| 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/opt/modeling_opt.py |
| 6 | # Copyright 2023 The vLLM team. |
| 7 | # Copyright 2022 The Fairseq Authors and The HuggingFace Inc. team. All rights |
| 8 | # reserved. |
| 9 | # |
| 10 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 11 | # you may not use this file except in compliance with the License. |
| 12 | # You may obtain a copy of the License at |
| 13 | # |
| 14 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 15 | # |
| 16 | # Unless required by applicable law or agreed to in writing, software |
| 17 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 18 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 19 | # See the License for the specific language governing permissions and |
| 20 | # limitations under the License. |
| 21 | """Inference-only OPT model compatible with HuggingFace weights.""" |
| 22 | |
| 23 | from collections.abc import Iterable |
| 24 | from itertools import islice |
| 25 | |
| 26 | import torch |
| 27 | from torch import nn |
| 28 | from transformers import OPTConfig |
| 29 | |
| 30 | from vllm.compilation.decorators import support_torch_compile |
| 31 | from vllm.config import CacheConfig, VllmConfig |
| 32 | from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size |
| 33 | from vllm.model_executor.layers.activation import get_act_fn |
| 34 | from vllm.model_executor.layers.attention import Attention |
| 35 | from vllm.model_executor.layers.layernorm import LayerNorm |
| 36 | from vllm.model_executor.layers.linear import ( |
| 37 | ColumnParallelLinear, |
| 38 | QKVParallelLinear, |
| 39 | ReplicatedLinear, |
| 40 | RowParallelLinear, |
| 41 | ) |
| 42 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 43 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 44 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 45 | ParallelLMHead, |
| 46 | VocabParallelEmbedding, |
| 47 | ) |
| 48 | from vllm.sequence import IntermediateTensors |
| 49 | |
| 50 | from .interfaces import SupportsLoRA, SupportsPP |
| 51 | from .utils import ( |
| 52 | AutoWeightsLoader, |
| 53 | WeightsMapper, |
| 54 | make_empty_intermediate_tensors_factory, |
| 55 | make_layers, |
| 56 | maybe_prefix, |
| 57 | ) |
| 58 | |
| 59 | |
| 60 | class OPTLearnedPositionalEmbedding(nn.Embedding): |
| 61 | def __init__(self, num_embeddings: int, embedding_dim: int): |
| 62 | # OPT is set up so that if padding_idx is specified then offset the |
| 63 | # embedding ids by 2 and adjust num_embeddings appropriately. Other |
| 64 | # models don't have this hack |
| 65 | self.offset = 2 |
| 66 | super().__init__(num_embeddings + self.offset, embedding_dim) |
| 67 | |
| 68 | def forward(self, positions: torch.Tensor): |
| 69 | return super().forward(positions + self.offset) |
| 70 | |
| 71 | |
| 72 | class OPTAttention(nn.Module): |
| 73 | def __init__( |
| 74 | self, |
| 75 | embed_dim: int, |
| 76 | num_heads: int, |
| 77 | bias: bool = True, |
| 78 | cache_config: CacheConfig | None = None, |
| 79 | quant_config: QuantizationConfig | None = None, |
| 80 | prefix: str = "", |
| 81 | ) -> None: |
| 82 | super().__init__() |
| 83 | self.embed_dim = embed_dim |
| 84 | tensor_model_parallel_world_size = get_tensor_model_parallel_world_size() |
| 85 | total_num_heads = num_heads |
| 86 | assert num_heads % tensor_model_parallel_world_size == 0 |
| 87 | self.num_heads = total_num_heads // tensor_model_parallel_world_size |
| 88 | self.head_dim = embed_dim // total_num_heads |
| 89 | self.scaling = self.head_dim**-0.5 |
| 90 | |
| 91 | self.qkv_proj = QKVParallelLinear( |
| 92 | embed_dim, |
| 93 | self.head_dim, |
| 94 | total_num_heads, |
| 95 | bias=bias, |
| 96 | quant_config=quant_config, |
| 97 | prefix=f"{prefix}.qkv_proj", |
| 98 | ) |
| 99 | self.out_proj = RowParallelLinear( |
| 100 | embed_dim, |
| 101 | embed_dim, |
| 102 | bias=bias, |
| 103 | quant_config=quant_config, |
| 104 | prefix=f"{prefix}.out_proj", |
| 105 | ) |
| 106 | self.attn = Attention( |
| 107 | self.num_heads, |
| 108 | self.head_dim, |
| 109 | scale=self.scaling, |
| 110 | cache_config=cache_config, |
| 111 | quant_config=quant_config, |
| 112 | prefix=f"{prefix}.attn", |
| 113 | ) |
| 114 | |
| 115 | def forward( |
| 116 | self, |
| 117 | hidden_states: torch.Tensor, |
| 118 | ) -> torch.Tensor: |
| 119 | qkv, _ = self.qkv_proj(hidden_states) |
| 120 | q, k, v = qkv.chunk(chunks=3, dim=-1) |
| 121 | attn_output = self.attn(q, k, v) |
| 122 | output, _ = self.out_proj(attn_output) |
| 123 | return output |
| 124 | |
| 125 | |
| 126 | class OPTDecoderLayer(nn.Module): |
| 127 | def __init__( |
| 128 | self, |
| 129 | config: OPTConfig, |
| 130 | cache_config: CacheConfig | None = None, |
| 131 | quant_config: QuantizationConfig | None = None, |
| 132 | prefix: str = "", |
| 133 | ): |
| 134 | super().__init__() |
| 135 | self.config = config |
| 136 | self.embed_dim = config.hidden_size |
| 137 | self.self_attn = OPTAttention( |
| 138 | embed_dim=self.embed_dim, |
| 139 | num_heads=config.num_attention_heads, |
| 140 | bias=config.enable_bias, |
| 141 | cache_config=cache_config, |
| 142 | quant_config=quant_config, |
| 143 | prefix=f"{prefix}.self_attn", |
| 144 | ) |
| 145 | self.do_layer_norm_before = config.do_layer_norm_before |
| 146 | |
| 147 | self.self_attn_layer_norm = LayerNorm( |
| 148 | self.embed_dim, elementwise_affine=config.layer_norm_elementwise_affine |
| 149 | ) |
| 150 | self.fc1 = ColumnParallelLinear( |
| 151 | self.embed_dim, |
| 152 | config.ffn_dim, |
| 153 | bias=config.enable_bias, |
| 154 | quant_config=quant_config, |
| 155 | prefix=f"{prefix}.fc1", |
| 156 | ) |
| 157 | self.activation_fn = get_act_fn(config.activation_function) |
| 158 | self.fc2 = RowParallelLinear( |
| 159 | config.ffn_dim, |
| 160 | self.embed_dim, |
| 161 | bias=config.enable_bias, |
| 162 | quant_config=quant_config, |
| 163 | prefix=f"{prefix}.fc2", |
| 164 | ) |
| 165 | self.final_layer_norm = LayerNorm( |
| 166 | self.embed_dim, elementwise_affine=config.layer_norm_elementwise_affine |
| 167 | ) |
| 168 | |
| 169 | def forward( |
| 170 | self, |
| 171 | hidden_states: torch.Tensor, |
| 172 | ) -> torch.Tensor: |
| 173 | # Self Attention |
| 174 | residual = hidden_states |
| 175 | # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention |
| 176 | if self.do_layer_norm_before: |
| 177 | hidden_states = self.self_attn_layer_norm(hidden_states) |
| 178 | hidden_states = self.self_attn(hidden_states=hidden_states) |
| 179 | hidden_states = residual + hidden_states |
| 180 | # 350m applies layer norm AFTER attention |
| 181 | if not self.do_layer_norm_before: |
| 182 | hidden_states = self.self_attn_layer_norm(hidden_states) |
| 183 | |
| 184 | # Fully Connected |
| 185 | residual = hidden_states |
| 186 | # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention |
| 187 | if self.do_layer_norm_before: |
| 188 | hidden_states = self.final_layer_norm(hidden_states) |
| 189 | hidden_states, _ = self.fc1(hidden_states) |
| 190 | hidden_states = self.activation_fn(hidden_states) |
| 191 | hidden_states, _ = self.fc2(hidden_states) |
| 192 | hidden_states = residual + hidden_states |
| 193 | # 350m applies layer norm AFTER attention |
| 194 | if not self.do_layer_norm_before: |
| 195 | hidden_states = self.final_layer_norm(hidden_states) |
| 196 | return hidden_states |
| 197 | |
| 198 | |
