paged_attn.py

from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple

import torch

from vllm._C import cache_ops, ops
from vllm.attention.ops.prefix_prefill import context_attention_fwd

# Should be the same as PARTITION_SIZE in `paged_attention_v2_launcher`.
_PARTITION_SIZE = 512


@dataclass
class PagedAttentionMetadata:
    """Metadata for PagedAttention."""
    # (num_tokens,). The indices of the token slots that input tokens will be
    # stored into. E.g., if `slot_mapping` is [35, 2, 17] and the block size
    # is 16, the three tokens are stored in the 3rd slot in block 2, 2nd slot
    # in block 0, and 1st slot in block 1, respectively.
    slot_mapping: torch.Tensor
    # (batch_size,). The length of context (tokens stored in KV cache) per
    # sequence. WARNING: When it is a prefill request, it doesn't include new
    # tokens. When it is for decoding, it includes a new token.
    context_lens: Optional[torch.Tensor]
    # Maximum context length in the batch.
    max_context_len: Optional[int]
    # (batch_size, max_blocks_per_seq).
    # Block addresses per sequence. (Seq id -> list of physical block)
    # E.g., [0, 1, 2] means tokens are stored in 0th, 1st, and 2nd blocks
    # in the kv cache. Each block can contain up to block_size tokens.
    # 2nd dimensions are padded up to max_blocks_per_seq if it is cuda-graph
    # captured.
    block_tables: Optional[torch.Tensor]
    kv_cache_dtype: str


class PagedAttention:

    @staticmethod
    def get_supported_head_sizes() -> List[int]:
        return [64, 80, 96, 112, 128, 256]

    @staticmethod
    def get_kv_cache_shape(
        num_blocks: int,
        block_size: int,
        num_kv_heads: int,
        head_size: int,
    ) -> Tuple[int, ...]:
        return (2, num_blocks, block_size * num_kv_heads * head_size)

    @staticmethod
    def split_kv_cache(
        kv_cache: torch.Tensor,
        num_kv_heads: int,
        head_size: int,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        x = 16 // kv_cache.element_size()
        num_blocks = kv_cache.shape[1]

        key_cache = kv_cache[0]
        key_cache = key_cache.view(num_blocks, num_kv_heads, head_size // x,
                                   -1, x)
        value_cache = kv_cache[1]
        value_cache = value_cache.view(num_blocks, num_kv_heads, head_size, -1)
        return key_cache, value_cache

    @staticmethod
    def write_to_paged_cache(
        key: torch.Tensor,
        value: torch.Tensor,
        key_cache: torch.Tensor,
        value_cache: torch.Tensor,
        slot_mapping: torch.Tensor,
        kv_cache_dtype: str,
    ) -> None:
        cache_ops.reshape_and_cache(
            key,
            value,
            key_cache,
            value_cache,
            slot_mapping.flatten(),
            kv_cache_dtype,
        )

    @staticmethod
    def forward_decode(
        query: torch.Tensor,
        key_cache: torch.Tensor,
        value_cache: torch.Tensor,
        block_tables: torch.Tensor,
        context_lens: torch.Tensor,
        max_context_len: int,
        kv_cache_dtype: str,
        num_kv_heads: int,
        scale: float,
        alibi_slopes: Optional[torch.Tensor],
    ) -> torch.Tensor:
        output = torch.empty_like(query)

        block_size = value_cache.shape[3]
        num_seqs, num_heads, head_size = query.shape
        max_num_partitions = ((max_context_len + _PARTITION_SIZE - 1) //
                              _PARTITION_SIZE)
        # NOTE(woosuk): We use a simple heuristic to decide whether to use
        # PagedAttention V1 or V2. If the number of partitions is 1, we use
        # V1 to avoid the overhead of reduction. Also, if the number of
        # sequences or heads is large, we use V1 since there is enough work
        # to parallelize.
        # TODO(woosuk): Tune this heuristic.
        # For context len > 8192, use V2 kernel to avoid shared memory shortage.
        use_v1 = (max_context_len <= 8192
                  and (max_num_partitions == 1 or num_seqs * num_heads > 512))
        if use_v1:
            # Run PagedAttention V1.
            ops.paged_attention_v1(
                output,
                query,
                key_cache,
                value_cache,
                num_kv_heads,
                scale,
                block_tables,
                context_lens,
                block_size,
                max_context_len,
                alibi_slopes,
                kv_cache_dtype,
            )
        else:
            # Run PagedAttention V2.
            assert _PARTITION_SIZE % block_size == 0
            tmp_output = torch.empty(
                size=(num_seqs, num_heads, max_num_partitions, head_size),
                dtype=output.dtype,
                device=output.device,
            )
            exp_sums = torch.empty(
                size=(num_seqs, num_heads, max_num_partitions),
                dtype=torch.float32,
                device=output.device,
            )
            max_logits = torch.empty_like(exp_sums)
            ops.paged_attention_v2(
                output,
                exp_sums,
                max_logits,
                tmp_output,
                query,
                key_cache,
                value_cache,
                num_kv_heads,
                scale,
                block_tables,
                context_lens,
                block_size,
                max_context_len,
                alibi_slopes,
                kv_cache_dtype,
            )
        return output

    @staticmethod
    def forward_prefix(
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        key_cache: torch.Tensor,
        value_cache: torch.Tensor,
        block_tables: torch.Tensor,
        subquery_start_loc: torch.Tensor,
        prompt_lens_tensor: torch.Tensor,
        context_lens: torch.Tensor,
        max_subquery_len: int,
        alibi_slopes: Optional[torch.Tensor],
    ) -> torch.Tensor:
        output = torch.empty_like(query)
        context_attention_fwd(
            query,
            key,
            value,
            output,
            key_cache,
            value_cache,
            block_tables,
            # subquery_start_loc is (batch_size + 1,)
            subquery_start_loc[:-1],
            prompt_lens_tensor,
            context_lens,
            max_subquery_len,
            alibi_slopes,
        )
        return output

    @staticmethod
    def swap_blocks(
        src_kv_cache: torch.Tensor,
        dst_kv_cache: torch.Tensor,
        src_to_dst: Dict[int, int],
    ) -> None:
        src_key_cache = src_kv_cache[0]
        dst_key_cache = dst_kv_cache[0]
        cache_ops.swap_blocks(src_key_cache, dst_key_cache, src_to_dst)

        src_value_cache = src_kv_cache[1]
        dst_value_cache = dst_kv_cache[1]
        cache_ops.swap_blocks(src_value_cache, dst_value_cache, src_to_dst)

    @staticmethod
    def copy_blocks(
        kv_caches: List[torch.Tensor],
        src_to_dists: Dict[int, List[int]],
    ) -> None:
        key_caches = [kv_cache[0] for kv_cache in kv_caches]
        value_caches = [kv_cache[1] for kv_cache in kv_caches]
        cache_ops.copy_blocks(key_caches, value_caches, src_to_dists)