| 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/modeling_qwen2.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 Qwen2 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 | from torch import nn |
| 34 | from transformers import Qwen2Config |
| 35 | |
| 36 | from vllm.compilation.decorators import support_torch_compile |
| 37 | from vllm.config import CacheConfig, VllmConfig |
| 38 | from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size |
| 39 | from vllm.model_executor.layers.activation import SiluAndMul |
| 40 | from vllm.model_executor.layers.attention import ( |
| 41 | Attention, |
| 42 | EncoderOnlyAttention, |
| 43 | ) |
| 44 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 45 | from vllm.model_executor.layers.linear import ( |
| 46 | MergedColumnParallelLinear, |
| 47 | QKVParallelLinear, |
| 48 | RowParallelLinear, |
| 49 | ) |
| 50 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 51 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 52 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 53 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 54 | ParallelLMHead, |
| 55 | VocabParallelEmbedding, |
| 56 | ) |
| 57 | from vllm.sequence import IntermediateTensors |
| 58 | from vllm.transformers_utils.config import set_default_rope_theta |
| 59 | from vllm.v1.attention.backend import AttentionType |
| 60 | |
| 61 | from .interfaces import ( |
| 62 | EagleModelMixin, |
| 63 | SupportsEagle, |
| 64 | SupportsEagle3, |
| 65 | SupportsLoRA, |
| 66 | SupportsPP, |
| 67 | SupportsQuant, |
| 68 | ) |
| 69 | from .utils import ( |
| 70 | AutoWeightsLoader, |
| 71 | PPMissingLayer, |
| 72 | WeightsMapper, |
| 73 | extract_layer_index, |
| 74 | make_empty_intermediate_tensors_factory, |
| 75 | make_layers, |
| 76 | maybe_prefix, |
| 77 | spec_decode_needs_target_embed, |
| 78 | ) |
| 79 | |
| 80 | |
| 81 | class Qwen2MLP(nn.Module): |
| 82 | def __init__( |
| 83 | self, |
| 84 | hidden_size: int, |
| 85 | intermediate_size: int, |
| 86 | hidden_act: str, |
| 87 | quant_config: QuantizationConfig | None = None, |
| 88 | prefix: str = "", |
| 89 | ) -> None: |
| 90 | super().__init__() |
| 91 | self.gate_up_proj = MergedColumnParallelLinear( |
| 92 | hidden_size, |
| 93 | [intermediate_size] * 2, |
| 94 | bias=False, |
| 95 | quant_config=quant_config, |
| 96 | prefix=f"{prefix}.gate_up_proj", |
| 97 | ) |
| 98 | self.down_proj = RowParallelLinear( |
| 99 | intermediate_size, |
| 100 | hidden_size, |
| 101 | bias=False, |
| 102 | quant_config=quant_config, |
| 103 | prefix=f"{prefix}.down_proj", |
| 104 | ) |
| 105 | if hidden_act != "silu": |
| 106 | raise ValueError( |
| 107 | f"Unsupported activation: {hidden_act}. Only silu is supported for now." |
| 108 | ) |
| 109 | self.act_fn = SiluAndMul() |
| 110 | |
| 111 | def forward(self, x): |
| 112 | gate_up, _ = self.gate_up_proj(x) |
| 113 | x = self.act_fn(gate_up) |
| 114 | x, _ = self.down_proj(x) |
| 115 | return x |
| 116 | |
| 117 | |
| 118 | class Qwen2Attention(nn.Module): |
| 119 | def __init__( |
| 120 | self, |
| 121 | hidden_size: int, |
| 122 | num_heads: int, |
| 123 | num_kv_heads: int, |
| 124 | rope_parameters: dict[str, Any], |
| 125 | max_position: int = 4096 * 32, |
| 126 | cache_config: CacheConfig | None = None, |
| 127 | quant_config: QuantizationConfig | None = None, |
| 128 | prefix: str = "", |
| 129 | attn_type: str = AttentionType.DECODER, |
| 130 | dual_chunk_attention_config: dict[str, Any] | None = None, |
| 131 | qk_norm: bool = False, |
| 132 | rms_norm_eps: float = 1e-6, |
| 133 | ) -> None: |
| 134 | super().__init__() |
| 135 | self.hidden_size = hidden_size |
| 136 | tp_size = get_tensor_model_parallel_world_size() |
| 137 | self.total_num_heads = num_heads |
| 138 | assert self.total_num_heads % tp_size == 0 |
| 139 | self.num_heads = self.total_num_heads // tp_size |
| 140 | self.total_num_kv_heads = num_kv_heads |
| 141 | if self.total_num_kv_heads >= tp_size: |
| 142 | # Number of KV heads is greater than TP size, so we partition |
| 143 | # the KV heads across multiple tensor parallel GPUs. |
| 144 | assert self.total_num_kv_heads % tp_size == 0 |
| 145 | else: |
| 146 | # Number of KV heads is less than TP size, so we replicate |
| 147 | # the KV heads across multiple tensor parallel GPUs. |
| 148 | assert tp_size % self.total_num_kv_heads == 0 |
| 149 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 150 | self.head_dim = hidden_size // self.total_num_heads |
| 151 | self.q_size = self.num_heads * self.head_dim |
| 152 | self.kv_size = self.num_kv_heads * self.head_dim |
| 153 | self.scaling = self.head_dim**-0.5 |
| 154 | self.dual_chunk_attention_config = dual_chunk_attention_config |
| 155 | self.qk_norm = qk_norm |
| 156 | |
| 157 | self.qkv_proj = QKVParallelLinear( |
| 158 | hidden_size, |
| 159 | self.head_dim, |
| 160 | self.total_num_heads, |
| 161 | self.total_num_kv_heads, |
| 162 | bias=True, |
| 163 | quant_config=quant_config, |
| 164 | prefix=f"{prefix}.qkv_proj", |
| 165 | ) |
| 166 | self.o_proj = RowParallelLinear( |
| 167 | self.total_num_heads * self.head_dim, |
| 168 | hidden_size, |
| 169 | bias=False, |
| 170 | quant_config=quant_config, |
| 171 | prefix=f"{prefix}.o_proj", |
| 172 | ) |
| 173 | |
| 174 | # QK Normalization support (used in BAGEL and some other models) |
| 175 | if self.qk_norm: |
| 176 | self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps) |
| 177 | self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps) |
| 178 | |
| 179 | self.rotary_emb = get_rope( |
| 180 | self.head_dim, |
| 181 | max_position=max_position, |
| 182 | rope_parameters=rope_parameters, |
| 183 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 184 | ) |
| 185 | attn_cls = ( |
| 186 | EncoderOnlyAttention |
| 187 | if attn_type == AttentionType.ENCODER_ONLY |
| 188 | else Attention |
| 189 | ) |
| 190 | attention_kwargs: dict[str, Any] = {} |
| 191 | if dual_chunk_attention_config: |
| 192 | attention_kwargs = { |
| 193 | "layer_idx": extract_layer_index(prefix), |
| 194 | "dual_chunk_attention_config": dual_chunk_attention_config, |
| 195 | } |
| 196 | self.attn = attn_cls( |
| 197 | self.num_heads, |
| 198 | self.head_dim, |
| 199 | self.scaling, |
| 200 | num_kv_heads=self.num_kv_heads, |
| 201 | cache_config=cache_config, |
| 202 | quant_config=quant_config, |
| 203 | attn_type=attn_type, |
| 204 | prefix=f"{prefix}.attn", |
| 205 | **attention_kwargs, |
| 206 | ) |
| 207 | |
| 208 | def forward( |
| 209 | self, |
| 210 | positions: torch.Tensor, |
| 211 | hidden_states: torch.Tensor, |
| 212 | ) -> torch.Tensor: |
| 213 | qkv, _ = self.qkv_proj(hidden_states) |
| 214 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 215 | |
| 216 | # Apply QK normalization if enabled (before RoPE) |
| 217 | if self.qk_norm: |
| 218 | # Reshape to apply per-head normalization |
| 219 | # q shape: (total_tokens, q_size) -> (total_tokens, num_heads, head_dim) |
| 220 | total_tokens = q.shape[0] |
| 221 | q = q.view(total_tokens, self.num_heads, self.head_dim) |
| 222 | k = k.view(total_tokens, self.num_kv_heads, self.head_dim) |
| 223 | |
| 224 | # Apply normalization |
| 225 | q = self.q_norm(q) |
| 226 | k = self.k_norm(k) |
| 227 | |
