vllm.inputs.engine ¶

class EmbedsInput(_InputOptions):
    """Represents embeddings-based input to the engine."""

    type: Literal["embeds"]
    """The type of input."""

    prompt_embeds: "torch.Tensor"
    """The embeddings of the prompt."""

    prompt: NotRequired[str]
    """The prompt text corresponding to the token IDs, if available."""

    prompt_token_ids: NotRequired[list[int]]
    """Token IDs of the rendered prompt. Only set for mixed-mode inputs
    (chat completion with `prompt_embeds` content parts). When present,
    `is_token_ids` MUST also be present and have the same length. 
    For pure-embeds inputs this field is absent."""

    is_token_ids: NotRequired[list[bool]]
    """Per-position mask for mixed-mode inputs. `True` means the position
    is a real token ID (use the model's embedding layer); `False` means
    the position uses a pre-computed embedding row from `prompt_embeds`.
    Length MUST equal `len(prompt_token_ids)`.
    For pure-embeds inputs this field is absent."""

class EncoderDecoderInput(TypedDict):
    """
    A rendered [`EncoderDecoderPrompt`][vllm.inputs.llm.EncoderDecoderPrompt]
    which can be passed to `LLMEngine.add_request` or `AsyncLLM.add_request`.
    """

    type: Literal["enc_dec"]

    encoder_prompt: EncoderInput
    """The inputs for the encoder portion."""

    decoder_prompt: DecoderEngineInput
    """The inputs for the decoder portion."""

    arrival_time: NotRequired[float]
    """The time when the input was received (before rendering)."""

class MultiModalEncDecInput(MultiModalInput):
    """
    Represents multi-modal input to the engine for encoder-decoder models.

    Note:
        Even text-only encoder-decoder models are currently implemented
        as multi-modal models for convenience.
        (Example: https://github.com/vllm-project/bart-plugin)
    """

    encoder_prompt_token_ids: list[int]
    """The processed token IDs of the encoder prompt."""

    encoder_prompt: NotRequired[str]
    """The prompt text corresponding to the encoder token IDs, if available."""

class MultiModalInput(_InputOptions):
    """Represents multi-modal input to the engine."""

    type: Literal["multimodal"]
    """The type of input."""

    prompt_token_ids: list[int]
    """The processed token IDs which includes placeholder tokens."""

    prompt: NotRequired[str]
    """The prompt text corresponding to the token IDs, if available."""

    mm_kwargs: "MultiModalKwargsOptionalItems"
    """Keyword arguments to be directly passed to the model after batching."""

    mm_hashes: MultiModalHashes
    """The hashes of the multi-modal data."""

    mm_placeholders: MultiModalPlaceholders
    """
    For each modality, information about the placeholder tokens in
    `prompt_token_ids`.
    """

class TokensInput(_InputOptions):
    """Represents token-based input to the engine."""

    type: Literal["token"]
    """The type of input."""

    prompt_token_ids: list[int]
    """The token IDs of the prompt."""

    prompt: NotRequired[str]
    """The prompt text corresponding to the token IDs, if available."""

class _InputOptions(TypedDict):
    """
    Additional options available to all
    [`SingletonInput`][vllm.inputs.engine.SingletonInput] types.
    """

    arrival_time: NotRequired[float]
    """The time when the input was received (before rendering)."""

    cache_salt: NotRequired[str]
    """Optional cache salt to be used for prefix caching."""