vllm.entrypoints.chat_utils ¶

class AsyncMultiModalContentParser(BaseMultiModalContentParser):
    def __init__(self, tracker: AsyncMultiModalItemTracker) -> None:
        super().__init__()

        self._tracker = tracker
        multimodal_config = self._tracker.model_config.multimodal_config
        media_io_kwargs = getattr(multimodal_config, "media_io_kwargs", None)
        self._connector = MediaConnector(
            media_io_kwargs=media_io_kwargs,
            allowed_local_media_path=tracker.allowed_local_media_path,
        )

    def parse_image(
        self, image_url: Optional[str], uuid: Optional[str] = None
    ) -> None:
        image_coro = (
            self._connector.fetch_image_async(image_url) if image_url else None
        )

        placeholder = self._tracker.add("image", image_coro, uuid)
        self._add_placeholder("image", placeholder)

    def parse_image_embeds(
        self,
        image_embeds: Union[str, dict[str, str], None],
        uuid: Optional[str] = None,
    ) -> None:
        future: asyncio.Future[Union[str, dict[str, str], None]] = (
            asyncio.Future()
        )

        if isinstance(image_embeds, dict):
            embeds = {
                k: self._connector.fetch_image_embedding(v)
                for k, v in image_embeds.items()
            }
            future.set_result(embeds)

        if isinstance(image_embeds, str):
            embedding = self._connector.fetch_image_embedding(image_embeds)
            future.set_result(embedding)

        if image_embeds is None:
            future.set_result(None)

        placeholder = self._tracker.add("image_embeds", future, uuid)
        self._add_placeholder("image", placeholder)

    def parse_image_pil(
        self, image_pil: Optional[Image.Image], uuid: Optional[str] = None
    ) -> None:
        future: asyncio.Future[Optional[Image.Image]] = asyncio.Future()
        if image_pil:
            future.set_result(image_pil)
        else:
            future.set_result(None)

        placeholder = self._tracker.add("image", future, uuid)
        self._add_placeholder("image", placeholder)

    def parse_audio(
        self, audio_url: Optional[str], uuid: Optional[str] = None
    ) -> None:
        audio_coro = (
            self._connector.fetch_audio_async(audio_url) if audio_url else None
        )

        placeholder = self._tracker.add("audio", audio_coro, uuid)
        self._add_placeholder("audio", placeholder)

    def parse_input_audio(
        self, input_audio: Optional[InputAudio], uuid: Optional[str] = None
    ) -> None:
        if input_audio:
            audio_data = input_audio.get("data", "")
            audio_format = input_audio.get("format", "")
            if audio_data:
                audio_url = f"data:audio/{audio_format};base64,{audio_data}"
            else:
                # If a UUID is provided, audio data may be empty.
                audio_url = None
        else:
            audio_url = None

        return self.parse_audio(audio_url, uuid)

    def parse_video(
        self, video_url: Optional[str], uuid: Optional[str] = None
    ) -> None:
        video = (
            self._connector.fetch_video_async(video_url=video_url)
            if video_url
            else None
        )

        placeholder = self._tracker.add("video", video, uuid)
        self._add_placeholder("video", placeholder)

class AsyncMultiModalItemTracker(BaseMultiModalItemTracker[Awaitable[object]]):
    async def all_mm_data(self) -> Optional[MultiModalDataDict]:
        if not self._items_by_modality:
            return None
        mm_inputs = {}
        items_by_modality = {}
        for modality, items in self._items_by_modality.items():
            coros = []
            for item in items:
                if item is not None:
                    coros.append(item)
                else:
                    coros.append(asyncio.sleep(0))
            items_by_modality[modality] = await asyncio.gather(*coros)

        if "image" in items_by_modality and "image_embeds" in items_by_modality:
            raise ValueError(
                "Mixing raw image and embedding inputs is not allowed"
            )

        if "image_embeds" in items_by_modality:
            image_embeds_lst = items_by_modality["image_embeds"]
            if len(image_embeds_lst) > 1:
                raise ValueError(
                    "Only one message can have {'type': 'image_embeds'}"
                )
            mm_inputs["image"] = image_embeds_lst[0]
        if "image" in items_by_modality:
            mm_inputs["image"] = items_by_modality["image"]  # A list of images
        if "audio" in items_by_modality:
            mm_inputs["audio"] = items_by_modality["audio"]  # A list of audios
        if "video" in items_by_modality:
            mm_inputs["video"] = items_by_modality["video"]  # A list of videos
        return mm_inputs

    def create_parser(self) -> "BaseMultiModalContentParser":
        return AsyncMultiModalContentParser(self)

class AudioURL(TypedDict, total=False):
    url: Required[str]
    """
    Either a URL of the audio or a data URL with base64 encoded audio data.
    """

class BaseMultiModalContentParser(ABC):
    def __init__(self) -> None:
        super().__init__()

        # stores model placeholders list with corresponding
        # general MM placeholder:
        # {
        #   "<##IMAGE##>": ["<image>", "<image>", "<image>"],
        #   "<##AUDIO##>": ["<audio>", "<audio>"]
        # }
        self._placeholder_storage: dict[str, list] = defaultdict(list)

    def _add_placeholder(
        self, modality: ModalityStr, placeholder: Optional[str]
    ):
        mod_placeholder = MODALITY_PLACEHOLDERS_MAP[modality]
        if placeholder:
            self._placeholder_storage[mod_placeholder].append(placeholder)

    def mm_placeholder_storage(self) -> dict[str, list]:
        return dict(self._placeholder_storage)

    @abstractmethod
    def parse_image(
        self, image_url: Optional[str], uuid: Optional[str] = None) -> None:
        raise NotImplementedError

    @abstractmethod
    def parse_image_embeds(
        self,
        image_embeds: Union[str, dict[str, str], None],
        uuid: Optional[str] = None,
    ) -> None:
        raise NotImplementedError

    @abstractmethod
    def parse_image_pil(
        self, image_pil: Optional[Image.Image], uuid: Optional[str] = None
    ) -> None:
        raise NotImplementedError

    @abstractmethod
    def parse_audio(
        self, audio_url: Optional[str], uuid: Optional[str] = None
    ) -> None:
        raise NotImplementedError

    @abstractmethod
    def parse_input_audio(
        self, input_audio: Optional[InputAudio], uuid: Optional[str] = None
    ) -> None:
        raise NotImplementedError

    @abstractmethod
    def parse_video(
        self, video_url: Optional[str], uuid: Optional[str] = None
    ) -> None:
        raise NotImplementedError

class BaseMultiModalItemTracker(ABC, Generic[_T]):
    """
    Tracks multi-modal items in a given request and ensures that the number
    of multi-modal items in a given request does not exceed the configured
    maximum per prompt.
    """

    def __init__(self, model_config: ModelConfig, tokenizer: AnyTokenizer):
        super().__init__()

        self._model_config = model_config
        self._tokenizer = tokenizer

        self._items_by_modality = defaultdict[str, list[Optional[_T]]](list)
        self._uuids_by_modality = defaultdict[str, list[Optional[str]]](list)

    @property
    def model_config(self) -> ModelConfig:
        return self._model_config

    @cached_property
    def model_cls(self) -> type[SupportsMultiModal]:
        from vllm.model_executor.model_loader import get_model_cls

        model_cls = get_model_cls(self.model_config)
        return cast(type[SupportsMultiModal], model_cls)

    @property
    def allowed_local_media_path(self):
        return self._model_config.allowed_local_media_path

    @property
    def mm_registry(self):
        return MULTIMODAL_REGISTRY

    @cached_property
    def mm_processor(self):
        return self.mm_registry.create_processor(self.model_config)

    def add(
        self,
        modality: ModalityStr,
        item: Optional[_T],
        uuid: Optional[str] = None,
    ) -> Optional[str]:
        """
        Add a multi-modal item to the current prompt and returns the
        placeholder string to use, if any.

        An optional uuid can be added which serves as a unique identifier of the
        media.
        """
        input_modality = modality.replace("_embeds", "")
        num_items = len(self._items_by_modality[modality]) + 1

        self.mm_processor.validate_num_items(input_modality, num_items)

        self._items_by_modality[modality].append(item)
        self._uuids_by_modality[modality].append(uuid)

        return self.model_cls.get_placeholder_str(modality, num_items)

    def all_mm_uuids(self) -> Optional[MultiModalUUIDDict]:
        if not self._items_by_modality:
            return None
        mm_uuids = {}
        uuids_by_modality = dict(self._uuids_by_modality)
        if "image" in uuids_by_modality and "image_embeds" in uuids_by_modality:
            raise ValueError(
                "Mixing raw image and embedding inputs is not allowed"
            )

        if "image_embeds" in uuids_by_modality:
            image_embeds_uuids = uuids_by_modality["image_embeds"]
            if len(image_embeds_uuids) > 1:
                raise ValueError(
                    "Only one message can have {'type': 'image_embeds'}"
                )
            mm_uuids["image"] = uuids_by_modality["image_embeds"]
        if "image" in uuids_by_modality:
            mm_uuids["image"] = uuids_by_modality["image"]  # UUIDs of images
        if "audio" in uuids_by_modality:
            mm_uuids["audio"] = uuids_by_modality["audio"]  # UUIDs of audios
        if "video" in uuids_by_modality:
            mm_uuids["video"] = uuids_by_modality["video"]  # UUIDs of videos
        return mm_uuids

    @abstractmethod
    def create_parser(self) -> "BaseMultiModalContentParser":
        raise NotImplementedError

class ChatCompletionContentPartAudioParam(TypedDict, total=False):
    audio_url: Required[AudioURL]

    type: Required[Literal["audio_url"]]
    """The type of the content part."""

class ChatCompletionContentPartImageEmbedsParam(TypedDict, total=False):
    image_embeds: Optional[Union[str, dict[str, str]]]
    """
    The image embeddings. It can be either:
    - A single base64 string.
    - A dictionary where each value is a base64 string.
    """
    type: Required[Literal["image_embeds"]]
    """The type of the content part."""
    uuid: Optional[str]
    """
    User-provided UUID of a media. User must guarantee that it is properly
    generated and unique for different medias.
    """

class ChatCompletionContentPartVideoParam(TypedDict, total=False):
    video_url: Required[VideoURL]

    type: Required[Literal["video_url"]]
    """The type of the content part."""

class ConversationMessage(TypedDict, total=False):
    role: Required[str]
    """The role of the message's author."""

    content: Union[Optional[str], list[dict[str, str]]]
    """The contents of the message"""

    tool_call_id: Optional[str]
    """Tool call that this message is responding to."""

    name: Optional[str]
    """The name of the function to call"""

    tool_calls: Optional[Iterable[ChatCompletionMessageToolCallParam]]
    """The tool calls generated by the model, such as function calls."""