vllm.transformers_utils.processor ¶

def _transformers_v4_compatibility_import():
    """Some remote code processors still import `ChatTemplateLoadKwargs` which was a
    subset of `ProcessorChatTemplateKwargs` as defined in Transformers v4.
    In Transformers v5 these were merged into `ProcessorChatTemplateKwargs` and
    `ChatTemplateLoadKwargs` was removed. For backward compatibility, we add an alias
    for `ChatTemplateLoadKwargs` if it doesn't exist.

    This can be removed if `HCXVisionForCausalLM` is upstreamed to Transformers."""
    old_import = getattr(processing_utils, "ChatTemplateLoadKwargs", None)
    new_import = getattr(processing_utils, "ProcessorChatTemplateKwargs", None)
    if old_import is None and new_import is not None:
        processing_utils.ChatTemplateLoadKwargs = new_import

def get_feature_extractor(
    processor_name: str,
    *args: Any,
    revision: str | None = None,
    trust_remote_code: bool = False,
    **kwargs: Any,
):
    """Load an audio feature extractor for the given model name
    via HuggingFace."""
    try:
        processor_name = convert_model_repo_to_path(processor_name)
        feature_extractor = AutoFeatureExtractor.from_pretrained(
            processor_name,
            *args,
            revision=revision,
            trust_remote_code=trust_remote_code,
            **kwargs,
        )
    except ValueError as e:
        # If the error pertains to the processor class not existing or not
        # currently being imported, suggest using the --trust-remote-code flag.
        # Unlike AutoTokenizer, AutoImageProcessor does not separate such errors
        if not trust_remote_code:
            err_msg = (
                "Failed to load the feature extractor. If the feature "
                "extractor is a custom extractor not yet available in the "
                "HuggingFace transformers library, consider setting "
                "`trust_remote_code=True` in LLM or using the "
                "`--trust-remote-code` flag in the CLI."
            )
            raise RuntimeError(err_msg) from e
        else:
            raise e
    return cast(FeatureExtractionMixin, feature_extractor)

def get_image_processor(
    processor_name: str,
    *args: Any,
    revision: str | None = None,
    trust_remote_code: bool = False,
    **kwargs: Any,
):
    """Load an image processor for the given model name via HuggingFace."""
    try:
        processor_name = convert_model_repo_to_path(processor_name)
        processor = AutoImageProcessor.from_pretrained(
            processor_name,
            *args,
            revision=revision,
            trust_remote_code=trust_remote_code,
            **kwargs,
        )
    except ValueError as e:
        # If the error pertains to the processor class not existing or not
        # currently being imported, suggest using the --trust-remote-code flag.
        # Unlike AutoTokenizer, AutoImageProcessor does not separate such errors
        if not trust_remote_code:
            err_msg = (
                "Failed to load the image processor. If the image processor is "
                "a custom processor not yet available in the HuggingFace "
                "transformers library, consider setting "
                "`trust_remote_code=True` in LLM or using the "
                "`--trust-remote-code` flag in the CLI."
            )
            raise RuntimeError(err_msg) from e
        else:
            raise e

    return cast(BaseImageProcessor, processor)

def get_processor(
    processor_name: str,
    *args: Any,
    revision: str | None = None,
    trust_remote_code: bool = False,
    processor_cls: type[_P] | tuple[type[_P], ...] = ProcessorMixin,
    **kwargs: Any,
) -> _P:
    """Load a processor for the given model name via HuggingFace."""
    if revision is None:
        revision = "main"
    try:
        processor_name = convert_model_repo_to_path(processor_name)
        if isinstance(processor_cls, tuple) or processor_cls == ProcessorMixin:
            processor = AutoProcessor.from_pretrained(
                processor_name,
                *args,
                revision=revision,
                trust_remote_code=trust_remote_code,
                **kwargs,
            )
        elif issubclass(processor_cls, ProcessorMixin):
            processor = processor_cls.from_pretrained(
                processor_name,
                *args,
                revision=revision,
                trust_remote_code=trust_remote_code,
                **kwargs,
            )
        else:
            # Processors that are standalone classes unrelated to HF
            processor = processor_cls(*args, **kwargs)
    except ValueError as e:
        # If the error pertains to the processor class not existing or not
        # currently being imported, suggest using the --trust-remote-code flag.
        # Unlike AutoTokenizer, AutoProcessor does not separate such errors
        if not trust_remote_code:
            err_msg = (
                "Failed to load the processor. If the processor is "
                "a custom processor not yet available in the HuggingFace "
                "transformers library, consider setting "
                "`trust_remote_code=True` in LLM or using the "
                "`--trust-remote-code` flag in the CLI."
            )
            raise RuntimeError(err_msg) from e
        else:
            raise e

    if not isinstance(processor, processor_cls):
        raise TypeError(
            "Invalid type of HuggingFace processor. "
            f"Expected type: {processor_cls}, but "
            f"found type: {type(processor)}"
        )

    return processor

def get_video_processor(
    processor_name: str,
    *args: Any,
    revision: str | None = None,
    trust_remote_code: bool = False,
    processor_cls_overrides: type[_V] | None = None,
    **kwargs: Any,
):
    """Load a video processor for the given model name via HuggingFace."""
    try:
        processor_name = convert_model_repo_to_path(processor_name)
        processor_cls = processor_cls_overrides or AutoVideoProcessor
        processor = processor_cls.from_pretrained(
            processor_name,
            *args,
            revision=revision,
            trust_remote_code=trust_remote_code,
            **kwargs,
        )
    except ValueError as e:
        # If the error pertains to the processor class not existing or not
        # currently being imported, suggest using the --trust-remote-code flag.
        # Unlike AutoTokenizer, AutoVideoProcessor does not separate such errors
        if not trust_remote_code:
            err_msg = (
                "Failed to load the video processor. If the video processor is "
                "a custom processor not yet available in the HuggingFace "
                "transformers library, consider setting "
                "`trust_remote_code=True` in LLM or using the "
                "`--trust-remote-code` flag in the CLI."
            )
            raise RuntimeError(err_msg) from e
        else:
            raise e

    return cast(BaseVideoProcessor, processor)