| 199 | class OPTDecoder(nn.Module): |
| 200 | def __init__( |
| 201 | self, |
| 202 | config: OPTConfig, |
| 203 | cache_config: CacheConfig | None = None, |
| 204 | quant_config: QuantizationConfig | None = None, |
| 205 | prefix: str = "", |
| 206 | ): |
| 207 | super().__init__() |
| 208 | self.config = config |
| 209 | self.max_target_positions = config.max_position_embeddings |
| 210 | self.vocab_size = config.vocab_size |
| 211 | |
| 212 | self.embed_tokens = VocabParallelEmbedding( |
| 213 | config.vocab_size, |
| 214 | config.word_embed_proj_dim, |
| 215 | ) |
| 216 | # Positional embeddings are replicated (not sharded). |
| 217 | self.embed_positions = OPTLearnedPositionalEmbedding( |
| 218 | config.max_position_embeddings, config.hidden_size |
| 219 | ) |
| 220 | |
| 221 | # Project out & in will be replicated if they exist. |
| 222 | self.project_out: ReplicatedLinear | None |
| 223 | if config.word_embed_proj_dim != config.hidden_size: |
| 224 | self.project_out = ReplicatedLinear( |
| 225 | config.hidden_size, |
| 226 | config.word_embed_proj_dim, |
| 227 | bias=False, |
| 228 | quant_config=quant_config, |
| 229 | prefix=f"{prefix}.project_out", |
| 230 | ) |
| 231 | else: |
| 232 | self.project_out = None |
| 233 | |
| 234 | self.project_in: ReplicatedLinear | None |
| 235 | if config.word_embed_proj_dim != config.hidden_size: |
| 236 | self.project_in = ReplicatedLinear( |
| 237 | config.word_embed_proj_dim, |
| 238 | config.hidden_size, |
| 239 | bias=False, |
| 240 | quant_config=quant_config, |
| 241 | prefix=f"{prefix}.project_in", |
| 242 | ) |
| 243 | else: |
| 244 | self.project_in = None |
| 245 | |
| 246 | # Note that the only purpose of `config._remove_final_layer_norm` is to |
| 247 | # keep backward compatibility with checkpoints that have been fine-tuned |
| 248 | # before transformers v4.20.1 |
| 249 | # see https://github.com/facebookresearch/metaseq/pull/164 |
| 250 | self.final_layer_norm: LayerNorm | None |
| 251 | if config.do_layer_norm_before and not config._remove_final_layer_norm: |
| 252 | self.final_layer_norm = LayerNorm( |
| 253 | config.hidden_size, |
| 254 | elementwise_affine=config.layer_norm_elementwise_affine, |
| 255 | ) |
| 256 | else: |
| 257 | self.final_layer_norm = None |
| 258 | |
| 259 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 260 | config.num_hidden_layers, |
| 261 | lambda prefix: OPTDecoderLayer( |
| 262 | config, cache_config, quant_config, prefix=prefix |
| 263 | ), |
| 264 | prefix=f"{prefix}.layers", |
| 265 | ) |
| 266 | |
| 267 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 268 | return self.embed_tokens(input_ids) |
| 269 | |
| 270 | def forward( |
| 271 | self, |
| 272 | input_ids: torch.Tensor | None, |
| 273 | positions: torch.Tensor, |
| 274 | intermediate_tensors: IntermediateTensors | None, |
| 275 | inputs_embeds: torch.Tensor | None = None, |
| 276 | ) -> torch.Tensor | IntermediateTensors: |
| 277 | if get_pp_group().is_first_rank: |
| 278 | if inputs_embeds is None: |
| 279 | inputs_embeds = self.embed_input_ids(input_ids) |
| 280 | pos_embeds = self.embed_positions(positions) |
| 281 | if self.project_in is not None: |
| 282 | inputs_embeds, _ = self.project_in(inputs_embeds) |
| 283 | hidden_states = inputs_embeds + pos_embeds |
| 284 | else: |
| 285 | assert intermediate_tensors is not None |
| 286 | hidden_states = intermediate_tensors["hidden_states"] |
| 287 | |
| 288 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 289 | hidden_states = layer(hidden_states) |
| 290 | |
| 291 | if not get_pp_group().is_last_rank: |
| 292 | return IntermediateTensors({"hidden_states": hidden_states}) |
| 293 | if self.final_layer_norm is not None: |
