vllm.forward_context

class BatchDescriptor(NamedTuple):
    """
    Batch descriptor for cudagraph dispatching. We should keep the num of
    items as minimal as possible to properly and uniquely describe the padded
    batch for cudagraph.
    """
    num_tokens: int
    uniform_decode: bool = False
    """
    False can also be used for an uniform decode batch to dispatch to the 
    cudagraph supporting non-uniform batches.
    """

    @property
    def non_uniform(self) -> "BatchDescriptor":
        """
        Return a non-uniform version of current batch descriptor.
        """
        return BatchDescriptor(self.num_tokens, uniform_decode=False)

@dataclass
class DPMetadata:
    max_tokens_across_dp_cpu: torch.Tensor
    cu_tokens_across_dp_cpu: torch.Tensor
    local_sizes: Optional[list[int]] = None

    @staticmethod
    def num_tokens_across_dp(num_tokens: int, dp_size: int,
                             dp_rank: int) -> torch.Tensor:
        """
        Gather the num_tokens across all DP ranks and return results in a
        CPU tensor of size dp_size.
        """
        num_tokens_across_dp = [0] * dp_size
        num_tokens_across_dp[dp_rank] = num_tokens
        num_tokens_tensor = torch.tensor(num_tokens_across_dp,
                                         device="cpu",
                                         dtype=torch.int32)
        from vllm.distributed.parallel_state import get_dp_group
        dist.all_reduce(num_tokens_tensor, group=get_dp_group().cpu_group)
        return num_tokens_tensor

    @staticmethod
    def make(
            parallel_config: ParallelConfig,
            attn_metadata: Any,
            num_tokens: int,
            num_tokens_across_dp: Optional[torch.Tensor] = None
    ) -> "DPMetadata":

        assert parallel_config.data_parallel_size > 1
        dp_size = parallel_config.data_parallel_size
        dp_rank = parallel_config.data_parallel_rank
        if attn_metadata is not None and hasattr(attn_metadata,
                                                 "num_prefill_tokens"):
            # for v0 attention backends
            batchsize = attn_metadata.num_prefill_tokens + \
                attn_metadata.num_decode_tokens
        else:
            # for v1 attention backends or no attn_metadata
            batchsize = num_tokens

        # If num_tokens_across_dp is None, it will be computed by all_reduce
        # Otherwise, num_tokens_across_dp[dp_rank] should be equal to batchsize
        assert (num_tokens_across_dp is None
                or num_tokens_across_dp[dp_rank] == batchsize)
        if num_tokens_across_dp is None:
            num_tokens_across_dp = DPMetadata.num_tokens_across_dp(
                batchsize, dp_size, dp_rank)
        max_tokens_across_dp_cpu = torch.max(num_tokens_across_dp)
        cu_tokens_across_dp_cpu = torch.cumsum(num_tokens_across_dp, dim=0)
        return DPMetadata(max_tokens_across_dp_cpu, cu_tokens_across_dp_cpu)

    @contextmanager
    def chunked_sizes(self, max_chunk_size_per_rank: int, chunk_idx: int):
        """
        Context manager to compute and temporarily set the per-rank local token
        sizes for a specific chunk during chunked forward execution.

        This is necessary to ensure each DP (data parallel) rank processes its
        designated portion of tokens in lockstep with others, even when the
        token counts are uneven or some ranks have completed their input early.

        For chunked execution, we break up the total tokens on each rank into
        multiple chunks (of at most `max_chunk_size_per_rank`), and for a given
        `chunk_idx`, this context manager sets `self.local_sizes` to the number
        of tokens to process in that chunk on each rank.

        It uses cumulative sizes (`cu_tokens_across_dp_cpu`) to derive the
        number of tokens per rank, and calls `_compute_chunked_local_num_tokens`
        to determine the chunk-wise split.

        `self.local_sizes` is only valid inside the context.

        Args:
            max_chunk_size_per_rank: The max number of tokens each rank is 
                                     allowed to process in this chunk.
            chunk_idx: The index of the chunk to compute sizes for.
        """
        cu_sizes = self.cu_tokens_across_dp_cpu
        num_tokens_across_dp_cpu = [
            (cu_sizes[i] -
             cu_sizes[i - 1]).item() if i > 0 else cu_sizes[0].item()
            for i in range(len(cu_sizes))
        ]
        self.local_sizes = _compute_chunked_local_num_tokens(
            num_tokens_across_dp_cpu, max_chunk_size_per_rank, chunk_idx)
        try:
            yield self.local_sizes
        finally:
            self.local_sizes = None

    def get_chunk_sizes_across_dp_rank(self) -> Optional[list[int]]:
        return self.local_sizes

@dataclass
class ForwardContext:
    # copy from vllm_config.compilation_config.static_forward_context
    no_compile_layers: dict[str, Any]
    """
    Type AttentionMetadata for v0, 
    Type Dict[str, AttentionMetadata] for v1, map from layer_name of each 
    attention layer to its attention metadata
    set dynamically for each forward pass
    """
    attn_metadata: Union["AttentionMetadata", dict[str, "AttentionMetadata"]]
    # TODO: remove after making all virtual_engines share the same kv cache
    virtual_engine: int  # set dynamically for each forward pass
    # set dynamically for each forward pass
    dp_metadata: Optional[DPMetadata] = None
    # determine the cudagraph style at runtime to be FULL, PIECEWISE, or NONE.
    # by default NONE, no cudagraph is used.
    cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE
    batch_descriptor: Optional[BatchDescriptor] = None

    def __post_init__(self):
        assert self.cudagraph_runtime_mode in [
            CUDAGraphMode.NONE, CUDAGraphMode.PIECEWISE, CUDAGraphMode.FULL], \
            f"Invalid cudagraph runtime mode: {self.cudagraph_runtime_mode}"