| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Copyright 2024 Cohere and the HuggingFace Inc. team. All rights reserved. |
| 5 | # |
| 6 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 7 | # and OPT implementations in this library. It has been modified from its |
| 8 | # original forms to accommodate minor architectural differences compared |
| 9 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 10 | # |
| 11 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 12 | # you may not use this file except in compliance with the License. |
| 13 | # You may obtain a copy of the License at |
| 14 | # |
| 15 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 16 | # |
| 17 | # Unless required by applicable law or agreed to in writing, software |
| 18 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 19 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 20 | # See the License for the specific language governing permissions and |
| 21 | # limitations under the License. |
| 22 | |
| 23 | # This file is based on the LLama model definition file in transformers |
| 24 | """PyTorch Cohere model.""" |
| 25 | |
| 26 | from collections.abc import Iterable |
| 27 | from itertools import islice |
| 28 | |
| 29 | import torch |
| 30 | from torch import nn |
| 31 | from transformers import Cohere2Config, CohereConfig |
| 32 | |
| 33 | from vllm.compilation.decorators import support_torch_compile |
| 34 | from vllm.config import CacheConfig, VllmConfig |
| 35 | from vllm.distributed import ( |
| 36 | get_pp_group, |
| 37 | get_tensor_model_parallel_world_size, |
| 38 | tensor_model_parallel_all_reduce, |
| 39 | ) |
| 40 | from vllm.model_executor.layers.activation import SiluAndMul |
| 41 | from vllm.model_executor.layers.attention import Attention |
| 42 | from vllm.model_executor.layers.linear import ( |
| 43 | MergedColumnParallelLinear, |
| 44 | QKVParallelLinear, |
| 45 | RowParallelLinear, |
| 46 | ) |
| 47 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 48 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 49 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 50 | from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding |
| 51 | from vllm.model_executor.model_loader.weight_utils import row_parallel_weight_loader |
| 52 | from vllm.model_executor.utils import set_weight_attrs |
| 53 | from vllm.platforms import current_platform |
| 54 | from vllm.sequence import IntermediateTensors |
| 55 | |
| 56 | from .interfaces import SupportsLoRA, SupportsPP, SupportsQuant |
| 57 | from .utils import ( |
| 58 | AutoWeightsLoader, |
| 59 | WeightsMapper, |
| 60 | extract_layer_index, |
| 61 | make_empty_intermediate_tensors_factory, |
| 62 | make_layers, |
| 63 | maybe_prefix, |
| 64 | ) |
| 65 | |
| 66 | |
| 67 | @torch.compile(backend=current_platform.simple_compile_backend) |
| 68 | def layer_norm_func(hidden_states, weight, variance_epsilon): |
| 69 | input_dtype = hidden_states.dtype |
| 70 | hidden_states = hidden_states.to(torch.float32) |
| 71 | mean = hidden_states.mean(-1, keepdim=True) |
| 72 | variance = (hidden_states - mean).pow(2).mean(-1, keepdim=True) |
| 73 | hidden_states = (hidden_states - mean) * torch.rsqrt(variance + variance_epsilon) |
| 74 | hidden_states = weight.to(torch.float32) * hidden_states |
| 75 | return hidden_states.to(input_dtype) |
| 76 | |
| 77 | |
| 78 | class LayerNorm(nn.Module): |
| 79 | def __init__(self, param_shape=None, eps=1e-5): |
| 80 | super().__init__() |
| 81 | self.weight = nn.Parameter(torch.ones(param_shape)) |
| 82 | self.variance_epsilon = eps |
| 83 | set_weight_attrs(self.weight, {"weight_loader": row_parallel_weight_loader}) |
| 84 | |
| 85 | def forward(self, hidden_states, residuals=None): |
| 86 | hidden_states = layer_norm_func( |
| 87 | hidden_states, self.weight, self.variance_epsilon |
| 88 | ) |
| 89 | return hidden_states, residuals |
| 90 | |
| 91 | |
| 92 | @torch.compile(backend=current_platform.simple_compile_backend) |
| 93 | def rms_norm_func(hidden_states, weight, variance_epsilon): |
| 94 | input_dtype = hidden_states.dtype |
| 95 | hidden_states = hidden_states.to(torch.float32) |
| 96 | |
| 97 | variance = hidden_states.pow(2).mean(-1, keepdim=True) |
| 98 | hidden_states = hidden_states * torch.rsqrt(variance + variance_epsilon) |
| 99 | |
| 100 | hidden_states = weight.to(torch.float32) * hidden_states |
| 101 | return hidden_states.to(input_dtype) |
| 102 | |
| 103 | |
| 104 | class RMSNorm(nn.Module): |
| 105 | def __init__(self, param_shape=None, eps=1e-6): |
| 106 | super().__init__() |
| 107 | self.weight = nn.Parameter(torch.ones(param_shape)) |
| 108 | self.variance_epsilon = eps |
| 109 | set_weight_attrs(self.weight, {"weight_loader": row_parallel_weight_loader}) |
| 110 | |
| 111 | def forward(self, hidden_states, residuals=None): |
| 112 | hidden_states = rms_norm_func(hidden_states, self.weight, self.variance_epsilon) |
| 113 | return hidden_states, residuals |
| 114 | |
| 115 | |
| 116 | def select_norm_impl(config: CohereConfig) -> tuple[type[nn.Module], float]: |
| 117 | """Returns the normalization layer class and epsilon value to use. |
| 118 | If `config.rms_norm_eps` is present, use RMSNorm. |
| 119 | Otherwise default to LayerNorm with `config.layer_norm_eps`. |
| 120 | """ |
| 121 | rms_eps = getattr(config, "rms_norm_eps", None) |
| 122 | if rms_eps is not None: |
| 123 | return RMSNorm, rms_eps |
| 124 | |
| 125 | return LayerNorm, config.layer_norm_eps |
| 126 | |
| 127 | |
| 128 | # Copied from transformers.models.llama.modeling_llama.LlamaMLP Llama->Cohere |
| 129 | class CohereMLP(nn.Module): |
| 130 | def __init__( |
| 131 | self, |
| 132 | config: CohereConfig, |
| 133 | intermediate_size: int | None = None, |
| 134 | quant_config: QuantizationConfig | None = None, |
| 135 | # so we can reduce the attention and MLP outputs together |
| 136 | reduce_results: bool = False, |
| 137 | prefix: str = "", |
| 138 | ): |
| 139 | super().__init__() |
| 140 | self.config = config |
| 141 | self.hidden_size = config.hidden_size |
| 142 | if intermediate_size is None: |
| 143 | self.intermediate_size = config.intermediate_size |
| 144 | else: |
| 145 | self.intermediate_size = intermediate_size |
| 146 | self.gate_up_proj = MergedColumnParallelLinear( |
| 147 | self.hidden_size, |
| 148 | [self.intermediate_size] * 2, |
| 149 | bias=False, |
| 150 | quant_config=quant_config, |
| 151 | prefix=f"{prefix}.gate_up_proj", |
| 152 | ) |
| 153 | self.down_proj = RowParallelLinear( |
| 154 | self.intermediate_size, |
| 155 | self.hidden_size, |
| 156 | bias=False, |
| 157 | quant_config=quant_config, |
| 158 | reduce_results=reduce_results, |
| 159 | prefix=f"{prefix}.down_proj", |
| 160 | ) |
| 161 | self.act_fn = SiluAndMul() |
| 162 | |
| 163 | def forward(self, x): |
| 164 | gate_up, _ = self.gate_up_proj(x) |
| 165 | x = self.act_fn(gate_up) |
| 166 | x, _ = self.down_proj(x) |
| 167 | return x |
| 168 | |
| 169 | |
| 170 | class CohereAttention(nn.Module): |
| 171 | def __init__( |
| 172 | self, |
| 173 | config: CohereConfig | Cohere2Config, |
| 174 | cache_config: CacheConfig | None = None, |
| 175 | quant_config: QuantizationConfig | None = None, |
| 176 | prefix: str = "", |
| 177 | ): |
| 178 | super().__init__() |
| 179 | tp_size = get_tensor_model_parallel_world_size() |
| 180 | self.config = config |
| 181 | self.layer_idx = extract_layer_index(prefix) |
| 182 | self.attention_dropout = config.attention_dropout |
| 183 | self.hidden_size = config.hidden_size |
| 184 | self.total_num_heads = config.num_attention_heads |
| 185 | self.num_heads = self.total_num_heads // tp_size |
| 186 | if hasattr(config, "head_dim"): |
| 187 | self.head_dim = config.head_dim |
| 188 | else: |
| 189 | self.head_dim = self.hidden_size // self.total_num_heads |
| 190 | self.total_num_kv_heads = config.num_key_value_heads |
| 191 | if self.total_num_kv_heads >= tp_size: |
| 192 | # Number of KV heads is greater than TP size, so we partition |
| 193 | # the KV heads across multiple tensor parallel GPUs. |
| 194 | assert self.total_num_kv_heads % tp_size == 0 |
| 195 | else: |
| 196 | # Number of KV heads is less than TP size, so we replicate |
| 197 | # the KV heads across multiple tensor parallel GPUs. |
| 198 | assert tp_size % self.total_num_kv_heads == 0 |
| 199 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 200 | self.q_size = self.num_heads * self.head_dim |
| 201 | self.kv_size = self.num_kv_heads * self.head_dim |
| 202 | self.scaling = self.head_dim**-0.5 |
| 203 | self.max_position_embeddings = getattr( |
| 204 | config, "model_max_length", None |
| 205 | ) or getattr(config, "max_position_embeddings", 8192) |
| 206 | assert isinstance(self.max_position_embeddings, int) |
| 207 | self.use_qk_norm = getattr(config, "use_qk_norm", False) |
| 208 | self.qkv_proj = QKVParallelLinear( |
| 209 | self.hidden_size, |
| 210 | self.head_dim, |
| 211 | self.total_num_heads, |
| 212 | self.total_num_kv_heads, |
| 213 | bias=False, |
| 214 | quant_config=quant_config, |
| 215 | prefix=f"{prefix}.qkv_proj", |
| 216 | ) |
| 217 | self.o_proj = RowParallelLinear( |
| 218 | self.total_num_heads * self.head_dim, |
| 219 | self.hidden_size, |
| 220 | bias=False, |
| 221 | quant_config=quant_config, |
| 222 | prefix=f"{prefix}.o_proj", |
| 223 | # NOTE: reduction will happen in the decoder layer forward |
| 224 | # so we can combine the attention and MLP outputs together |
| 225 | reduce_results=False, |
| 226 | ) |
| 227 | self.rotary_emb = get_rope( |
| 228 | self.head_dim, |
| 229 | max_position=self.max_position_embeddings, |
| 230 | rope_parameters=config.rope_parameters, |
| 231 | is_neox_style=False, |
| 232 | ) |
| 233 | |
| 234 | # Model v2 has interleaved sliding windows, v1 does not |
| 235 | self.v1 = isinstance(config, CohereConfig) |
| 236 | |
| 237 | # cohere swa layer sees [pos - sliding_window, pos], i.e. sliding_window + 1 |
| 238 | # tokens. vLLM's FlashAttention backend does (value - 1, 0), so we pass |
| 239 | # sliding_window + 1 here to match the training convention. The same +1 |
| 240 | # propagates into the KV-cache eviction formula |
| 241 | # (single_type_kv_cache_manager.py), keeping both paths consistent. |
| 242 | self.sliding_window = None |
| 243 | if not self.v1 and config.layer_types[self.layer_idx] == "sliding_attention": |
| 244 | self.sliding_window = config.sliding_window + 1 |
| 245 | |
| 246 | self.attn = Attention( |
| 247 | self.num_heads, |
| 248 | self.head_dim, |
| 249 | self.scaling, |
| 250 | num_kv_heads=self.num_kv_heads, |
| 251 | cache_config=cache_config, |
| 252 | quant_config=quant_config, |
| 253 | per_layer_sliding_window=self.sliding_window, |
| 254 | prefix=f"{prefix}.attn", |
| 255 | ) |
| 256 | norm_cls, norm_eps = select_norm_impl(config) |
| 257 | if self.use_qk_norm: |
| 258 | self.q_norm = norm_cls( |
| 259 | param_shape=(self.num_heads, self.head_dim), eps=norm_eps |
| 260 | ) |
| 261 | self.k_norm = norm_cls( |
| 262 | param_shape=(self.num_kv_heads, self.head_dim), eps=norm_eps |
| 263 | ) |
| 264 | |
| 265 | def _apply_qk_norm(self, q, k): |
| 266 | q = q.view(*q.shape[:-1], -1, self.head_dim) |
| 267 | k = k.view(*k.shape[:-1], -1, self.head_dim) |
| 268 | q, _ = self.q_norm(q) |
| 269 | k, _ = self.k_norm(k) |
| 270 | q = q.view(*q.shape[:-2], -1) |
| 271 | k = k.view(*k.shape[:-2], -1) |
| 272 | return q, k |
| 273 | |
| 274 | def forward( |
| 275 | self, |
| 276 | positions: torch.Tensor, |
| 277 | hidden_states: torch.Tensor, |
| 278 | ) -> torch.Tensor: |
| 279 | qkv, _ = self.qkv_proj(hidden_states) |
| 280 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 281 | if self.use_qk_norm: |
| 282 | q, k = self._apply_qk_norm(q, k) |
| 283 | if self.v1 or self.sliding_window: |
| 284 | q, k = self.rotary_emb(positions, q, k) |
| 285 | attn_output = self.attn(q, k, v) |
| 286 | output, _ = self.o_proj(attn_output) |
| 287 | return output |
| 288 | |
| 289 | |
| 290 | class CohereDecoderLayer(nn.Module): |
| 291 | def __init__( |
| 292 | self, |
| 293 | config: CohereConfig | Cohere2Config, |
| 294 | cache_config: CacheConfig | None = None, |
| 295 | quant_config: QuantizationConfig | None = None, |
| 296 | prefix: str = "", |
| 297 | ): |
| 298 | super().__init__() |
| 299 | self.hidden_size = config.hidden_size |
| 300 | self.tp_size = get_tensor_model_parallel_world_size() |
| 301 | self.self_attn = CohereAttention( |
| 302 | config, |
| 303 | cache_config, |
| 304 | quant_config=quant_config, |
| 305 | prefix=f"{prefix}.self_attn", |
| 306 | ) |
| 307 | self.mlp = CohereMLP(config, quant_config=quant_config, prefix=f"{prefix}.mlp") |
| 308 | norm_cls, norm_eps = select_norm_impl(config) |
| 309 | self.input_layernorm = norm_cls(param_shape=(config.hidden_size), eps=norm_eps) |
| 310 | |
| 311 | def forward( |
| 312 | self, |
| 313 | positions: torch.Tensor, |
| 314 | hidden_states: torch.Tensor, |
| 315 | residual: torch.Tensor | None, |
| 316 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 317 | # Self Attention |
| 318 | residual = hidden_states |
| 319 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 320 | hidden_states_attention = self.self_attn( |
| 321 | positions=positions, |
| 322 | hidden_states=hidden_states, |
| 323 | ) |
| 324 | hidden_states_mlp = self.mlp(hidden_states) |
| 325 | parallel_block_output = hidden_states_attention + hidden_states_mlp |
| 326 | # Add everything together |
| 327 | if self.tp_size > 1: |
| 328 | # do the reduction in 1 shot instead of 2 separate all reduce |
| 329 | parallel_block_output = tensor_model_parallel_all_reduce( |
| 330 | parallel_block_output |
| 331 | ) |
| 332 | hidden_states = residual + parallel_block_output |
| 333 | |
| 334 | return hidden_states, residual |
| 335 | |
| 336 | |
| 337 | @support_torch_compile |
| 338 | class CohereModel(nn.Module): |
| 339 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 340 | super().__init__() |
| 341 | |
| 342 | config = vllm_config.model_config.hf_config |
| 343 | cache_config = vllm_config.cache_config |
| 344 | quant_config = vllm_config.quant_config |
| 345 | self.quant_config = quant_config |
| 346 | |
| 347 | self.config = config |
| 348 | |
| 349 | self.vocab_size = config.vocab_size |
| 350 | |
| 351 | self.embed_tokens = VocabParallelEmbedding( |
| 352 | config.vocab_size, config.hidden_size |
| 353 | ) |
| 354 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 355 | config.num_hidden_layers, |
| 356 | lambda prefix: CohereDecoderLayer( |
| 357 | config, cache_config, quant_config, prefix=prefix |
| 358 | ), |
| 359 | prefix=f"{prefix}.layers", |
| 360 | ) |
| 361 | |
| 362 | norm_cls, norm_eps = select_norm_impl(config) |
| 363 | self.norm = norm_cls(param_shape=(config.hidden_size), eps=norm_eps) |
| 364 | |
| 365 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 366 | ["hidden_states", "residual"], config.hidden_size |
| 367 | ) |
| 368 | |
| 369 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 370 | return self.embed_tokens(input_ids) |
| 371 | |
| 372 | def forward( |
| 373 | self, |
| 374 | input_ids: torch.Tensor | None, |
| 375 | positions: torch.Tensor, |
| 376 | intermediate_tensors: IntermediateTensors | None, |
| 377 | inputs_embeds: torch.Tensor | None = None, |
| 378 | ) -> torch.Tensor | IntermediateTensors: |
| 379 | if get_pp_group().is_first_rank: |
| 380 | if inputs_embeds is not None: |
| 381 | hidden_states = inputs_embeds |
| 382 | else: |
| 383 | hidden_states = self.embed_input_ids(input_ids) |
| 384 | residual = None |
| 385 | else: |
| 386 | assert intermediate_tensors is not None |
| 387 | hidden_states = intermediate_tensors["hidden_states"] |
| 388 | residual = intermediate_tensors["residual"] |
| 389 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 390 | hidden_states, residual = layer( |
| 391 | positions, |
| 392 | hidden_states, |
| 393 | residual, |
| 394 | ) |
| 395 | if not get_pp_group().is_last_rank: |
| 396 | return IntermediateTensors( |
| 397 | {"hidden_states": hidden_states, "residual": residual} |
| 398 | ) |
| 399 | hidden_states, _ = self.norm(hidden_states, residual) |
| 400 | return hidden_states |
| 401 | |
| 402 | |
| 403 | class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant): |
| 404 | hf_to_vllm_mapper = WeightsMapper( |
| 405 | orig_to_new_stacked={ |
| 406 | # weight_name: (param_name, shard_id) |
| 407 | ".q_proj": (".qkv_proj", "q"), |
| 408 | ".k_proj": (".qkv_proj", "k"), |
| 409 | ".v_proj": (".qkv_proj", "v"), |
| 410 | ".gate_proj": (".gate_up_proj", 0), |
| 411 | ".up_proj": (".gate_up_proj", 1), |
| 412 | }, |
| 413 | # ModelOpt NVFP4 checkpoints carry raw quantizer-module state |
| 414 | # (e.g. "*.weight_quantizer._double_scale"); drop them before loading. |
| 415 | # See #41925. |
| 416 | orig_to_new_substr={"_quantizer.": None}, |
| 417 | orig_to_new_prefix={"lm_head": None}, |
| 418 | ) |
| 419 | packed_modules_mapping: dict[str, list[str]] = { |
| 420 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 421 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 422 | } |
| 423 | # LoRA specific attributes |
| 424 | embedding_modules = {"embed_tokens": "input_embeddings"} |
| 425 | |
| 426 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 427 | super().__init__() |
| 428 | config = vllm_config.model_config.hf_config |
| 429 | quant_config = vllm_config.quant_config |
| 430 | |
| 431 | self.config = config |
| 432 | |
| 433 | self.quant_config = quant_config |
| 434 | self.logits_processor = LogitsProcessor( |
| 435 | config.vocab_size, scale=config.logit_scale |
| 436 | ) |
| 437 | self.model = CohereModel( |
| 438 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 439 | ) |
| 440 | self.make_empty_intermediate_tensors = ( |
| 441 | self.model.make_empty_intermediate_tensors |
| 442 | ) |
| 443 | |
| 444 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 445 | return self.model.embed_input_ids(input_ids) |
| 446 | |
| 447 | @torch.no_grad() |
| 448 | def forward( |
| 449 | self, |
| 450 | input_ids: torch.Tensor | None, |
| 451 | positions: torch.Tensor, |
| 452 | intermediate_tensors: IntermediateTensors | None = None, |
| 453 | inputs_embeds: torch.Tensor | None = None, |
| 454 | ) -> torch.Tensor | IntermediateTensors: |
| 455 | hidden_states = self.model( |
| 456 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 457 | ) |
| 458 | return hidden_states |
| 459 | |
| 460 | def compute_logits( |
| 461 | self, |
| 462 | hidden_states: torch.Tensor, |
| 463 | ) -> torch.Tensor | None: |
| 464 | is_not_lora = hasattr(self.model.embed_tokens, "weight") |
| 465 | if is_not_lora: |
| 466 | logits = self.logits_processor(self.model.embed_tokens, hidden_states) |
| 467 | else: |
| 468 | logits = self.logits_processor( |
| 469 | self.model.embed_tokens.base_layer, hidden_states |
| 470 | ) |
| 471 | |
| 472 | return logits |
| 473 | |
| 474 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 475 | loader = AutoWeightsLoader(self) |
| 476 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |