| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Copyright 2026 The Qwen team. |
| 4 | # Copyright 2023 The vLLM team. |
| 5 | # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. |
| 6 | # |
| 7 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 8 | # and OPT implementations in this library. It has been modified from its |
| 9 | # original forms to accommodate minor architectural differences compared |
| 10 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 11 | # |
| 12 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 13 | # you may not use this file except in compliance with the License. |
| 14 | # You may obtain a copy of the License at |
| 15 | # |
| 16 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 17 | # |
| 18 | # Unless required by applicable law or agreed to in writing, software |
| 19 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 20 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 21 | # See the License for the specific language governing permissions and |
| 22 | # limitations under the License. |
| 23 | """Inference-only Qwen3-ASR model.""" |
| 24 | |
| 25 | from collections.abc import Iterable, Mapping, Sequence |
| 26 | from typing import Any |
| 27 | |
| 28 | import regex as re |
| 29 | import torch |
| 30 | import torch.nn as nn |
| 31 | from transformers.feature_extraction_utils import BatchFeature |
| 32 | from transformers.models.whisper import WhisperFeatureExtractor |
| 33 | |
| 34 | from vllm.config import ModelConfig, SpeechToTextConfig, VllmConfig |
| 35 | from vllm.config.multimodal import MultiModalDummyOptions |
| 36 | from vllm.config.speech_to_text import SpeechToTextParams |
| 37 | from vllm.inputs import ModalityData, MultiModalDataDict, PromptType, TokensPrompt |
| 38 | from vllm.logger import init_logger |
| 39 | from vllm.model_executor.models.interfaces import ( |
| 40 | MultiModalEmbeddings, |
| 41 | StreamingTranscriptionPostProcessor, |
| 42 | SupportsLoRA, |
| 43 | SupportsMRoPE, |
| 44 | SupportsMultiModal, |
| 45 | SupportsPP, |
| 46 | SupportsTranscription, |
| 47 | ) |
| 48 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 49 | from vllm.model_executor.models.qwen2_5_omni_thinker import ( |
| 50 | Qwen2_5OmniAudioFeatureInputs, |
| 51 | unpad_and_flat_audio_features, |
| 52 | ) |
| 53 | from vllm.model_executor.models.qwen3 import Qwen3ForCausalLM |
| 54 | from vllm.model_executor.models.qwen3_omni_moe_thinker import ( |
| 55 | Qwen3OmniMoeAudioEncoder, |
| 56 | Qwen3OmniMoeThinkerMultiModalProcessor, |
| 57 | ) |
| 58 | from vllm.model_executor.models.utils import ( |
| 59 | AutoWeightsLoader, |
| 60 | WeightsMapper, |
| 61 | _merge_multimodal_embeddings, |
| 62 | maybe_prefix, |
| 63 | ) |
| 64 | from vllm.model_executor.models.whisper import ISO639_1_SUPPORTED_LANGS |
| 65 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 66 | from vllm.multimodal.inputs import ( |
| 67 | AudioItem, |
| 68 | MultiModalFeatureSpec, |
| 69 | MultiModalFieldConfig, |
| 70 | MultiModalKwargsItem, |
| 71 | MultiModalKwargsItems, |
| 72 | ) |
| 73 | from vllm.multimodal.parse import ( |
| 74 | AudioProcessorItems, |
| 75 | DictEmbeddingItems, |
| 76 | ModalityDataItems, |
| 77 | MultiModalDataItems, |
| 78 | MultiModalDataParser, |
| 79 | ) |
| 80 | from vllm.multimodal.processing import ( |
| 81 | BaseDummyInputsBuilder, |
| 82 | BaseProcessingInfo, |
| 83 | PromptReplacement, |
| 84 | PromptUpdate, |
| 85 | cached_encode, |
| 86 | ) |
| 87 | from vllm.sequence import IntermediateTensors |
| 88 | from vllm.tokenizers import cached_tokenizer_from_config |
| 89 | from vllm.transformers_utils.configs.qwen3_asr import ( |
| 90 | Qwen3ASRConfig, |
| 91 | Qwen3ASRThinkerConfig, |
| 92 | ) |
| 93 | from vllm.transformers_utils.processor import cached_processor_from_config |
| 94 | from vllm.transformers_utils.processors.qwen3_asr import ( |
| 95 | Qwen3ASRProcessor, |
| 96 | ) |
| 97 | from vllm.utils.gpu_sync_debug import gpu_sync_allowed |
| 98 | from vllm.utils.torch_utils import async_tensor_h2d |
| 99 | |
| 100 | logger = init_logger(__name__) |
| 101 | _ASR_TEXT_TAG = "<asr_text>" |
| 102 | # User-supplied `prompt` / `response_prefix` must not inject extra ChatML turns. |
| 103 | _CHATML_LIKE_TOKEN = re.compile(r"<\|[^|]+\|>") |
| 104 | _LANGUAGE_PREFIX = "language " |
| 105 | _MAX_STREAMING_PREFIX_CHARS = 50 |
| 106 | |
| 107 | |
| 108 | def _post_process_qwen3_asr_output(text: str) -> str: |
| 109 | if not text or _ASR_TEXT_TAG not in text: |
| 110 | return text |
| 111 | |
| 112 | _, text_part = text.rsplit(_ASR_TEXT_TAG, 1) |
| 113 | return text_part |
| 114 | |
| 115 | |
| 116 | class Qwen3ASRStreamingPostProcessor(StreamingTranscriptionPostProcessor): |
| 117 | def __init__(self) -> None: |
| 118 | self.raw_text = "" |
| 119 | self.emitted_text = "" |
| 120 | self.is_structured_output: bool | None = None |
| 121 | |
| 122 | def process_delta(self, text_delta: str, finished: bool) -> str: |
| 123 | self.raw_text += text_delta |
| 124 | |
| 125 | if self.is_structured_output is None: |
| 126 | text_without_leading_space = self.raw_text.lstrip() |
| 127 | maybe_structured_output = ( |
| 128 | text_without_leading_space == "" |
| 129 | or _LANGUAGE_PREFIX.startswith(text_without_leading_space) |
| 130 | or ( |
| 131 | text_without_leading_space.startswith(_LANGUAGE_PREFIX) |
| 132 | and len(text_without_leading_space) < _MAX_STREAMING_PREFIX_CHARS |
| 133 | and "\n" not in text_without_leading_space |
| 134 | ) |
| 135 | ) |
| 136 | if maybe_structured_output and _ASR_TEXT_TAG not in self.raw_text: |
| 137 | if not finished: |
| 138 | return "" |
| 139 | self.is_structured_output = False |
| 140 | else: |
| 141 | self.is_structured_output = _ASR_TEXT_TAG in self.raw_text |
| 142 | |
| 143 | processed_text = ( |
| 144 | _post_process_qwen3_asr_output(self.raw_text) |
| 145 | if self.is_structured_output |
| 146 | else self.raw_text |
| 147 | ) |
| 148 | new_text = processed_text[len(self.emitted_text) :] |
| 149 | self.emitted_text = processed_text |
| 150 | return new_text |
| 151 | |
| 152 | |
| 153 | def _sanitize_transcription_user_text(text: str) -> str: |
| 154 | """Strip ChatML-style special tokens from user-controlled transcription fields. |
| 155 | |
| 156 | Applies the regex / ``<asr_text>`` substitutions to a fixpoint so nested |
| 157 | payloads cannot reconstruct a valid token after a single pass: |
| 158 | |
| 159 | - ``<|im<|x|>_end|>`` would, with a single ``re.sub``, leave ``<|im_end|>`` |
| 160 | (a real ChatML control token). |
| 161 | - ``<asr_te<asr_text>xt>`` would, with a single ``str.replace``, leave |
| 162 | ``<asr_text>`` (the model-significant assistant-prefix delimiter). |
| 163 | |
| 164 | Looping both substitutions until the string stabilises eliminates these |
| 165 | reconstruction attacks. |
| 166 | """ |
| 167 | if not text: |
| 168 | return "" |
| 169 | prev = None |
| 170 | while prev != text: |
| 171 | prev = text |
| 172 | text = _CHATML_LIKE_TOKEN.sub("", text).replace(_ASR_TEXT_TAG, "") |
| 173 | return text |
| 174 | |
| 175 | |
| 176 | def _get_feat_extract_output_lengths(input_lengths: torch.Tensor): |
| 177 | input_lengths_leave = input_lengths % 100 |
| 178 | feat_lengths = (input_lengths_leave - 1) // 2 + 1 |
| 179 | output_lengths = ( |
| 180 | ((feat_lengths - 1) // 2 + 1 - 1) // 2 + 1 + (input_lengths // 100) * 13 |
| 181 | ) |
| 182 | return output_lengths |
| 183 | |
| 184 | |
| 185 | class Qwen3ASRProcessingInfo(BaseProcessingInfo): |
| 186 | def get_hf_config(self): |
| 187 | return self.ctx.get_hf_config(Qwen3ASRConfig).thinker_config |
| 188 | |
| 189 | def get_hf_processor(self, **kwargs: object) -> Qwen3ASRProcessor: |
| 190 | processor = self.ctx.get_hf_processor( |
| 191 | Qwen3ASRProcessor, |
| 192 | **kwargs, |
| 193 | ) |
| 194 | if not hasattr(processor, "audio_token"): |
| 195 | processor.audio_token = "<|audio_pad|>" |
| 196 | return processor |
| 197 | |
| 198 | def get_feature_extractor(self, **kwargs: object) -> WhisperFeatureExtractor: |
| 199 | hf_processor = self.get_hf_processor(**kwargs) |
| 200 | feature_extractor = hf_processor.feature_extractor |
| 201 | assert isinstance(feature_extractor, WhisperFeatureExtractor) |
| 202 | return feature_extractor |
| 203 | |
| 204 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 205 | return {"audio": None} |
| 206 | |
| 207 | def get_data_parser(self) -> MultiModalDataParser: |
| 208 | feature_extractor = self.get_feature_extractor() |
| 209 | return Qwen3ASRMultiModalDataParser( |
| 210 | target_sr=feature_extractor.sampling_rate, |
| 211 | expected_hidden_size=self._get_expected_hidden_size(), |
| 212 | ) |
| 213 | |
| 214 | |
| 215 | class Qwen3ASRDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3ASRProcessingInfo]): |
| 216 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 217 | num_audios = mm_counts.get("audio", 0) |
| 218 | |
| 219 | hf_processor = self.info.get_hf_processor() |
| 220 | audio_token = hf_processor.audio_token |
| 221 | |
| 222 | return audio_token * num_audios |
| 223 | |
| 224 | def get_dummy_mm_data( |
| 225 | self, |
| 226 | seq_len: int, |
| 227 | mm_counts: Mapping[str, int], |
| 228 | mm_options: MultiModalDummyOptions, |
| 229 | ) -> MultiModalDataDict: |
| 230 | feature_extractor = self.info.get_feature_extractor() |
| 231 | |
| 232 | target_audio_length = ( |
| 233 | min( |
| 234 | feature_extractor.chunk_length, |
| 235 | 30, |
| 236 | ) |
| 237 | * feature_extractor.sampling_rate |
| 238 | ) |
| 239 | |
| 240 | return { |
| 241 | "audio": self._get_dummy_audios( |
| 242 | length=target_audio_length, |
| 243 | num_audios=mm_counts.get("audio", 0), |
| 244 | overrides=mm_options.get("audio"), |
| 245 | ), |
| 246 | } |
| 247 | |
| 248 | |
| 249 | def _qwen3asr_field_config(hf_inputs: Mapping[str, torch.Tensor]): |
| 250 | audio_feature_lengths = hf_inputs.get("audio_feature_lengths", torch.empty((0,))) |
| 251 | return dict( |
| 252 | input_audio_features=MultiModalFieldConfig.flat_from_sizes( |
| 253 | "audio", audio_feature_lengths, dim=1 |
| 254 | ), |
| 255 | feature_attention_mask=MultiModalFieldConfig.batched("audio", keep_on_cpu=True), |
| 256 | audio_feature_lengths=MultiModalFieldConfig.batched("audio", keep_on_cpu=True), |
| 257 | ) |
| 258 | |
| 259 | |
| 260 | class Qwen3ASRMultiModalDataParser(MultiModalDataParser): |
| 261 | def _parse_audio_data( |
| 262 | self, |
| 263 | data: dict[str, torch.Tensor] | ModalityData[AudioItem], |
| 264 | ) -> ModalityDataItems[Any, Any] | None: |
| 265 | if isinstance(data, dict): |
| 266 | return DictEmbeddingItems( |
| 267 | data, |
| 268 | modality="audio", |
| 269 | required_fields={"input_audio_features", "audio_feature_lengths"}, |
| 270 | fields_factory=_qwen3asr_field_config, |
| 271 | ) |
| 272 | |
| 273 | return super()._parse_audio_data(data) |
| 274 | |
| 275 | |
| 276 | class Qwen3ASRMultiModalProcessor( |
| 277 | Qwen3OmniMoeThinkerMultiModalProcessor[Qwen3ASRProcessingInfo] |
| 278 | ): |
| 279 | def _get_mm_fields_config( |
| 280 | self, |
| 281 | hf_inputs: BatchFeature, |
| 282 | hf_processor_mm_kwargs: Mapping[str, object], |
| 283 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 284 | return _qwen3asr_field_config(hf_inputs) |
| 285 | |
| 286 | def _get_prompt_updates( |
| 287 | self, |
| 288 | mm_items: MultiModalDataItems, |
| 289 | hf_processor_mm_kwargs: Mapping[str, Any], |
| 290 | out_mm_kwargs: MultiModalKwargsItems, |
| 291 | ) -> Sequence[PromptUpdate]: |
| 292 | processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 293 | tokenizer = self.info.get_tokenizer() |
| 294 | vocab = tokenizer.get_vocab() |
| 295 | |
| 296 | audio_token = processor.audio_token |
| 297 | audio_token_id = vocab[audio_token] |
| 298 | |
| 299 | out_mm_data = out_mm_kwargs.get_data() |
| 300 | audio_feature_lengths = out_mm_data.get("audio_feature_lengths") |
| 301 | feature_attention_mask = out_mm_data.get("feature_attention_mask") |
| 302 | if audio_feature_lengths is None and feature_attention_mask is None: |
| 303 | audio_output_lengths = [] |
| 304 | elif audio_feature_lengths is not None: |
| 305 | audio_output_lens = _get_feat_extract_output_lengths(audio_feature_lengths) |
| 306 | audio_output_lengths = audio_output_lens.tolist() |
| 307 | elif feature_attention_mask is not None: |
| 308 | assert isinstance(feature_attention_mask, torch.Tensor) |
| 309 | audio_output_lens = _get_feat_extract_output_lengths( |
| 310 | feature_attention_mask.sum(-1) |
| 311 | ) |
| 312 | audio_output_lengths = audio_output_lens.tolist() |
| 313 | |
| 314 | def get_replacement_qwen2_audio(item_idx: int): |
| 315 | num_features = audio_output_lengths[item_idx] |
| 316 | if num_features == 0: |
| 317 | audios = mm_items.get_items("audio", AudioProcessorItems) |
| 318 | audio = audios.get(item_idx) |
| 319 | assert audio is not None |
| 320 | raise ValueError( |
| 321 | f"The audio {audio} (len={len(audio)}) is too short " |
| 322 | "to be represented inside the model" |
| 323 | ) |
| 324 | |
| 325 | return [audio_token_id] * num_features |
| 326 | |
| 327 | return [ |
| 328 | PromptReplacement( |
| 329 | modality="audio", |
| 330 | target=[audio_token_id], |
| 331 | replacement=get_replacement_qwen2_audio, |
| 332 | ), |
| 333 | ] |
| 334 | |
| 335 | |
| 336 | @MULTIMODAL_REGISTRY.register_processor( |
| 337 | Qwen3ASRMultiModalProcessor, |
| 338 | info=Qwen3ASRProcessingInfo, |
| 339 | dummy_inputs=Qwen3ASRDummyInputsBuilder, |
| 340 | ) |
| 341 | class Qwen3ASRForConditionalGeneration( |
| 342 | nn.Module, |
| 343 | SupportsMultiModal, |
| 344 | SupportsPP, |
| 345 | SupportsMRoPE, |
| 346 | SupportsTranscription, |
| 347 | SupportsLoRA, |
| 348 | ): |
| 349 | # LoRA support |
| 350 | packed_modules_mapping = { |
| 351 | "qkv_proj": [ |
| 352 | "q_proj", |
| 353 | "k_proj", |
| 354 | "v_proj", |
| 355 | ], |
| 356 | "gate_up_proj": [ |
| 357 | "gate_proj", |
| 358 | "up_proj", |
| 359 | ], |
| 360 | } |
| 361 | |
| 362 | supported_languages = ISO639_1_SUPPORTED_LANGS |
| 363 | supports_tower_connector_lora = True |
| 364 | |
| 365 | hf_to_vllm_mapper = WeightsMapper( |
| 366 | orig_to_new_prefix={ |
| 367 | "thinker.lm_head.": "language_model.lm_head.", |
| 368 | "thinker.model.": "language_model.model.", |
| 369 | "thinker.": "", |
| 370 | # HF format mapper |
| 371 | "model.audio_tower.": "audio_tower.", |
| 372 | "model.language_model.": "language_model.model.", |
| 373 | "model.multi_modal_projector.linear_1.": "audio_tower.proj1.", |
| 374 | "model.multi_modal_projector.linear_2.": "audio_tower.proj2.", |
| 375 | "talker.": None, |
| 376 | "code2wav.": None, |
| 377 | } |
| 378 | ) |
| 379 | |
| 380 | @classmethod |
| 381 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 382 | if modality.startswith("audio"): |
| 383 | return "<|audio_start|><|audio_pad|><|audio_end|>" |
| 384 | |
| 385 | raise ValueError("Only audio modality is supported") |
| 386 | |
| 387 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 388 | super().__init__() |
| 389 | self.vllm_config = vllm_config # needed for torch compile forward context |
| 390 | thinker_config: Qwen3ASRThinkerConfig = ( |
| 391 | vllm_config.model_config.hf_config.thinker_config |
| 392 | ) |
| 393 | quant_config = vllm_config.quant_config |
| 394 | multimodal_config = vllm_config.model_config.multimodal_config |
| 395 | self.config = thinker_config |
| 396 | self.multimodal_config = multimodal_config |
| 397 | self.quant_config = quant_config |
| 398 | |
| 399 | with self._mark_tower_model(vllm_config, "audio"): |
| 400 | self.audio_tower = Qwen3OmniMoeAudioEncoder( |
| 401 | thinker_config.audio_config, |
| 402 | prefix=maybe_prefix(prefix, "audio_tower"), |
| 403 | ) |
| 404 | |
| 405 | with self._mark_language_model(vllm_config): |
| 406 | self.language_model = Qwen3ForCausalLM( |
| 407 | vllm_config=vllm_config.with_hf_config( |
| 408 | thinker_config.text_config, architectures=["Qwen3ForCausalLM"] |
| 409 | ), |
| 410 | prefix=maybe_prefix(prefix, "language_model"), |
| 411 | ) |
| 412 | |
| 413 | self.make_empty_intermediate_tensors = ( |
| 414 | self.language_model.make_empty_intermediate_tensors |
| 415 | ) |
| 416 | |
| 417 | def _parse_and_validate_audio_input( |
| 418 | self, **kwargs: object |
| 419 | ) -> Qwen2_5OmniAudioFeatureInputs | None: |
| 420 | input_audio_features = kwargs.pop("input_audio_features", None) |
| 421 | audio_feature_lengths = kwargs.pop("audio_feature_lengths", None) |
| 422 | feature_attention_mask = kwargs.pop("feature_attention_mask", None) |
| 423 | if input_audio_features is None: |
| 424 | return None |
| 425 | |
| 426 | # inputs features from rust frontend is batched and padded |
| 427 | # with shape [batch_size, n_mels, padded_seq_len], different |
| 428 | # from python's shape [n_mels, batch_size * seq_len] |
| 429 | if ( |
| 430 | isinstance(input_audio_features, torch.Tensor) |
| 431 | and input_audio_features.dim() == 3 |
| 432 | ): |
| 433 | input_audio_features = unpad_and_flat_audio_features( |
| 434 | input_audio_features, audio_feature_lengths |
| 435 | ) |
| 436 | |
| 437 | return Qwen2_5OmniAudioFeatureInputs( |
| 438 | type="audio_features", |
| 439 | input_features=input_audio_features, |
| 440 | audio_feature_lengths=audio_feature_lengths, |
| 441 | feature_attention_mask=feature_attention_mask, |
| 442 | ) |
| 443 | |
| 444 | def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict: |
| 445 | mm_input_by_modality = {} |
| 446 | |
| 447 | # Preserve the order of modalities if there are multiple of them |
| 448 | # from the order of kwargs. |
| 449 | for input_key in kwargs: |
| 450 | if ( |
| 451 | input_key in ("input_audio_features") |
| 452 | and "audio" not in mm_input_by_modality |
| 453 | ): |
| 454 | mm_input_by_modality["audio"] = self._parse_and_validate_audio_input( |
| 455 | **kwargs |
| 456 | ) |
| 457 | return mm_input_by_modality |
| 458 | |
| 459 | def _process_audio_input( |
| 460 | self, |
| 461 | audio_input: Qwen2_5OmniAudioFeatureInputs, |
| 462 | ) -> torch.Tensor: |
| 463 | input_features = audio_input["input_features"] |
| 464 | # audio_feature_lengths is keep_on_cpu; the audio tower derives |
| 465 | # device placement from feature_lens, so move it explicitly. |
| 466 | audio_feature_lengths = async_tensor_h2d( |
| 467 | audio_input["audio_feature_lengths"], input_features.device |
| 468 | ) |
| 469 | |
| 470 | audio_output_lengths = _get_feat_extract_output_lengths(audio_feature_lengths) |
| 471 | |
| 472 | audio_features = self.audio_tower( |
| 473 | input_features.to(self.audio_tower.dtype), |
| 474 | feature_lens=audio_feature_lengths, |
| 475 | aftercnn_lens=audio_output_lengths, |
| 476 | ) |
| 477 | with gpu_sync_allowed(): |
| 478 | split_sizes = audio_output_lengths.tolist() |
| 479 | return audio_features.split(split_sizes) |
| 480 | |
| 481 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None: |
| 482 | mm_input_by_modality = self._parse_and_validate_multimodal_inputs(**kwargs) |
| 483 | if not mm_input_by_modality: |
| 484 | return [] |
| 485 | |
| 486 | # The result multimodal_embeddings is tuple of tensors, with each |
| 487 | # tensor correspoending to a multimodal data item (image or video). |
| 488 | multimodal_embeddings: tuple[torch.Tensor, ...] = () |
| 489 | |
| 490 | # NOTE: It is important to iterate over the keys in this dictionary |
| 491 | # to preserve the order of the modalities. |
| 492 | for modality in mm_input_by_modality: |
| 493 | multimodal_input = mm_input_by_modality[modality] |
| 494 | if modality == "audio": |
| 495 | audio_embeddings = self._process_audio_input(multimodal_input) |
| 496 | multimodal_embeddings += tuple(audio_embeddings) |
| 497 | return multimodal_embeddings |
| 498 | |
| 499 | def embed_input_ids( |
| 500 | self, |
| 501 | input_ids: torch.Tensor, |
| 502 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 503 | *, |
| 504 | is_multimodal: torch.Tensor | None = None, |
| 505 | ) -> torch.Tensor: |
| 506 | inputs_embeds = self._embed_text_input_ids( |
| 507 | input_ids, |
| 508 | self.language_model.embed_input_ids, |
| 509 | is_multimodal=is_multimodal, |
| 510 | ) |
| 511 | |
| 512 | if multimodal_embeddings is None or len(multimodal_embeddings) == 0: |
| 513 | return inputs_embeds |
| 514 | |
| 515 | inputs_embeds = _merge_multimodal_embeddings( |
| 516 | inputs_embeds=inputs_embeds, |
| 517 | multimodal_embeddings=multimodal_embeddings, |
| 518 | is_multimodal=is_multimodal, |
| 519 | ) |
| 520 | |
| 521 | return inputs_embeds |
| 522 | |
| 523 | def forward( |
| 524 | self, |
| 525 | input_ids: torch.Tensor, |
| 526 | positions: torch.Tensor, |
| 527 | intermediate_tensors: IntermediateTensors | None = None, |
| 528 | inputs_embeds: torch.Tensor | None = None, |
| 529 | **kwargs: object, |
| 530 | ) -> torch.Tensor | IntermediateTensors: |
| 531 | if intermediate_tensors is not None: |
| 532 | inputs_embeds = None |
| 533 | |
| 534 | hidden_states = self.language_model.model( |
| 535 | input_ids, |
| 536 | positions, |
| 537 | intermediate_tensors, |
| 538 | inputs_embeds=inputs_embeds, |
| 539 | ) |
| 540 | |
| 541 | return hidden_states |
| 542 | |
| 543 | def compute_logits( |
| 544 | self, |
| 545 | hidden_states: torch.Tensor, |
| 546 | ) -> torch.Tensor | None: |
| 547 | return self.language_model.compute_logits(hidden_states) |
| 548 | |
| 549 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 550 | loader = AutoWeightsLoader(self) |
| 551 | loaded_weights = loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 552 | |
| 553 | return loaded_weights |
| 554 | |
| 555 | def get_mrope_input_positions( |
| 556 | self, |
| 557 | input_tokens: list[int], |
| 558 | mm_features: list[MultiModalFeatureSpec], |
| 559 | ) -> tuple[torch.Tensor, int]: |
| 560 | seq_len = len(input_tokens) |
| 561 | |
| 562 | if not mm_features: |
| 563 | # No audio features, just return linear positions |
| 564 | llm_positions = ( |
| 565 | torch.arange(seq_len, dtype=torch.long).view(1, -1).expand(3, -1) |
| 566 | ) |
| 567 | return llm_positions.clone(), 0 |
| 568 | |
| 569 | llm_pos_ids_list: list[torch.Tensor] = [] |
| 570 | st = 0 |
| 571 | |
| 572 | for mm_feature in sorted(mm_features, key=lambda f: f.mm_position.offset): |
| 573 | offset = mm_feature.mm_position.offset |
| 574 | |
| 575 | # The placeholder already contains one token per audio embedding. |
| 576 | # Use its length so M-RoPE can be reconstructed even when a covered |
| 577 | # prefix no longer carries the audio processor tensors. |
| 578 | audio_len = mm_feature.mm_position.length |
| 579 | |
| 580 | # Text segment before audio (includes audio_start token) |
| 581 | text_len = offset - st |
| 582 | st_idx = llm_pos_ids_list[-1].max() + 1 if llm_pos_ids_list else 0 |
| 583 | text_positions = ( |
| 584 | torch.arange(text_len, dtype=torch.long).view(1, -1).expand(3, -1) |
| 585 | + st_idx |
| 586 | ) |
| 587 | llm_pos_ids_list.append(text_positions) |
| 588 | st_idx = st_idx + text_len |
| 589 | |
| 590 | # Audio token segment |
| 591 | audio_positions = ( |
| 592 | torch.arange(audio_len, dtype=torch.long).view(1, -1).expand(3, -1) |
| 593 | + st_idx |
| 594 | ) |
| 595 | llm_pos_ids_list.append(audio_positions) |
| 596 | |
| 597 | st = offset + audio_len |
| 598 | |
| 599 | # Handle remaining text (includes audio_end and any trailing text) |
| 600 | if st < seq_len: |
| 601 | st_idx = llm_pos_ids_list[-1].max() + 1 if llm_pos_ids_list else 0 |
| 602 | text_len = seq_len - st |
| 603 | final_text_positions = ( |
| 604 | torch.arange(text_len, dtype=torch.long).view(1, -1).expand(3, -1) |
| 605 | + st_idx |
| 606 | ) |
| 607 | llm_pos_ids_list.append(final_text_positions) |
| 608 | |
| 609 | llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1) |
| 610 | if llm_positions.shape[1] != seq_len: |
| 611 | raise RuntimeError("Position ids length mismatch with input ids length") |
| 612 | |
| 613 | mrope_position_delta = (llm_positions.max() + 1 - seq_len).item() |
| 614 | return llm_positions, mrope_position_delta |
| 615 | |
| 616 | def get_mm_mapping(self) -> MultiModelKeys: |
| 617 | """Get the module prefix in multimodal models.""" |
| 618 | return MultiModelKeys.from_string_field( |
| 619 | language_model="language_model", |
| 620 | tower_model=["audio_tower."], |
| 621 | ) |
| 622 | |
| 623 | def get_mm_lora_token_counts( |
| 624 | self, |
| 625 | *, |
| 626 | modality: str, |
| 627 | mm_kwargs: MultiModalKwargsItem | None, |
| 628 | num_mm_embeds: int, |
| 629 | ) -> tuple[int, int | None]: |
| 630 | del modality, mm_kwargs |
| 631 | # For Qwen3-ASR, the audio tower produces one embedding per audio |
| 632 | # placeholder token inserted into the prompt (no additional |
| 633 | # merge/downsample step like vision towers). Therefore, the encoder |
| 634 | # token budget is identity. |
| 635 | return num_mm_embeds, None |
| 636 | |
| 637 | @classmethod |
| 638 | def get_speech_to_text_config( |
| 639 | cls, model_config: ModelConfig, task_type: str |
| 640 | ) -> SpeechToTextConfig: |
| 641 | processor = cached_processor_from_config(model_config) |
| 642 | feature_extractor: WhisperFeatureExtractor = processor.feature_extractor |
| 643 | return SpeechToTextConfig( |
| 644 | max_audio_clip_s=feature_extractor.chunk_length, |
| 645 | sample_rate=feature_extractor.sampling_rate, |
| 646 | ) |
| 647 | |
| 648 | @classmethod |
| 649 | def get_generation_prompt(cls, stt_params: SpeechToTextParams) -> PromptType: |
| 650 | """Get the generation prompt to be used for transcription requests. |
| 651 | |
| 652 | Matches the official Qwen3-ASR SDK prompt format. The ``system`` turn |
| 653 | is only emitted when the caller supplied a ``prompt``, mirroring the |
| 654 | SDK's ``_build_messages`` (which omits the system role when context is |
| 655 | empty) and preserving the prior no-prompt behavior: |
| 656 | |
| 657 | [system: {context}] # only when prompt given |
| 658 | user: {audio} |
| 659 | assistant: [language {Lang}<asr_text>] # when language is forced |
| 660 | """ |
| 661 | audio = stt_params.audio |
| 662 | model_config = stt_params.model_config |
| 663 | language = stt_params.language |
| 664 | task_type = stt_params.task_type |
| 665 | request_prompt = stt_params.request_prompt |
| 666 | to_language = stt_params.to_language |
| 667 | |
| 668 | tokenizer = cached_tokenizer_from_config(model_config) |
| 669 | audio_placeholder = cls.get_placeholder_str("audio", 0) |
| 670 | |
| 671 | if task_type not in ("transcribe", "translate"): |
| 672 | raise ValueError( |
| 673 | f"Unsupported task_type '{task_type}'. " |
| 674 | "Supported task types are 'transcribe' and 'translate'." |
| 675 | ) |
| 676 | |
| 677 | context = _sanitize_transcription_user_text(request_prompt) |
| 678 | system_turn = f"<|im_start|>system\n{context}<|im_end|>\n" if context else "" |
| 679 | |
| 680 | prompt = ( |
| 681 | f"{system_turn}" |
| 682 | f"<|im_start|>user\n{audio_placeholder}<|im_end|>\n" |
| 683 | f"<|im_start|>assistant\n" |
| 684 | ) |
| 685 | |
| 686 | lang_code = to_language if task_type == "translate" else language |
| 687 | if lang_code is not None: |
| 688 | full_lang_name = cls.supported_languages.get(lang_code, lang_code) |
| 689 | prompt += f"language {full_lang_name}{_ASR_TEXT_TAG}" |
| 690 | |
| 691 | prompt_token_ids = cached_encode(tokenizer, prompt) |
| 692 | |
| 693 | return TokensPrompt( |
| 694 | prompt_token_ids=prompt_token_ids, |
| 695 | multi_modal_data={"audio": audio}, |
| 696 | ) |
| 697 | |
| 698 | @classmethod |
| 699 | def post_process_output(cls, text: str) -> str: |
| 700 | """Post-process Qwen3-ASR raw output to extract clean transcription. |
| 701 | |
| 702 | The model outputs in format: "language {lang}<asr_text>{transcription}" |
| 703 | This method strips the language prefix and asr_text tags. |
| 704 | """ |
| 705 | return _post_process_qwen3_asr_output(text) |
| 706 | |
| 707 | @classmethod |
| 708 | def get_streaming_post_processor_cls( |
| 709 | cls, |
| 710 | ) -> type[StreamingTranscriptionPostProcessor]: |
| 711 | return Qwen3ASRStreamingPostProcessor |