| 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/qwen2_moe/modeling_qwen2_moe.py |
| 6 | # Copyright 2024 The Qwen team. |
| 7 | # Copyright 2023 The vLLM team. |
| 8 | # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. |
| 9 | # |
| 10 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 11 | # and OPT implementations in this library. It has been modified from its |
| 12 | # original forms to accommodate minor architectural differences compared |
| 13 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 14 | # |
| 15 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 16 | # you may not use this file except in compliance with the License. |
| 17 | # You may obtain a copy of the License at |
| 18 | # |
| 19 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 20 | # |
| 21 | # Unless required by applicable law or agreed to in writing, software |
| 22 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 23 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 24 | # See the License for the specific language governing permissions and |
| 25 | # limitations under the License. |
| 26 | """Inference-only Qwen2MoE model compatible with HuggingFace weights.""" |
| 27 | |
| 28 | from collections.abc import Iterable |
| 29 | from itertools import islice |
| 30 | from typing import Any |
| 31 | |
| 32 | import torch |
| 33 | import torch.nn.functional as F |
| 34 | from torch import nn |
| 35 | from transformers import Qwen2MoeConfig |
| 36 | |
| 37 | from vllm.compilation.decorators import support_torch_compile |
| 38 | from vllm.config import CacheConfig, VllmConfig |
| 39 | from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size |
| 40 | from vllm.logger import init_logger |
| 41 | from vllm.model_executor.layers.activation import SiluAndMul |
| 42 | from vllm.model_executor.layers.attention import Attention |
| 43 | from vllm.model_executor.layers.fused_moe import ( |
| 44 | FusedMoEFactory, |
| 45 | GateLinear, |
| 46 | ) |
| 47 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 48 | from vllm.model_executor.layers.linear import ( |
| 49 | MergedColumnParallelLinear, |
| 50 | QKVParallelLinear, |
| 51 | ReplicatedLinear, |
| 52 | RowParallelLinear, |
| 53 | ) |
| 54 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 55 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 56 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 57 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 58 | ParallelLMHead, |
| 59 | VocabParallelEmbedding, |
| 60 | ) |
| 61 | from vllm.sequence import IntermediateTensors |
| 62 | |
| 63 | from .interfaces import SupportsLoRA, SupportsPP |
| 64 | from .utils import ( |
| 65 | AutoWeightsLoader, |
| 66 | WeightsMapper, |
| 67 | extract_layer_index, |
| 68 | make_empty_intermediate_tensors_factory, |
| 69 | make_layers, |
| 70 | maybe_prefix, |
| 71 | ) |
| 72 | |
| 73 | logger = init_logger(__name__) |
| 74 | |
| 75 | |
| 76 | class Qwen2MoeMLP(nn.Module): |
| 77 | def __init__( |
| 78 | self, |
| 79 | hidden_size: int, |
| 80 | intermediate_size: int, |
| 81 | hidden_act: str, |
| 82 | quant_config: QuantizationConfig | None = None, |
| 83 | reduce_results: bool = True, |
| 84 | expert_gate: torch.nn.Linear | None = None, |
| 85 | is_sequence_parallel: bool = False, |
| 86 | disable_tp: bool = False, |
| 87 | prefix: str = "", |
| 88 | ) -> None: |
| 89 | super().__init__() |
| 90 | disable_tp = disable_tp or is_sequence_parallel |
| 91 | self.gate_up_proj = MergedColumnParallelLinear( |
| 92 | hidden_size, |
| 93 | [intermediate_size] * 2, |
| 94 | bias=False, |
| 95 | quant_config=quant_config, |
| 96 | disable_tp=disable_tp, |
| 97 | prefix=f"{prefix}.gate_up_proj", |
| 98 | ) |
| 99 | self.down_proj = RowParallelLinear( |
| 100 | intermediate_size, |
| 101 | hidden_size, |
| 102 | bias=False, |
| 103 | quant_config=quant_config, |
| 104 | reduce_results=reduce_results, |
| 105 | disable_tp=disable_tp, |
| 106 | prefix=f"{prefix}.down_proj", |
| 107 | ) |
| 108 | if hidden_act != "silu": |
| 109 | raise ValueError( |
| 110 | f"Unsupported activation: {hidden_act}. Only silu is supported for now." |
| 111 | ) |
| 112 | self.act_fn = SiluAndMul() |
| 113 | self.expert_gate = expert_gate |
| 114 | |
| 115 | def forward(self, x): |
| 116 | gate_up, _ = self.gate_up_proj(x) |
| 117 | out = self.act_fn(gate_up) |
| 118 | out, _ = self.down_proj(out) |
| 119 | |
| 120 | if self.expert_gate is not None: |
| 121 | out = F.sigmoid(self.expert_gate(x)[0]) * out |
| 122 | |
| 123 | return out |
| 124 | |
| 125 | |
| 126 | class Qwen2MoeSparseMoeBlock(nn.Module): |
| 127 | def __init__( |
| 128 | self, |
| 129 | config: Qwen2MoeConfig, |
| 130 | quant_config: QuantizationConfig | None = None, |
| 131 | prefix: str = "", |
| 132 | ): |
| 133 | super().__init__() |
| 134 | self.tp_size = get_tensor_model_parallel_world_size() |
| 135 | |
| 136 | if self.tp_size > config.num_experts: |
| 137 | raise ValueError( |
| 138 | f"Tensor parallel size {self.tp_size} is greater than " |
| 139 | f"the number of experts {config.num_experts}." |
| 140 | ) |
| 141 | |
| 142 | self.gate = GateLinear( |
| 143 | config.hidden_size, |
| 144 | config.num_experts, |
| 145 | prefix=f"{prefix}.gate", |
| 146 | ) |
| 147 | |
| 148 | self.shared_expert_gate = ReplicatedLinear( |
| 149 | config.hidden_size, |
| 150 | 1, |
| 151 | bias=False, |
| 152 | quant_config=None, |
| 153 | prefix=f"{prefix}.shared_expert_gate", |
| 154 | ) |
| 155 | |
| 156 | self.shared_expert: Qwen2MoeMLP | None |
| 157 | if config.shared_expert_intermediate_size > 0: |
| 158 | self.shared_expert = Qwen2MoeMLP( |
| 159 | hidden_size=config.hidden_size, |
| 160 | intermediate_size=config.shared_expert_intermediate_size, |
| 161 | hidden_act=config.hidden_act, |
| 162 | quant_config=quant_config, |
| 163 | reduce_results=False, |
| 164 | expert_gate=self.shared_expert_gate, |
| 165 | prefix=f"{prefix}.shared_expert", |
| 166 | ) |
| 167 | else: |
| 168 | self.shared_expert = None |
| 169 | |
| 170 | self.experts = FusedMoEFactory( |
| 171 | shared_experts=self.shared_expert, |
| 172 | num_experts=config.num_experts, |
| 173 | top_k=config.num_experts_per_tok, |
| 174 | hidden_size=config.hidden_size, |
| 175 | intermediate_size=config.moe_intermediate_size, |
| 176 | renormalize=config.norm_topk_prob, |
| 177 | quant_config=quant_config, |
| 178 | prefix=f"{prefix}.experts", |
| 179 | ) |
| 180 | |
| 181 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 182 | # NOTE: hidden_states can have either 1D or 2D shape. |
| 183 | orig_shape = hidden_states.shape |
| 184 | hidden_dim = hidden_states.shape[-1] |
| 185 | hidden_states = hidden_states.view(-1, hidden_dim) |
| 186 | |
| 187 | # router_logits: (num_tokens, n_experts) |
| 188 | router_logits, _ = self.gate(hidden_states) |
| 189 | final_hidden_states = self.experts( |
| 190 | hidden_states=hidden_states, router_logits=router_logits |
| 191 | ) |
| 192 | |
| 193 | return final_hidden_states.view(orig_shape) |
| 194 | |
| 195 | |
| 196 | class Qwen2MoeAttention(nn.Module): |
| 197 | def __init__( |
| 198 | self, |
| 199 | hidden_size: int, |
| 200 | num_heads: int, |
| 201 | num_kv_heads: int, |
| 202 | rope_parameters: dict[str, Any] | None = None, |
| 203 | max_position_embeddings: int = 8192, |
| 204 | cache_config: CacheConfig | None = None, |
| 205 | quant_config: QuantizationConfig | None = None, |
| 206 | prefix: str = "", |
| 207 | dual_chunk_attention_config: dict[str, Any] | None = None, |
| 208 | ) -> None: |
| 209 | super().__init__() |
| 210 | self.hidden_size = hidden_size |
| 211 | tp_size = get_tensor_model_parallel_world_size() |
| 212 | self.total_num_heads = num_heads |
| 213 | assert self.total_num_heads % tp_size == 0 |
| 214 | self.num_heads = self.total_num_heads // tp_size |
| 215 | self.total_num_kv_heads = num_kv_heads |
| 216 | if self.total_num_kv_heads >= tp_size: |
| 217 | # Number of KV heads is greater than TP size, so we partition |
| 218 | # the KV heads across multiple tensor parallel GPUs. |
| 219 | assert self.total_num_kv_heads % tp_size == 0 |
| 220 | else: |
| 221 | # Number of KV heads is less than TP size, so we replicate |
| 222 | # the KV heads across multiple tensor parallel GPUs. |
| 223 | assert tp_size % self.total_num_kv_heads == 0 |
| 224 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 225 | self.head_dim = hidden_size // self.total_num_heads |
| 226 | self.q_size = self.num_heads * self.head_dim |
| 227 | self.kv_size = self.num_kv_heads * self.head_dim |
| 228 | self.scaling = self.head_dim**-0.5 |
| 229 | self.max_position_embeddings = max_position_embeddings |
| 230 | self.dual_chunk_attention_config = dual_chunk_attention_config |
| 231 | |
| 232 | self.qkv_proj = QKVParallelLinear( |
| 233 | hidden_size, |
| 234 | self.head_dim, |
| 235 | self.total_num_heads, |
| 236 | self.total_num_kv_heads, |
| 237 | bias=True, |
| 238 | quant_config=quant_config, |
| 239 | prefix=f"{prefix}.qkv_proj", |
| 240 | ) |
| 241 | |
| 242 | self.o_proj = RowParallelLinear( |
| 243 | self.total_num_heads * self.head_dim, |
| 244 | hidden_size, |
| 245 | bias=False, |
| 246 | quant_config=quant_config, |
| 247 | prefix=f"{prefix}.o_proj", |
| 248 | ) |
| 249 | |
| 250 | self.rotary_emb = get_rope( |
| 251 | self.head_dim, |
| 252 | max_position=max_position_embeddings, |
| 253 | rope_parameters=rope_parameters, |
| 254 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 255 | ) |
| 256 | attention_kwargs: dict[str, Any] = {} |
| 257 | if dual_chunk_attention_config: |
| 258 | attention_kwargs = { |
| 259 | "layer_idx": extract_layer_index(prefix), |
| 260 | "dual_chunk_attention_config": dual_chunk_attention_config, |
| 261 | } |
| 262 | self.attn = Attention( |
| 263 | self.num_heads, |
| 264 | self.head_dim, |
| 265 | self.scaling, |
| 266 | num_kv_heads=self.num_kv_heads, |
| 267 | cache_config=cache_config, |
| 268 | quant_config=quant_config, |
| 269 | prefix=f"{prefix}.attn", |
| 270 | **attention_kwargs, |
| 271 | ) |
| 272 | |
| 273 | def forward( |
| 274 | self, |
| 275 | positions: torch.Tensor, |
| 276 | hidden_states: torch.Tensor, |
| 277 | ) -> torch.Tensor: |
| 278 | qkv, _ = self.qkv_proj(hidden_states) |
| 279 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 280 | q, k = self.rotary_emb(positions, q, k) |
| 281 | attn_output = self.attn(q, k, v) |
| 282 | output, _ = self.o_proj(attn_output) |
| 283 | return output |
| 284 | |
| 285 | |
| 286 | class Qwen2MoeDecoderLayer(nn.Module): |
| 287 | def __init__( |
| 288 | self, |
| 289 | config: Qwen2MoeConfig, |
| 290 | cache_config: CacheConfig | None = None, |
| 291 | quant_config: QuantizationConfig | None = None, |
| 292 | prefix: str = "", |
| 293 | ) -> None: |
| 294 | super().__init__() |
| 295 | self.hidden_size = config.hidden_size |
| 296 | dual_chunk_attention_config = getattr( |
| 297 | config, "dual_chunk_attention_config", None |
| 298 | ) |
| 299 | max_position_embeddings = getattr(config, "max_position_embeddings", 8192) |
| 300 | self.self_attn = Qwen2MoeAttention( |
| 301 | hidden_size=self.hidden_size, |
| 302 | num_heads=config.num_attention_heads, |
| 303 | num_kv_heads=config.num_key_value_heads, |
| 304 | rope_parameters=config.rope_parameters, |
| 305 | max_position_embeddings=max_position_embeddings, |
| 306 | cache_config=cache_config, |
| 307 | quant_config=quant_config, |
| 308 | prefix=f"{prefix}.self_attn", |
| 309 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 310 | ) |
| 311 | |
| 312 | # Note: Qwen/Qwen2-57B-A14B-Instruct does not have |
| 313 | # `mlp_only_layers` in the config. |
| 314 | layer_idx = extract_layer_index(prefix) |
| 315 | mlp_only_layers = ( |
| 316 | [] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers |
| 317 | ) |
| 318 | if (layer_idx not in mlp_only_layers) and ( |
| 319 | config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0 |
| 320 | ): |
| 321 | self.mlp = Qwen2MoeSparseMoeBlock( |
| 322 | config=config, quant_config=quant_config, prefix=f"{prefix}.mlp" |
| 323 | ) |
| 324 | else: |
| 325 | self.mlp = Qwen2MoeMLP( |
| 326 | hidden_size=config.hidden_size, |
| 327 | intermediate_size=config.intermediate_size, |
| 328 | hidden_act=config.hidden_act, |
| 329 | quant_config=quant_config, |
| 330 | prefix=f"{prefix}.mlp", |
| 331 | ) |
| 332 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 333 | self.post_attention_layernorm = RMSNorm( |
| 334 | config.hidden_size, eps=config.rms_norm_eps |
| 335 | ) |
| 336 | |
| 337 | def forward( |
| 338 | self, |
| 339 | positions: torch.Tensor, |
| 340 | hidden_states: torch.Tensor, |
| 341 | residual: torch.Tensor | None, |
| 342 | ) -> torch.Tensor: |
| 343 | # Self Attention |
| 344 | if residual is None: |
| 345 | residual = hidden_states |
| 346 | hidden_states = self.input_layernorm(hidden_states) |
| 347 | else: |
| 348 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 349 | hidden_states = self.self_attn( |
| 350 | positions=positions, |
| 351 | hidden_states=hidden_states, |
| 352 | ) |
| 353 | |
| 354 | # Fully Connected |
| 355 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 356 | hidden_states = self.mlp(hidden_states) |
| 357 | return hidden_states, residual |
| 358 | |
| 359 | |
| 360 | @support_torch_compile |
| 361 | class Qwen2MoeModel(nn.Module): |
| 362 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 363 | super().__init__() |
| 364 | |
| 365 | config = vllm_config.model_config.hf_config |
| 366 | cache_config = vllm_config.cache_config |
| 367 | quant_config = vllm_config.quant_config |
| 368 | |
| 369 | self.vocab_size = config.vocab_size |
| 370 | self.config = config |
| 371 | |
| 372 | self.embed_tokens = VocabParallelEmbedding( |
| 373 | config.vocab_size, |
| 374 | config.hidden_size, |
| 375 | quant_config=quant_config, |
| 376 | prefix=f"{prefix}.embed_tokens", |
| 377 | ) |
| 378 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 379 | config.num_hidden_layers, |
| 380 | lambda prefix: Qwen2MoeDecoderLayer( |
| 381 | config=config, |
| 382 | cache_config=cache_config, |
| 383 | quant_config=quant_config, |
| 384 | prefix=prefix, |
| 385 | ), |
| 386 | prefix=f"{prefix}.layers", |
| 387 | ) |
| 388 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 389 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 390 | ["hidden_states", "residual"], config.hidden_size |
| 391 | ) |
| 392 | |
| 393 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 394 | return self.embed_tokens(input_ids) |
| 395 | |
| 396 | def forward( |
| 397 | self, |
| 398 | input_ids: torch.Tensor | None, |
| 399 | positions: torch.Tensor, |
| 400 | intermediate_tensors: IntermediateTensors | None = None, |
| 401 | inputs_embeds: torch.Tensor | None = None, |
| 402 | ) -> torch.Tensor | IntermediateTensors: |
| 403 | if get_pp_group().is_first_rank: |
| 404 | if inputs_embeds is not None: |
| 405 | hidden_states = inputs_embeds |
| 406 | else: |
| 407 | hidden_states = self.embed_input_ids(input_ids) |
| 408 | residual = None |
| 409 | else: |
| 410 | assert intermediate_tensors is not None |
| 411 | hidden_states = intermediate_tensors["hidden_states"] |
| 412 | residual = intermediate_tensors["residual"] |
| 413 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 414 | hidden_states, residual = layer(positions, hidden_states, residual) |
| 415 | if not get_pp_group().is_last_rank: |
| 416 | return IntermediateTensors( |
| 417 | {"hidden_states": hidden_states, "residual": residual} |
| 418 | ) |
| 419 | hidden_states, _ = self.norm(hidden_states, residual) |
| 420 | return hidden_states |
| 421 | |
| 422 | |
| 423 | class Qwen2MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA): |
| 424 | fall_back_to_pt_during_load = False |
| 425 | hf_to_vllm_mapper = WeightsMapper( |
| 426 | orig_to_new_stacked={ |
| 427 | # weight_name: (param_name, shard_id) |
| 428 | ".q_proj": (".qkv_proj", "q"), |
| 429 | ".k_proj": (".qkv_proj", "k"), |
| 430 | ".v_proj": (".qkv_proj", "v"), |
| 431 | # .experts.gate_up_proj must be handled by MoERunner.load_weights for EP |
| 432 | ".mlp.gate_proj": (".mlp.gate_up_proj", 0), |
| 433 | ".mlp.up_proj": (".mlp.gate_up_proj", 1), |
| 434 | ".shared_expert.gate_proj": (".shared_expert.gate_up_proj", 0), |
| 435 | ".shared_expert.up_proj": (".shared_expert.gate_up_proj", 1), |
| 436 | } |
| 437 | ) |
| 438 | packed_modules_mapping = { |
| 439 | "qkv_proj": [ |
| 440 | "q_proj", |
| 441 | "k_proj", |
| 442 | "v_proj", |
| 443 | ] |
| 444 | } |
| 445 | |
| 446 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 447 | super().__init__() |
| 448 | config = vllm_config.model_config.hf_config |
| 449 | quant_config = vllm_config.quant_config |
| 450 | self.config = config |
| 451 | self.quant_config = quant_config |
| 452 | # Only perform the following mapping when Qwen2MoeMLP exists |
| 453 | if ( |
| 454 | getattr(config, "mlp_only_layers", []) |
| 455 | or config.shared_expert_intermediate_size > 0 |
| 456 | ): |
| 457 | self.packed_modules_mapping["gate_up_proj"] = ["gate_proj", "up_proj"] |
| 458 | |
| 459 | self.model = Qwen2MoeModel( |
| 460 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 461 | ) |
| 462 | self.lm_head = ParallelLMHead( |
| 463 | config.vocab_size, |
| 464 | config.hidden_size, |
| 465 | quant_config=quant_config, |
| 466 | prefix=maybe_prefix(prefix, "lm_head"), |
| 467 | ) |
| 468 | if self.config.tie_word_embeddings: |
| 469 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 470 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 471 | self.make_empty_intermediate_tensors = ( |
| 472 | self.model.make_empty_intermediate_tensors |
| 473 | ) |
| 474 | |
| 475 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 476 | return self.model.embed_input_ids(input_ids) |
| 477 | |
| 478 | def forward( |
| 479 | self, |
| 480 | input_ids: torch.Tensor | None, |
| 481 | positions: torch.Tensor, |
| 482 | intermediate_tensors: IntermediateTensors | None = None, |
| 483 | inputs_embeds: torch.Tensor | None = None, |
| 484 | ) -> torch.Tensor | IntermediateTensors: |
| 485 | hidden_states = self.model( |
| 486 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 487 | ) |
| 488 | return hidden_states |
| 489 | |
| 490 | def compute_logits( |
| 491 | self, |
| 492 | hidden_states: torch.Tensor, |
| 493 | ) -> torch.Tensor | None: |
| 494 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 495 | return logits |
| 496 | |
| 497 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 498 | def _maybe_reshape( |
| 499 | weights: Iterable[tuple[str, torch.Tensor]], |
| 500 | ) -> Iterable[tuple[str, torch.Tensor]]: |
| 501 | for name, loaded_weight in weights: |
| 502 | # GGUF: make sure that shared_expert_gate is a 2D tensor. |
| 503 | if "mlp.shared_expert_gate" in name and loaded_weight.dim() == 1: |
| 504 | loaded_weight = loaded_weight[None, :] |
| 505 | yield name, loaded_weight |
| 506 | |
| 507 | loader = AutoWeightsLoader(self) |
| 508 | return loader.load_weights( |
| 509 | _maybe_reshape(weights), mapper=self.hf_to_vllm_mapper |
| 510 | ) |