vllm.v1.worker.gpu.buffer_utils

Classes:

• FusedStagedWriter –

  Applies the staged writes of several StagedWriteTensors at once.

• NonUvaBuffer –

  Explicit-copy fallback for platforms without pinned memory.

• UvaBufferPool –

  Preallocate each slot at size, growing its first dimension as needed.

FusedStagedWriter

Applies the staged writes of several StagedWriteTensors at once.

Methods:

• apply –

  Apply and clear the staged writes of tensors with one kernel.

Source code in vllm/v1/worker/gpu/buffer_utils.py

class FusedStagedWriter:
    """Applies the staged writes of several `StagedWriteTensor`s at once."""

    def __init__(
        self, device: torch.device, max_writes: int, max_concurrency: int | None = None
    ):
        new_pool = partial(
            UvaBufferPool, dtype=torch.int32, max_concurrency=max_concurrency
        )
        self.group_ids = new_pool(max_writes)
        self.indices = new_pool(max_writes)
        self.starts = new_pool(max_writes)
        self.cu_lens = new_pool(max_writes)
        self.device = device

    def apply(
        self,
        tensors: Sequence[StagedWriteTensor],
        output_ptrs: torch.Tensor,
        output_strides: torch.Tensor,
    ) -> None:
        """Apply and clear the staged writes of `tensors` with one kernel."""
        group_ids: list[int] = []
        indices: list[int] = []
        starts: list[int] = []
        contents: list[int | float] = []
        cu_lens: list[int] = []

        for group_id, t in enumerate(tensors):
            n = len(t._staged_write_indices)
            if n == 0:
                continue

            group_ids.extend([group_id] * n)
            indices.extend(t._staged_write_indices)
            starts.extend(t._staged_write_starts)
            content_base = len(contents)
            contents.extend(t._staged_write_contents)
            cu_lens.extend(content_base + cu_len for cu_len in t._staged_write_cu_lens)

        if not group_ids:
            return

        group_ids_uva = self.group_ids.copy_to_uva(group_ids)
        indices_uva = self.indices.copy_to_uva(indices)
        starts_uva = self.starts.copy_to_uva(starts)
        cu_lens_uva = self.cu_lens.copy_to_uva(cu_lens)
        contents_gpu = async_tensor_h2d(contents, device=self.device, dtype=torch.int32)

        _apply_write_kernel[(len(group_ids),)](
            output_ptrs,
            output_strides,
            indices_uva,
            starts_uva,
            contents_gpu,
            cu_lens_uva,
            group_ids_uva,
            BLOCK_SIZE=1024,
            MULTI_GROUP=True,
        )
        for t in tensors:
            t.clear_staged_writes()

apply(tensors, output_ptrs, output_strides)

Apply and clear the staged writes of tensors with one kernel.

Source code in vllm/v1/worker/gpu/buffer_utils.py

def apply(
    self,
    tensors: Sequence[StagedWriteTensor],
    output_ptrs: torch.Tensor,
    output_strides: torch.Tensor,
) -> None:
    """Apply and clear the staged writes of `tensors` with one kernel."""
    group_ids: list[int] = []
    indices: list[int] = []
    starts: list[int] = []
    contents: list[int | float] = []
    cu_lens: list[int] = []

    for group_id, t in enumerate(tensors):
        n = len(t._staged_write_indices)
        if n == 0:
            continue

        group_ids.extend([group_id] * n)
        indices.extend(t._staged_write_indices)
        starts.extend(t._staged_write_starts)
        content_base = len(contents)
        contents.extend(t._staged_write_contents)
        cu_lens.extend(content_base + cu_len for cu_len in t._staged_write_cu_lens)

    if not group_ids:
        return

    group_ids_uva = self.group_ids.copy_to_uva(group_ids)
    indices_uva = self.indices.copy_to_uva(indices)
    starts_uva = self.starts.copy_to_uva(starts)
    cu_lens_uva = self.cu_lens.copy_to_uva(cu_lens)
    contents_gpu = async_tensor_h2d(contents, device=self.device, dtype=torch.int32)

    _apply_write_kernel[(len(group_ids),)](
        output_ptrs,
        output_strides,
        indices_uva,
        starts_uva,
        contents_gpu,
        cu_lens_uva,
        group_ids_uva,
        BLOCK_SIZE=1024,
        MULTI_GROUP=True,
    )
    for t in tensors:
        t.clear_staged_writes()

NonUvaBuffer

Explicit-copy fallback for platforms without pinned memory.

Source code in vllm/v1/worker/gpu/buffer_utils.py

class NonUvaBuffer:
    """Explicit-copy fallback for platforms without pinned memory."""

    def __init__(self, size: int | Sequence[int], dtype: torch.dtype):
        from vllm.platforms import current_platform

        logger.warning_once(
            "Pinned memory is not available on this platform; falling back "
            "to device memory for UVA buffers."
        )
        self.cpu = torch.zeros(size, dtype=dtype, device="cpu")
        self.np = self.cpu.numpy()
        self._uva = torch.zeros(size, dtype=dtype, device=current_platform.device_type)

    def uva(self, n: int | None = None) -> torch.Tensor:
        if n is None:
            return self._uva.copy_(self.cpu)
        return self._uva[:n].copy_(self.cpu[:n])

UvaBufferPool

Preallocate each slot at size, growing its first dimension as needed.

Callers must retire a slot's GPU readers before reuse, including growth.

Source code in vllm/v1/worker/gpu/buffer_utils.py

class UvaBufferPool:
    """Preallocate each slot at size, growing its first dimension as needed.

    Callers must retire a slot's GPU readers before reuse, including growth.
    """

    def __init__(
        self,
        size: int | Sequence[int],
        dtype: torch.dtype,
        max_concurrency: int | None = None,
    ):
        if max_concurrency is None:
            max_concurrency = _DEFAULT_MAX_CONCURRENCY
        self.size = size
        self.dtype = dtype
        self.max_concurrency = max_concurrency

        # UVA buffers for concurrency
        self._buffer_cls = UvaBuffer if is_uva_available() else NonUvaBuffer
        self._uva_bufs = [self._buffer_cls(size, dtype) for _ in range(max_concurrency)]
        # Current buffer index
        self._curr = 0

    def copy_to_uva(self, x: torch.Tensor | np.ndarray | list) -> torch.Tensor:
        # Round robin to the next buffer.
        self._curr = (self._curr + 1) % self.max_concurrency
        buf = self._uva_bufs[self._curr]
        n = len(x)
        if n > buf.cpu.shape[0]:
            capacity = 1 << (n - 1).bit_length()
            buf = self._buffer_cls((capacity, *buf.cpu.shape[1:]), self.dtype)
            self._uva_bufs[self._curr] = buf
        # CPU-to-CPU copy
        dst = buf.cpu if isinstance(x, torch.Tensor) else buf.np
        dst[:n] = x
        return buf.uva(n)

    def copy_to_gpu(
        self,
        x: torch.Tensor | np.ndarray,
        out: torch.Tensor | None = None,
    ) -> torch.Tensor:
        uva = self.copy_to_uva(x)
        # CPU-to-GPU copy
        return uva.clone() if out is None else out.copy_(uva, non_blocking=True)