| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Iterable |
| 5 | from functools import partial |
| 6 | from itertools import islice |
| 7 | from typing import Any |
| 8 | |
| 9 | import torch |
| 10 | from torch import nn |
| 11 | from transformers import PreTrainedConfig |
| 12 | |
| 13 | from vllm.compilation.decorators import support_torch_compile |
| 14 | from vllm.config import CacheConfig, VllmConfig |
| 15 | from vllm.distributed import ( |
| 16 | get_pp_group, |
| 17 | get_tensor_model_parallel_rank, |
| 18 | get_tensor_model_parallel_world_size, |
| 19 | split_tensor_along_last_dim, |
| 20 | tensor_model_parallel_all_gather, |
| 21 | ) |
| 22 | from vllm.model_executor.layers.activation import SiluAndMul |
| 23 | from vllm.model_executor.layers.attention import Attention |
| 24 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 25 | from vllm.model_executor.layers.linear import ( |
| 26 | MergedColumnParallelLinear, |
| 27 | QKVParallelLinear, |
| 28 | RowParallelLinear, |
| 29 | ) |
| 30 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 31 | from vllm.model_executor.layers.pooler.tokwise import pooler_for_token_classify |
| 32 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 33 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 34 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 35 | ParallelLMHead, |
| 36 | VocabParallelEmbedding, |
| 37 | ) |
| 38 | from vllm.sequence import IntermediateTensors |
| 39 | |
| 40 | from .interfaces import SupportsLoRA, SupportsPP, SupportsQuant |
| 41 | from .interfaces_base import default_pooling_type |
| 42 | from .utils import ( |
| 43 | AutoWeightsLoader, |
| 44 | StageMissingLayer, |
| 45 | WeightsMapper, |
| 46 | make_empty_intermediate_tensors_factory, |
| 47 | make_layers, |
| 48 | maybe_prefix, |
| 49 | no_init_weights, |
| 50 | ) |
| 51 | |
| 52 | |
| 53 | class InternLM2MLP(nn.Module): |
| 54 | def __init__( |
| 55 | self, |
| 56 | hidden_size: int, |
| 57 | intermediate_size: int, |
| 58 | hidden_act: str, |
| 59 | quant_config: QuantizationConfig | None = None, |
| 60 | prefix: str = "", |
| 61 | ) -> None: |
| 62 | super().__init__() |
| 63 | self.gate_up_proj = MergedColumnParallelLinear( |
| 64 | hidden_size, |
| 65 | [intermediate_size] * 2, |
| 66 | bias=False, |
| 67 | quant_config=quant_config, |
| 68 | prefix=f"{prefix}.gate_up_proj", |
| 69 | ) |
| 70 | self.w2 = RowParallelLinear( |
| 71 | intermediate_size, |
| 72 | hidden_size, |
| 73 | bias=False, |
| 74 | quant_config=quant_config, |
| 75 | prefix=f"{prefix}.w2", |
| 76 | ) |
| 77 | if hidden_act != "silu": |
| 78 | raise ValueError( |
| 79 | f"Unsupported activation: {hidden_act}. Only silu is supported for now." |
| 80 | ) |
| 81 | self.act_fn = SiluAndMul() |
| 82 | |
| 83 | def forward(self, x): |
| 84 | gate_up, _ = self.gate_up_proj(x) |
| 85 | x = self.act_fn(gate_up) |
| 86 | x, _ = self.w2(x) |
| 87 | return x |
| 88 | |
| 89 | |
| 90 | class InternLM2Attention(nn.Module): |
| 91 | def __init__( |
| 92 | self, |
| 93 | hidden_size: int, |
| 94 | num_heads: int, |
| 95 | num_kv_heads: int, |
| 96 | rope_parameters: dict[str, Any] | None = None, |
| 97 | max_position_embeddings: int = 8192, |
| 98 | cache_config: CacheConfig | None = None, |
| 99 | quant_config: QuantizationConfig | None = None, |
| 100 | prefix: str = "", |
| 101 | ) -> None: |
| 102 | super().__init__() |
| 103 | self.hidden_size = hidden_size |
| 104 | self.tp_size = get_tensor_model_parallel_world_size() |
| 105 | self.tp_rank = get_tensor_model_parallel_rank() |
| 106 | self.total_num_heads = num_heads |
| 107 | assert self.total_num_heads % self.tp_size == 0 |
| 108 | self.num_heads = self.total_num_heads // self.tp_size |
| 109 | self.total_num_kv_heads = num_kv_heads |
| 110 | if self.total_num_kv_heads >= self.tp_size: |
| 111 | # Number of KV heads is greater than TP size, so we partition |
| 112 | # the KV heads across multiple tensor parallel GPUs. |
| 113 | assert self.total_num_kv_heads % self.tp_size == 0 |
