vllm.model_executor.layers.pooler ¶

class DispatchPooler(Pooler):
    """Dispatches calls to a sub-pooler based on the pooling task."""

    @classmethod
    def for_embedding(cls, pooler_config: PoolerConfig):
        return cls(
            {
                "token_embed": pooler_for_token_embed(pooler_config),
                "embed": pooler_for_embed(pooler_config),
            },
        )

    @classmethod
    def for_seq_cls(
        cls,
        pooler_config: PoolerConfig,
        *,
        pooling: SequencePoolingMethod | SequencePoolingFn | None = None,
        classifier: ClassifierFn | None = None,
    ):
        return cls(
            {
                "token_classify": pooler_for_token_classify(
                    pooler_config,
                    pooling=AllPool(),
                    classifier=classifier,
                ),
                "classify": pooler_for_classify(
                    pooler_config,
                    pooling=pooling,
                    classifier=classifier,
                    act_fn="classify",
                ),
                "score": pooler_for_classify(
                    pooler_config,
                    pooling=pooling,
                    classifier=classifier,
                    act_fn="score",
                ),
            }
        )

    def __init__(self, poolers_by_task: Mapping[PoolingTask, Pooler]) -> None:
        super().__init__()

        for task, pooler in poolers_by_task.items():
            if task not in pooler.get_supported_tasks():
                raise ValueError(
                    f"{pooler=} does not support {task=}. "
                    f"Supported tasks: {pooler.get_supported_tasks()}"
                )

        self.poolers_by_task = poolers_by_task

    def get_supported_tasks(self) -> Set[PoolingTask]:
        return set(self.poolers_by_task)

    def get_pooling_updates(self, task: PoolingTask) -> PoolingParamsUpdate:
        return self.poolers_by_task[task].get_pooling_updates(task)

    def forward(
        self,
        hidden_states: torch.Tensor,
        pooling_metadata: PoolingMetadata,
    ) -> PoolerOutput:
        poolers_by_task = self.poolers_by_task

        outputs = list[torch.Tensor | None]()
        offset = 0
        for task, group in groupby(pooling_metadata.tasks):
            if not (pooler := poolers_by_task.get(task)):
                raise ValueError(
                    f"Unsupported task: {task!r} "
                    f"Supported tasks: {self.get_supported_tasks()}"
                )

            num_items = len(list(group))
            group_output: PoolerOutput = pooler(
                hidden_states,
                pooling_metadata[offset : offset + num_items],
            )

            outputs.extend(group_output)
            offset += num_items

        return outputs

    def extra_repr(self) -> str:
        s = f"supported_task={self.get_supported_tasks()}"
        return s

class IdentityPooler(Pooler):
    def get_supported_tasks(self) -> Set[PoolingTask]:
        return {"plugin", "score"}

    def forward(
        self,
        hidden_states: torch.Tensor,
        pooling_metadata: PoolingMetadata,
    ) -> PoolerOutput:
        return hidden_states

class Pooler(nn.Module, ABC):
    """The interface required for all poolers used in pooling models in vLLM."""

    @abstractmethod
    def get_supported_tasks(self) -> Set[PoolingTask]:
        """Determine which pooling tasks are supported."""
        raise NotImplementedError

    def get_pooling_updates(self, task: PoolingTask) -> PoolingParamsUpdate:
        """
        Construct the updated pooling parameters to use for a supported task.
        """
        return PoolingParamsUpdate()

    @abstractmethod
    def forward(
        self,
        hidden_states: torch.Tensor,
        pooling_metadata: PoolingMetadata,
    ) -> PoolerOutput:
        raise NotImplementedError

@dataclass(frozen=True)
class PoolingParamsUpdate:
    requires_token_ids: bool = False
    """Set this flag to enable `get_prompt_token_ids` for your pooler."""

    def __or__(self, other: "PoolingParamsUpdate") -> "PoolingParamsUpdate":
        return PoolingParamsUpdate(
            requires_token_ids=self.requires_token_ids or other.requires_token_ids,
        )

    def apply(self, params: PoolingParams) -> None:
        params.requires_token_ids = self.requires_token_ids