| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | # you may not use this file except in compliance with the License. |
| 6 | # You may obtain a copy of the License at |
| 7 | # |
| 8 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | # |
| 10 | # Unless required by applicable law or agreed to in writing, software |
| 11 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | # See the License for the specific language governing permissions and |
| 14 | # limitations under the License. |
| 15 | """Inference-only OLMoE model compatible with HuggingFace weights.""" |
| 16 | |
| 17 | from collections.abc import Iterable |
| 18 | from functools import partial |
| 19 | from itertools import islice |
| 20 | |
| 21 | import torch |
| 22 | from torch import nn |
| 23 | |
| 24 | from vllm.compilation.decorators import support_torch_compile |
| 25 | from vllm.config import VllmConfig |
| 26 | from vllm.distributed import ( |
| 27 | get_pp_group, |
| 28 | get_tensor_model_parallel_rank, |
| 29 | get_tensor_model_parallel_world_size, |
| 30 | tensor_model_parallel_all_gather, |
| 31 | ) |
| 32 | from vllm.distributed.utils import split_tensor_along_last_dim |
| 33 | from vllm.logger import init_logger |
| 34 | from vllm.model_executor.layers.attention import Attention |
| 35 | from vllm.model_executor.layers.fused_moe import ( |
| 36 | FusedMoEFactory, |
| 37 | GateLinear, |
| 38 | ) |
| 39 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 40 | from vllm.model_executor.layers.linear import ( |
| 41 | QKVParallelLinear, |
| 42 | RowParallelLinear, |
| 43 | ) |
| 44 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 45 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 46 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 47 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 48 | ParallelLMHead, |
| 49 | VocabParallelEmbedding, |
| 50 | ) |
| 51 | from vllm.sequence import IntermediateTensors |
| 52 | |
| 53 | from .interfaces import SupportsLoRA, SupportsPP |
| 54 | from .utils import ( |
| 55 | AutoWeightsLoader, |
| 56 | WeightsMapper, |
| 57 | make_empty_intermediate_tensors_factory, |
| 58 | make_layers, |
| 59 | maybe_prefix, |
| 60 | ) |
| 61 | |
| 62 | logger = init_logger(__name__) |
| 63 | |
| 64 | |
| 65 | class OlmoeMoE(nn.Module): |
| 66 | """A tensor-parallel MoE implementation for Olmoe that shards each expert |
| 67 | across all ranks. |
| 68 | |
| 69 | Each expert's weights are sharded across all ranks and a fused MoE |
| 70 | kernel is used for the forward pass, and finally we reduce the outputs |
| 71 | across ranks. |
| 72 | """ |
| 73 | |
| 74 | def __init__( |
| 75 | self, |
| 76 | num_experts: int, |
| 77 | top_k: int, |
| 78 | hidden_size: int, |
| 79 | intermediate_size: int, |
| 80 | params_dtype: torch.dtype | None = None, |
| 81 | quant_config: QuantizationConfig | None = None, |
| 82 | tp_size: int | None = None, |
| 83 | prefix: str = "", |
| 84 | ): |
| 85 | super().__init__() |
| 86 | self.hidden_size = hidden_size |
| 87 | |
| 88 | # Gate always runs at half / full precision for now. |
| 89 | self.gate = GateLinear( |
| 90 | hidden_size, |
| 91 | num_experts, |
| 92 | prefix=f"{prefix}.gate", |
| 93 | ) |
| 94 | |
| 95 | self.experts = FusedMoEFactory( |
| 96 | num_experts=num_experts, |
| 97 | top_k=top_k, |
| 98 | hidden_size=hidden_size, |
| 99 | intermediate_size=intermediate_size, |
| 100 | renormalize=False, |
| 101 | quant_config=quant_config, |
| 102 | tp_size=tp_size, |
| 103 | prefix=f"{prefix}.experts", |
| 104 | ) |
| 105 | |
| 106 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 107 | # NOTE: hidden_states can have either 1D or 2D shape. |
| 108 | orig_shape = hidden_states.shape |
| 109 | hidden_dim = hidden_states.shape[-1] |
| 110 | hidden_states = hidden_states.view(-1, hidden_dim) |
| 111 | # router_logits: (num_tokens, n_experts) |
| 112 | router_logits, _ = self.gate(hidden_states) |
| 113 | final_hidden_states = self.experts( |
| 114 | hidden_states=hidden_states, router_logits=router_logits |
| 115 | ) |
| 116 | return final_hidden_states.view(orig_shape) |
| 117 | |
| 118 | |
| 119 | class OlmoeAttention(nn.Module): |
| 120 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 121 | super().__init__() |
| 122 | |
| 123 | config = vllm_config.model_config.hf_config |
| 124 | cache_config = vllm_config.cache_config |
| 125 | quant_config = vllm_config.quant_config |
| 126 | |
| 127 | self.hidden_size = config.hidden_size |
| 128 | max_position_embeddings = getattr(config, "max_position_embeddings", 4096) |
| 129 | |
| 130 | num_heads = config.num_attention_heads |
| 131 | num_kv_heads = config.num_key_value_heads |
| 132 | |
| 133 | tp_size = get_tensor_model_parallel_world_size() |
| 134 | self.total_num_heads = num_heads |
| 135 | assert self.total_num_heads % tp_size == 0 |
| 136 | self.num_heads = self.total_num_heads // tp_size |
| 137 | self.total_num_kv_heads = num_kv_heads |
| 138 | if self.total_num_kv_heads >= tp_size: |
| 139 | # Number of KV heads is greater than TP size, so we partition |
| 140 | # the KV heads across multiple tensor parallel GPUs. |
| 141 | assert self.total_num_kv_heads % tp_size == 0 |
| 142 | else: |
| 143 | # Number of KV heads is less than TP size, so we replicate |
| 144 | # the KV heads across multiple tensor parallel GPUs. |
| 145 | assert tp_size % self.total_num_kv_heads == 0 |
| 146 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 147 | self.head_dim = self.hidden_size // self.total_num_heads |
| 148 | self.q_size = self.num_heads * self.head_dim |
| 149 | self.kv_size = self.num_kv_heads * self.head_dim |
| 150 | self.scaling = self.head_dim**-0.5 |
| 151 | self.max_position_embeddings = max_position_embeddings |
| 152 | |
| 153 | self.qkv_proj = QKVParallelLinear( |
| 154 | self.hidden_size, |
| 155 | self.head_dim, |
| 156 | self.total_num_heads, |
| 157 | self.total_num_kv_heads, |
| 158 | bias=False, |
| 159 | quant_config=quant_config, |
| 160 | prefix=f"{prefix}.qkv_proj", |
| 161 | ) |
| 162 | self.tp_size = tp_size |
| 163 | self.tp_rank = get_tensor_model_parallel_rank() |
| 164 | self.q_norm = RMSNorm(self.total_num_heads * self.head_dim, eps=1e-5) |
| 165 | self.k_norm = RMSNorm(self.total_num_kv_heads * self.head_dim, eps=1e-5) |
| 166 | self.o_proj = RowParallelLinear( |
| 167 | self.total_num_heads * self.head_dim, |
| 168 | self.hidden_size, |
| 169 | bias=False, |
| 170 | quant_config=quant_config, |
| 171 | prefix=f"{prefix}.o_proj", |
| 172 | ) |
| 173 | |
| 174 | self.rotary_emb = get_rope( |
| 175 | self.head_dim, |
| 176 | max_position=max_position_embeddings, |
| 177 | rope_parameters=config.rope_parameters, |
| 178 | is_neox_style=True, |
| 179 | ) |
| 180 | self.attn = Attention( |
| 181 | self.num_heads, |
| 182 | self.head_dim, |
| 183 | self.scaling, |
| 184 | num_kv_heads=self.num_kv_heads, |
| 185 | cache_config=cache_config, |
| 186 | quant_config=quant_config, |
| 187 | prefix=f"{prefix}.attn", |
| 188 | ) |
| 189 | |
| 190 | def _apply_qk_norm( |
| 191 | self, q: torch.Tensor, k: torch.Tensor |
| 192 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 193 | if self.tp_size > 1: |
| 194 | q = tensor_model_parallel_all_gather(q.contiguous()) |
| 195 | k = tensor_model_parallel_all_gather(k.contiguous()) |
| 196 | q = self.q_norm(q) |
| 197 | k = self.k_norm(k) |
| 198 | if self.tp_size > 1: |
| 199 | splitter = partial(split_tensor_along_last_dim, num_partitions=self.tp_size) |
| 200 | q = splitter(q)[self.tp_rank] |
| 201 | k = splitter(k)[self.tp_rank] |
| 202 | return q, k |
| 203 | |
| 204 | def forward( |
| 205 | self, |
| 206 | positions: torch.Tensor, |
| 207 | hidden_states: torch.Tensor, |
| 208 | ) -> torch.Tensor: |
| 209 | qkv, _ = self.qkv_proj(hidden_states) |
| 210 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 211 | q, k = self._apply_qk_norm(q, k) |
| 212 | q, k = self.rotary_emb(positions, q, k) |
| 213 | attn_output = self.attn(q, k, v) |
| 214 | output, _ = self.o_proj(attn_output) |
| 215 | return output |
| 216 | |
| 217 | |
| 218 | class OlmoeDecoderLayer(nn.Module): |
| 219 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 220 | super().__init__() |
| 221 | config = vllm_config.model_config.hf_config |
| 222 | quant_config = vllm_config.quant_config |
| 223 | |
| 224 | self.hidden_size = config.hidden_size |
| 225 | |
| 226 | self.self_attn = OlmoeAttention( |
| 227 | vllm_config=vllm_config, |
| 228 | prefix=f"{prefix}.self_attn", |
| 229 | ) |
| 230 | |
| 231 | self.mlp = OlmoeMoE( |
| 232 | num_experts=config.num_experts, |
| 233 | top_k=config.num_experts_per_tok, |
| 234 | hidden_size=config.hidden_size, |
| 235 | intermediate_size=config.intermediate_size, |
| 236 | quant_config=quant_config, |
| 237 | prefix=f"{prefix}.mlp", |
| 238 | ) |
| 239 | self.input_layernorm = RMSNorm(config.hidden_size, eps=1e-5) |
| 240 | self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=1e-5) |
| 241 | |
| 242 | def forward( |
| 243 | self, |
| 244 | positions: torch.Tensor, |
| 245 | hidden_states: torch.Tensor, |
| 246 | residual: torch.Tensor | None, |
| 247 | ) -> torch.Tensor: |
| 248 | # Self Attention |
| 249 | if residual is None: |
| 250 | residual = hidden_states |
| 251 | hidden_states = self.input_layernorm(hidden_states) |
| 252 | else: |
| 253 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 254 | |
| 255 | hidden_states = self.self_attn( |
| 256 | positions=positions, |
| 257 | hidden_states=hidden_states, |
| 258 | ) |
| 259 | |
| 260 | # Fully Connected |
| 261 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 262 | hidden_states = self.mlp(hidden_states) |
| 263 | return hidden_states, residual |
| 264 | |
| 265 | |
| 266 | @support_torch_compile |
| 267 | class OlmoeModel(nn.Module): |
| 268 | def __init__( |
| 269 | self, |
| 270 | *, |
| 271 | vllm_config: VllmConfig, |
| 272 | prefix: str = "", |
| 273 | layer_type: type[nn.Module] = OlmoeDecoderLayer, |
| 274 | ): |
| 275 | super().__init__() |
| 276 | |
| 277 | config = vllm_config.model_config.hf_config |
| 278 | quant_config = vllm_config.quant_config |
| 279 | |
| 280 | self.vocab_size = config.vocab_size |
| 281 | self.config = config |
| 282 | self.embed_tokens = VocabParallelEmbedding( |
| 283 | config.vocab_size, |
| 284 | config.hidden_size, |
| 285 | quant_config=quant_config, |
| 286 | ) |
| 287 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 288 | config.num_hidden_layers, |
| 289 | lambda prefix: layer_type(vllm_config=vllm_config, prefix=prefix), |
| 290 | prefix=f"{prefix}.layers", |
| 291 | ) |
| 292 | self.norm = RMSNorm(config.hidden_size, eps=1e-5) |
| 293 | |
| 294 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 295 | ["hidden_states", "residual"], config.hidden_size |
| 296 | ) |
| 297 | |
| 298 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 299 | return self.embed_tokens(input_ids) |
| 300 | |
| 301 | def forward( |
| 302 | self, |
| 303 | input_ids: torch.Tensor | None, |
| 304 | positions: torch.Tensor, |
| 305 | intermediate_tensors: IntermediateTensors | None, |
| 306 | inputs_embeds: torch.Tensor | None = None, |
| 307 | ) -> torch.Tensor | IntermediateTensors: |
| 308 | if get_pp_group().is_first_rank: |
| 309 | if inputs_embeds is not None: |
| 310 | hidden_states = inputs_embeds |
| 311 | else: |
| 312 | hidden_states = self.embed_input_ids(input_ids) |
| 313 | residual = None |
| 314 | else: |
| 315 | assert intermediate_tensors is not None |
| 316 | hidden_states = intermediate_tensors["hidden_states"] |
| 317 | residual = intermediate_tensors["residual"] |
| 318 | |
| 319 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 320 | hidden_states, residual = layer( |
| 321 | positions, |
| 322 | hidden_states, |
| 323 | residual, |
| 324 | ) |
| 325 | |
| 326 | if not get_pp_group().is_last_rank: |
| 327 | return IntermediateTensors( |
| 328 | {"hidden_states": hidden_states, "residual": residual} |
| 329 | ) |
| 330 | |
| 331 | if residual is not None: |
| 332 | hidden_states, _ = self.norm(hidden_states, residual) |
| 333 | else: |
| 334 | hidden_states = self.norm(hidden_states) |
| 335 | return hidden_states |
| 336 | |
| 337 | |
| 338 | class OlmoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA): |
| 339 | hf_to_vllm_mapper = WeightsMapper( |
| 340 | orig_to_new_stacked={ |
| 341 | # weight_name: (param_name, shard_id) |
| 342 | ".q_proj": (".qkv_proj", "q"), |
| 343 | ".k_proj": (".qkv_proj", "k"), |
| 344 | ".v_proj": (".qkv_proj", "v"), |
| 345 | } |
| 346 | ) |
| 347 | packed_modules_mapping = { |
| 348 | "qkv_proj": [ |
| 349 | "q_proj", |
| 350 | "k_proj", |
| 351 | "v_proj", |
| 352 | ] |
| 353 | } |
| 354 | |
| 355 | def __init__( |
| 356 | self, |
| 357 | *, |
| 358 | vllm_config: VllmConfig, |
| 359 | prefix: str = "", |
| 360 | layer_type: type[nn.Module] = OlmoeDecoderLayer, |
| 361 | ): |
| 362 | super().__init__() |
| 363 | config = vllm_config.model_config.hf_config |
| 364 | quant_config = vllm_config.quant_config |
| 365 | self.config = config |
| 366 | self.quant_config = quant_config |
| 367 | self.model = OlmoeModel( |
| 368 | vllm_config=vllm_config, |
| 369 | prefix=maybe_prefix(prefix, "model"), |
| 370 | layer_type=layer_type, |
| 371 | ) |
| 372 | self.lm_head = ParallelLMHead( |
| 373 | config.vocab_size, |
| 374 | config.hidden_size, |
| 375 | quant_config=quant_config, |
| 376 | prefix=maybe_prefix(prefix, "lm_head"), |
| 377 | ) |
| 378 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 379 | |
| 380 | self.make_empty_intermediate_tensors = ( |
| 381 | self.model.make_empty_intermediate_tensors |
| 382 | ) |
| 383 | |
| 384 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 385 | return self.model.embed_input_ids(input_ids) |
| 386 | |
| 387 | def forward( |
| 388 | self, |
| 389 | input_ids: torch.Tensor | None, |
| 390 | positions: torch.Tensor, |
| 391 | intermediate_tensors: IntermediateTensors | None = None, |
| 392 | inputs_embeds: torch.Tensor | None = None, |
| 393 | ) -> torch.Tensor | IntermediateTensors: |
| 394 | hidden_states = self.model( |
| 395 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 396 | ) |
| 397 | return hidden_states |
| 398 | |
| 399 | def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 400 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 401 | return logits |
| 402 | |
| 403 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 404 | loader = AutoWeightsLoader(self) |
| 405 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |