vllm.v1.outputs ¶

class AsyncModelRunnerOutput(ABC):
    @abstractmethod
    def get_output(self) -> ModelRunnerOutput:
        """Get the ModelRunnerOutput for this async output.

        This is a blocking call that waits until the results are ready, which
        might involve copying device tensors to the host.
        This method should only be called once per AsyncModelRunnerOutput.
        """
        pass

class LogprobsTensors(NamedTuple):
    # [num_reqs x num_generated_tokens, max_num_logprobs + 1]
    logprob_token_ids: torch.Tensor
    # [num_reqs x num_generated_tokens, max_num_logprobs + 1]
    logprobs: torch.Tensor
    # [num_reqs x num_generated_tokens]
    selected_token_ranks: torch.Tensor
    # [num_reqs]
    cu_num_generated_tokens: list[int] | None = None

    def tolists(self, cu_num_generated_tokens: list[int] | None = None):
        return LogprobsLists(
            self.logprob_token_ids.cpu().numpy(),
            self.logprobs.cpu().numpy(),
            self.selected_token_ranks.cpu().numpy(),
            cu_num_generated_tokens
            if cu_num_generated_tokens is not None
            else self.cu_num_generated_tokens,
        )

    def to_cpu_nonblocking(self) -> "LogprobsTensors":
        if self.logprob_token_ids.device.type == "cpu":
            return self
        return LogprobsTensors(
            self.logprob_token_ids.to("cpu", non_blocking=True),
            self.logprobs.to("cpu", non_blocking=True),
            self.selected_token_ranks.to("cpu", non_blocking=True),
            self.cu_num_generated_tokens,
        )

    def filter(self, mask: torch.Tensor) -> "LogprobsTensors":
        """Filter the logprobs tensors with the given bool mask."""
        assert self.cu_num_generated_tokens is None, (
            "filter can't be used with cu_num_generated_tokens"
        )
        return LogprobsTensors(
            self.logprob_token_ids[mask],
            self.logprobs[mask],
            self.selected_token_ranks[mask],
        )

    @staticmethod
    def empty_cpu(
        num_positions: int, num_tokens_per_position: int
    ) -> "LogprobsTensors":
        """Create empty LogprobsTensors on CPU."""

        logprob_token_ids = torch.empty(
            (num_positions, num_tokens_per_position), dtype=torch.int32, device="cpu"
        )
        logprobs = torch.empty_like(logprob_token_ids, dtype=torch.float32)
        selected_token_ranks = torch.empty(
            num_positions, dtype=torch.int32, device="cpu"
        )
        return LogprobsTensors(
            logprob_token_ids=logprob_token_ids,
            logprobs=logprobs,
            selected_token_ranks=selected_token_ranks,
        )

@dataclass
class ModelRunnerOutput:
    # [num_reqs]
    req_ids: list[str]
    # req_id -> index
    req_id_to_index: dict[str, int]

    # num_reqs x num_generated_tokens
    # num_generated_tokens is the number of tokens
    # generated in the current step. It can be different for
    # each request due to speculative/jump decoding.
    sampled_token_ids: list[list[int]] = field(default_factory=list)

    # [num_reqs, max_num_logprobs + 1]
    # [num_reqs, max_num_logprobs + 1]
    # [num_reqs]
    logprobs: LogprobsLists | None = None

    # req_id -> (token_ids, logprobs, ranks)
    # [prompt_len, num_prompt_logprobs]
    # [prompt_len, num_prompt_logprobs]
    # [prompt_len]
    prompt_logprobs_dict: dict[str, LogprobsTensors | None] = field(
        default_factory=dict
    )

    # [num_reqs, hidden_size]
    pooler_output: list[torch.Tensor | None] | None = None

    kv_connector_output: KVConnectorOutput | None = None

    ec_connector_output: ECConnectorOutput | None = None

    # req_id -> num_nans_in_logits
    num_nans_in_logits: dict[str, int] | None = None

    # information related to cudagraph execution
    cudagraph_stats: CUDAGraphStat | None = None

    # Per-step routed experts data captured by the worker.
    # ``routing_data`` shape: (num_scheduled_tokens, num_layers,
    #                         num_experts_per_tok); expert IDs as uint8/uint16.
    # ``slot_mapping`` shape: (num_scheduled_tokens,); physical KV-cache
    #                         slot for each row of routing_data.
    # ``num_scheduled_tokens`` is step-level (total across all requests
    # in this step), not per-request. The scheduler persists this into
    # its slot buffer via ``slot_buffer[slot_mapping] = routing_data``.
    # ``None`` when ``enable_return_routed_experts`` is off.
    routed_experts: RoutedExpertsLists | None = None

    @staticmethod
    def with_kv_conn_output_only(
        kv_connector_output: KVConnectorOutput | None,
    ) -> "ModelRunnerOutput":
        """Return ModelRunnerOutput containing the provided KVConnectorOutput,
        otherwise empty. Returns None if kv_connector_output is passed as None.
        """
        if kv_connector_output is None or kv_connector_output.is_empty():
            return EMPTY_MODEL_RUNNER_OUTPUT
        output = copy(EMPTY_MODEL_RUNNER_OUTPUT)
        output.kv_connector_output = kv_connector_output
        return output

class RoutedExpertsLists(NamedTuple):
    """CPU-side routed experts, the form :meth:`RoutedExpertsManager.store_batch`
    consumes.

    Batched per scheduler step: the leading dim is the number of tokens
    scheduled across all requests in this step (``total_num_scheduled_tokens``),
    not per-request tokens. ``slot_mapping[i]`` tells the scheduler which
    physical KV-cache slot row ``i`` of ``routing_data`` belongs to.
    """

    # (num_scheduled_tokens, num_layers, num_experts_per_tok)
    routing_data: np.ndarray
    # (num_scheduled_tokens,)
    slot_mapping: np.ndarray

class RoutedExpertsTensors(NamedTuple):
    """Device-side snapshot of routed experts data, pending async D2H.

    Produced by :class:`GPUModelRunner` at the end of each async-scheduled
    step. The copy stream waits on the default stream, then issues
    non-blocking D2H via :meth:`to_cpu_nonblocking` into a pinned CPU
    buffer; :class:`AsyncGPUModelRunnerOutput.get_output` synchronizes
    the copy before the scheduler reads it.

    Sliced to ``total_num_scheduled_tokens`` (step-level, across all
    requests — NOT per-request). Both ``routing_data`` and
    ``slot_mapping`` must be private clones when sourced from shared
    capturer / prepare-input buffers, so the next forward pass /
    ``_prepare_inputs`` on the default stream does not race with a
    D2H still pending on the copy stream.
    """

    # (num_scheduled_tokens, num_layers, num_experts_per_tok)
    routing_data: torch.Tensor
    # (num_scheduled_tokens,)
    slot_mapping: torch.Tensor

    def to_cpu_nonblocking(self) -> "RoutedExpertsTensors":
        """Issue non-blocking D2H on the current stream.

        NOTE: ``non_blocking=True`` only delivers true overlap when the
        CPU target is pinned. The current fallback here allocates a
        new pageable CPU tensor per call, which silently degrades to a
        synchronous copy; acceptable because the sync happens on the
        dedicated copy stream, not the default stream.
        """
        if self.routing_data.device.type == "cpu":
            return self
        return RoutedExpertsTensors(
            self.routing_data.to("cpu", non_blocking=True),
            self.slot_mapping.to("cpu", non_blocking=True),
        )

    def tolists(self) -> "RoutedExpertsLists":
        """Convert to the numpy-backed form consumed by the scheduler.

        ``.cpu()`` is a no-op when the tensor is already on CPU, so this
        is cheap for the post-D2H case; for raw device tensors it will
        synchronously block, which is only reached in tests.
        """
        return RoutedExpertsLists(
            self.routing_data.cpu().numpy(),
            self.slot_mapping.cpu().numpy(),
        )