| 228 | # Reshape back |
| 229 | q = q.view(total_tokens, self.q_size) |
| 230 | k = k.view(total_tokens, self.kv_size) |
| 231 | |
| 232 | q, k = self.rotary_emb(positions, q, k) |
| 233 | attn_output = self.attn(q, k, v) |
| 234 | output, _ = self.o_proj(attn_output) |
| 235 | return output |
| 236 | |
| 237 | |
| 238 | class Qwen2DecoderLayer(nn.Module): |
| 239 | def __init__( |
| 240 | self, |
| 241 | config: Qwen2Config, |
| 242 | cache_config: CacheConfig | None = None, |
| 243 | quant_config: QuantizationConfig | None = None, |
| 244 | prefix: str = "", |
| 245 | ) -> None: |
| 246 | super().__init__() |
| 247 | self.hidden_size = config.hidden_size |
| 248 | set_default_rope_theta(config, default_theta=1000000) |
| 249 | dual_chunk_attention_config = getattr( |
| 250 | config, "dual_chunk_attention_config", None |
| 251 | ) |
| 252 | |
| 253 | # By default, Qwen2 uses causal attention as it is a decoder-only model. |
| 254 | # You can override the HF config with `is_causal=False` to enable |
| 255 | # bidirectional attention, which is used in some embedding models |
| 256 | # (e.g. Alibaba-NLP/gte-Qwen2-7B-instruct) |
| 257 | if getattr(config, "is_causal", True): |
| 258 | attn_type = AttentionType.DECODER |
| 259 | else: |
| 260 | attn_type = AttentionType.ENCODER_ONLY |
| 261 | |
| 262 | # Check if QK normalization is enabled (used in BAGEL and some other models) |
| 263 | qk_norm = getattr(config, "qk_norm", False) |
| 264 | |
| 265 | self.self_attn = Qwen2Attention( |
| 266 | hidden_size=self.hidden_size, |
| 267 | num_heads=config.num_attention_heads, |
| 268 | max_position=config.max_position_embeddings, |
| 269 | num_kv_heads=config.num_key_value_heads, |
| 270 | cache_config=cache_config, |
| 271 | quant_config=quant_config, |
| 272 | rope_parameters=config.rope_parameters, |
| 273 | prefix=f"{prefix}.self_attn", |
| 274 | attn_type=attn_type, |
| 275 | dual_chunk_attention_config=dual_chunk_attention_config, |
| 276 | qk_norm=qk_norm, |
| 277 | rms_norm_eps=config.rms_norm_eps, |
| 278 | ) |
| 279 | self.mlp = Qwen2MLP( |
| 280 | hidden_size=self.hidden_size, |
| 281 | intermediate_size=config.intermediate_size, |
| 282 | hidden_act=config.hidden_act, |
| 283 | quant_config=quant_config, |
| 284 | prefix=f"{prefix}.mlp", |
| 285 | ) |
| 286 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 287 | self.post_attention_layernorm = RMSNorm( |
| 288 | config.hidden_size, eps=config.rms_norm_eps |
| 289 | ) |
| 290 | |
| 291 | def forward( |
| 292 | self, |
| 293 | positions: torch.Tensor, |
| 294 | hidden_states: torch.Tensor, |
| 295 | residual: torch.Tensor | None, |
| 296 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 297 | # Self Attention |
| 298 | if residual is None: |
| 299 | residual = hidden_states |
| 300 | hidden_states = self.input_layernorm(hidden_states) |
| 301 | else: |
| 302 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 303 | hidden_states = self.self_attn( |
| 304 | positions=positions, |
| 305 | hidden_states=hidden_states, |
| 306 | ) |
| 307 | |
| 308 | # Fully Connected |
| 309 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 310 | hidden_states = self.mlp(hidden_states) |
| 311 | return hidden_states, residual |
| 312 | |
| 313 | |
| 314 | @support_torch_compile( |
| 315 | dynamic_arg_dims={ |
| 316 | "input_ids": {0: "b"}, |
| 317 | # positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl, |
| 318 | # otherwise (seq_len, ). |
| 319 | "positions": {-1: "b"}, |
| 320 | "intermediate_tensors": {0: "b"}, |
| 321 | "inputs_embeds": {0: "b"}, |
| 322 | } |
| 323 | ) |