| 294 | hidden_states = self.final_layer_norm(hidden_states) |
| 295 | if self.project_out is not None: |
| 296 | hidden_states, _ = self.project_out(hidden_states) |
| 297 | return hidden_states |
| 298 | |
| 299 | |
| 300 | @support_torch_compile |
| 301 | class OPTModel(nn.Module): |
| 302 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 303 | super().__init__() |
| 304 | |
| 305 | config = vllm_config.model_config.hf_config |
| 306 | cache_config = vllm_config.cache_config |
| 307 | quant_config = vllm_config.quant_config |
| 308 | |
| 309 | self.decoder = OPTDecoder( |
| 310 | config, cache_config, quant_config, prefix=f"{prefix}.decoder" |
| 311 | ) |
| 312 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 313 | ["hidden_states"], config.hidden_size |
| 314 | ) |
| 315 | |
| 316 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 317 | return self.decoder.embed_input_ids(input_ids) |
| 318 | |
| 319 | def forward( |
| 320 | self, |
| 321 | input_ids: torch.Tensor | None, |
| 322 | positions: torch.Tensor, |
| 323 | intermediate_tensors: IntermediateTensors | None, |
| 324 | inputs_embeds: torch.Tensor | None = None, |
| 325 | ) -> torch.Tensor | IntermediateTensors: |
| 326 | return self.decoder( |
| 327 | input_ids, positions, intermediate_tensors, inputs_embeds=inputs_embeds |
| 328 | ) |
| 329 | |
| 330 | |
| 331 | class OPTForCausalLM(nn.Module, SupportsPP, SupportsLoRA): |
| 332 | hf_to_vllm_mapper = WeightsMapper( |
| 333 | orig_to_new_stacked={ |
| 334 | # weight_name: (param_name, shard_id) |
| 335 | ".q_proj": (".qkv_proj", "q"), |
| 336 | ".k_proj": (".qkv_proj", "k"), |
| 337 | ".v_proj": (".qkv_proj", "v"), |
| 338 | }, |
| 339 | orig_to_new_prefix={ |
| 340 | "decoder.": "model.decoder.", |
| 341 | }, |
| 342 | ) |
| 343 | packed_modules_mapping = { |
| 344 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 345 | } |
| 346 | |
| 347 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 348 | super().__init__() |
| 349 | config = vllm_config.model_config.hf_config |
| 350 | quant_config = vllm_config.quant_config |
| 351 | self.config = config |
| 352 | self.quant_config = quant_config |
| 353 | self.model = OPTModel( |
| 354 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 355 | ) |
| 356 | self.lm_head = ParallelLMHead( |
| 357 | config.vocab_size, |
| 358 | config.word_embed_proj_dim, |
| 359 | prefix=maybe_prefix(prefix, "lm_head"), |
| 360 | ) |
| 361 | if self.config.tie_word_embeddings: |
| 362 | self.lm_head = self.lm_head.tie_weights(self.model.decoder.embed_tokens) |
| 363 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 364 | self.make_empty_intermediate_tensors = ( |
| 365 | self.model.make_empty_intermediate_tensors |
| 366 | ) |
| 367 | |
| 368 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 369 | return self.model.embed_input_ids(input_ids) |
| 370 | |
| 371 | def forward( |
| 372 | self, |
| 373 | input_ids: torch.Tensor | None, |
| 374 | positions: torch.Tensor, |
| 375 | intermediate_tensors: IntermediateTensors | None = None, |
| 376 | inputs_embeds: torch.Tensor | None = None, |
| 377 | ) -> torch.Tensor | IntermediateTensors: |
| 378 | hidden_states = self.model( |
| 379 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 380 | ) |
| 381 | return hidden_states |
| 382 | |
| 383 | def compute_logits( |
| 384 | self, |
| 385 | hidden_states: torch.Tensor, |
| 386 | ) -> torch.Tensor | None: |
| 387 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 388 | return logits |
| 389 | |
| 390 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 391 | loader = AutoWeightsLoader(self) |
| 392 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |