vllm.v1.worker.gpu.block_table ¶

class BlockTables:
    def __init__(
        self,
        block_sizes: list[int],
        max_num_reqs: int,
        max_num_batched_tokens: int,
        max_model_len: int,
        device: torch.device,
    ):
        self.block_sizes = block_sizes
        self.max_num_reqs = max_num_reqs
        self.max_num_batched_tokens = max_num_batched_tokens
        self.max_model_len = max_model_len
        self.device = device

        if not is_uva_available():
            raise RuntimeError("UVA is not available")

        self.num_kv_cache_groups = len(self.block_sizes)
        # num_kv_cache_groups x [max_num_reqs, max_num_blocks]
        self.block_tables: list[UvaBuffer] = []
        for i in range(self.num_kv_cache_groups):
            block_size = self.block_sizes[i]
            max_num_blocks = cdiv(self.max_model_len, block_size)
            block_table = UvaBuffer(
                self.max_num_reqs,
                max_num_blocks,
                dtype=torch.int32,
            )
            self.block_tables.append(block_table)
        self.block_table_ptrs = self._make_ptr_tensor(
            [b.gpu for b in self.block_tables]
        )
        self.block_table_strides = torch.tensor(
            [b.gpu.stride(0) for b in self.block_tables],
            dtype=torch.int64,
            device=self.device,
        )

        self.block_sizes_tensor = torch.tensor(
            self.block_sizes, dtype=torch.int32, device=self.device
        )
        self.num_blocks = UvaBuffer(
            self.num_kv_cache_groups,
            self.max_num_reqs,
            dtype=torch.int32,
        )

        # Block tables used for model's forward pass.
        # num_kv_cache_groups x [max_num_reqs, max_num_blocks]
        self.input_block_tables: list[torch.Tensor] = [
            torch.zeros_like(b.gpu) for b in self.block_tables
        ]
        self.input_block_table_ptrs = self._make_ptr_tensor(self.input_block_tables)

        self.slot_mappings = torch.zeros(
            self.num_kv_cache_groups,
            self.max_num_batched_tokens,
            dtype=torch.int64,
            device=self.device,
        )

    def _make_ptr_tensor(self, x: Iterable[torch.Tensor]) -> torch.Tensor:
        # NOTE(woosuk): Use uint64 instead of int64 to cover all possible addresses.
        ptrs_tensor_cpu = torch.tensor(
            [t.data_ptr() for t in x],
            dtype=torch.uint64,
            device="cpu",
            pin_memory=True,
        )
        return ptrs_tensor_cpu.to(self.device, non_blocking=True)

    def append_block_ids(
        self,
        req_index: int,
        new_block_ids: tuple[list[int], ...],
        overwrite: bool,
    ) -> None:
        for i in range(self.num_kv_cache_groups):
            block_ids = new_block_ids[i]
            num_new_blocks = len(block_ids)
            if num_new_blocks == 0:
                continue

            # TODO(woosuk): Too many Numpy invocations. Optimize this.
            start = self.num_blocks.np[i, req_index] if not overwrite else 0
            end = start + num_new_blocks
            if num_new_blocks == 1:
                self.block_tables[i].np[req_index, start] = block_ids[0]
            else:
                self.block_tables[i].np[req_index, start:end] = block_ids
            self.num_blocks.np[i, req_index] = end

    def gather_block_tables(
        self,
        idx_mapping: torch.Tensor,
    ) -> tuple[torch.Tensor, ...]:
        num_reqs = idx_mapping.shape[0]
        _gather_block_tables_kernel[(self.num_kv_cache_groups, num_reqs)](
            idx_mapping,
            self.block_table_ptrs,
            self.input_block_table_ptrs,
            self.block_table_strides,
            self.num_blocks.gpu,
            self.num_blocks.gpu.stride(0),
            BLOCK_SIZE=1024,  # type: ignore
        )
        return tuple(block_table[:num_reqs] for block_table in self.input_block_tables)

    def get_dummy_block_tables(self, num_reqs: int) -> tuple[torch.Tensor, ...]:
        return tuple(block_table[:num_reqs] for block_table in self.input_block_tables)

    def compute_slot_mappings(
        self,
        query_start_loc: torch.Tensor,
        positions: torch.Tensor,
    ) -> torch.Tensor:
        num_reqs = query_start_loc.shape[0] - 1
        num_tokens = positions.shape[0]
        num_groups = self.num_kv_cache_groups
        _compute_slot_mappings_kernel[(num_groups, num_reqs + 1)](
            num_tokens,
            self.max_num_batched_tokens,
            query_start_loc,
            positions,
            self.input_block_table_ptrs,
            self.block_table_strides,
            self.block_sizes_tensor,
            self.slot_mappings,
            self.slot_mappings.stride(0),
            PAD_ID=PAD_SLOT_ID,
            BLOCK_SIZE=1024,  # type: ignore
        )
        return self.slot_mappings[:, :num_tokens]

    def get_dummy_slot_mappings(self, num_tokens: int) -> torch.Tensor:
        self.slot_mappings.fill_(PAD_SLOT_ID)
        return self.slot_mappings[:, :num_tokens]

class UvaBuffer:
    def __init__(self, *size, dtype: torch.dtype):
        self.cpu = torch.zeros(*size, dtype=dtype, device="cpu", pin_memory=True)
        self.np = self.cpu.numpy()
        self.gpu = get_cuda_view_from_cpu_tensor(self.cpu)