn |
| 1158 | # layers (layer_id >= num_hidden_layers) always build a full |
| 1159 | # indexer: they compute indices at draft step 0 and toggle |
| 1160 | # at runtime via set_skip_topk |
| 1161 | # (index_share_for_mtp_iteration). |
| 1162 | _num_hidden_layers = getattr(config, "num_hidden_layers", None) |
| 1163 | is_mtp_layer = ( |
| 1164 | _num_hidden_layers is not None and layer_id >= _num_hidden_layers |
| 1165 | ) |
| 1166 | |
| 1167 | if self.is_v32 and (not _skip_topk or is_mtp_layer): |
| 1168 | assert q_lora_rank is not None |
| 1169 | assert cache_config is not None |
| 1170 | self.indexer_rope_emb: nn.Module | None |
| 1171 | self.indexer: Indexer | None |
| 1172 | self.indexer_rope_emb = get_rope( |
| 1173 | qk_rope_head_dim, |
| 1174 | max_position=max_position_embeddings, |
| 1175 | rope_parameters=config.rope_parameters, |
| 1176 | is_neox_style=not getattr(config, "indexer_rope_interleave", False), |
| 1177 | ) |
| 1178 | self.indexer = Indexer( |
| 1179 | vllm_config, |
| 1180 | config, |
| 1181 | hidden_size, |
| 1182 | q_lora_rank, |
| 1183 | quant_config, |
| 1184 | cache_config, |
| 1185 | topk_indices_buffer, |
| 1186 | f"{prefix}.indexer", |
| 1187 | is_inplace_rope=self.indexer_rope_emb.enabled(), |
| 1188 | ) |
| 1189 | else: |
| 1190 | self.indexer_rope_emb = None |
| 1191 | self.indexer = None |
| 1192 | |
| 1193 | mla_modules = MLAModules( |
| 1194 | kv_a_layernorm=self.kv_a_layernorm, |
| 1195 | kv_b_proj=self.kv_b_proj, |
| 1196 | rotary_emb=self.rotary_emb, |
| 1197 | o_proj=self.o_proj, |
| 1198 | fused_qkv_a_proj=self.fused_qkv_a_proj |
| 1199 | if self.q_lora_rank is not None |
| 1200 | else None, |
| 1201 | kv_a_proj_with_mqa=self.kv_a_proj_with_mqa |
| 1202 | if self.q_lora_rank is None |
| 1203 | else None, |
| 1204 | q_a_layernorm=self.q_a_layernorm if self.q_lora_rank is not None else None, |
| 1205 | q_b_proj=self.q_b_proj if self.q_lora_rank is not None else None, |
| 1206 | q_proj=self.q_proj if self.q_lora_rank is None else None, |
| 1207 | indexer=self.indexer, |
| 1208 | indexer_rotary_emb=self.indexer_rope_emb, |
| 1209 | is_sparse=self.is_v32, |
| 1210 | topk_indices_buffer=topk_indices_buffer, |
| 1211 | index_group_builder=index_group_builder, |
| 1212 | ) |
| 1213 | |
| 1214 | self.mla_attn = MultiHeadLatentAttentionWrapper( |
| 1215 | self.hidden_size, |
| 1216 | self.num_local_heads, |
| 1217 | self.scaling, |
| 1218 | self.qk_nope_head_dim, |
| 1219 | self.qk_rope_head_dim, |
| 1220 | self.v_head_dim, |
| 1221 | self.q_lora_rank, |
| 1222 | self.kv_lora_rank, |
| 1223 | mla_modules, |
| 1224 | cache_config, |
| 1225 | quant_config, |
| 1226 | prefix, |
| 1227 | # MTP layers must never start with skip_topk=True: their indexer |
| 1228 | # computes indices at draft step 0, and the runtime toggle |
| 1229 | # (set_skip_topk, index_share_for_mtp_iteration) only exists in |
| 1230 | # the V1 proposer. A frozen True would leave the draft reading a |
| 1231 | # never-written topk buffer. |
| 1232 | skip_topk=_skip_topk and not is_mtp_layer, |
| 1233 | non_causal_multi_token_decode=non_causal_multi_token_decode, |
| 1234 | # Do not skip scoring for MTP layers: their top-k buffer may be |
| 1235 | # reused by later draft iterations through index sharing. |
| 1236 | allow_short_prefill_indexer_scoring_skip=not is_mtp_layer, |
| 1237 | ) |
| 1238 | |
| 1239 | def forward( |
| 1240 | self, |
| 1241 | positions: torch.Tensor, |
| 1242 | hidden_states: torch.Tensor, |
| 1243 | llama_4_scaling: torch.Tensor | None, |
| 1244 | ) -> torch.Tensor: |
| 1245 | return self.mla_attn(positions, hidden_states, llama_4_scaling) |
| 1246 | |
| 1247 | |
| 1248 | class DeepseekV2DecoderLayer(nn.Module): |
| 1249 | def __init__( |
| 1250 | self, |
| 1251 | vllm_config: VllmConfig, |
| 1252 | prefix: str, |
| 1253 | config: DeepseekV2Config | None = None, |
| 1254 | topk_indices_buffer: torch.Tensor | None = None, |
| 1255 | index_group_builder: SparseMLAIndexGroupBuilder | None = None, |
| 1256 | ) -> None: |
| 1257 | super().__init__() |
| 1258 | |
| 1259 | if config is None: |
| 1260 | config = vllm_config.model_config.hf_config |
| 1261 | model_config = vllm_config.model_config |
| 1262 | cache_config = vllm_config.cache_config |
| 1263 | quant_config = vllm_config.quant_config |
| 1264 | parallel_config = vllm_config.parallel_config |
| 1265 | |
| 1266 | self.hidden_size = config.hidden_size |
| 1267 | max_position_embeddings = getattr(config, "max_position_embeddings", 8192) |
| 1268 | moe_layer_freq = getattr(config, "moe_layer_freq", 1) |
| 1269 | # DecoderLayers are created with `make_layers` which passes the prefix |
| 1270 | # with the layer's index. |
| 1271 | layer_idx = int(prefix.split(sep=".")[-1]) |
| 1272 | self.layer_idx = layer_idx |
| 1273 | |
| 1274 | # verify MLA attention specific fields |
| 1275 | qk_nope_head_dim = getattr(config, "qk_nope_head_dim", 0) |
| 1276 | qk_rope_head_dim = getattr(config, "qk_rope_head_dim", 0) |
| 1277 | v_head_dim = getattr(config, "v_head_dim", 0) |
| 1278 | kv_lora_rank = getattr(config, "kv_lora_rank", 0) |
| 1279 | use_mha = config.model_type == "deepseek" or all( |
| 1280 | dim == 0 for dim in (qk_nope_head_dim, qk_rope_head_dim) |
| 1281 | ) |
| 1282 | |
| 1283 | self.use_mha = use_mha |
| 1284 | |
| 1285 | attn_cls: typing.Any |
| 1286 | if use_mha: |
| 1287 | attn_cls = DeepseekAttention |
| 1288 | elif model_config.use_mla: |
| 1289 | attn_cls = DeepseekV2MLAAttention |
| 1290 | else: |
| 1291 | attn_cls = DeepseekV2Attention |
| 1292 | is_moe_layer = ( |
| 1293 | config.n_routed_experts is not None |
| 1294 | and layer_idx >= config.first_k_dense_replace |
| 1295 | and layer_idx % moe_layer_freq == 0 |
| 1296 | ) |
| 1297 | # TODO(wentao): enable SP MoE with PP after the PP boundary logic can safely |
| 1298 | # send/receive sequence-parallel hidden_states across stages. |
| 1299 | self.use_sequence_parallel_moe = ( |
| 1300 | parallel_config.use_sequence_parallel_moe |
| 1301 | and parallel_config.pipeline_parallel_size == 1 |
| 1302 | and is_moe_layer |
| 1303 | ) |
| 1304 | attn_kwargs = ( |
| 1305 | {"index_group_builder": index_group_builder} |
| 1306 | if attn_cls is DeepseekV2MLAAttention |
| 1307 | else {} |
| 1308 | ) |
| 1309 | self.self_attn = attn_cls( |
| 1310 | vllm_config=vllm_config, |
| 1311 | config=config, |
| 1312 | hidden_size=self.hidden_size, |
| 1313 | num_heads=config.num_attention_heads, |
| 1314 | qk_nope_head_dim=qk_nope_head_dim, |
| 1315 | qk_rope_head_dim=qk_rope_head_dim, |
| 1316 | v_head_dim=v_head_dim, |
| 1317 | q_lora_rank=config.q_lora_rank if hasattr(config, "q_lora_rank") else None, |
| 1318 | kv_lora_rank=kv_lora_rank, |
| 1319 | max_position_embeddings=max_position_embeddings, |
| 1320 | cache_config=cache_config, |
| 1321 | quant_config=quant_config, |
| 1322 | prefix=f"{prefix}.self_attn", |
| 1323 | topk_indices_buffer=topk_indices_buffer, |
| 1324 | reduce_results=not self.use_sequence_parallel_moe, |
| 1325 | **attn_kwargs, |
| 1326 | ) |
| 1327 | |
| 1328 | if is_moe_layer: |
| 1329 | self.mlp = DeepseekV2MoE( |
| 1330 | config=config, |
| 1331 | parallel_config=parallel_config, |
| 1332 | quant_config=quant_config, |
| 1333 | prefix=f"{prefix}.mlp", |
| 1334 | # aiter applies routed_scaling_factor internally |
| 1335 | apply_routed_scale_to_output=not rocm_aiter_ops.is_fused_moe_enabled(), |
| 1336 | ) |
| 1337 | else: |
| 1338 | self.mlp = DeepseekV2MLP( |
| 1339 | hidden_size=config.hidden_size, |
| 1340 | intermediate_size=config.intermediate_size, |
| 1341 | hidden_act=config.hidden_act, |
| 1342 | quant_config=quant_config, |
| 1343 | prefix=f"{prefix}.mlp", |
| 1344 | ) |
| 1345 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 1346 | self.post_attention_layernorm = RMSNorm( |
| 1347 | config.hidden_size, eps=config.rms_norm_eps |
| 1348 | ) |
| 1349 | self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0) |
| 1350 | |
| 1351 | def forward( |
| 1352 | self, |
| 1353 | positions: torch.Tensor, |
| 1354 | hidden_states: torch.Tensor, |
| 1355 | residual: torch.Tensor | None, |
| 1356 | llama_4_scaling: torch.Tensor | None = None, |
| 1357 | ) -> torch.Tensor: |
| 1358 | full_num_tokens = positions.shape[0] |
| 1359 | input_is_sequence_parallel = ( |
| 1360 | self.use_sequence_parallel_moe |
| 1361 | and residual is not None |
| 1362 | and hidden_states.shape[0] != full_num_tokens |
| 1363 | ) |
| 1364 | |
| 1365 | # Self Attention |
| 1366 | if residual is None: |
| 1367 | residual = hidden_states |
| 1368 | hidden_states = self.input_layernorm(hidden_states) |
| 1369 | else: |
| 1370 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 1371 | |
| 1372 | if input_is_sequence_parallel: |
| 1373 | hidden_states = tensor_model_parallel_all_gather(hidden_states, 0) |
| 1374 | hidden_states = hidden_states[:full_num_tokens] |
| 1375 | |
| 1376 | if self.use_mha: |
| 1377 | hidden_states = self.self_attn(positions, hidden_states) |
| 1378 | else: |
| 1379 | hidden_states = self.self_attn(positions, hidden_states, llama_4_scaling) |
| 1380 | |
| 1381 | if ( |
| 1382 | not isinstance(self.self_attn, DeepseekAttention) |
| 1383 | and hidden_states.dtype == torch.float16 |
| 1384 | ): |
| 1385 | # Fix FP16 overflow |
| 1386 | # We scale both hidden_states and residual before |
| 1387 | # rmsnorm, and rmsnorm result would not affect by scale. |
| 1388 | hidden_states *= 1.0 / self.routed_scaling_factor |
| 1389 | if self.layer_idx == 0: |
| 1390 | # The residual is shared by all layers, we only scale it on |
| 1391 | # first layer. |
| 1392 | residual *= 1.0 / self.routed_scaling_factor |
| 1393 | |
| 1394 | if self.use_sequence_parallel_moe: |
| 1395 | tp_world_size = get_tensor_model_parallel_world_size() |
| 1396 | # small trick using minus, eg. -17 % 8 = 7 |
| 1397 | sp_pad = (-hidden_states.shape[0]) % tp_world_size |
| 1398 | # pad if not divisible by world size |
| 1399 | hidden_states = torch.nn.functional.pad(hidden_states, (0, 0, 0, sp_pad)) |
| 1400 | hidden_states = tensor_model_parallel_reduce_scatter(hidden_states, 0) |
| 1401 | if not input_is_sequence_parallel: |
| 1402 | residual = sequence_parallel_chunk(residual) |
| 1403 | |
| 1404 | # Fully Connected |
| 1405 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 1406 | if self.use_sequence_parallel_moe: |
| 1407 | hidden_states = self.mlp( |
| 1408 | hidden_states, |
| 1409 | already_sequence_parallel=True, |
| 1410 | ) |
| 1411 | else: |
| 1412 | hidden_states = self.mlp(hidden_states) |
| 1413 | |
| 1414 | if isinstance(self.mlp, DeepseekV2MLP) and hidden_states.dtype == torch.float16: |
| 1415 | # Fix FP16 overflow |
| 1416 | # Scaling the DeepseekV2MLP output, it is the input of |
| 1417 | # input_layernorm of next decoder layer. |
| 1418 | # The scaling of DeepseekV2MOE output would be done in the forward |
| 1419 | # of DeepseekV2MOE |
| 1420 | hidden_states *= 1.0 / self.routed_scaling_factor |
| 1421 | |
| 1422 | return hidden_states, residual |
| 1423 | |
| 1424 | |
| 1425 | @support_torch_compile |
| 1426 | class DeepseekV2Model(nn.Module): |
| 1427 | fall_back_to_pt_during_load = False |
| 1428 | |
| 1429 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1430 | super().__init__() |
| 1431 | |
| 1432 | config = vllm_config.model_config.hf_config |
| 1433 | quant_config = vllm_config.quant_config |
| 1434 | self.config = config |
| 1435 | self.device = current_platform.device_type |
| 1436 | self.hidden_size = config.hidden_size |
| 1437 | self.vocab_size = config.vocab_size |
| 1438 | self.is_v32 = hasattr(config, "index_topk") |
| 1439 | if self.is_v32: |
| 1440 | topk_tokens = config.index_topk |
| 1441 | topk_indices_buffer = torch.empty( |
| 1442 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 1443 | topk_tokens, |
| 1444 | dtype=torch.int32, |
| 1445 | device=self.device, |
| 1446 | ) |
| 1447 | else: |
| 1448 | topk_indices_buffer = None |
| 1449 | index_group_builder = ( |
| 1450 | SparseMLAIndexGroupBuilder( |
| 1451 | topk_indices_buffer, |
| 1452 | get_sparse_mla_index_group_max_rows(vllm_config), |
| 1453 | ) |
| 1454 | if topk_indices_buffer is not None |
| 1455 | else None |
| 1456 | ) |
| 1457 | |
| 1458 | if get_pp_group().is_first_rank: |
| 1459 | self.embed_tokens = VocabParallelEmbedding( |
| 1460 | config.vocab_size, |
| 1461 | self.hidden_size, |
| 1462 | quant_config=quant_config, |
| 1463 | prefix=f"{prefix}.embed_tokens", |
| 1464 | ) |
| 1465 | else: |
| 1466 | self.embed_tokens = PPMissingLayer() |
| 1467 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 1468 | config.num_hidden_layers, |
| 1469 | lambda prefix: DeepseekV2DecoderLayer( |
| 1470 | vllm_config=vllm_config, |
| 1471 | prefix=prefix, |
| 1472 | topk_indices_buffer=topk_indices_buffer, |
| 1473 | index_group_builder=index_group_builder, |
| 1474 | ), |
| 1475 | prefix=f"{prefix}.layers", |
| 1476 | ) |
| 1477 | |
| 1478 | self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible( |
| 1479 | self.layers, |
| 1480 | DeepseekV2MoE, |
| 1481 | "mlp", |
| 1482 | ) |
| 1483 | |
| 1484 | if get_pp_group().is_last_rank: |
| 1485 | self.norm = RMSNorm(self.hidden_size, eps=config.rms_norm_eps) |
| 1486 | else: |
| 1487 | self.norm = PPMissingLayer() |
| 1488 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 1489 | ["hidden_states", "residual"], self.hidden_size |
| 1490 | ) |
| 1491 | |
| 1492 | self.aux_hidden_state_layers = tuple[int, ...]() |
| 1493 | |
| 1494 | # Needed by load_weights |
| 1495 | qk_nope_head_dim = getattr(config, "qk_nope_head_dim", 0) |
| 1496 | qk_rope_head_dim = getattr(config, "qk_rope_head_dim", 0) |
| 1497 | self.use_mha = config.model_type == "deepseek" or all( |
| 1498 | dim == 0 for dim in (qk_nope_head_dim, qk_rope_head_dim) |
| 1499 | ) |
| 1500 | self.num_redundant_experts = ( |
| 1501 | vllm_config.parallel_config.eplb_config.num_redundant_experts |
| 1502 | ) |
| 1503 | |
| 1504 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1505 | return self.embed_tokens(input_ids) |
| 1506 | |
| 1507 | def forward( |
| 1508 | self, |
| 1509 | input_ids: torch.Tensor | None, |
| 1510 | positions: torch.Tensor, |
| 1511 | intermediate_tensors: IntermediateTensors | None, |
| 1512 | inputs_embeds: torch.Tensor | None = None, |
| 1513 | ) -> torch.Tensor | IntermediateTensors: |
| 1514 | if get_pp_group().is_first_rank: |
| 1515 | if inputs_embeds is not None: |
| 1516 | hidden_states = inputs_embeds |
| 1517 | else: |
| 1518 | if input_ids is None: |
| 1519 | raise ValueError( |
| 1520 | "Either input_ids or inputs_embeds must be provided " |
| 1521 | "to DeepseekV2Model.forward" |
| 1522 | ) |
| 1523 | hidden_states = self.embed_input_ids(input_ids) |
| 1524 | residual = None |
| 1525 | else: |
| 1526 | assert intermediate_tensors is not None |
| 1527 | hidden_states = intermediate_tensors["hidden_states"] |
| 1528 | residual = intermediate_tensors["residual"] |
| 1529 | |
| 1530 | # Compute llama 4 scaling once per forward pass if enabled |
| 1531 | llama_4_scaling_config = getattr(self.config, "llama_4_scaling", None) |
| 1532 | llama_4_scaling: torch.Tensor | None |
| 1533 | if llama_4_scaling_config is not None: |
| 1534 | llama_4_scaling = _get_llama_4_scaling( |
| 1535 | original_max_position_embeddings=llama_4_scaling_config[ |
| 1536 | "original_max_position_embeddings" |
| 1537 | ], |
| 1538 | scaling_beta=llama_4_scaling_config["beta"], |
| 1539 | positions=positions, |
| 1540 | ) |
| 1541 | else: |
| 1542 | llama_4_scaling = None |
| 1543 | |
| 1544 | aux_hidden_states = [] |
| 1545 | for idx, layer in enumerate( |
| 1546 | islice(self.layers, self.start_layer, self.end_layer), |
| 1547 | start=self.start_layer, |
| 1548 | ): |
| 1549 | # all gather if we need to use the whole states |
| 1550 | if ( |
| 1551 | hidden_states.shape[0] != positions.shape[0] |
| 1552 | and not layer.use_sequence_parallel_moe |
| 1553 | ): |
| 1554 | combined_states = torch.cat([hidden_states, residual], dim=-1) |
| 1555 | combined_states = tensor_model_parallel_all_gather(combined_states, 0) |
| 1556 | combined_states = combined_states[: positions.shape[0]] |
| 1557 | hidden_states, residual = combined_states.split( |
| 1558 | [self.hidden_size, self.hidden_size], dim=-1 |
| 1559 | ) |
| 1560 | if idx in self.aux_hidden_state_layers: |
| 1561 | aux_hidden_state = hidden_states + residual |
| 1562 | if aux_hidden_state.shape[0] != positions.shape[0]: |
| 1563 | aux_hidden_state = tensor_model_parallel_all_gather( |
| 1564 | aux_hidden_state, 0 |
| 1565 | ) |
| 1566 | aux_hidden_state = aux_hidden_state[: positions.shape[0]] |
| 1567 | aux_hidden_states.append(aux_hidden_state) |
| 1568 | hidden_states, residual = layer( |
| 1569 | positions, hidden_states, residual, llama_4_scaling |
| 1570 | ) |
| 1571 | |
| 1572 | if not get_pp_group().is_last_rank: |
| 1573 | return IntermediateTensors( |
| 1574 | {"hidden_states": hidden_states, "residual": residual} |
| 1575 | ) |
| 1576 | |
| 1577 | if hidden_states.shape[0] != positions.shape[0]: |
| 1578 | combined_states = torch.cat([hidden_states, residual], dim=-1) |
| 1579 | combined_states = tensor_model_parallel_all_gather(combined_states, 0) |
| 1580 | combined_states = combined_states[: positions.shape[0]] |
| 1581 | hidden_states, residual = combined_states.split( |
| 1582 | [self.hidden_size, self.hidden_size], dim=-1 |
| 1583 | ) |
| 1584 | |
| 1585 | if self.end_layer in self.aux_hidden_state_layers: |
| 1586 | aux_hidden_states.append(hidden_states + residual) |
| 1587 | |
| 1588 | hidden_states, _ = self.norm(hidden_states, residual) |
| 1589 | if len(aux_hidden_states) > 0: |
| 1590 | return hidden_states, aux_hidden_states |
| 1591 | return hidden_states |
| 1592 | |
| 1593 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1594 | stacked_params_mapping: list[tuple[str, str, int | str]] = [ |
| 1595 | # (param_name, shard_name, shard_id) |
| 1596 | ("gate_up_proj", "gate_proj", 0), |
| 1597 | ("gate_up_proj", "up_proj", 1), |
| 1598 | ] |
| 1599 | mla_params_mapping = [ |
| 1600 | ("fused_qkv_a_proj", "q_a_proj", 0), |
| 1601 | ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1), |
| 1602 | ] |
| 1603 | mha_params_mapping = [ |
| 1604 | ("qkv_proj", "q_proj", "q"), |
| 1605 | ("qkv_proj", "k_proj", "k"), |
| 1606 | ("qkv_proj", "v_proj", "v"), |
| 1607 | ] |
| 1608 | # Fused indexer wk + weights_proj (shard 0 = wk, shard 1 = weights_proj) |
| 1609 | _pending_wk_fp8 = getattr(self, "_pending_indexer_wk_fp8", None) |
| 1610 | if _pending_wk_fp8 is None: |
| 1611 | self._pending_indexer_wk_fp8 = _pending_wk_fp8 = {} |
| 1612 | |
| 1613 | indexer_fused_mapping = [ |
| 1614 | ("wk_weights_proj", "wk", 0), |
| 1615 | ("wk_weights_proj", "weights_proj", 1), |
| 1616 | ] |
| 1617 | stacked_params_mapping.extend(indexer_fused_mapping) |
| 1618 | |
| 1619 | if self.use_mha: |
| 1620 | stacked_params_mapping.extend(mha_params_mapping) |
| 1621 | else: |
| 1622 | stacked_params_mapping.extend(mla_params_mapping) |
| 1623 | |
| 1624 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 1625 | # (param_name, weight_name, expert_id, shard_id) |
| 1626 | expert_params_mapping = fused_moe_make_expert_params_mapping( |
| 1627 | self, |
| 1628 | ckpt_gate_proj_name="gate_proj", |
| 1629 | ckpt_down_proj_name="down_proj", |
| 1630 | ckpt_up_proj_name="up_proj", |
| 1631 | num_experts=self.config.n_routed_experts |
| 1632 | + ( |
| 1633 | self.config.n_shared_experts |
| 1634 | if self.is_fused_shared_expert_enabled |
| 1635 | else 0 |
| 1636 | ), |
| 1637 | num_redundant_experts=self.num_redundant_experts, |
| 1638 | ) |
| 1639 | |
| 1640 | pp_missing_layer_names = get_pp_missing_layer_names(self) |
| 1641 | params_dict = dict(self.named_parameters()) |
| 1642 | loaded_params: set[str] = set() |
| 1643 | # With index_topk_freq>1 only some layers build an indexer, yet the |
| 1644 | # checkpoint ships indexer weights for all of them; track the built ones. |
| 1645 | indexer_present_prefixes = { |
| 1646 | n.rsplit(".indexer.", 1)[0] for n in params_dict if ".indexer." in n |
| 1647 | } |
| 1648 | for name, loaded_weight in weights: |
| 1649 | if "rotary_emb.inv_freq" in name: |
| 1650 | continue |
| 1651 | |
| 1652 | spec_layer = get_spec_layer_idx_from_weight_name(self.config, name) |
| 1653 | if spec_layer is not None: |
| 1654 | continue # skip spec decode layers for main model |
| 1655 | |
| 1656 | if ".indexer." in name and ( |
| 1657 | name.rsplit(".indexer.", 1)[0] not in indexer_present_prefixes |
| 1658 | ): |
| 1659 | continue # this layer has no indexer; drop its checkpoint weights |
| 1660 | |
| 1661 | is_fusion_moe_shared_experts_layer = ( |
| 1662 | self.is_fused_shared_expert_enabled and ("mlp.shared_experts" in name) |
| 1663 | ) |
| 1664 | |
| 1665 | if _try_load_fp8_indexer_wk( |
| 1666 | name, |
| 1667 | loaded_weight, |
| 1668 | _pending_wk_fp8, |
| 1669 | params_dict, |
| 1670 | loaded_params, |
| 1671 | pp_missing_layer_names, |
| 1672 | ): |
| 1673 | continue |
| 1674 | |
| 1675 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 1676 | # Skip non-stacked layers and experts (experts handled below). |
| 1677 | if weight_name not in name: |
| 1678 | continue |
| 1679 | # We have mlp.experts[0].gate_proj in the checkpoint. |
| 1680 | # Since we handle the experts below in expert_params_mapping, |
| 1681 | # we need to skip here BEFORE we update the name, otherwise |
| 1682 | # name will be updated to mlp.experts[0].gate_up_proj, which |
