@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / a2dd48c386

Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 15, 21:00 UTC

[VLM] Deprecate legacy input mapper for OOT multimodal models (#13979)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

Cyrus Leung committed Feb 28, 2025 at 03:14 UTC a2dd48c386f787cdfbe35099e0d273f19748dbb6

2 files changed +32 -27

vllm/config.py

+23 -22

@@ -400,7 +400,7 @@ class ModelConfig:

400 else:

401 self.override_neuron_config = None

402

403 - supported_tasks, task = self._resolve_task(task, self.hf_config)

403 + supported_tasks, task = self._resolve_task(task)

404 self.supported_tasks = supported_tasks

405 self.task: Final = task

406 if self.task in ("draft", "generate"):

@@ -418,6 +418,14 @@ class ModelConfig:

418 self._verify_cuda_graph()

419 self._verify_bnb_config()

420

421 + @property

422 + def registry(self):

423 + return ModelRegistry

424 +

425 + @property

426 + def architectures(self) -> list[str]:

427 + return getattr(self.hf_config, "architectures", [])

428 +

429 def maybe_pull_model_tokenizer_for_s3(self, model: str,

430 tokenizer: str) -> None:

431 """

@@ -446,8 +454,7 @@ class ModelConfig:

454 def _init_multimodal_config(

455 self, limit_mm_per_prompt: Optional[Mapping[str, int]]

456 ) -> Optional["MultiModalConfig"]:

449 - architectures = getattr(self.hf_config, "architectures", [])

450 - if ModelRegistry.is_multimodal_model(architectures):

457 + if self.registry.is_multimodal_model(self.architectures):

458 return MultiModalConfig(limit_per_prompt=limit_mm_per_prompt or {})

459

460 if limit_mm_per_prompt:

@@ -480,16 +487,13 @@ class ModelConfig:

487 return None

488

489 def _init_attention_free(self) -> bool:

483 - architectures = getattr(self.hf_config, "architectures", [])

484 - return ModelRegistry.is_attention_free_model(architectures)

490 + return self.registry.is_attention_free_model(self.architectures)

491

492 def _init_is_hybrid(self) -> bool:

487 - architectures = getattr(self.hf_config, "architectures", [])

488 - return ModelRegistry.is_hybrid_model(architectures)

493 + return self.registry.is_hybrid_model(self.architectures)

494

495 def _init_has_inner_state(self) -> bool:

491 - architectures = getattr(self.hf_config, "architectures", [])

492 - return ModelRegistry.model_has_inner_state(architectures)

496 + return self.registry.model_has_inner_state(self.architectures)

497

498 def _verify_tokenizer_mode(self) -> None:

499 tokenizer_mode = self.tokenizer_mode.lower()

@@ -507,9 +511,9 @@ class ModelConfig:

511 model_id = self.model

512 if get_pooling_config(model_id, self.revision):

513 return "embed"

510 - if ModelRegistry.is_cross_encoder_model(architectures):

514 + if self.registry.is_cross_encoder_model(architectures):

515 return "score"

512 - if ModelRegistry.is_transcription_model(architectures):

516 + if self.registry.is_transcription_model(architectures):

517 return "transcription"

518

519 suffix_to_preferred_task: List[Tuple[str, _ResolvedTask]] = [

@@ -522,7 +526,7 @@ class ModelConfig:

526 ("EmbeddingModel", "embed"),

527 ("RewardModel", "reward"),

528 ]

525 - _, arch = ModelRegistry.inspect_model_cls(architectures)

529 + _, arch = self.registry.inspect_model_cls(architectures)

530

531 for suffix, pref_task in suffix_to_preferred_task:

532 if arch.endswith(suffix) and pref_task in supported_tasks:

@@ -533,20 +537,19 @@ class ModelConfig:

537 def _resolve_task(

538 self,

539 task_option: Union[TaskOption, Literal["draft"]],

536 - hf_config: PretrainedConfig,

540 ) -> Tuple[Set[_ResolvedTask], _ResolvedTask]:

541 if task_option == "draft":

542 return {"draft"}, "draft"

543

541 - architectures = getattr(hf_config, "architectures", [])

544 + registry = self.registry

545 + architectures = self.architectures

546

547 runner_support: Dict[RunnerType, bool] = {

548 # NOTE: Listed from highest to lowest priority,

549 # in case the model supports multiple of them

546 - "transcription":

547 - ModelRegistry.is_transcription_model(architectures),

548 - "generate": ModelRegistry.is_text_generation_model(architectures),

549 - "pooling": ModelRegistry.is_pooling_model(architectures),

550 + "transcription": registry.is_transcription_model(architectures),

551 + "generate": registry.is_text_generation_model(architectures),

552 + "pooling": registry.is_pooling_model(architectures),

553 }

554 supported_runner_types_lst: List[RunnerType] = [

555 runner_type

@@ -755,8 +758,7 @@ class ModelConfig:

758

759 pipeline_parallel_size = parallel_config.pipeline_parallel_size

760 if pipeline_parallel_size > 1:

758 - architectures = getattr(self.hf_config, "architectures", [])

759 - if not ModelRegistry.is_pp_supported_model(architectures):

761 + if not self.registry.is_pp_supported_model(self.architectures):

762 raise NotImplementedError(

763 "Pipeline parallelism is not supported for this model. "

764 "Supported models implement the `SupportsPP` interface.")

@@ -1023,8 +1025,7 @@ class ModelConfig:

1025

1026 @property

1027 def is_cross_encoder(self) -> bool:

1026 - architectures = getattr(self.hf_config, "architectures", [])

1027 - return ModelRegistry.is_cross_encoder_model(architectures)

1028 + return self.registry.is_cross_encoder_model(self.architectures)

1029

1030 @property

1031 def use_mla(self) -> bool:

vllm/inputs/preprocess.py

+9 -5

@@ -236,11 +236,15 @@ class InputPreprocessor:

236 # updated to use the new multi-modal processor

237 can_process_multimodal = self.mm_registry.has_processor(model_config)

238 if not can_process_multimodal:

239 - logger.info_once(

240 - "Your model uses the legacy input pipeline instead of the new "

241 - "multi-modal processor. Please note that the legacy pipeline "

242 - "will be removed in a future release. For more details, see: "

243 - "https://github.com/vllm-project/vllm/issues/10114")

239 + from vllm.model_executor.models.registry import _VLLM_MODELS

240 + if not any(arch in _VLLM_MODELS

241 + for arch in model_config.architectures):

242 + logger.warning_once(

243 + "Your model uses the legacy input pipeline, which will be "

244 + "removed in an upcoming release. "

245 + "Please upgrade to the new multi-modal processing pipeline "

246 + "(https://docs.vllm.ai/en/latest/design/mm_processing.html)"

247 + )

248

249 return can_process_multimodal

250