| 114 | else: |
| 115 | # Number of KV heads is less than TP size, so we replicate |
| 116 | # the KV heads across multiple tensor parallel GPUs. |
| 117 | assert self.tp_size % self.total_num_kv_heads == 0 |
| 118 | self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size) |
| 119 | self.head_dim = hidden_size // self.total_num_heads |
| 120 | self.q_size = self.num_heads * self.head_dim |
| 121 | self.kv_size = self.num_kv_heads * self.head_dim |
| 122 | self.key_value_groups = int(self.num_heads / self.num_kv_heads) |
| 123 | self.scaling = self.head_dim**-0.5 |
| 124 | self.max_position_embeddings = max_position_embeddings |
| 125 | |
| 126 | self.wqkv = QKVParallelLinear( |
| 127 | hidden_size, |
| 128 | self.head_dim, |
| 129 | self.total_num_heads, |
| 130 | self.total_num_kv_heads, |
| 131 | bias=False, |
| 132 | quant_config=quant_config, |
| 133 | prefix=f"{prefix}.wqkv", |
| 134 | ) |
| 135 | self.wo = RowParallelLinear( |
| 136 | self.total_num_heads * self.head_dim, |
| 137 | hidden_size, |
| 138 | bias=False, |
| 139 | quant_config=quant_config, |
| 140 | prefix=f"{prefix}.wo", |
| 141 | ) |
| 142 | |
| 143 | self.rotary_emb = get_rope( |
| 144 | self.head_dim, |
| 145 | max_position=max_position_embeddings, |
| 146 | rope_parameters=rope_parameters, |
| 147 | ) |
| 148 | self.attn = Attention( |
| 149 | self.num_heads, |
| 150 | self.head_dim, |
| 151 | self.scaling, |
| 152 | num_kv_heads=self.num_kv_heads, |
| 153 | cache_config=cache_config, |
| 154 | quant_config=quant_config, |
| 155 | prefix=f"{prefix}.attn", |
| 156 | ) |
| 157 | |
| 158 | def split_qkv(self, qkv: torch.Tensor): |
| 159 | seq_len = qkv.shape[0] |
| 160 | if self.tp_size > 1: |
| 161 | qkv_map = [self.q_size, self.kv_size, self.kv_size] * self.tp_size |
| 162 | qkv = tensor_model_parallel_all_gather(qkv) |
| 163 | qkv = torch.split(qkv, qkv_map, dim=-1) |
| 164 | qkv = qkv[::3] + qkv[1::3] + qkv[2::3] |
| 165 | qkv = torch.cat(qkv, dim=-1) |
| 166 | |
| 167 | qkv = qkv.view( |
| 168 | seq_len, self.total_num_kv_heads, self.key_value_groups + 2, self.head_dim |
| 169 | ) |
| 170 | q, k, v = torch.split(qkv, [self.key_value_groups, 1, 1], dim=-2) |
| 171 | q = q.reshape(seq_len, self.q_size * self.tp_size) |
| 172 | k = k.reshape(seq_len, self.kv_size * self.tp_size) |
| 173 | v = v.reshape(seq_len, self.kv_size * self.tp_size) |
| 174 | |
| 175 | if self.tp_size > 1: |
| 176 | splitter = partial(split_tensor_along_last_dim, num_partitions=self.tp_size) |
| 177 | q = splitter(q)[self.tp_rank] |
| 178 | k = splitter(k)[self.tp_rank] |
| 179 | v = splitter(v)[self.tp_rank] |
| 180 | return q, k, v |
| 181 | |
| 182 | def forward( |
| 183 | self, |
| 184 | positions: torch.Tensor, |
| 185 | hidden_states: torch.Tensor, |
| 186 | ) -> torch.Tensor: |
| 187 | qkv, _ = self.wqkv(hidden_states) |
| 188 | q, k, v = self.split_qkv(qkv) |
| 189 | q, k = self.rotary_emb(positions, q, k) |
| 190 | attn_output = self.attn(q, k, v) |
| 191 | output, _ = self.wo(attn_output) |
| 192 | return output |
| 193 | |
| 194 | |
| 195 | class InternLMDecoderLayer(nn.Module): |
| 196 | def __init__( |
| 197 | self, |
| 198 | config: PreTrainedConfig, |
| 199 | cache_config: CacheConfig | None = None, |
| 200 | quant_config: QuantizationConfig | None = None, |
| 201 | prefix: str = "", |
| 202 | ) -> None: |
| 203 | super().__init__() |
| 204 | self.hidden_size = config.hidden_size |
| 205 | max_position_embeddings = getattr(config, "max_position_embeddings", 8192) |
| 206 | self.attention = InternLM2Attention( |
| 207 | hidden_size=self.hidden_size, |
| 208 | num_heads=config.num_attention_heads, |
| 209 | num_kv_heads=config.num_key_value_heads, |
| 210 | rope_parameters=config.rope_parameters, |
| 211 | max_position_embeddings=max_position_embeddings, |
| 212 | cache_config=cache_config, |
| 213 | quant_config=quant_config, |
| 214 | prefix=f"{prefix}.attention", |
| 215 | ) |
| 216 | self.feed_forward = InternLM2MLP( |
| 217 | hidden_size=self.hidden_size, |
| 218 | intermediate_size=config.intermediate_size, |
| 219 | hidden_act=config.hidden_act, |
| 220 | quant_config=quant_config, |
| 221 | prefix=f"{prefix}.feed_forward", |
| 222 | ) |
| 223 | self.attention_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 224 | self.ffn_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 225 | |
| 226 | def forward( |
| 227 | self, |
| 228 | positions: torch.Tensor, |
| 229 | hidden_states: torch.Tensor, |
| 230 | residual: torch.Tensor | None, |
| 231 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 232 | # Self Attention |
| 233 | if residual is None: |
| 234 | residual = hidden_states |
| 235 | hidden_states = self.attention_norm(hidden_states) |
| 236 | else: |
| 237 | hidden_states, residual = self.attention_norm(hidden_states, residual) |
| 238 | hidden_states = self.attention( |
| 239 | positions=positions, |
| 240 | hidden_states=hidden_states, |
| 241 | ) |
| 242 | |
| 243 | # Fully Connected |
| 244 | hidden_states, residual = self.ffn_norm(hidden_states, residual) |
| 245 | hidden_states = self.feed_forward(hidden_states) |
| 246 | return hidden_states, residual |
| 247 | |
| 248 | |
| 249 | @support_torch_compile |
| 250 | class InternLM2Model(nn.Module): |
| 251 | hf_to_vllm_mapper = WeightsMapper( |
| 252 | orig_to_new_stacked={ |
| 253 | # weight_name: (param_name, shard_id) |
| 254 | ".w1": (".gate_up_proj", 0), |
| 255 | ".w3": (".gate_up_proj", 1), |
| 256 | } |
| 257 | ) |
| 258 | |
| 259 | def __init__( |
| 260 | self, |
| 261 | *, |
| 262 | vllm_config: VllmConfig, |
| 263 | prefix: str = "", |
| 264 | layer_type: type[InternLMDecoderLayer] = InternLMDecoderLayer, |
| 265 | ): |
| 266 | super().__init__() |
| 267 | |
| 268 | config = vllm_config.model_config.hf_config |
| 269 | cache_config = vllm_config.cache_config |
| 270 | quant_config = vllm_config.quant_config |
| 271 | |
| 272 | self.config = config |
| 273 | self.vocab_size = config.vocab_size |
| 274 | self.tok_embeddings = VocabParallelEmbedding( |
| 275 | config.vocab_size, |
| 276 | config.hidden_size, |
| 277 | ) |
| 278 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 279 | config.num_hidden_layers, |
| 280 | lambda prefix: layer_type( |
| 281 | config, cache_config, quant_config, prefix=prefix |
| 282 | ), |
| 283 | prefix=f"{prefix}.layers", |
| 284 | ) |
| 285 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 286 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 287 | ["hidden_states", "residual"], config.hidden_size |
| 288 | ) |
| 289 | |
| 290 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 291 | return self.tok_embeddings(input_ids) |
| 292 | |
| 293 | def forward( |
| 294 | self, |
| 295 | input_ids: torch.Tensor | None, |
| 296 | positions: torch.Tensor, |
| 297 | intermediate_tensors: IntermediateTensors | None = None, |
| 298 | inputs_embeds: torch.Tensor | None = None, |
| 299 | ) -> torch.Tensor | IntermediateTensors: |
| 300 | if get_pp_group().is_first_rank: |
| 301 | if inputs_embeds is not None: |
| 302 | hidden_states = inputs_embeds |
| 303 | else: |
| 304 | hidden_states = self.embed_input_ids(input_ids) |
| 305 | residual = None |
| 306 | else: |
| 307 | assert intermediate_tensors is not None |
| 308 | hidden_states = intermediate_tensors["hidden_states"] |
| 309 | residual = intermediate_tensors["residual"] |
| 310 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 311 | hidden_states, residual = layer(positions, hidden_states, residual) |
| 312 | if not get_pp_group().is_last_rank: |
| 313 | return IntermediateTensors( |
| 314 | {"hidden_states": hidden_states, "residual": residual} |
| 315 | ) |
| 316 | hidden_states, _ = self.norm(hidden_states, residual) |
| 317 | return hidden_states |
| 318 | |
| 319 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 320 | loader = AutoWeightsLoader(self) |
| 321 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 322 | |
| 323 | |
| 324 | class InternLM2ForCausalLM(nn.Module, SupportsPP, SupportsLoRA, SupportsQuant): |
| 325 | hf_to_vllm_mapper = InternLM2Model.hf_to_vllm_mapper |
| 326 | packed_modules_mapping: dict[str, list[str]] = { |
| 327 | "wqkv": ["wqkv"], |
| 328 | "gate_up_proj": ["w1", "w3"], |
| 329 | } |
| 330 | |
| 331 | def __init__( |
| 332 | self, |
| 333 | *, |
| 334 | vllm_config: VllmConfig, |
| 335 | prefix: str = "", |
| 336 | model_type: type[InternLM2Model] = InternLM2Model, |
| 337 | ): |
| 338 | super().__init__() |
| 339 | config = vllm_config.model_config.hf_config |
| 340 | quant_config = vllm_config.quant_config |
| 341 | |
| 342 | self.config = config |
| 343 | self.quant_config = quant_config |
| 344 | |
| 345 | self.model = model_type( |
| 346 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 347 | ) |
| 348 | self.output = ParallelLMHead( |
| 349 | config.vocab_size, |
| 350 | config.hidden_size, |
| 351 | quant_config=quant_config, |
| 352 | prefix=maybe_prefix(prefix, "output"), |
| 353 | ) |
| 354 | if self.config.tie_word_embeddings: |
| 355 | self.output = self.output.tie_weights(self.model.tok_embeddings) |
| 356 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 357 | self.make_empty_intermediate_tensors = ( |
| 358 | self.model.make_empty_intermediate_tensors |
| 359 | ) |
| 360 | |
| 361 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 362 | return self.model.embed_input_ids(input_ids) |
| 363 | |
| 364 | def forward( |
| 365 | self, |
| 366 | input_ids: torch.Tensor | None, |
| 367 | positions: torch.Tensor, |
| 368 | intermediate_tensors: IntermediateTensors | None, |
| 369 | inputs_embeds: torch.Tensor | None = None, |
| 370 | ) -> torch.Tensor: |
| 371 | hidden_states = self.model( |
| 372 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 373 | ) |
| 374 | return hidden_states |
| 375 | |
| 376 | def compute_logits( |
| 377 | self, |
| 378 | hidden_states: torch.Tensor, |
| 379 | ) -> torch.Tensor | None: |
| 380 | logits = self.logits_processor(self.output, hidden_states) |
| 381 | return logits |
| 382 | |
| 383 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 384 | loader = AutoWeightsLoader(self) |
| 385 | return loader.load_weights(weights) |
| 386 | |
| 387 | |
| 388 | @default_pooling_type(tok_pooling_type="ALL") |
| 389 | class InternLM2ForRewardModel(InternLM2ForCausalLM): |
| 390 | is_pooling_model = True |
| 391 | |
| 392 | def __init__( |
| 393 | self, |
| 394 | *, |
| 395 | vllm_config: VllmConfig, |
| 396 | prefix: str = "", |
| 397 | model_type: type[InternLM2Model] = InternLM2Model, |
| 398 | ): |
| 399 | with no_init_weights( |
| 400 | self, |
| 401 | lambda mod: StageMissingLayer("output", mod), |
| 402 | targets=(LogitsProcessor, ParallelLMHead), |
| 403 | ): |
| 404 | super().__init__( |
| 405 | vllm_config=vllm_config, |
| 406 | prefix=prefix, |
| 407 | model_type=model_type, |
| 408 | ) |
| 409 | |
| 410 | config = vllm_config.model_config.hf_config |
| 411 | self.head_dtype = vllm_config.model_config.head_dtype |
| 412 | |
| 413 | self.v_head = RowParallelLinear( |
| 414 | config.hidden_size, |
| 415 | 1, |
| 416 | bias=False, |
| 417 | input_is_parallel=False, |
| 418 | params_dtype=self.head_dtype, |
| 419 | prefix=maybe_prefix(prefix, "v_head"), |
| 420 | return_bias=False, |
| 421 | ) |
| 422 | |
| 423 | pooler_config = vllm_config.model_config.pooler_config |
| 424 | assert pooler_config is not None |
| 425 | |
| 426 | self.pooler = pooler_for_token_classify(pooler_config) |
| 427 | |
| 428 | def forward( |
| 429 | self, |
| 430 | input_ids: torch.Tensor | None, |
| 431 | positions: torch.Tensor, |
| 432 | intermediate_tensors: IntermediateTensors | None = None, |
| 433 | inputs_embeds: torch.Tensor | None = None, |
| 434 | ) -> torch.Tensor | IntermediateTensors: |
| 435 | hidden_states = self.model( |
| 436 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 437 | ) |
| 438 | hidden_states = hidden_states.to(self.head_dtype) |
| 439 | logits = self.v_head(hidden_states) |
| 440 | return logits |