| 1683 | # will then be updated below in expert_params_mapping |
| 1684 | # for mlp.experts[0].gate_gate_up_proj, which breaks load. |
| 1685 | if ("mlp.experts." in name) and name not in params_dict: |
| 1686 | continue |
| 1687 | if is_fusion_moe_shared_experts_layer: |
| 1688 | continue |
| 1689 | name_mapped = name.replace(weight_name, param_name) |
| 1690 | |
| 1691 | # QKV fusion is optional, fall back to normal |
| 1692 | # weight loading if it's not enabled |
| 1693 | # if go with fusion option, then update name |
| 1694 | if ( |
| 1695 | param_name == "fused_qkv_a_proj" |
| 1696 | ) and name_mapped not in params_dict: |
| 1697 | continue |
| 1698 | else: |
| 1699 | name = name_mapped |
| 1700 | # Skip loading extra bias for GPTQ models. |
| 1701 | if name.endswith(".bias") and name not in params_dict: |
| 1702 | continue |
| 1703 | |
| 1704 | if is_pp_missing_parameter(name, self): |
| 1705 | continue |
| 1706 | |
| 1707 | param = params_dict[name] |
| 1708 | weight_loader = param.weight_loader |
| 1709 | weight_loader(param, loaded_weight, shard_id) |
| 1710 | break |
| 1711 | else: |
| 1712 | is_expert_weight = False |
| 1713 | |
| 1714 | # Special handling: when AITER fusion_shared_experts is enabled, |
| 1715 | # checkpoints may provide a single widened shared_experts tensor |
| 1716 | # without explicit expert indices |
| 1717 | # (e.g. ...mlp.shared_experts.gate_proj.weight). |
| 1718 | # For models with multiple shared experts, split that tensor |
| 1719 | # evenly into per-shared-expert slices and load them into |
| 1720 | # appended expert slots mlp.experts.{n_routed_experts + j}.* |
| 1721 | # accordingly. |
| 1722 | num_chunks = 1 |
| 1723 | if is_fusion_moe_shared_experts_layer: |
| 1724 | num_chunks = getattr(self.config, "n_shared_experts", 1) or 1 |
| 1725 | # Determine split axis based on op type |
| 1726 | # gate/up: ColumnParallel → split along dim 0 |
| 1727 | # down: RowParallel → split along dim 1 |
| 1728 | split_dim = ( |
| 1729 | 1 |
| 1730 | if ("down_proj.weight" in name and loaded_weight.ndim > 1) |
| 1731 | else 0 |
| 1732 | ) |
| 1733 | total = loaded_weight.shape[split_dim] |
| 1734 | assert total % num_chunks == 0, ( |
| 1735 | f"Shared expert weight dim {total} " |
| 1736 | f"not divisible by num_chunks {num_chunks}" |
| 1737 | ) |
| 1738 | chunk_size = total // num_chunks |
| 1739 | |
| 1740 | for j in range(num_chunks): |
| 1741 | chunk_name = name |
| 1742 | weight_to_load = loaded_weight |
| 1743 | |
| 1744 | if is_fusion_moe_shared_experts_layer: |
| 1745 | chunk_slice = slice(j * chunk_size, (j + 1) * chunk_size) |
| 1746 | if loaded_weight.ndim == 1: |
| 1747 | weight_to_load = loaded_weight[chunk_slice] |
| 1748 | elif split_dim == 0: |
| 1749 | weight_to_load = loaded_weight[chunk_slice, :] |
| 1750 | else: |
| 1751 | weight_to_load = loaded_weight[:, chunk_slice] |
| 1752 | # Synthesize an expert-style name so expert mapping |
| 1753 | # can route it |
| 1754 | chunk_name = name.replace( |
| 1755 | "mlp.shared_experts", |
| 1756 | f"mlp.experts.{self.config.n_routed_experts + j}", |
| 1757 | ) |
| 1758 | |
| 1759 | # Use expert_params_mapping to locate the destination |
| 1760 | # param and delegate to its expert-aware weight_loader |
| 1761 | # with expert_id. |
| 1762 | for mapping in expert_params_mapping: |
| 1763 | param_name, weight_name, expert_id, expert_shard_id = mapping |
| 1764 | if weight_name not in chunk_name: |
| 1765 | continue |
| 1766 | |
| 1767 | # Anyway, this is an expert weight and should not be |
| 1768 | # attempted to load as other weights later |
| 1769 | is_expert_weight = True |
| 1770 | |
| 1771 | # Do not modify `name` since the loop may continue here |
| 1772 | # Instead, create a new variable |
| 1773 | name_mapped = chunk_name.replace(weight_name, param_name) |
| 1774 | |
| 1775 | if is_pp_missing_parameter(name_mapped, self): |
| 1776 | continue |
| 1777 | |
| 1778 | param = params_dict[name_mapped] |
| 1779 | # We should ask the weight loader to return success or |
| 1780 | # not here since otherwise we may skip experts with |
| 1781 | # other available replicas. |
| 1782 | weight_loader = typing.cast( |
| 1783 | Callable[..., bool], param.weight_loader |
| 1784 | ) |
| 1785 | success = weight_loader( |
| 1786 | param, |
| 1787 | weight_to_load, |
| 1788 | name_mapped, |
| 1789 | shard_id=expert_shard_id, |
| 1790 | expert_id=expert_id, |
| 1791 | return_success=True, |
| 1792 | ) |
| 1793 | if success: |
| 1794 | if not is_fusion_moe_shared_experts_layer: |
| 1795 | name = name_mapped |
| 1796 | else: |
| 1797 | loaded_params.add(name_mapped) |
| 1798 | break |
| 1799 | else: |
| 1800 | if is_expert_weight: |
| 1801 | # We've checked that this is an expert weight |
| 1802 | # However it's not mapped locally to this rank |
| 1803 | # So we simply skip it |
| 1804 | continue |
| 1805 | |
| 1806 | # Skip loading extra bias for GPTQ models. |
| 1807 | if name.endswith(".bias") and name not in params_dict: |
| 1808 | continue |
| 1809 | |
| 1810 | # Remapping the name of FP8 kv-scale. |
| 1811 | remapped_name = maybe_remap_kv_scale_name(name, params_dict) |
| 1812 | if remapped_name is None: |
| 1813 | continue |
| 1814 | name = remapped_name |
| 1815 | |
| 1816 | if is_pp_missing_parameter(name, self): |
| 1817 | continue |
| 1818 | |
| 1819 | param = params_dict[name] |
| 1820 | weight_loader = getattr( |
| 1821 | param, "weight_loader", default_weight_loader |
| 1822 | ) |
| 1823 | weight_loader(param, loaded_weight) |
| 1824 | if name is not None and not is_fusion_moe_shared_experts_layer: |
| 1825 | loaded_params.add(name) |
| 1826 | |
| 1827 | return loaded_params |
| 1828 | |
| 1829 | |
| 1830 | class DeepseekV2MixtureOfExperts(MixtureOfExperts): |
| 1831 | moe_mlp_layers: list[DeepseekV2MoE] |
| 1832 | """ |
| 1833 | List of MoE MLP layers in the model. |
| 1834 | """ |
| 1835 | |
| 1836 | def extract_moe_parameters(self, example_moe: DeepseekV2MoE | None): |
| 1837 | if example_moe is None: |
| 1838 | self.num_moe_layers = 0 |
| 1839 | self.num_expert_groups = 0 |
| 1840 | self.num_logical_experts = 0 |
| 1841 | self.num_physical_experts = 0 |
| 1842 | self.num_local_physical_experts = 0 |
| 1843 | self.num_routed_experts = 0 |
| 1844 | self.num_shared_experts = 0 |
| 1845 | self.num_redundant_experts = 0 |
| 1846 | logger.warning("DeepSeekV2: No DeepseekV2MoE layer found in model.layers.") |
| 1847 | else: |
| 1848 | self.num_logical_experts = example_moe.n_logical_experts |
| 1849 | self.num_physical_experts = example_moe.n_physical_experts |
| 1850 | self.num_local_physical_experts = example_moe.n_local_physical_experts |
| 1851 | self.num_routed_experts = example_moe.n_routed_experts |
| 1852 | self.num_shared_experts = example_moe.n_shared_experts |
| 1853 | self.num_redundant_experts = example_moe.n_redundant_experts |
| 1854 | |
| 1855 | def update_physical_experts_metadata( |
| 1856 | self, |
| 1857 | num_physical_experts: int, |
| 1858 | num_local_physical_experts: int, |
| 1859 | ) -> None: |
| 1860 | assert self.num_local_physical_experts == num_local_physical_experts |
| 1861 | self.num_physical_experts = num_physical_experts |
| 1862 | self.num_local_physical_experts = num_local_physical_experts |
| 1863 | self.num_redundant_experts = num_physical_experts - self.num_logical_experts |
| 1864 | for moe in self.moe_mlp_layers: |
| 1865 | moe.n_local_physical_experts = num_local_physical_experts |
| 1866 | moe.n_physical_experts = num_physical_experts |
| 1867 | moe.n_redundant_experts = self.num_redundant_experts |
| 1868 | moe.experts.update_expert_map() |
| 1869 | |
| 1870 | |
| 1871 | class DeepseekV2ForCausalLM( |
| 1872 | nn.Module, |
| 1873 | SupportsPP, |
| 1874 | DeepseekV2MixtureOfExperts, |
| 1875 | SupportsLoRA, |
| 1876 | SupportsEagle, |
| 1877 | SupportsEagle3, |
| 1878 | ): |
| 1879 | packed_modules_mapping = { |
| 1880 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 1881 | } |
| 1882 | model_cls = DeepseekV2Model |
| 1883 | |
| 1884 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1885 | super().__init__() |
| 1886 | config = vllm_config.model_config.hf_config |
| 1887 | quant_config = vllm_config.quant_config |
| 1888 | self.config = config |
| 1889 | self.quant_config = quant_config |
| 1890 | |
| 1891 | qk_nope_head_dim = getattr(config, "qk_nope_head_dim", 0) |
| 1892 | qk_rope_head_dim = getattr(config, "qk_rope_head_dim", 0) |
| 1893 | self.use_mha = config.model_type == "deepseek" or all( |
| 1894 | dim == 0 for dim in (qk_nope_head_dim, qk_rope_head_dim) |
| 1895 | ) |
| 1896 | |
| 1897 | if self.use_mha: |
