vllm.distributed.kv_transfer.kv_connector.v1 ¶

class DecodeBenchConnector(KVConnectorBase_V1):
    """
    A KV Connector for decode instance performance testing.

    This connector fills the KV cache with dummy (non-zero) values to
    emulate a prefill-decode disaggregated setting, enabling performance
    testing of the decoder with larger input sequence lengths.
    """

    def __init__(
        self,
        vllm_config: "VllmConfig",
        role: KVConnectorRole,
        kv_cache_config: Optional["KVCacheConfig"] = None,
    ):
        super().__init__(vllm_config, role, kv_cache_config)

        self.connector_scheduler: DecodeBenchConnectorScheduler | None = None
        self.connector_worker: DecodeBenchConnectorWorker | None = None

        if role == KVConnectorRole.SCHEDULER:
            self.connector_scheduler = DecodeBenchConnectorScheduler(vllm_config)
        elif role == KVConnectorRole.WORKER:
            self.connector_worker = DecodeBenchConnectorWorker(vllm_config)

    # ==============================
    # Worker-side methods
    # ==============================

    def register_kv_caches(self, kv_caches: dict[str, torch.Tensor]):
        assert self.connector_worker is not None
        self.connector_worker.register_kv_caches(kv_caches)

    def start_load_kv(self, forward_context: "ForwardContext", **kwargs: Any) -> None:
        assert self.connector_worker is not None
        assert isinstance(self._connector_metadata, DecodeBenchConnectorMetadata)
        self.connector_worker.start_fill_kv(self._connector_metadata)

    def wait_for_layer_load(self, layer_name: str) -> None:
        # All operations are synchronous, so nothing to wait for
        pass

    def save_kv_layer(
        self,
        layer_name: str,
        kv_layer: torch.Tensor,
        attn_metadata: "AttentionMetadata",
        **kwargs: Any,
    ) -> None:
        # This connector doesn't save KV cache (benchmarking only)
        pass

    def wait_for_save(self):
        # This connector doesn't save KV cache (benchmarking only)
        pass

    # ==============================
    # Scheduler-side methods
    # ==============================

    def get_num_new_matched_tokens(
        self,
        request: "Request",
        num_computed_tokens: int,
    ) -> tuple[int | None, bool]:
        assert self.connector_scheduler is not None
        return self.connector_scheduler.get_num_new_matched_tokens(
            request, num_computed_tokens
        )

    def update_state_after_alloc(
        self, request: "Request", blocks: "KVCacheBlocks", num_external_tokens: int
    ):
        assert self.connector_scheduler is not None
        return self.connector_scheduler.update_state_after_alloc(
            request, blocks, num_external_tokens
        )

    def build_connector_meta(
        self, scheduler_output: "SchedulerOutput"
    ) -> KVConnectorMetadata:
        assert self.connector_scheduler is not None
        return self.connector_scheduler.build_connector_meta(scheduler_output)

    def request_finished(
        self,
        request: "Request",
        block_ids: list[int],
    ) -> tuple[bool, dict[str, Any] | None]:
        assert self.connector_scheduler is not None
        self.connector_scheduler.request_finished(request)
        return False, None

class KVConnectorBase_V1(ABC):
    def __init__(
        self,
        vllm_config: "VllmConfig",
        role: KVConnectorRole,
        kv_cache_config: Optional["KVCacheConfig"] = None,
    ):
        logger.warning(
            "Initializing KVConnectorBase_V1. This API is experimental and "
            "subject to change in the future as we iterate the design."
        )
        self._connector_metadata: KVConnectorMetadata | None = None
        self._vllm_config = vllm_config
        if vllm_config.kv_transfer_config is not None:
            self._kv_transfer_config = vllm_config.kv_transfer_config
        else:
            raise ValueError("kv_transfer_config must be set for KVConnectorBase_V1")
        self._kv_cache_config = kv_cache_config
        if self._kv_cache_config is None:
            logger.warning(
                "KVConnectorBase_V1 initialized without kv_cache_config. "
                "This is deprecated - please update your connector to accept "
                "kv_cache_config as the third constructor argument and pass it "
                "to super().__init__()."
            )
        self._role = role

    @property
    def role(self) -> KVConnectorRole:
        return self._role

    # ==============================
    # Worker-side methods
    # ==============================

    def bind_connector_metadata(self, connector_metadata: KVConnectorMetadata) -> None:
        """Set the connector metadata from the scheduler.

        This function should be called by the model runner every time
        before the model execution. The metadata will be used for runtime
        KV cache loading and saving.

        Args:
            connector_metadata (dict): the connector metadata.
        """
        self._connector_metadata = connector_metadata

    def clear_connector_metadata(self) -> None:
        """Clear the connector metadata.

        This function should be called by the model runner every time
        after the model execution.
        """
        self._connector_metadata = None

    def _get_connector_metadata(self) -> KVConnectorMetadata:
        """Get the connector metadata.

        This function should only be called inside the connector.

        Returns:
            ConnectorMetadata: the connector metadata.
        """

        # Should only be called while set to valid metadata.
        assert self._connector_metadata is not None
        return self._connector_metadata

    def register_kv_caches(self, kv_caches: dict[str, torch.Tensor]):
        """
        Initialize with the KV caches. Useful for pre-registering the
        KV Caches in the KVConnector (e.g. for NIXL).

        Args:
            kv_caches: dictionary of layer names, kv cache
        """
        return

    def set_host_xfer_buffer_ops(self, copy_operation: CopyBlocksOp):
        """
        Set the xPU-specific ops for copying KV between host and device.
        Needed when host buffer is used for kv transfer (e.g., in NixlConnector)
        """
        return

    @abstractmethod
    def start_load_kv(self, forward_context: "ForwardContext", **kwargs: Any) -> None:
        """
        Start loading the KV cache from the connector to vLLM's paged
        KV buffer. This is called from the forward context before the
        forward pass to enable async loading during model execution.

        Args:
            forward_context (ForwardContext): the forward context.
            **kwargs: additional arguments for the load operation

        Note:
            The number of elements in kv_caches and layer_names should be
            the same.

        """
        pass

    @abstractmethod
    def wait_for_layer_load(self, layer_name: str) -> None:
        """
        Block until the KV for a specific layer is loaded into vLLM's
        paged buffer. This is called from within attention layer to ensure
        async copying from start_load_kv is complete.

        This interface will be useful for layer-by-layer pipelining.

        Args:
            layer_name: the name of that layer
        """
        pass

    @abstractmethod
    def save_kv_layer(
        self,
        layer_name: str,
        kv_layer: torch.Tensor,
        attn_metadata: "AttentionMetadata",
        **kwargs: Any,
    ) -> None:
        """
        Start saving a layer of KV cache from vLLM's paged buffer
        to the connector. This is called from within attention layer to
        enable async copying during execution.

        Args:
            layer_name (str): the name of the layer.
            kv_layer (torch.Tensor): the paged KV buffer of the current
                layer in vLLM.
            attn_metadata (AttentionMetadata): the attention metadata.
            **kwargs: additional arguments for the save operation.
        """
        pass

    @abstractmethod
    def wait_for_save(self):
        """
        Block until all the save operations is done. This is called
        as the forward context exits to ensure that the async saving
        from save_kv_layer is complete before finishing the forward.

        This prevents overwrites of paged KV buffer before saving done.
        """
        pass

    def get_finished(
        self, finished_req_ids: set[str]
    ) -> tuple[set[str] | None, set[str] | None]:
        """
        Notifies worker-side connector ids of requests that have
        finished generating tokens on the worker.
        The scheduler process (via the Executors) will use this output
        to track which workers are done.

        Returns:
            ids of requests that have finished asynchronous transfer
            (requests that previously returned True from request_finished()),
            tuple of (sending/saving ids, recving/loading ids).
            The finished saves/sends req ids must belong to a set provided in a
            call to this method (this call or a prior one).
        """
        return None, None

    def get_block_ids_with_load_errors(self) -> set[int]:
        """
        Get the set of block IDs that failed to load.

        Returns:
            Set of block IDs that encountered load errors.
            Empty set if no load errors occurred.

        Notes:
            - Applies to both sync- and async-loading requests.
            - Async loading: failed blocks may be reported in any forward pass
              up to and including the pass where the request ID is returned by
              `get_finished()`. Even if failures occur, the request must still
              be reported via `get_finished()`, and the failed block IDs must
              appear here no later than that same pass.
            - Sync loading: failed blocks should be reported in the forward
              pass in which they are detected.
        """
        return set()

    def shutdown(self):
        """
        Shutdown the connector. This is called when the worker process
        is shutting down to ensure that all the async operations are
        completed and the connector is cleaned up properly.
        """
        return None

    def get_kv_connector_stats(self) -> Optional["KVConnectorStats"]:
        """
        Get the KV connector stats collected during the last interval.
        """
        return None

    def get_handshake_metadata(self) -> KVConnectorHandshakeMetadata | None:
        """
        Get the KVConnector handshake metadata for this connector.
        This metadata is used for out-of-band connector handshake
        between P/D workers.

        Returns:
            KVConnectorHandshakeMetadata: the handshake metadata.
            None if no handshake metadata is available.
        """
        return None

    # ==============================
    # Scheduler-side methods
    # ==============================

    @abstractmethod
    def get_num_new_matched_tokens(
        self,
        request: "Request",
        num_computed_tokens: int,
    ) -> tuple[int | None, bool]:
        """
        Get number of new tokens that can be loaded from the
        external KV cache beyond the num_computed_tokens.

