vllm.model_executor.models.interfaces_base ¶

@runtime_checkable
class VllmModel(Protocol[T_co]):
    """The interface required for all models in vLLM."""

    def __init__(self, vllm_config: VllmConfig, prefix: str = "") -> None: ...

    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
        """Apply token embeddings to `input_ids`."""
        ...

    def forward(self, input_ids: torch.Tensor, positions: torch.Tensor) -> T_co: ...

@runtime_checkable
class VllmModelForPooling(VllmModel[T_co], Protocol[T_co]):
    """The interface required for all pooling models in vLLM."""

    is_pooling_model: ClassVar[Literal[True]] = True
    """
    A flag that indicates this model supports pooling.

    Note:
        There is no need to redefine this flag if this class is in the
        MRO of your model class.
    """

    default_seq_pooling_type: ClassVar[SequencePoolingType] = "LAST"
    """
    Indicates the [vllm.config.pooler.PoolerConfig.seq_pooling_type][]
    to use by default.

    You can use the
    [vllm.model_executor.models.interfaces_base.default_pooling_type][]
    decorator to conveniently set this field.
    """

    default_tok_pooling_type: ClassVar[TokenPoolingType] = "ALL"
    """
    Indicates the [vllm.config.pooler.PoolerConfig.tok_pooling_type][]
    to use by default.

    You can use the
    [vllm.model_executor.models.interfaces_base.default_pooling_type][]
    decorator to conveniently set this field.
    """

    attn_type: ClassVar[AttnTypeStr] = "decoder"
    """
    Indicates the
    [vllm.config.model.ModelConfig.attn_type][]
    to use by default.

    You can use the
    [vllm.model_executor.models.interfaces_base.attn_type][]
    decorator to conveniently set this field.
    """

    score_type: ClassVar[ScoreType] = "bi-encoder"
    """
    Indicates the
    [vllm.config.model.ModelConfig.score_type][]
    to use by default.

    Score API handles score/rerank for:
    - "score" task (score_type: cross-encoder models)
    - "embed" task (score_type: bi-encoder models)
    - "token_embed" task (score_type: late interaction models)

    score_type defaults to bi-encoder, then the Score API uses the "embed" task.
    If you set score_type to cross-encoder via 
    [vllm.model_executor.models.interfaces.SupportsCrossEncoding][], 
    then the Score API uses the "score" task.
    If you set score_type to late-interaction via 
    [vllm.model_executor.models.interfaces.SupportsLateInteraction][], 
    then the Score API uses the "token_embed" task.    
    """

    pooler: Pooler
    """The pooler is only called on TP rank 0."""

@runtime_checkable
class VllmModelForTextGeneration(VllmModel[T], Protocol[T]):
    """The interface required for all generative models in vLLM."""

    def compute_logits(
        self,
        hidden_states: T,
    ) -> T | None:
        """Return `None` if TP rank > 0."""
        ...