| 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/main/src/transformers/models/olmo_hybrid/modeling_olmo_hybrid.py |
| 6 | # Copyright 2026 The vLLM team. |
| 7 | # |
| 8 | # This code combines OLMo2/OLMo3 attention with Gated DeltaNet linear attention |
| 9 | # for the OLMo Hybrid architecture. |
| 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 | """Inference-only OLMo Hybrid model compatible with HuggingFace weights.""" |
| 23 | |
| 24 | from collections.abc import Iterable |
| 25 | from functools import partial |
| 26 | from itertools import islice |
| 27 | |
| 28 | import torch |
| 29 | from torch import nn |
| 30 | |
| 31 | from vllm.compilation.decorators import support_torch_compile |
| 32 | from vllm.config import ( |
| 33 | VllmConfig, |
| 34 | ) |
| 35 | from vllm.distributed import ( |
| 36 | get_pp_group, |
| 37 | get_tensor_model_parallel_rank, |
| 38 | get_tensor_model_parallel_world_size, |
| 39 | tensor_model_parallel_all_gather, |
| 40 | ) |
| 41 | from vllm.distributed.utils import split_tensor_along_last_dim |
| 42 | from vllm.logger import init_logger |
| 43 | from vllm.model_executor.layers.activation import SiluAndMul |
| 44 | from vllm.model_executor.layers.attention import Attention |
| 45 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 46 | from vllm.model_executor.layers.linear import ( |
| 47 | MergedColumnParallelLinear, |
| 48 | QKVParallelLinear, |
| 49 | RowParallelLinear, |
| 50 | ) |
| 51 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 52 | from vllm.model_executor.layers.mamba.gdn.olmo_gdn_linear_attn import ( |
| 53 | OlmoHybridGatedDeltaNetAttention, |
| 54 | ) |
| 55 | from vllm.model_executor.layers.mamba.mamba_utils import ( |
| 56 | MambaStateCopyFunc, |
| 57 | MambaStateCopyFuncCalculator, |
| 58 | MambaStateDtypeCalculator, |
| 59 | MambaStateShapeCalculator, |
| 60 | ) |
| 61 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 62 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 63 | ParallelLMHead, |
| 64 | VocabParallelEmbedding, |
| 65 | ) |
| 66 | from vllm.sequence import IntermediateTensors |
| 67 | |
| 68 | from .interfaces import HasInnerState, IsHybrid, SupportsLoRA, SupportsPP |
| 69 | from .utils import ( |
| 70 | AutoWeightsLoader, |
| 71 | WeightsMapper, |
| 72 | extract_layer_index, |
| 73 | make_empty_intermediate_tensors_factory, |
| 74 | make_layers, |
| 75 | maybe_prefix, |
| 76 | ) |
| 77 | |
| 78 | logger = init_logger(__name__) |
| 79 | |
| 80 | |
| 81 | class OlmoHybridAttention(nn.Module): |
| 82 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 83 | super().__init__() |
| 84 | self.config = vllm_config.model_config.hf_config |
| 85 | |
| 86 | hidden_size = self.config.hidden_size |
| 87 | self.tp_size = get_tensor_model_parallel_world_size() |
| 88 | self.total_num_heads = self.config.num_attention_heads |
| 89 | |
| 90 | assert hidden_size % self.total_num_heads == 0 |
| 91 | assert self.total_num_heads % self.tp_size == 0 |
| 92 | |
| 93 | self.num_heads = self.total_num_heads // self.tp_size |
| 94 | self.total_num_kv_heads = ( |
| 95 | self.config.num_key_value_heads or self.total_num_heads |
| 96 | ) |
| 97 | if self.total_num_kv_heads >= self.tp_size: |
| 98 | assert self.total_num_kv_heads % self.tp_size == 0 |
| 99 | else: |
| 100 | assert self.tp_size % self.total_num_kv_heads == 0 |
| 101 | |
| 102 | self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size) |
| 103 | self.head_dim = hidden_size // self.total_num_heads |
| 104 | self.q_size = self.num_heads * self.head_dim |
| 105 | self.kv_size = self.num_kv_heads * self.head_dim |
| 106 | self.max_position_embeddings = self.config.max_position_embeddings |
| 107 | |
| 108 | self.qkv_proj = QKVParallelLinear( |
| 109 | hidden_size, |
| 110 | self.head_dim, |
| 111 | self.total_num_heads, |
| 112 | self.total_num_kv_heads, |
| 113 | bias=False, |
| 114 | quant_config=vllm_config.quant_config, |
| 115 | prefix=f"{prefix}.qkv_proj", |
| 116 | ) |
| 117 | |
| 118 | self.tp_rank = get_tensor_model_parallel_rank() |
| 119 | |
| 120 | self.k_norm = RMSNorm( |
| 121 | self.total_num_kv_heads * self.head_dim, |
| 122 | eps=self.config.rms_norm_eps, |
| 123 | ) |
| 124 | self.q_norm = RMSNorm( |
| 125 | self.config.hidden_size, |
| 126 | eps=self.config.rms_norm_eps, |
| 127 | ) |
| 128 | |
| 129 | self.scaling = self.head_dim**-0.5 |
| 130 | |
| 131 | self.attn = Attention( |
| 132 | self.num_heads, |
| 133 | self.head_dim, |
| 134 | self.scaling, |
| 135 | num_kv_heads=self.num_kv_heads, |
| 136 | cache_config=vllm_config.cache_config, |
| 137 | quant_config=vllm_config.quant_config, |
| 138 | prefix=f"{prefix}.attn", |
| 139 | ) |
| 140 | |
| 141 | rope_parameters = getattr(self.config, "rope_parameters", None) |
| 142 | self._use_rope = (rope_parameters is not None) and ( |
| 143 | rope_parameters["rope_theta"] is not None |
| 144 | ) |
| 145 | |
| 146 | self.rotary_emb: nn.Module | None |
| 147 | if self._use_rope: |
| 148 | self.rotary_emb = get_rope( |
| 149 | self.head_dim, |
| 150 | max_position=self.max_position_embeddings, |
| 151 | rope_parameters=rope_parameters, |
| 152 | ) |
| 153 | else: |
| 154 | self.rotary_emb = None |
| 155 | |
| 156 | self.o_proj = RowParallelLinear( |
| 157 | self.total_num_heads * self.head_dim, |
| 158 | hidden_size, |
| 159 | bias=False, |
| 160 | quant_config=vllm_config.quant_config, |
| 161 | prefix=f"{prefix}.o_proj", |
| 162 | ) |
| 163 | |
| 164 | def _apply_qk_norm( |
| 165 | self, q: torch.Tensor, k: torch.Tensor |
| 166 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 167 | if self.tp_size > 1: |
| 168 | q = tensor_model_parallel_all_gather(q.contiguous()) |
| 169 | k = tensor_model_parallel_all_gather(k.contiguous()) |
| 170 | q = self.q_norm(q) |
| 171 | k = self.k_norm(k) |
| 172 | if self.tp_size > 1: |
| 173 | splitter = partial(split_tensor_along_last_dim, num_partitions=self.tp_size) |
| 174 | q = splitter(q)[self.tp_rank] |
| 175 | k = splitter(k)[self.tp_rank] |
| 176 | return q, k |
| 177 | |
| 178 | def forward( |
| 179 | self, |
| 180 | positions: torch.Tensor, |
| 181 | hidden_states: torch.Tensor, |
| 182 | ) -> torch.Tensor: |
| 183 | qkv, _ = self.qkv_proj(hidden_states) |
| 184 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 185 | q, k = self._apply_qk_norm(q, k) |
| 186 | if self._use_rope: |
| 187 | assert self.rotary_emb is not None |
| 188 | q, k = self.rotary_emb(positions, q, k) |
| 189 | attn_output = self.attn(q, k, v) |
| 190 | output, _ = self.o_proj(attn_output) |
| 191 | return output |
| 192 | |
| 193 | |
| 194 | class OlmoHybridMLP(nn.Module): |
| 195 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 196 | super().__init__() |
| 197 | config = vllm_config.model_config.hf_config |
| 198 | hidden_size = config.hidden_size |
| 199 | intermediate_size = config.intermediate_size |
| 200 | |
| 201 | self.gate_up_proj = MergedColumnParallelLinear( |
| 202 | hidden_size, |
| 203 | [intermediate_size] * 2, |
| 204 | bias=False, |
| 205 | quant_config=vllm_config.quant_config, |
| 206 | prefix=f"{prefix}.gate_up_proj", |
| 207 | ) |
| 208 | |
| 209 | self.act_fn = SiluAndMul() |
| 210 | |
| 211 | self.down_proj = RowParallelLinear( |
| 212 | intermediate_size, |
| 213 | hidden_size, |
| 214 | bias=False, |
| 215 | quant_config=vllm_config.quant_config, |
| 216 | prefix=f"{prefix}.down_proj", |
| 217 | ) |
| 218 | |
| 219 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 220 | gate_up, _ = self.gate_up_proj(x) |
| 221 | x = self.act_fn(gate_up) |
| 222 | x, _ = self.down_proj(x) |
| 223 | return x |
| 224 | |
| 225 | |
| 226 | class OlmoHybridDecoderLayer(nn.Module): |
| 227 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 228 | super().__init__() |
| 229 | config = vllm_config.model_config.hf_config |
| 230 | |
| 231 | layer_idx = extract_layer_index(prefix) |
| 232 | self.layer_type = config.layer_types[layer_idx] |
| 233 | self.layer_idx = layer_idx |
| 234 | |
| 235 | if self.layer_type == "linear_attention": |
| 236 | self.linear_attn = OlmoHybridGatedDeltaNetAttention( |
| 237 | config, |
| 238 | vllm_config, |
| 239 | prefix=f"{prefix}.linear_attn", |
| 240 | ) |
| 241 | self.input_layernorm = RMSNorm( |
| 242 | config.hidden_size, |
| 243 | eps=config.rms_norm_eps, |
| 244 | ) |
| 245 | self.post_attention_layernorm = RMSNorm( |
| 246 | config.hidden_size, |
| 247 | eps=config.rms_norm_eps, |
| 248 | ) |
| 249 | else: |
| 250 | self.self_attn = OlmoHybridAttention( |
| 251 | vllm_config=vllm_config, |
| 252 | prefix=f"{prefix}.self_attn", |
| 253 | ) |
| 254 | # Attention layers use these norm names |
| 255 | self.post_attention_layernorm = RMSNorm( |
| 256 | config.hidden_size, |
| 257 | eps=config.rms_norm_eps, |
| 258 | ) |
| 259 | self.post_feedforward_layernorm = RMSNorm( |
| 260 | config.hidden_size, |
| 261 | eps=config.rms_norm_eps, |
| 262 | ) |
| 263 | |
| 264 | self.mlp = OlmoHybridMLP( |
| 265 | vllm_config=vllm_config, |
| 266 | prefix=f"{prefix}.mlp", |
| 267 | ) |
| 268 | |
| 269 | def forward( |
| 270 | self, |
| 271 | positions: torch.Tensor, |
| 272 | hidden_states: torch.Tensor, |
| 273 | ) -> torch.Tensor: |
| 274 | if self.layer_type == "linear_attention": |
| 275 | residual = hidden_states |
| 276 | hidden_states = self.input_layernorm(hidden_states) |
| 277 | |
| 278 | attn_output = torch.empty_like(hidden_states) |
| 279 | self.linear_attn( |
| 280 | hidden_states=hidden_states, |
| 281 | output=attn_output, |
| 282 | ) |
| 283 | hidden_states = residual + attn_output |
| 284 | |
| 285 | residual = hidden_states |
| 286 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 287 | hidden_states = self.mlp(hidden_states) |
| 288 | hidden_states = residual + hidden_states |
| 289 | else: |
| 290 | residual = hidden_states |
| 291 | hidden_states = self.self_attn(positions, hidden_states) |
| 292 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 293 | hidden_states = residual + hidden_states |
| 294 | |
| 295 | residual = hidden_states |
| 296 | hidden_states = self.mlp(hidden_states) |
| 297 | hidden_states = self.post_feedforward_layernorm(hidden_states) |
| 298 | hidden_states = residual + hidden_states |
| 299 | return hidden_states |
| 300 | |
| 301 | |
| 302 | @support_torch_compile |
| 303 | class OlmoHybridModel(nn.Module): |
| 304 | hf_to_vllm_mapper = WeightsMapper( |
| 305 | orig_to_new_stacked={ |
| 306 | ".self_attn.q_proj": (".self_attn.qkv_proj", "q"), |
| 307 | ".self_attn.k_proj": (".self_attn.qkv_proj", "k"), |
| 308 | ".self_attn.v_proj": (".self_attn.qkv_proj", "v"), |
| 309 | ".linear_attn.q_proj": (".linear_attn.in_proj_qkvg", 0), |
| 310 | ".linear_attn.k_proj": (".linear_attn.in_proj_qkvg", 1), |
| 311 | ".linear_attn.v_proj": (".linear_attn.in_proj_qkvg", 2), |
| 312 | ".linear_attn.g_proj": (".linear_attn.in_proj_qkvg", 3), |
| 313 | ".q_conv1d": (".conv1d", 0), |
| 314 | ".k_conv1d": (".conv1d", 1), |
| 315 | ".v_conv1d": (".conv1d", 2), |
| 316 | ".gate_proj": (".gate_up_proj", 0), |
| 317 | ".up_proj": (".gate_up_proj", 1), |
| 318 | } |
| 319 | ) |
| 320 | |
| 321 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 322 | super().__init__() |
| 323 | self.config = vllm_config.model_config.hf_config |
| 324 | |
| 325 | self.embed_tokens = VocabParallelEmbedding( |
| 326 | self.config.vocab_size, |
| 327 | self.config.hidden_size, |
| 328 | prefix=f"{prefix}.embed_tokens", |
| 329 | ) |
| 330 | |
| 331 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 332 | self.config.num_hidden_layers, |
| 333 | lambda prefix: OlmoHybridDecoderLayer( |
| 334 | vllm_config=vllm_config, prefix=prefix |
| 335 | ), |
| 336 | prefix=f"{prefix}.layers", |
| 337 | ) |
| 338 | |
| 339 | self.norm = RMSNorm( |
| 340 | self.config.hidden_size, |
| 341 | eps=self.config.rms_norm_eps, |
| 342 | ) |
| 343 | |
| 344 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 345 | ["hidden_states"], self.config.hidden_size |
| 346 | ) |
| 347 | |
| 348 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 349 | return self.embed_tokens(input_ids) |
| 350 | |
| 351 | def forward( |
| 352 | self, |
| 353 | input_ids: torch.Tensor, |
| 354 | positions: torch.Tensor, |
| 355 | intermediate_tensors: IntermediateTensors | None, |
| 356 | inputs_embeds: torch.Tensor | None = None, |
| 357 | ) -> torch.Tensor | IntermediateTensors: |
| 358 | if get_pp_group().is_first_rank: |
| 359 | if inputs_embeds is not None: |
| 360 | hidden_states = inputs_embeds |
| 361 | else: |
| 362 | hidden_states = self.embed_tokens(input_ids) |
| 363 | else: |
| 364 | assert intermediate_tensors is not None |
| 365 | hidden_states = intermediate_tensors["hidden_states"] |
| 366 | assert isinstance(hidden_states, torch.Tensor) |
| 367 | |
| 368 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 369 | hidden_states = layer(positions, hidden_states) |
| 370 | |
| 371 | if not get_pp_group().is_last_rank: |
| 372 | return IntermediateTensors({"hidden_states": hidden_states}) |
| 373 | |
| 374 | hidden_states = self.norm(hidden_states) |
| 375 | return hidden_states |
| 376 | |
| 377 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 378 | loader = AutoWeightsLoader(self) |
| 379 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 380 | |
| 381 | |
| 382 | class OlmoHybridForCausalLM( |
| 383 | nn.Module, HasInnerState, SupportsPP, SupportsLoRA, IsHybrid |
| 384 | ): |
| 385 | packed_modules_mapping = { |
| 386 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 387 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 388 | "in_proj_qkvg": ["q_proj", "k_proj", "v_proj", "g_proj"], |
| 389 | } |
| 390 | |
| 391 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 392 | super().__init__() |
| 393 | config = vllm_config.model_config.hf_config |
| 394 | self.config = config |
| 395 | self.vllm_config = vllm_config |
| 396 | self.model_config = vllm_config.model_config |
| 397 | |
| 398 | self.model = OlmoHybridModel( |
| 399 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 400 | ) |
| 401 | |
| 402 | self.lm_head = ParallelLMHead( |
| 403 | config.vocab_size, |
| 404 | config.hidden_size, |
| 405 | quant_config=vllm_config.quant_config, |
| 406 | prefix=maybe_prefix(prefix, "lm_head"), |
| 407 | ) |
| 408 | if config.tie_word_embeddings: |
| 409 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 410 | |
| 411 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 412 | self.make_empty_intermediate_tensors = ( |
| 413 | self.model.make_empty_intermediate_tensors |
| 414 | ) |
| 415 | |
| 416 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 417 | return self.model.embed_input_ids(input_ids) |
| 418 | |
| 419 | def forward( |
| 420 | self, |
| 421 | input_ids: torch.Tensor, |
| 422 | positions: torch.Tensor, |
| 423 | intermediate_tensors: IntermediateTensors | None = None, |
| 424 | inputs_embeds: torch.Tensor | None = None, |
| 425 | ) -> torch.Tensor | IntermediateTensors: |
| 426 | hidden_states = self.model( |
| 427 | input_ids=input_ids, |
| 428 | positions=positions, |
| 429 | intermediate_tensors=intermediate_tensors, |
| 430 | inputs_embeds=inputs_embeds, |
| 431 | ) |
| 432 | return hidden_states |
| 433 | |
| 434 | def compute_logits( |
| 435 | self, |
| 436 | hidden_states: torch.Tensor, |
| 437 | ) -> torch.Tensor | None: |
| 438 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 439 | return logits |
| 440 | |
| 441 | @classmethod |
| 442 | def get_mamba_state_dtype_from_config( |
| 443 | cls, |
| 444 | vllm_config: "VllmConfig", |
| 445 | ) -> tuple[torch.dtype, torch.dtype]: |
| 446 | return MambaStateDtypeCalculator.gated_delta_net_state_dtype( |
| 447 | vllm_config.model_config.dtype, |
| 448 | vllm_config.cache_config.mamba_cache_dtype, |
| 449 | vllm_config.cache_config.mamba_ssm_cache_dtype, |
| 450 | ) |
| 451 | |
| 452 | @classmethod |
| 453 | def get_mamba_state_shape_from_config( |
| 454 | cls, vllm_config: "VllmConfig" |
| 455 | ) -> tuple[tuple[int, int], tuple[int, int]]: |
| 456 | parallel_config = vllm_config.parallel_config |
| 457 | hf_config = vllm_config.model_config.hf_config |
| 458 | tp_size = parallel_config.tensor_parallel_size |
| 459 | num_spec = ( |
| 460 | vllm_config.speculative_config.num_speculative_tokens |
| 461 | if vllm_config.speculative_config |
| 462 | else 0 |
| 463 | ) |
| 464 | return MambaStateShapeCalculator.gated_delta_net_state_shape( |
| 465 | tp_size, |
| 466 | hf_config.linear_num_key_heads, |
| 467 | hf_config.linear_num_value_heads, |
| 468 | hf_config.linear_key_head_dim, |
| 469 | hf_config.linear_value_head_dim, |
| 470 | hf_config.linear_conv_kernel_dim, |
| 471 | num_spec, |
| 472 | ) |
| 473 | |
| 474 | @classmethod |
| 475 | def get_mamba_state_copy_func(cls) -> tuple[MambaStateCopyFunc, MambaStateCopyFunc]: |
| 476 | return MambaStateCopyFuncCalculator.gated_delta_net_state_copy_func() |
| 477 | |
| 478 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 479 | loader = AutoWeightsLoader(self) |
| 480 | return loader.load_weights(weights) |