| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Iterable |
| 5 | |
| 6 | import torch |
| 7 | import torch.nn as nn |
| 8 | from transformers import Qwen3Config |
| 9 | |
| 10 | from vllm.compilation.decorators import support_torch_compile |
| 11 | from vllm.config import VllmConfig, get_current_vllm_config |
| 12 | from vllm.logger import init_logger |
| 13 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 14 | from vllm.model_executor.layers.linear import QKVParallelLinear, ReplicatedLinear |
| 15 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 16 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 17 | ParallelLMHead, |
| 18 | VocabParallelEmbedding, |
| 19 | ) |
| 20 | from vllm.model_executor.models.qwen3 import Qwen3DecoderLayer, Qwen3ForCausalLM |
| 21 | from vllm.multimodal.inputs import NestedTensors |
| 22 | |
| 23 | from .utils import ( |
| 24 | AutoWeightsLoader, |
| 25 | WeightsMapper, |
| 26 | get_draft_quant_config, |
| 27 | maybe_prefix, |
| 28 | process_eagle_weight, |
| 29 | ) |
| 30 | |
| 31 | logger = init_logger(__name__) |
| 32 | |
| 33 | |
| 34 | class Qwen3Eagle3DecoderLayer(Qwen3DecoderLayer): |
| 35 | def __init__( |
| 36 | self, |
| 37 | vllm_config: VllmConfig, |
| 38 | prefix: str = "", |
| 39 | config: Qwen3Config | None = None, |
| 40 | layer_idx: int = 0, |
| 41 | ) -> None: |
| 42 | config = config or vllm_config.model_config.hf_config |
| 43 | cache_config = vllm_config.cache_config |
| 44 | quant_config = get_draft_quant_config(vllm_config) |
| 45 | |
| 46 | # Resolve per-layer sliding window from draft config |
| 47 | sliding_window = None |
| 48 | layer_types = getattr(config, "layer_types", None) |
| 49 | if ( |
| 50 | layer_types |
| 51 | and layer_idx < len(layer_types) |
| 52 | and layer_types[layer_idx] == "sliding_attention" |
| 53 | ): |
| 54 | sliding_window = getattr(config, "sliding_window", None) |
| 55 | |
| 56 | super().__init__( |
| 57 | config=config, |
| 58 | cache_config=cache_config, |
| 59 | quant_config=quant_config, |
| 60 | prefix=prefix, |
| 61 | per_layer_sliding_window=sliding_window, |
| 62 | ) |
| 63 | |
| 64 | # First layer uses 2*hidden_size (embeds + hidden_states concatenated) |
| 65 | # Subsequent layers use hidden_size (only hidden_states, no embeds) |
| 66 | qkv_input_size = 2 * self.hidden_size if layer_idx == 0 else self.hidden_size |
| 67 | |
| 68 | # Parallel drafting checkpoints may have attention bias enabled |
| 69 | qkv_bias = getattr(config, "attention_bias", False) |
| 70 | |
| 71 | # Override qkv_proj with correct input size and bias setting |
| 72 | self.self_attn.qkv_proj = QKVParallelLinear( |
| 73 | qkv_input_size, |
| 74 | self.self_attn.head_dim, |
| 75 | self.self_attn.total_num_heads, |
| 76 | self.self_attn.total_num_kv_heads, |
| 77 | bias=qkv_bias, |
| 78 | quant_config=quant_config, |
| 79 | prefix=maybe_prefix(prefix, "qkv_proj"), |
| 80 | ) |
| 81 | |
| 82 | self.hidden_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 83 | self.layer_idx = layer_idx |
| 84 | |
| 85 | if getattr(config, "norm_before_residual", False): |
| 86 | self._residual_norm = self._norm_before_residual |
| 87 | else: |
| 88 | self._residual_norm = self._norm_after_residual |
| 89 | |
| 90 | def _norm_before_residual( |
| 91 | self, hidden_states: torch.Tensor |
| 92 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 93 | hidden_states = self.hidden_norm(hidden_states) |
| 94 | residual = hidden_states |
| 95 | return hidden_states, residual |
| 96 | |
| 97 | def _norm_after_residual( |
| 98 | self, hidden_states: torch.Tensor |
| 99 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 100 | residual = hidden_states |
| 101 | hidden_states = self.hidden_norm(hidden_states) |
| 102 | return hidden_states, residual |
| 103 | |
| 104 | def forward( # type: ignore[override] |
| 105 | self, |
| 106 | positions: torch.Tensor, |
| 107 | embeds: torch.Tensor, |
| 108 | hidden_states: torch.Tensor, |
| 109 | residual: torch.Tensor | None, |
| 110 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 111 | if self.layer_idx == 0: |
| 112 | # First layer: concatenate embeds with hidden_states |
| 113 | embeds = self.input_layernorm(embeds) |
| 114 | hidden_states, residual = self._residual_norm(hidden_states=hidden_states) |
| 115 | hidden_states = torch.cat([embeds, hidden_states], dim=-1) |
| 116 | else: |
| 117 | # Subsequent layers: process hidden_states and residuals only |
| 118 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 119 | |
| 120 | # Self Attention |
| 121 | hidden_states = self.self_attn( |
| 122 | positions=positions, |
| 123 | hidden_states=hidden_states, |
| 124 | ) |
| 125 | |
| 126 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 127 | |
| 128 | # Fully Connected |
| 129 | hidden_states = self.mlp(hidden_states) |
| 130 | |
| 131 | return hidden_states, residual |
| 132 | |
| 133 | |
| 134 | @support_torch_compile( |
| 135 | dynamic_arg_dims={ |
| 136 | "input_ids": 0, |
| 137 | "positions": -1, |
| 138 | "hidden_states": 0, |
| 139 | "input_embeds": 0, |
| 140 | } |
| 141 | ) |
| 142 | class Qwen3Eagle3Model(nn.Module): |
| 143 | hf_to_vllm_mapper = WeightsMapper( |
| 144 | orig_to_new_substr={"midlayer.": "layers.0."}, |
| 145 | orig_to_new_stacked={ |
| 146 | ".q_proj": (".qkv_proj", "q"), |
| 147 | ".k_proj": (".qkv_proj", "k"), |
| 148 | ".v_proj": (".qkv_proj", "v"), |
| 149 | ".gate_proj": (".gate_up_proj", 0), |
| 150 | ".up_proj": (".gate_up_proj", 1), |
| 151 | }, |
| 152 | ) |
| 153 | |
| 154 | def __init__( |
| 155 | self, |
| 156 | *, |
| 157 | vllm_config: VllmConfig, |
| 158 | start_layer_id: int = 0, |
| 159 | prefix: str = "", |
| 160 | ) -> None: |
| 161 | super().__init__() |
| 162 | speculative_config = vllm_config.speculative_config |
| 163 | assert speculative_config is not None |
| 164 | self.config = speculative_config.draft_model_config.hf_config |
| 165 | self.vocab_size = self.config.vocab_size |
| 166 | |
| 167 | # Get drafter's quantization config |
| 168 | self.quant_config = get_draft_quant_config(vllm_config) |
| 169 | |
| 170 | eagle_config = getattr(self.config, "eagle_config", None) or {} |
| 171 | if "use_aux_hidden_state" in eagle_config: |
| 172 | self.use_aux_hidden_state = eagle_config["use_aux_hidden_state"] |
| 173 | else: |
| 174 | self.use_aux_hidden_state = True |
| 175 | self.norm_before_fc = bool( |
| 176 | eagle_config.get( |
| 177 | "norm_before_fc", getattr(self.config, "norm_before_fc", False) |
| 178 | ) |
| 179 | ) |
| 180 | self.fc_input_size = self.config.hidden_size |
| 181 | |
| 182 | current_vllm_config = get_current_vllm_config() |
| 183 | |
| 184 | self.embed_tokens = VocabParallelEmbedding( |
| 185 | self.config.vocab_size, |
| 186 | self.config.hidden_size, |
| 187 | prefix=maybe_prefix(prefix, "embed_tokens"), |
| 188 | ) |
| 189 | |
| 190 | self.layers = nn.ModuleList( |
| 191 | [ |
| 192 | Qwen3Eagle3DecoderLayer( |
| 193 | current_vllm_config, |
| 194 | prefix=maybe_prefix(prefix, f"layers.{layer_idx + start_layer_id}"), |
| 195 | config=self.config, |
| 196 | layer_idx=layer_idx, |
| 197 | ) |
| 198 | for layer_idx in range(self.config.num_hidden_layers) |
| 199 | ] |
| 200 | ) |
| 201 | if self.use_aux_hidden_state: |
| 202 | num_aux_features = getattr(self.config, "num_aux_layers", None) |
| 203 | if num_aux_features is None: |
| 204 | num_aux_features = getattr(self.config, "num_aux_hidden_states", None) |
| 205 | if num_aux_features is None: |
| 206 | aux_ids = getattr( |
| 207 | self.config, "eagle_aux_hidden_state_layer_ids", None |
| 208 | ) or eagle_config.get("eagle_aux_hidden_state_layer_ids") |
| 209 | num_aux_features = len(aux_ids) if aux_ids else 3 |
| 210 | self.num_aux_layers = num_aux_features |
| 211 | target_hidden_size = getattr( |
| 212 | self.config, "target_hidden_size", self.config.hidden_size |
| 213 | ) |
| 214 | self.fc_input_size = target_hidden_size * num_aux_features |
| 215 | self.input_norm: RMSNorm | None |
| 216 | if self.norm_before_fc: |
| 217 | self.input_norm = RMSNorm( |
| 218 | self.fc_input_size, |
| 219 | eps=self.config.rms_norm_eps, |
| 220 | ) |
| 221 | else: |
| 222 | self.input_norm = None |
| 223 | |
| 224 | use_fc_norm = getattr(self.config, "fc_norm", False) |
| 225 | if use_fc_norm: |
| 226 | self.fc_norm = nn.ModuleList( |
| 227 | [ |
| 228 | RMSNorm(target_hidden_size, eps=self.config.rms_norm_eps) |
| 229 | for _ in range(num_aux_features) |
| 230 | ] |
| 231 | ) |
| 232 | else: |
| 233 | self.fc_norm = None |
| 234 | |
| 235 | self.fc = ReplicatedLinear( |
| 236 | input_size=self.fc_input_size, |
| 237 | output_size=self.config.hidden_size, |
| 238 | bias=False, |
| 239 | params_dtype=vllm_config.model_config.dtype, |
| 240 | quant_config=self.quant_config, |
| 241 | prefix=maybe_prefix(prefix, "fc"), |
| 242 | return_bias=False, |
| 243 | ) |
| 244 | |
| 245 | self.norm_output = getattr(self.config, "norm_output", False) |
| 246 | self.norm = RMSNorm( |
| 247 | self.config.hidden_size, |
| 248 | eps=self.config.rms_norm_eps, |
| 249 | ) |
| 250 | |
| 251 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 252 | return self.embed_tokens(input_ids) |
| 253 | |
| 254 | def forward( |
| 255 | self, |
| 256 | input_ids: torch.Tensor, |
| 257 | positions: torch.Tensor, |
| 258 | hidden_states: torch.Tensor, |
| 259 | input_embeds: torch.Tensor | None = None, |
| 260 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 261 | if input_embeds is None: |
| 262 | input_embeds = self.embed_input_ids(input_ids) |
| 263 | assert hidden_states.shape[-1] == input_embeds.shape[-1] |
| 264 | |
| 265 | residual = None |
| 266 | for layer in self.layers: |
| 267 | hidden_states, residual = layer( |
| 268 | positions=positions, |
| 269 | embeds=input_embeds, |
| 270 | hidden_states=hidden_states, |
| 271 | residual=residual, |
| 272 | ) |
| 273 | hidden_states, hidden_prenorm = self.norm(hidden_states, residual) |
| 274 | |
| 275 | # norm_output variant uses the post-norm hidden states. |
| 276 | aux_output = hidden_states if self.norm_output else hidden_prenorm |
| 277 | |
| 278 | return hidden_states, aux_output |
| 279 | |
| 280 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 281 | loader = AutoWeightsLoader(self) |
| 282 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 283 | |
| 284 | |
| 285 | class Eagle3Qwen3ForCausalLM(Qwen3ForCausalLM): |
| 286 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 287 | nn.Module.__init__(self) |
| 288 | speculative_config = vllm_config.speculative_config |
| 289 | assert speculative_config is not None |
| 290 | self.config = speculative_config.draft_model_config.hf_config |
| 291 | # Ensure draft_vocab_size is set |
| 292 | # default to the base vocab size when absent |
| 293 | if getattr(self.config, "draft_vocab_size", None) is None: |
| 294 | base_vocab_size = getattr(self.config, "vocab_size", None) |
| 295 | self.config.draft_vocab_size = base_vocab_size |
| 296 | target_layer_num = vllm_config.model_config.get_total_num_hidden_layers() |
| 297 | |
| 298 | # Store target layer count in draft config for |
| 299 | # proper layer_types indexing in draft models |
| 300 | self.config.target_layer_count = target_layer_num |
| 301 | self.model = Qwen3Eagle3Model( |
| 302 | vllm_config=vllm_config, |
| 303 | prefix=maybe_prefix(prefix, "model"), |
| 304 | start_layer_id=target_layer_num, |
| 305 | ) |
| 306 | |
| 307 | logit_scale = getattr(self.config, "logit_scale", 1.0) |
| 308 | self.lm_head = ParallelLMHead( |
| 309 | self.config.draft_vocab_size, |
| 310 | self.config.hidden_size, |
| 311 | quant_config=get_draft_quant_config(vllm_config), |
| 312 | prefix=maybe_prefix(prefix, "lm_head"), |
| 313 | ) |
| 314 | self.logits_processor = LogitsProcessor( |
| 315 | self.config.draft_vocab_size, scale=logit_scale |
| 316 | ) |
| 317 | target_vocab_size = vllm_config.model_config.get_vocab_size() |
| 318 | if self.config.draft_vocab_size != target_vocab_size: |
| 319 | self.draft_id_to_target_id = nn.Parameter( |
| 320 | torch.zeros(self.config.draft_vocab_size, dtype=torch.long), |
| 321 | requires_grad=False, |
| 322 | ) |
| 323 | else: |
| 324 | self.draft_id_to_target_id = None |
| 325 | |
| 326 | self.use_parallel_drafting = speculative_config.parallel_drafting |
| 327 | |
| 328 | if self.use_parallel_drafting: |
| 329 | self.register_buffer( |
| 330 | "mask_hidden", |
| 331 | torch.zeros(1, self.model.fc_input_size), |
| 332 | persistent=False, |
| 333 | ) |
| 334 | |
| 335 | def embed_input_ids( |
| 336 | self, |
| 337 | input_ids: torch.Tensor, |
| 338 | multimodal_embeddings: NestedTensors | None = None, |
| 339 | is_multimodal: torch.Tensor | None = None, |
| 340 | ) -> torch.Tensor: |
| 341 | return self.model.embed_input_ids(input_ids) |
| 342 | |
| 343 | def forward( # type: ignore[override] |
| 344 | self, |
| 345 | input_ids: torch.Tensor, |
| 346 | positions: torch.Tensor, |
| 347 | hidden_states: torch.Tensor, |
| 348 | inputs_embeds: torch.Tensor | None = None, |
| 349 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 350 | return self.model(input_ids, positions, hidden_states, inputs_embeds) |
| 351 | |
| 352 | def compute_logits( |
| 353 | self, |
| 354 | hidden_states: torch.Tensor, |
| 355 | ) -> torch.Tensor | None: |
| 356 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 357 | if self.draft_id_to_target_id is None: |
| 358 | assert logits.shape[1] == self.config.vocab_size, ( |
| 359 | "Expected logits to have shape " |
| 360 | f"(*, {self.config.vocab_size}), but got {logits.shape}" |
| 361 | ) |
| 362 | return logits |
| 363 | |
| 364 | base = torch.arange(self.config.draft_vocab_size, device=logits.device) |
| 365 | targets = base + self.draft_id_to_target_id |
| 366 | logits_new = logits.new_full( |
| 367 | ( |
| 368 | logits.shape[0], |
| 369 | self.config.vocab_size, |
| 370 | ), |
| 371 | float("-inf"), |
| 372 | ) |
| 373 | logits_new[:, targets] = logits |
| 374 | return logits_new |
| 375 | |
| 376 | def combine_hidden_states( |
| 377 | self, |
| 378 | hidden_states: torch.Tensor, |
| 379 | ) -> torch.Tensor: |
| 380 | if not self.model.use_aux_hidden_state: |
| 381 | return hidden_states |
| 382 | # combine multiple auxiliary hidden states returned by eagle3 |
| 383 | |
| 384 | if self.model.norm_before_fc: |
| 385 | hidden_states = self.model.input_norm(hidden_states) |
| 386 | |
| 387 | # `norm_before_fc` adds a single RMSNorm before the FC layer, whereas `fc_norm` |
| 388 | # applies separate RMSNorms to each chunk of the hidden states. |
| 389 | if self.model.fc_norm is not None: |
| 390 | chunks = hidden_states.chunk(self.model.num_aux_layers, dim=-1) |
| 391 | hidden_states = torch.cat( |
| 392 | [norm(chunk) for norm, chunk in zip(self.model.fc_norm, chunks)], |
| 393 | dim=-1, |
| 394 | ) |
| 395 | |
| 396 | return self.model.fc(hidden_states) |
| 397 | |
| 398 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 399 | model_weights = {} |
| 400 | includes_draft_id_mapping = False |
| 401 | includes_embed_tokens = False |
| 402 | includes_mask_hidden = False |
| 403 | for name, loaded_weight in weights: |
| 404 | if "t2d" in name: |
| 405 | continue |
| 406 | if "d2t" in name: |
| 407 | name = name.replace("d2t", "draft_id_to_target_id") |
| 408 | includes_draft_id_mapping = True |
| 409 | elif "mask_hidden" in name: |
| 410 | # Load mask_hidden directly into buffer |
| 411 | if not self.use_parallel_drafting: |
| 412 | logger.warning( |
| 413 | "mask_hidden found in weights but " |
| 414 | "model is not configured for parallel drafting. " |
| 415 | "Skipping loading mask_hidden." |
| 416 | ) |
| 417 | continue |
| 418 | self.mask_hidden.copy_(loaded_weight.view(1, -1)) |
| 419 | includes_mask_hidden = True |
| 420 | continue |
| 421 | elif "lm_head" not in name: |
| 422 | name = "model." + name |
| 423 | if "embed_tokens" in name: |
| 424 | includes_embed_tokens = True |
| 425 | model_weights[name] = loaded_weight |
| 426 | process_eagle_weight(self, name) |
| 427 | |
| 428 | if not includes_mask_hidden and self.use_parallel_drafting: |
| 429 | raise ValueError( |
| 430 | "mask_hidden not found in weights but " |
| 431 | "model is configured for parallel drafting. " |
| 432 | "Please provide mask_hidden in the weights." |
| 433 | ) |
| 434 | |
| 435 | orig_to_new_substr = {"mask_hidden": None} |
| 436 | if not includes_draft_id_mapping: |
| 437 | orig_to_new_substr["draft_id_to_target_id"] = None |
| 438 | if not includes_embed_tokens: |
| 439 | orig_to_new_substr["embed_tokens"] = None |
| 440 | if not self.model.use_aux_hidden_state: |
| 441 | orig_to_new_substr["fc."] = None |
| 442 | if not self.model.norm_before_fc: |
| 443 | orig_to_new_substr["input_norm."] = None |
| 444 | mapper = WeightsMapper(orig_to_new_substr=orig_to_new_substr) |
| 445 | loader = AutoWeightsLoader(self) |
| 446 | loader.load_weights(model_weights.items(), mapper=mapper) |