| 1898 | self.packed_modules_mapping["qkv_proj"] = ["q_proj", "k_proj", "v_proj"] |
| 1899 | |
| 1900 | # `packed_modules_mapping` needs to be modified before |
| 1901 | # initializing DeepseekV2Model, as it is passed inplace to |
| 1902 | # quantization config init and may be used to select the |
| 1903 | # quant_method for relevant layers during initialization. |
| 1904 | self.fuse_qkv_a_proj = ( |
| 1905 | hasattr(config, "q_lora_rank") and config.q_lora_rank is not None |
| 1906 | ) |
| 1907 | if self.fuse_qkv_a_proj: |
| 1908 | self.packed_modules_mapping["fused_qkv_a_proj"] = [ |
| 1909 | "q_a_proj", |
| 1910 | "kv_a_proj_with_mqa", |
| 1911 | ] |
| 1912 | |
| 1913 | self.model = self.model_cls( |
| 1914 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 1915 | ) |
| 1916 | if get_pp_group().is_last_rank: |
| 1917 | self.lm_head = ParallelLMHead( |
| 1918 | config.vocab_size, |
| 1919 | config.hidden_size, |
| 1920 | quant_config=quant_config, |
| 1921 | prefix=maybe_prefix(prefix, "lm_head"), |
| 1922 | ) |
| 1923 | else: |
| 1924 | self.lm_head = PPMissingLayer() |
| 1925 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 1926 | self.make_empty_intermediate_tensors = ( |
| 1927 | self.model.make_empty_intermediate_tensors |
| 1928 | ) |
| 1929 | # Set MoE hyperparameters |
| 1930 | self.num_moe_layers = ( |
| 1931 | self.config.num_hidden_layers - self.config.first_k_dense_replace |
| 1932 | ) |
| 1933 | self.set_moe_parameters() |
| 1934 | |
| 1935 | def set_moe_parameters(self): |
| 1936 | self.num_expert_groups = getattr(self.config, "n_group", 1) |
| 1937 | |
| 1938 | self.moe_layers = [] |
| 1939 | self.moe_mlp_layers = [] |
| 1940 | example_moe = None |
| 1941 | for layer in self.model.layers: |
| 1942 | if isinstance(layer, PPMissingLayer): |
| 1943 | continue |
| 1944 | |
| 1945 | assert isinstance(layer, DeepseekV2DecoderLayer) |
| 1946 | if isinstance(layer.mlp, DeepseekV2MoE): |
| 1947 | # Pick last one layer since the first ones may be dense layers. |
| 1948 | example_moe = layer.mlp |
| 1949 | self.moe_mlp_layers.append(layer.mlp) |
| 1950 | self.moe_layers.append(layer.mlp.experts) |
| 1951 | |
| 1952 | self.extract_moe_parameters(example_moe) |
| 1953 | |
| 1954 | def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None: |
| 1955 | self.model.aux_hidden_state_layers = layers |
| 1956 | |
| 1957 | def get_eagle3_aux_hidden_state_layers(self) -> tuple[int, ...]: |
| 1958 | num_layers = len(self.model.layers) |
| 1959 | return (2, num_layers // 2, num_layers - 3) |
| 1960 | |
| 1961 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1962 | return self.model.embed_input_ids(input_ids) |
| 1963 | |
| 1964 | def forward( |
| 1965 | self, |
| 1966 | input_ids: torch.Tensor | None, |
| 1967 | positions: torch.Tensor, |
| 1968 | intermediate_tensors: IntermediateTensors | None = None, |
| 1969 | inputs_embeds: torch.Tensor | None = None, |
| 1970 | ) -> torch.Tensor | IntermediateTensors: |
| 1971 | hidden_states = self.model( |
| 1972 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 1973 | ) |
| 1974 | return hidden_states |
| 1975 | |
| 1976 | def compute_logits( |
| 1977 | self, |
| 1978 | hidden_states: torch.Tensor, |
| 1979 | ) -> torch.Tensor | None: |
| 1980 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 1981 | return logits |
| 1982 | |
| 1983 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 1984 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 1985 | # (param_name, weight_name, expert_id, shard_id) |
| 1986 | return fused_moe_make_expert_params_mapping( |
| 1987 | self, |
| 1988 | ckpt_gate_proj_name="gate_proj", |
| 1989 | ckpt_down_proj_name="down_proj", |
| 1990 | ckpt_up_proj_name="up_proj", |
| 1991 | num_experts=self.config.n_routed_experts, |
| 1992 | num_redundant_experts=0, |
| 1993 | ) |
| 1994 | |
| 1995 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1996 | loader = AutoWeightsLoader(self) |
| 1997 | return loader.load_weights(weights) |
| 1998 | |
| 1999 | |
| 2000 | class DeepseekForCausalLM(DeepseekV2ForCausalLM): |
| 2001 | pass |
| 2002 | |
| 2003 | |
| 2004 | class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM): |
| 2005 | pass |
| 2006 | |
| 2007 | |
| 2008 | class GlmMoeDsaForCausalLM(DeepseekV2ForCausalLM): |
| 2009 | pass |