| 324 | class Qwen2Model(nn.Module, EagleModelMixin): |
| 325 | hf_to_vllm_mapper = WeightsMapper( |
| 326 | orig_to_new_stacked={ |
| 327 | # weight_name: (param_name, shard_id) |
| 328 | ".q_proj": (".qkv_proj", "q"), |
| 329 | ".k_proj": (".qkv_proj", "k"), |
| 330 | ".v_proj": (".qkv_proj", "v"), |
| 331 | ".gate_proj": (".gate_up_proj", 0), |
| 332 | ".up_proj": (".gate_up_proj", 1), |
| 333 | } |
| 334 | ) |
| 335 | supports_aux_hidden_states_over_pp = True |
| 336 | |
| 337 | def __init__( |
| 338 | self, |
| 339 | *, |
| 340 | vllm_config: VllmConfig, |
| 341 | prefix: str = "", |
| 342 | decoder_layer_type: type[nn.Module] = Qwen2DecoderLayer, |
| 343 | ): |
| 344 | super().__init__() |
| 345 | |
| 346 | config = vllm_config.model_config.hf_config.get_text_config() |
| 347 | cache_config = vllm_config.cache_config |
| 348 | quant_config = vllm_config.quant_config |
| 349 | |
| 350 | # TODO (@robertgshaw2): see if this can be moved out |
| 351 | if len(set(getattr(config, "layer_types", []))) > 1: |
| 352 | assert config.max_window_layers == config.num_hidden_layers, ( |
| 353 | "Sliding window for some but all layers is not supported. " |
| 354 | "This model uses sliding window but `max_window_layers` = {} " |
| 355 | "is less than `num_hidden_layers` = {}. Please open an issue " |
| 356 | "to discuss this feature.".format( |
| 357 | config.max_window_layers, |
| 358 | config.num_hidden_layers, |
| 359 | ) |
| 360 | ) |
| 361 | |
| 362 | self.config = config |
| 363 | self.quant_config = quant_config |
| 364 | self.vocab_size = config.vocab_size |
| 365 | |
| 366 | if ( |
| 367 | get_pp_group().is_first_rank |
| 368 | or (config.tie_word_embeddings and get_pp_group().is_last_rank) |
| 369 | or spec_decode_needs_target_embed(vllm_config) |
| 370 | ): |
| 371 | self.embed_tokens = VocabParallelEmbedding( |
| 372 | config.vocab_size, |
| 373 | config.hidden_size, |
| 374 | quant_config=quant_config, |
| 375 | prefix=f"{prefix}.embed_tokens", |
| 376 | ) |
| 377 | else: |
| 378 | self.embed_tokens = PPMissingLayer() |
| 379 | |
| 380 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 381 | config.num_hidden_layers, |
| 382 | lambda prefix: decoder_layer_type( |
| 383 | config=config, |
| 384 | cache_config=cache_config, |
| 385 | quant_config=quant_config, |
| 386 | prefix=prefix, |
| 387 | ), |
| 388 | prefix=f"{prefix}.layers", |
| 389 | ) |
| 390 | |
| 391 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 392 | ["hidden_states", "residual"], config.hidden_size |
| 393 | ) |
| 394 | if get_pp_group().is_last_rank: |
| 395 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 396 | else: |
| 397 | self.norm = PPMissingLayer() |
| 398 | |
| 399 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 400 | return self.embed_tokens(input_ids) |
| 401 | |
| 402 | def forward( |
| 403 | self, |
| 404 | input_ids: torch.Tensor | None, |
| 405 | positions: torch.Tensor, |
| 406 | intermediate_tensors: IntermediateTensors | None = None, |
| 407 | inputs_embeds: torch.Tensor | None = None, |
| 408 | ) -> torch.Tensor | IntermediateTensors: |
| 409 | if get_pp_group().is_first_rank: |
| 410 | if inputs_embeds is not None: |
| 411 | hidden_states = inputs_embeds |
| 412 | else: |
| 413 | hidden_states = self.embed_input_ids(input_ids) |
| 414 | residual = None |
| 415 | else: |
| 416 | assert intermediate_tensors is not None |
| 417 | hidden_states = intermediate_tensors["hidden_states"] |
| 418 | residual = intermediate_tensors["residual"] |
| 419 | |
| 420 | remote_aux = self.collect_remote_aux_hidden_states(intermediate_tensors) |
| 421 | |
| 422 | aux_hidden_states: list[torch.Tensor] = [] |
| 423 | if get_pp_group().is_first_rank: |
| 424 | self._maybe_add_hidden_state( |
| 425 | aux_hidden_states, self.start_layer, hidden_states, residual |
| 426 | ) |
| 427 | for idx, layer in enumerate( |
| 428 | islice(self.layers, self.start_layer, self.end_layer), |
| 429 | start=self.start_layer, |
| 430 | ): |
| 431 | hidden_states, residual = layer(positions, hidden_states, residual) |
| 432 | self._maybe_add_hidden_state( |
| 433 | aux_hidden_states, idx + 1, hidden_states, residual |
| 434 | ) |
| 435 | |
| 436 | if not get_pp_group().is_last_rank: |
| 437 | return IntermediateTensors( |
| 438 | { |
| 439 | "hidden_states": hidden_states, |
| 440 | "residual": residual, |
| 441 | **self.pack_local_aux_hidden_states(aux_hidden_states), |
| 442 | } |
| 443 | ) |
| 444 | |
| 445 | hidden_states, _ = self.norm(hidden_states, residual) |
| 446 | |
| 447 | aux_hidden_states = remote_aux + aux_hidden_states |
| 448 | if len(aux_hidden_states) > 0: |
| 449 | return hidden_states, aux_hidden_states |
| 450 | |
| 451 | return hidden_states |
| 452 | |
| 453 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 454 | loader = AutoWeightsLoader(self) |
| 455 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 456 | |
| 457 | |
| 458 | class Qwen2ForCausalLM( |
| 459 | nn.Module, SupportsLoRA, SupportsPP, SupportsEagle, SupportsEagle3, SupportsQuant |
| 460 | ): |
| 461 | hf_to_vllm_mapper = Qwen2Model.hf_to_vllm_mapper |
| 462 | packed_modules_mapping = { |
| 463 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 464 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 465 | } |
| 466 | |
| 467 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 468 | super().__init__() |
| 469 | config = vllm_config.model_config.hf_config.get_text_config() |
| 470 | quant_config = vllm_config.quant_config |
| 471 | |
| 472 | self.config = config |
| 473 | |
| 474 | self.quant_config = quant_config |
| 475 | self.model = Qwen2Model( |
| 476 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 477 | ) |
| 478 | |
| 479 | if get_pp_group().is_last_rank: |
| 480 | self.lm_head = ParallelLMHead( |
| 481 | config.vocab_size, |
| 482 | config.hidden_size, |
| 483 | quant_config=quant_config, |
| 484 | prefix=maybe_prefix(prefix, "lm_head"), |
| 485 | ) |
| 486 | if config.tie_word_embeddings: |
| 487 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 488 | else: |
| 489 | self.lm_head = PPMissingLayer() |
| 490 | |
| 491 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 492 | |
| 493 | self.make_empty_intermediate_tensors = ( |
| 494 | self.model.make_empty_intermediate_tensors |
| 495 | ) |
| 496 | |
| 497 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 498 | return self.model.embed_input_ids(input_ids) |
| 499 | |
| 500 | def forward( |
| 501 | self, |
| 502 | input_ids: torch.Tensor | None, |
| 503 | positions: torch.Tensor, |
| 504 | intermediate_tensors: IntermediateTensors | None = None, |
| 505 | inputs_embeds: torch.Tensor | None = None, |
| 506 | ) -> torch.Tensor | IntermediateTensors: |
| 507 | hidden_states = self.model( |
| 508 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 509 | ) |
| 510 | return hidden_states |
| 511 | |
| 512 | def compute_logits( |
| 513 | self, |
| 514 | hidden_states: torch.Tensor, |
| 515 | ) -> torch.Tensor | None: |
| 516 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 517 | return logits |
| 518 | |
| 519 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 520 | loader = AutoWeightsLoader(self) |
| 521 | return loader.load_weights(weights) |