        Args:
            request (Request): the request object.
            num_computed_tokens (int): the number of locally
                computed tokens for this request

        Returns:
            A tuple with the following elements:
                - An optional number of tokens that can be loaded from the
                  external KV cache beyond what is already computed.
                  If None, it means that the connector needs more time to
                  determine the number of matched tokens, and the scheduler
                  should query for this request again later.
                - `True` if external KV cache tokens will be loaded
                  asynchronously (between scheduler steps). Must be
                  'False' if the first element is 0.

        Notes:
            The connector should only consider the largest prefix of prompt-
            tokens for which KV cache is actually available at the time of the
            call. If the cache cannot be loaded for some tokens (e.g., due to
            connectivity issues or eviction), those tokens must not be taken
            into account.
        """
        pass

    @abstractmethod
    def update_state_after_alloc(
        self, request: "Request", blocks: "KVCacheBlocks", num_external_tokens: int
    ):
        """
        Update KVConnector state after block allocation.

        If get_num_new_matched_tokens previously returned True for a
        request, this function may be called twice for that same request -
        first when blocks are allocated for the connector tokens to be
        asynchronously loaded into, and second when any additional blocks
        are allocated, after the load/transfer is complete.

        Args:
            request (Request): the request object.
            blocks (KVCacheBlocks): the blocks allocated for the request.
            num_external_tokens (int): the number of tokens that will be
                loaded from the external KV cache.
        """
        pass

    @abstractmethod
    def build_connector_meta(
        self, scheduler_output: SchedulerOutput
    ) -> KVConnectorMetadata:
        """
        Build the connector metadata for this step.

        This function should NOT modify fields in the scheduler_output.
        Also, calling this function will reset the state of the connector.

        Args:
            scheduler_output (SchedulerOutput): the scheduler output object.
        """
        pass

    def update_connector_output(self, connector_output: KVConnectorOutput):
        """
        Update KVConnector state from worker-side connectors output.

        Args:
            connector_output (KVConnectorOutput): the worker-side
                connectors output.
        """
        return

    def request_finished(
        self,
        request: "Request",
        block_ids: list[int],
    ) -> tuple[bool, dict[str, Any] | None]:
        """
        Called exactly once when a request has finished, before its blocks are
        freed.

        The connector may assumes responsibility for freeing the blocks
        asynchronously by returning True.

        Returns:
            True if the request is being saved/sent asynchronously and blocks
            should not be freed until the request_id is returned from
            get_finished().
            Optional KVTransferParams to be included in the request outputs
            returned by the engine.
        """
        return False, None

    def take_events(self) -> Iterable["KVCacheEvent"]:
        """
        Take the KV cache events from the connector.

        Yields:
            New KV cache events since the last call.
        """
        return ()

    @classmethod
    def get_required_kvcache_layout(cls, vllm_config: "VllmConfig") -> str | None:
        """
        Get the required KV cache layout for this connector.
        Args:
            vllm_config (VllmConfig): the vllm config.

        Returns:
            str: the required KV cache layout. e.g. HND, or NHD.
            None if the connector does not require a specific layout.
        """

        if cls is KVConnectorBase_V1:
            raise TypeError(
                "get_required_kvcache_layout should not be called "
                "on the abstract base class"
            )
        return None

    def get_finished_count(self) -> int | None:
        """
        Get the count of requests expected to complete send/receive operations
        via this connector. This method is used to initialize the
        KVOutputAggregator, overwriting the default world_size.

        Returns:
            int: expected sending or receiving completion count.
        """

        return None

    @classmethod
    def build_kv_connector_stats(
        cls, data: dict[str, Any] | None = None
    ) -> Optional["KVConnectorStats"]:
        """
        KVConnectorStats resolution method. This method allows dynamically
        registered connectors to return their own KVConnectorStats object,
        which can implement custom aggregation logic on the data dict.
        """
        return None

    def set_xfer_handshake_metadata(
        self, metadata: dict[int, KVConnectorHandshakeMetadata]
    ) -> None:
        """
        Set the KV connector handshake metadata for this connector.

        Args:
            metadata (KVConnectorHandshakeMetadata): the handshake metadata to set.
        """
        return None

    @classmethod
    def build_prom_metrics(
        cls,
        vllm_config: "VllmConfig",
        metric_types: dict[type["PromMetric"], type["PromMetricT"]],
        labelnames: list[str],
        per_engine_labelvalues: dict[int, list[str]],
    ) -> Optional["KVConnectorPromMetrics"]:
        """
        Create a KVConnectorPromMetrics subclass which should register
        per-connector Prometheus metrics and implement observe() to
        expose connector transfer stats via Prometheus.
        """
        return None