| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import itertools |
| 5 | from collections.abc import Iterable |
| 6 | from contextlib import contextmanager |
| 7 | from typing import TYPE_CHECKING, Any, TypeVar |
| 8 | |
| 9 | import torch |
| 10 | import torch.nn as nn |
| 11 | |
| 12 | from vllm.config import VllmConfig |
| 13 | from vllm.logger import init_logger |
| 14 | from vllm.model_executor.layers.activation import get_act_fn |
| 15 | from vllm.model_executor.models.config import VerifyAndUpdateConfig |
| 16 | from vllm.transformers_utils.config import ( |
| 17 | try_get_dense_modules, |
| 18 | ) |
| 19 | from vllm.transformers_utils.repo_utils import get_hf_file_bytes |
| 20 | |
| 21 | from .interfaces import supports_multimodal |
| 22 | from .interfaces_base import VllmModelForPooling, is_pooling_model |
| 23 | |
| 24 | if TYPE_CHECKING: |
| 25 | from vllm.config import ModelConfig, VllmConfig |
| 26 | from vllm.model_executor.layers.pooler import Pooler |
| 27 | |
| 28 | _T = TypeVar("_T", bound=nn.Module) |
| 29 | |
| 30 | logger = init_logger(__name__) |
| 31 | |
| 32 | _GENERATE_SUFFIXES = [ |
| 33 | "ForCausalLM", |
| 34 | "ForConditionalGeneration", |
| 35 | "ChatModel", |
| 36 | "LMHeadModel", |
| 37 | ] |
| 38 | |
| 39 | |
| 40 | def _load_st_projector(model_config: "ModelConfig") -> nn.Module | None: |
| 41 | """Load Sentence-Transformers Dense projection layers.""" |
| 42 | dense_modules = try_get_dense_modules( |
| 43 | model_config.model, revision=model_config.revision |
| 44 | ) |
| 45 | |
| 46 | if dense_modules is None: |
| 47 | return None |
| 48 | |
| 49 | try: |
| 50 | layers = [] |
| 51 | for layer_config in dense_modules: |
| 52 | folder = layer_config["folder"] |
| 53 | linear = nn.Linear( |
| 54 | layer_config["in_features"], |
| 55 | layer_config["out_features"], |
| 56 | bias=layer_config.get("bias", True), |
| 57 | dtype=model_config.head_dtype, |
| 58 | ) |
| 59 | if not _load_dense_weights(linear, folder, model_config): |
| 60 | continue |
| 61 | layers.append(linear) |
| 62 | if act_name := layer_config.get("activation_function"): |
| 63 | layers.append(get_act_fn(act_name)) |
| 64 | return nn.Sequential(*layers).to(dtype=model_config.head_dtype) |
| 65 | except Exception: |
| 66 | logger.exception("ST projector loading failed") |
| 67 | |
| 68 | return None |
| 69 | |
| 70 | |
| 71 | def _load_dense_weights( |
| 72 | linear: nn.Linear, folder: str, model_config: "ModelConfig" |
| 73 | ) -> bool: |
| 74 | """Load weights using vLLM's weight_loader pattern.""" |
| 75 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 76 | |
| 77 | for filename in ["model.safetensors", "pytorch_model.bin"]: |
| 78 | file_path = f"{folder}/{filename}" if folder else filename |
| 79 | |
| 80 | try: |
| 81 | file_bytes = get_hf_file_bytes( |
| 82 | file_path, model_config.model, model_config.revision |
| 83 | ) |
| 84 | if not file_bytes: |
| 85 | continue |
| 86 | |
| 87 | if filename.endswith(".safetensors"): |
| 88 | from safetensors.torch import load as load_safetensors |
| 89 | |
| 90 | state_dict = load_safetensors(file_bytes) |
| 91 | else: |
| 92 | import io |
| 93 | |
| 94 | state_dict = torch.load( |
| 95 | io.BytesIO(file_bytes), map_location="cpu", weights_only=True |
| 96 | ) |
| 97 | |
| 98 | for weight_key in ["weight", "linear.weight", "dense.weight"]: |
| 99 | if weight_key in state_dict: |
| 100 | weight_loader = getattr( |
| 101 | linear.weight, "weight_loader", default_weight_loader |
| 102 | ) |
| 103 | weight_loader(linear.weight, state_dict[weight_key]) |
| 104 | |
| 105 | bias_key = weight_key.replace("weight", "bias") |
| 106 | if linear.bias is not None and bias_key in state_dict: |
| 107 | bias_loader = getattr( |
| 108 | linear.bias, "weight_loader", default_weight_loader |
| 109 | ) |
| 110 | bias_loader(linear.bias, state_dict[bias_key]) |
| 111 | return True |
| 112 | except Exception: |
| 113 | logger.exception("Failed to load %s", filename) |
| 114 | continue |
| 115 | |
| 116 | return False |
| 117 | |
| 118 | |
| 119 | def _get_pooling_model_name(orig_model_name: str, pooling_suffix: str) -> str: |
| 120 | model_name = orig_model_name |
| 121 | |
| 122 | for generate_suffix in _GENERATE_SUFFIXES: |
| 123 | model_name = model_name.removesuffix(generate_suffix) |
| 124 | |
| 125 | return model_name + pooling_suffix |
| 126 | |
| 127 | |
| 128 | def _create_pooling_model_cls(orig_cls: type[_T]) -> type[_T]: |
| 129 | # Lazy import |
| 130 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 131 | from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead |
| 132 | |
| 133 | from .utils import AutoWeightsLoader, StageMissingLayer, no_init_weights |
| 134 | |
| 135 | class ModelForPooling(orig_cls, VllmModelForPooling): # type: ignore[valid-type,misc] |
| 136 | is_pooling_model = True |
| 137 | |
| 138 | def __init__( |
| 139 | self, |
| 140 | *, |
| 141 | vllm_config: "VllmConfig", |
| 142 | prefix: str = "", |
| 143 | **kwargs: Any, |
| 144 | ) -> None: |
| 145 | with no_init_weights( |
| 146 | self, |
| 147 | lambda mod: StageMissingLayer("output", mod), |
| 148 | targets=(LogitsProcessor, ParallelLMHead), |
| 149 | ): |
| 150 | super().__init__( # type: ignore[safe-super] |
| 151 | vllm_config=vllm_config, prefix=prefix, **kwargs |
| 152 | ) |
| 153 | |
| 154 | # Used by SEQ_CLS_LOAD_METHODS |
| 155 | self.vllm_config = vllm_config |
| 156 | |
| 157 | # If the model already defines a pooler instance, don't overwrite it |
| 158 | pooler = getattr(self, "pooler", None) |
| 159 | multimodal_model: object = self |
| 160 | if not pooler and supports_multimodal(multimodal_model): |
| 161 | # Try to get the pooler from the LM backbone |
| 162 | language_model = multimodal_model.get_language_model() |
| 163 | if hasattr(language_model, "pooler"): |
| 164 | pooler = language_model.pooler |
| 165 | |
| 166 | if not pooler: |
| 167 | pooler = self._init_pooler(vllm_config, prefix=prefix) |
| 168 | |
| 169 | self.pooler = pooler |
| 170 | |
| 171 | def _init_pooler( |
| 172 | self, |
| 173 | vllm_config: "VllmConfig", |
| 174 | prefix: str = "", |
| 175 | ) -> "Pooler": |
| 176 | raise NotImplementedError |
| 177 | |
| 178 | def _load_pooling_model_weights( |
| 179 | self, weights: Iterable[tuple[str, torch.Tensor]] |
| 180 | ): |
| 181 | params_dict = dict(self.named_parameters()) |
| 182 | # Use model-owned remapping only for the probe membership check |
| 183 | # (e.g. Gemma4 model.language_model.* -> language_model.model.*) |
| 184 | # so the "" / "model." scan can early-exit. Forwarded names stay |
| 185 | # unmapped to avoid double-applying the mapper in the parent loader. |
| 186 | hf_to_vllm_mapper = getattr(self, "hf_to_vllm_mapper", None) |
| 187 | |
| 188 | # We support loading from both `*ForCausalLM` and `*Model` |
| 189 | candidate_prefixes = ["", "model."] |
| 190 | target_prefix = "" |
| 191 | |
| 192 | seen_weights = list[tuple[str, torch.Tensor]]() |
| 193 | for name, loaded_weight in weights: |
| 194 | # Clone because the iterator may reuse the tensor buffer |
| 195 | seen_weights.append((name, loaded_weight.clone())) |
| 196 | |
| 197 | lookup_name = name |
| 198 | if hf_to_vllm_mapper is not None: |
| 199 | mapped = hf_to_vllm_mapper._map_name(name) |
| 200 | if mapped is None: |
| 201 | continue |
| 202 | lookup_name = mapped |
| 203 | |
| 204 | try: |
| 205 | target_prefix = next( |
| 206 | prefix |
| 207 | for prefix in candidate_prefixes |
| 208 | if prefix + lookup_name in params_dict |
| 209 | ) |
| 210 | break |
| 211 | except StopIteration: |
| 212 | # The weight might not exist on the model |
| 213 | # (to be handled by AutoWeightsLoader) |
| 214 | pass |
| 215 | |
| 216 | if target_prefix: |
| 217 | target_model = self |
| 218 | for attr in target_prefix.split("."): |
| 219 | if attr: |
| 220 | target_model = getattr(self, attr) |
| 221 | |
| 222 | logger.info( |
| 223 | "Mapping weights to %s as they are " |
| 224 | "relative to this model instead of %s.", |
| 225 | target_model._get_name(), |
| 226 | self._get_name(), |
| 227 | ) |
| 228 | |
| 229 | # Lazy chain so buffer-reusing weight iterators (e.g. |
| 230 | # runai_streamer) are consumed one tensor at a time. |
| 231 | mapped_weights = ( |
| 232 | (target_prefix + name, weight) |
| 233 | for name, weight in itertools.chain(seen_weights, weights) |
| 234 | ) |
| 235 | |
| 236 | def default_load_weights(weights): |
| 237 | loader = AutoWeightsLoader(self) |
| 238 | mapper = getattr(self, "hf_to_vllm_mapper", None) |
| 239 | return loader.load_weights(weights, mapper=mapper) |
| 240 | |
| 241 | load_weights = getattr(super(), "load_weights", default_load_weights) |
| 242 | return load_weights(mapped_weights) |
| 243 | |
| 244 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 245 | return self._load_pooling_model_weights(weights) |
| 246 | |
| 247 | return ModelForPooling |
| 248 | |
| 249 | |
| 250 | def as_embedding_model(cls: type[_T]) -> type[_T]: |
| 251 | """Subclass an existing vLLM model to support embeddings. |
| 252 | |
| 253 | By default, the embeddings of the whole prompt are extracted from the |
| 254 | normalized hidden state corresponding to the last token. |
| 255 | |
| 256 | Note: |
| 257 | We assume that no extra layers are added to the original model; |
| 258 | please implement your own model if this is not the case. |
| 259 | |
| 260 | """ |
| 261 | # Avoid modifying existing embedding models |
| 262 | if is_pooling_model(cls): |
| 263 | return cls |
| 264 | |
| 265 | # Lazy import |
| 266 | from vllm.model_executor.layers.pooler import DispatchPooler |
| 267 | |
| 268 | class ModelForEmbedding(_create_pooling_model_cls(cls)): # type: ignore[misc] |
| 269 | def _init_pooler( |
| 270 | self, |
| 271 | vllm_config: "VllmConfig", |
| 272 | prefix: str = "", |
| 273 | ) -> "Pooler": |
| 274 | pooler_config = vllm_config.model_config.pooler_config |
| 275 | assert pooler_config is not None |
| 276 | |
| 277 | return DispatchPooler.for_embedding(pooler_config) |
| 278 | |
| 279 | ModelForEmbedding.__name__ = _get_pooling_model_name(cls.__name__, "ForEmbedding") |
| 280 | |
| 281 | return ModelForEmbedding |
| 282 | |
| 283 | |
| 284 | def _resolve_num_labels(hf_config: Any, text_config: Any) -> int: |
| 285 | """Resolve the label count for a sequence classification head. |
| 286 | |
| 287 | ``PreTrainedConfig.num_labels`` is derived from ``id2label``, which always |
| 288 | carries a default of two entries. Composite configs (such as multimodal |
| 289 | checkpoints) declare their label space on the top-level config, so reading |
| 290 | ``num_labels`` from ``get_text_config()`` silently returns that default and |
| 291 | builds a score head of the wrong size. |
| 292 | |
| 293 | Prefer the top-level config whenever it declares a label space of its own, |
| 294 | mirroring the ``classifier_from_token`` / ``method`` lookups in |
| 295 | ``as_seq_cls_model``. |
| 296 | """ |
| 297 | if text_config is hf_config: |
| 298 | return hf_config.num_labels |
| 299 | |
| 300 | from transformers import PreTrainedConfig |
| 301 | |
| 302 | if hf_config.num_labels != PreTrainedConfig().num_labels: |
| 303 | return hf_config.num_labels |
| 304 | return text_config.num_labels |
| 305 | |
| 306 | |
| 307 | def as_seq_cls_model(cls: type[_T]) -> type[_T]: |
| 308 | """Subclass an existing vLLM model to support classify and score tasks. |
| 309 | |
| 310 | By default, the class probabilities are extracted from the softmaxed |
| 311 | hidden state corresponding to the last token. |
| 312 | |
| 313 | Note: |
| 314 | We assume that the classification head is a single linear layer |
| 315 | stored as the attribute `score` of the top-level model; |
| 316 | please implement your own model if this is not the case. |
| 317 | |
| 318 | """ |
| 319 | # Avoid modifying existing classification models |
| 320 | if is_pooling_model(cls): |
| 321 | return cls |
| 322 | |
| 323 | # Lazy import |
| 324 | from vllm.model_executor.layers.linear import ReplicatedLinear |
| 325 | from vllm.model_executor.layers.pooler import DispatchPooler |
| 326 | from vllm.model_executor.models.interfaces import SupportsCrossEncoding |
| 327 | |
| 328 | from .utils import maybe_prefix |
| 329 | |
| 330 | class ModelForSequenceClassification( |
| 331 | _create_pooling_model_cls(cls), # type: ignore[misc] |
| 332 | SupportsCrossEncoding, |
| 333 | ): |
| 334 | def _init_pooler( |
| 335 | self, |
| 336 | vllm_config: "VllmConfig", |
| 337 | prefix: str = "", |
| 338 | ) -> "Pooler": |
| 339 | hf_config = vllm_config.model_config.hf_config |
| 340 | text_config = hf_config.get_text_config() |
| 341 | model_config = vllm_config.model_config |
| 342 | |
| 343 | # Check if score weights are derived online from LM head |
| 344 | # (same condition as load_weights branch) |
| 345 | tokens = getattr( |
| 346 | hf_config, |
| 347 | "classifier_from_token", |
| 348 | getattr(text_config, "classifier_from_token", None), |
| 349 | ) |
| 350 | method = getattr( |
| 351 | hf_config, |
| 352 | "method", |
| 353 | getattr(text_config, "method", None), |
| 354 | ) |
| 355 | |
| 356 | # Online conversion: no score weights in checkpoint, don't |
| 357 | # quantize (small output_dim breaks FP8/Marlin tile alignment). |
| 358 | # Checkpoint-based: respect the model's quant_config. |
| 359 | quant_config = ( |
| 360 | None |
| 361 | if (tokens is not None or method is not None) |
| 362 | else vllm_config.quant_config |
| 363 | ) |
| 364 | |
| 365 | self.score = ReplicatedLinear( |
| 366 | model_config.get_hidden_size(), |
| 367 | _resolve_num_labels(hf_config, text_config), |
| 368 | bias=False, |
| 369 | params_dtype=model_config.head_dtype, |
| 370 | quant_config=quant_config, |
| 371 | return_bias=False, |
| 372 | prefix=maybe_prefix(prefix, "score"), |
| 373 | ) |
| 374 | |
| 375 | pooler_config = vllm_config.model_config.pooler_config |
| 376 | assert pooler_config is not None |
| 377 | |
| 378 | return DispatchPooler.for_seq_cls(pooler_config, classifier=self.score) |
| 379 | |
| 380 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 381 | hf_config = self.config |
| 382 | text_config = hf_config.get_text_config() |
| 383 | tokens = getattr( |
| 384 | hf_config, |
| 385 | "classifier_from_token", |
| 386 | getattr(text_config, "classifier_from_token", None), |
| 387 | ) |
| 388 | method = getattr(hf_config, "method", getattr(text_config, "method", None)) |
| 389 | |
| 390 | def auto_set_score_bias(weights): |
| 391 | for name, weight in weights: |
| 392 | if name == "score.bias": |
| 393 | device = self.score.weight.device |
| 394 | dtype = self.score.weight.dtype |
| 395 | bias = weight.to(device).to(dtype) |
| 396 | self.score.bias = torch.nn.Parameter(bias) |
| 397 | self.score.skip_bias_add = False |
| 398 | else: |
| 399 | yield name, weight |
| 400 | |
| 401 | weights = auto_set_score_bias(weights) |
| 402 | if tokens is None and method is None: |
| 403 | return super().load_weights(weights) |
| 404 | else: |
| 405 | # Online convert ForCausalLM into |
| 406 | # ForSequenceClassification model. |
| 407 | return seq_cls_model_loader(self, weights) |
| 408 | |
| 409 | ModelForSequenceClassification.__name__ = _get_pooling_model_name( |
| 410 | cls.__name__, "ForSequenceClassification" |
| 411 | ) |
| 412 | |
| 413 | return ModelForSequenceClassification |
| 414 | |
| 415 | |
| 416 | class SequenceClassificationConfig(VerifyAndUpdateConfig): |
| 417 | @staticmethod |
| 418 | def verify_and_update_config(vllm_config: "VllmConfig") -> None: |
| 419 | hf_config = vllm_config.model_config.hf_config |
| 420 | text_config = hf_config.get_text_config() |
| 421 | method = getattr(hf_config, "method", getattr(text_config, "method", None)) |
| 422 | tokens = getattr( |
| 423 | hf_config, |
| 424 | "classifier_from_token", |
| 425 | getattr(text_config, "classifier_from_token", None), |
| 426 | ) |
| 427 | |
| 428 | if method is None: |
| 429 | return |
| 430 | |
| 431 | assert tokens is not None |
| 432 | assert method in SEQ_CLS_LOAD_METHODS, f"method {method} not supported" |
| 433 | |
| 434 | if method == "from_2_way_softmax": |
| 435 | assert len(tokens) == 2 |
| 436 | hf_config.num_labels = 1 |
| 437 | text_config.num_labels = 1 |
| 438 | else: |
| 439 | hf_config.num_labels = len(tokens) |
| 440 | text_config.num_labels = len(tokens) |
| 441 | |
| 442 | # `llm as reranker` defaults to not using separating token. |
| 443 | use_sep_token = getattr(text_config, "use_sep_token", False) |
| 444 | text_config.use_sep_token = use_sep_token |
| 445 | |
| 446 | |
| 447 | def _get_language_model_for_seq_cls(model: nn.Module) -> nn.Module: |
| 448 | """Get the language model component for sequence classification conversion. |
| 449 | For VLMs, returns the inner language model. For standard LLMs, returns model itself. |
| 450 | """ |
| 451 | multimodal_model: object = model |
| 452 | if supports_multimodal(multimodal_model): |
| 453 | try: |
| 454 | lm = multimodal_model.get_language_model() |
| 455 | if lm is not model: |
| 456 | return lm |
| 457 | except Exception: |
| 458 | pass |
| 459 | |
| 460 | for attr_name in ("language_model", "lm", "text_model"): |
| 461 | if hasattr(model, attr_name): |
| 462 | candidate = getattr(model, attr_name) |
| 463 | if ( |
| 464 | isinstance(candidate, nn.Module) |
| 465 | and candidate is not model |
| 466 | and hasattr(candidate, "model") |
| 467 | ): |
| 468 | return candidate |
| 469 | |
| 470 | for name, child in model.named_children(): |
| 471 | child_name = type(child).__name__ |
| 472 | if ("ForCausalLM" in child_name or "LMHead" in child_name) and hasattr( |
| 473 | child, "model" |
| 474 | ): |
| 475 | return child |
| 476 | |
| 477 | return model |
| 478 | |
| 479 | |
| 480 | @contextmanager |
| 481 | def _disable_seq_cls_loading_on_inner_model(language_model, is_vlm: bool): |
| 482 | """Context manager to temporarily disable sequence classification loading |
| 483 | on inner VLM models to prevent recursive seq_cls_model_loader calls. |
| 484 | """ |
| 485 | if not is_vlm: |
| 486 | yield |
| 487 | return |
| 488 | |
| 489 | inner_hf_config = getattr(language_model, "config", None) |
| 490 | if inner_hf_config is None: |
| 491 | yield |
| 492 | return |
| 493 | |
| 494 | inner_text_config = inner_hf_config.get_text_config() |
| 495 | original_method = getattr(inner_text_config, "method", None) |
| 496 | original_tokens = getattr(inner_text_config, "classifier_from_token", None) |
| 497 | original_hf_tokens = getattr(inner_hf_config, "classifier_from_token", None) |
| 498 | |
| 499 | try: |
| 500 | if original_method is not None: |
| 501 | inner_text_config.method = None |
| 502 | if original_tokens is not None: |
| 503 | inner_text_config.classifier_from_token = None |
| 504 | if original_hf_tokens is not None: |
| 505 | inner_hf_config.classifier_from_token = None |
| 506 | yield |
| 507 | finally: |
| 508 | if original_method is not None: |
| 509 | inner_text_config.method = original_method |
| 510 | if original_tokens is not None: |
| 511 | inner_text_config.classifier_from_token = original_tokens |
| 512 | if original_hf_tokens is not None: |
| 513 | inner_hf_config.classifier_from_token = original_hf_tokens |
| 514 | |
| 515 | |
| 516 | def load_weights_using_from_2_way_softmax( |
| 517 | model, weights: Iterable[tuple[str, torch.Tensor]] |
| 518 | ): |
| 519 | # refer to https://huggingface.co/Qwen/Qwen3-Reranker-0.6B/discussions/3 |
| 520 | from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead |
| 521 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 522 | |
| 523 | model_config = model.vllm_config.model_config |
| 524 | hf_config = model.config |
| 525 | text_config = hf_config.get_text_config() |
| 526 | |
| 527 | tokens: list[str] = getattr( |
| 528 | hf_config, |
| 529 | "classifier_from_token", |
| 530 | getattr(text_config, "classifier_from_token", []), |
| 531 | ) |
| 532 | assert len(tokens) == 2 |
| 533 | |
| 534 | language_model = _get_language_model_for_seq_cls(model) |
| 535 | is_vlm = language_model is not model |
| 536 | using_vlm_head = is_vlm and hasattr(language_model, "score") |
| 537 | |
| 538 | language_model.lm_head = ParallelLMHead( |
| 539 | text_config.vocab_size, |
| 540 | text_config.hidden_size, |
| 541 | ) |
| 542 | if text_config.tie_word_embeddings: |
| 543 | # embed_tokens is the assumed name for input embeddings. If the model does not |
| 544 | # have this attribute, we fall back to get_input_embeddings(), which is used by |
| 545 | # the Transformers modeling backend. |
| 546 | text_backbone = language_model.model |
| 547 | embed_tokens = ( |
| 548 | text_backbone.embed_tokens |
| 549 | if hasattr(text_backbone, "embed_tokens") |
| 550 | else text_backbone.get_input_embeddings() |
| 551 | ) |
| 552 | language_model.lm_head = language_model.lm_head.tie_weights(embed_tokens) |
| 553 | |
| 554 | with _disable_seq_cls_loading_on_inner_model(language_model, is_vlm): |
| 555 | loaded_weights = model._load_pooling_model_weights(weights) |
| 556 | |
| 557 | from vllm.tokenizers import get_tokenizer |
| 558 | |
| 559 | tokenizer = get_tokenizer( |
| 560 | model_config.tokenizer, |
| 561 | revision=model_config.tokenizer_revision, |
| 562 | tokenizer_mode=model_config.tokenizer_mode, |
| 563 | trust_remote_code=model_config.trust_remote_code, |
| 564 | ) |
| 565 | |
| 566 | false_id = tokenizer.convert_tokens_to_ids(tokens[0]) |
| 567 | true_id = tokenizer.convert_tokens_to_ids(tokens[1]) |
| 568 | lm_head_weight = language_model.lm_head.weight |
| 569 | score_weight = lm_head_weight.data[[true_id]].to( |
| 570 | torch.float32 |
| 571 | ) - lm_head_weight.data[[false_id]].to(torch.float32) |
| 572 | |
| 573 | score_layer = language_model.score if using_vlm_head else model.score |
| 574 | param = score_layer.weight |
| 575 | weight_loader = getattr(param, "weight_loader", default_weight_loader) |
| 576 | weight_loader(param, score_weight) |
| 577 | |
| 578 | del language_model.lm_head |
| 579 | |
| 580 | score_weight_name = ( |
| 581 | "language_model.score.weight" if using_vlm_head else "score.weight" |
| 582 | ) |
| 583 | loaded_weights.add(score_weight_name) |
| 584 | |
| 585 | lm_head_name: str | None = "lm_head.weight" |
| 586 | if hf_to_vllm_mapper := getattr(model, "hf_to_vllm_mapper", None): |
| 587 | lm_head_name = hf_to_vllm_mapper._map_name(lm_head_name) |
| 588 | if lm_head_name is not None: |
| 589 | loaded_weights.discard(lm_head_name) |
| 590 | return loaded_weights |
| 591 | |
| 592 | |
| 593 | def load_weights_no_post_processing(model, weights: Iterable[tuple[str, torch.Tensor]]): |
| 594 | from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead |
| 595 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 596 | |
| 597 | model_config = model.vllm_config.model_config |
| 598 | text_config = model.config.get_text_config() |
| 599 | |
| 600 | tokens: list[str] = getattr(text_config, "classifier_from_token", []) |
| 601 | assert len(tokens) > 0 |
| 602 | |
| 603 | language_model = _get_language_model_for_seq_cls(model) |
| 604 | is_vlm = language_model is not model |
| 605 | using_vlm_head = is_vlm and hasattr(language_model, "score") |
| 606 | |
| 607 | language_model.lm_head = ParallelLMHead( |
| 608 | text_config.vocab_size, |
| 609 | text_config.hidden_size, |
| 610 | ) |
| 611 | if text_config.tie_word_embeddings: |
| 612 | # embed_tokens is the assumed name for input embeddings. If the model does not |
| 613 | # have this attribute, we fall back to get_input_embeddings(), which is used by |
| 614 | # the Transformers modeling backend. |
| 615 | text_backbone = language_model.model |
| 616 | embed_tokens = ( |
| 617 | text_backbone.embed_tokens |
| 618 | if hasattr(text_backbone, "embed_tokens") |
| 619 | else text_backbone.get_input_embeddings() |
| 620 | ) |
| 621 | language_model.lm_head = language_model.lm_head.tie_weights(embed_tokens) |
| 622 | |
| 623 | with _disable_seq_cls_loading_on_inner_model(language_model, is_vlm): |
| 624 | # Bypass ModelForSequenceClassification to avoid infinite recursion |
| 625 | loaded_weights = model._load_pooling_model_weights(weights) |
| 626 | |
| 627 | from vllm.tokenizers import get_tokenizer |
| 628 | |
| 629 | tokenizer = get_tokenizer( |
| 630 | model_config.tokenizer, |
| 631 | revision=model_config.tokenizer_revision, |
| 632 | tokenizer_mode=model_config.tokenizer_mode, |
| 633 | trust_remote_code=model_config.trust_remote_code, |
| 634 | ) |
| 635 | |
| 636 | token_ids = [tokenizer.convert_tokens_to_ids(t) for t in tokens] |
| 637 | score_weight = language_model.lm_head.weight.data[token_ids] |
| 638 | |
| 639 | score_layer = language_model.score if using_vlm_head else model.score |
| 640 | param = score_layer.weight |
| 641 | weight_loader = getattr(param, "weight_loader", default_weight_loader) |
| 642 | weight_loader(param, score_weight) |
| 643 | |
| 644 | del language_model.lm_head |
| 645 | |
| 646 | score_weight_name = ( |
| 647 | "language_model.score.weight" if using_vlm_head else "score.weight" |
| 648 | ) |
| 649 | loaded_weights.add(score_weight_name) |
| 650 | |
| 651 | lm_head_name: str | None = "lm_head.weight" |
| 652 | if hf_to_vllm_mapper := getattr(model, "hf_to_vllm_mapper", None): |
| 653 | lm_head_name = hf_to_vllm_mapper._map_name(lm_head_name) |
| 654 | if lm_head_name is not None: |
| 655 | loaded_weights.discard(lm_head_name) |
| 656 | return loaded_weights |
| 657 | |
| 658 | |
| 659 | SEQ_CLS_LOAD_METHODS = { |
| 660 | "from_2_way_softmax": load_weights_using_from_2_way_softmax, |
| 661 | "no_post_processing": load_weights_no_post_processing, |
| 662 | } |
| 663 | |
| 664 | |
| 665 | def seq_cls_model_loader(model, weights: Iterable[tuple[str, torch.Tensor]]): |
| 666 | # Online convert ForCausalLM into ForSequenceClassification model. |
| 667 | # - from_2_way_softmax: |
| 668 | # - Qwen3ForCausalLM |
| 669 | # - Qwen3-Reranker |
| 670 | # - Qwen2ForCausalLM |
| 671 | # - mxbai-rerank-v2 |
| 672 | # - no_post_processing: |
| 673 | # - GemmaForCausalLM |
| 674 | # - bge-reranker-v2-gemma |
| 675 | |
| 676 | hf_config = model.vllm_config.model_config.hf_config |
| 677 | text_config = hf_config.get_text_config() |
| 678 | method = getattr(hf_config, "method", getattr(text_config, "method", None)) |
| 679 | assert method in SEQ_CLS_LOAD_METHODS, f"method {method} not supported" |
| 680 | return SEQ_CLS_LOAD_METHODS[method](model, weights) |