vllm.v1.kv_offload.base ¶

class BlockIDsLoadStoreSpec(LoadStoreSpec, ABC):
    """
    Spec for loading/storing KV blocks from given block numbers.
    """

    def __init__(self, block_ids: list[int]):
        self.block_ids = np.array(block_ids, dtype=np.int64)

    def __repr__(self) -> str:
        return repr(self.block_ids)

@dataclass
class CanonicalKVCacheRef:
    """
    Per-layer (or group of layers) reference to a specific (by index)
    CanonicalKVCacheTensor and records the un-padded page size used by that layer.
    """

    # Index into the list of CanonicalKVCacheTensor objects
    tensor_idx: int
    # The un-padded page size per block in bytes
    page_size_bytes: int

@dataclass
class CanonicalKVCacheTensor:
    """
    A canonicalized KV cache tensor whose first dimension is num_blocks.

    For attention backends where the raw tensor has num_blocks at a
    non-leading physical dimension (e.g. FlashAttention's
    (2, num_blocks, ...) layout), the tensor is split so that each
    resulting CanonicalKVCacheTensor starts with (num_blocks, ...).
    """

    # The KV cache tensor with shape (num_blocks, ...)
    tensor: torch.Tensor
    # The (possibly padded) page size per block in bytes
    page_size_bytes: int

@dataclass
class CanonicalKVCaches:
    """
    Canonicalized block-level representation of the KV caches.

    Composed of:
        - Unique list of KV cache data tensors,
          each with shape (num_blocks, page_size_in_bytes) and int8 dtype.
        - Per-group data references of the tensors.
          i.e. how each KV cache group maps to the tensors.
    """

    # Ordered list of unique block tensors, each with shape
    # (num_blocks, ...).
    tensors: list[CanonicalKVCacheTensor]
    # Per-KV-cache-group list of data references that map each layer
    # in the group to the appropriate entry in the tensors list.
    group_data_refs: list[list[CanonicalKVCacheRef]]

class GPULoadStoreSpec(BlockIDsLoadStoreSpec):
    """
    Spec for loading/storing a KV block to GPU memory.

    If there are multiple KV groups, the blocks are expected to be
    ordered by the group index.
    In that case, group_sizes[i] determines the number of blocks
    per the i-th KV group, and thus sum(group_sizes) == len(block_ids).
    group_sizes=None indicates a single KV group.

    If block_indices is given, each group (determined by group_sizes) of block IDs
    will correspond to logically contiguous blocks, e.g. blocks 5-10 of a some request.
    block_indices[i] will represent the block index of the first block in group #i.
    Thus, len(block_indices) == len(group_sizes) = number of KV cache groups.
    This information is required in order to support off/loading from offloaded blocks
    which are larger than GPU blocks.
    In such cases, the first GPU block per each group may be unaligned to the offloaded
    block size, and so knowing block_indices[i] allows the worker to correctly
    skip part of the first matching offloaded block.
    """

    def __init__(
        self,
        block_ids: list[int],
        group_sizes: Sequence[int],
        block_indices: Sequence[int],
    ):
        super().__init__(block_ids)
        assert sum(group_sizes) == len(block_ids)
        assert len(block_indices) == len(group_sizes)
        self.group_sizes: Sequence[int] = group_sizes
        self.block_indices: Sequence[int] = block_indices

    @staticmethod
    @override
    def medium() -> str:
        return "GPU"

class LoadStoreSpec(ABC):
    """
    Abstract metadata that encapsulates information allowing a worker
    to load, and optionally also to store, blocks of KV data.
    """

    @staticmethod
    @abstractmethod
    def medium() -> str:
        """
        Returns a string representation of the medium type
        this store/load targets.
        """
        pass

class LookupResult(Enum):
    """Result of OffloadingManager.lookup()."""

    MISS = auto()
    HIT = auto()
    HIT_PENDING = auto()
    RETRY = auto()

class OffloadingManager(ABC):
    @abstractmethod
    def lookup(self, key: OffloadKey, req_context: ReqContext) -> LookupResult:
        """
        Checks whether a single block is offloaded and ready to be read.

        Args:
            key: the key identifying the block to lookup.
            req_context: per-request context (e.g. kv_transfer_params).

        Returns:
            HIT if the block is offloaded and ready, MISS if not found,
            HIT_PENDING if found but not yet readable, or RETRY if the
            lookup should be retried later.
        """
        pass

    @abstractmethod
    def prepare_load(
        self,
        keys: Collection[OffloadKey],
        req_context: ReqContext,
    ) -> LoadStoreSpec:
        """
        Prepare the given blocks to be read.
        The given blocks will be protected from eviction until
        complete_load is called.
        It assumes all given blocks are offloaded.

        Args:
            keys: the keys identifying the blocks.
            req_context: per-request context (e.g. kv_transfer_params).

        Returns:
            A LoadStoreSpec that can be used by a worker to locate and load
            the actual offloaded KV data.
        """
        pass

    def touch(self, keys: Collection[OffloadKey], req_context: ReqContext):
        """
        Mark the given blocks as recently used.
        This could in practice mean moving them to the end of an LRU list.

        Args:
            keys: the keys identifying the blocks.
            req_context: per-request context (e.g. kv_transfer_params).
        """
        return

    def complete_load(self, keys: Collection[OffloadKey], req_context: ReqContext):
        """
        Marks previous blocks that were prepared to load as done loading.

        Args:
            keys: the keys identifying the blocks.
            req_context: per-request context (e.g. kv_transfer_params).
        """
        return

    @abstractmethod
    def prepare_store(
        self,
        keys: Collection[OffloadKey],
        req_context: ReqContext,
    ) -> PrepareStoreOutput | None:
        """
        Prepare the given blocks to be offloaded.
        The given blocks will be protected from eviction until
        complete_store is called.

        Args:
            keys: the keys identifying the blocks.
            req_context: per-request context (e.g. kv_transfer_params).

        Returns:
            A PrepareStoreOutput indicating which blocks need storing,
            where to store them (LoadStoreSpec), and list of blocks that
            were evicted as a result.
            None is returned if the blocks cannot be stored.
        """
        pass

    def complete_store(
        self,
        keys: Collection[OffloadKey],
        req_context: ReqContext,
        success: bool = True,
    ):
        """
        Marks blocks which were previously prepared to be stored, as stored.
        Following this call, the blocks become loadable.
        If success is False, blocks that were not marked as stored will be
        removed.

        Args:
            keys: the keys identifying the blocks.
            req_context: per-request context (e.g. kv_transfer_params).
            success: whether the blocks were stored successfully.
        """
        return

    @abstractmethod
    def on_new_request(self, req_context: ReqContext) -> RequestOffloadingContext:
        """
        Called when a new request is first seen by the scheduler.

        Returns a RequestOffloadingContext indicating how this request's
        blocks should be offloaded.

        Args:
            req_context: per-request context.
        """
        pass

    def on_request_finished(self, req_context: ReqContext) -> None:
        """
        Called when a request has finished.

        By the time this is called, the scheduler will issue no more
        submit-side calls for this request, such as prepare_store() and
        prepare_load(). Completion callbacks for already-submitted transfers
        (complete_store() and complete_load()) may still arrive afterward.

        This hook does NOT imply the data has been persisted. Asynchronous
        transfers already submitted for this request may still be in flight.
        Managers that cascade to lower tiers should delay those tiers'
        on_request_finished() calls until no more lower-tier submit calls can
        be issued for this request.

        Args:
            req_context: per-request context.
        """
        return

    def take_events(self) -> Iterable[OffloadingEvent]:
        """
        Take the offloading events from the manager.

        Yields:
            New OffloadingEvents collected since the last call.
        """
        return ()

    def on_schedule_end(self) -> None:
        """Called once at the end of each scheduler step.

        Managers may override this to flush deferred work accumulated
        during the step (e.g., batched promotions).
        """
        return

    def has_pending_work(self) -> bool:
        """Whether this manager needs the engine to keep stepping.

        While True, on_schedule_end() and get_finished_jobs() continue
        to be called even when no requests are scheduled.
        """
        return False

    def reset_cache(self) -> None:
        """Evict all tracked blocks and reset internal state."""
        return

    def get_stats(self) -> "OffloadingConnectorStats | None":
        """Return collected metrics since last call, or None if disabled."""
        return None

    def shutdown(self) -> None:
        """Shutdown the manager and release any resources."""
        return

class OffloadingSpec(ABC):
    """Spec for an offloading connector"""

    @classmethod
    def build_metric_definitions(
        cls, extra_config: dict[str, Any]
    ) -> dict[str, "OffloadingMetricMetadata"]:
        """Return Prometheus metric definitions emitted by this spec."""
        return {}

    def __init__(self, vllm_config: "VllmConfig", kv_cache_config: "KVCacheConfig"):
        logger.warning(
            "Initializing OffloadingSpec. This API is experimental and "
            "subject to change in the future as we iterate the design."
        )
        self.vllm_config = vllm_config
        self.kv_cache_config = kv_cache_config

        kv_transfer_config = vllm_config.kv_transfer_config
        assert kv_transfer_config is not None
        self.extra_config = kv_transfer_config.kv_connector_extra_config
        kv_events_config = vllm_config.kv_events_config
        self.kv_events_config = OffloadingKVEventsConfig(
            enable_kv_cache_events=(
                kv_events_config is not None and kv_events_config.enable_kv_cache_events
            ),
            self_describing_kv_events=bool(
                self.extra_config.get("self_describing_kv_events", False)
            ),
        )

        # When True, only prompt (prefill) blocks are offloaded; decode-phase
        # blocks (KV generated after the prompt) are skipped. Useful when prior
        # turns' generated tokens are dropped before the next turn (e.g.
        # reasoning models that strip thinking).
        self.offload_prompt_only: bool = bool(
            self.extra_config.get("offload_prompt_only", True)
        )

        parallel_config = vllm_config.parallel_config
        context_parallel_factor = (
            parallel_config.decode_context_parallel_size
            * parallel_config.prefill_context_parallel_size
        )

        # gpu block size per group
        self.gpu_block_size: tuple[int, ...] = tuple(
            kv_cache_group.kv_cache_spec.block_size * context_parallel_factor
            for kv_cache_group in kv_cache_config.kv_cache_groups
        )

        # hash_block_size must match what the scheduler uses for
        # Request.block_hashes (resolved via resolve_kv_cache_block_sizes).
        _, self.hash_block_size = resolve_kv_cache_block_sizes(
            kv_cache_config, vllm_config
        )

        for block_size in self.gpu_block_size:
            assert block_size % self.hash_block_size == 0, (
                f"gpu_block_size={block_size} not divisible by "
                f"hash_block_size={self.hash_block_size}. "
                f"Hybrid models (e.g. Mamba+Attention) need "
                f"--enable-prefix-caching to align block sizes."
            )

        # offloaded_block_size / gpu_block_size
        self.block_size_factor: int = 1

        offloaded_block_size = self.extra_config.get("block_size")
        if offloaded_block_size is not None:
            offloaded_block_size_int = int(offloaded_block_size)
            gpu_block_sizes = set(self.gpu_block_size)
            assert len(gpu_block_sizes) == 1, (
                "If 'block_size' is specified in kv_connector_extra_config, "
                "there must be at least one KV cache group, "
                "and all groups must have the same block size."
            )
            gpu_block_size = gpu_block_sizes.pop()

            assert offloaded_block_size_int % gpu_block_size == 0
            self.block_size_factor = offloaded_block_size_int // gpu_block_size

    @abstractmethod
    def get_manager(self) -> OffloadingManager:
        """
        Get an OffloadingManager that will be used
        by the scheduler-side offloading connector to track
        offloaded blocks and manage evictions.
        """
        pass

    @abstractmethod
    def get_worker(self, kv_caches: CanonicalKVCaches) -> OffloadingWorker:
        """
        Get an OffloadingWorker that handles async KV transfers for this spec.

        Args:
            kv_caches: Canonicalized KV caches.

        Returns:
            An OffloadingWorker instance for this medium.
        """
        pass

class OffloadingWorker(ABC):
    """Runs in the worker process. Performs async KV transfers for ONE
    offloaded medium (e.g. CPU). Direction is explicit via submit_store /
    submit_load, so there is no (src_medium, dst_medium) routing."""

    @abstractmethod
    def submit_store(
        self, job_id: int, src_spec: GPULoadStoreSpec, dst_spec: LoadStoreSpec
    ) -> bool:
        """Async GPU -> offloaded medium."""

    @abstractmethod
    def submit_load(
        self, job_id: int, src_spec: LoadStoreSpec, dst_spec: GPULoadStoreSpec
    ) -> bool:
        """Async offloaded medium -> GPU."""

    @abstractmethod
    def get_finished(self) -> list[TransferResult]: ...

    @abstractmethod
    def wait(self, job_ids: set[int]) -> None: ...

    def shutdown(self) -> None:
        return