flash_attn.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Attention layer with FlashAttention."""
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, ClassVar, Optional

import numpy as np
import torch

from vllm import _custom_ops as ops
from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                              AttentionMetadata, AttentionType,
                                              is_quantized_kv_cache)
from vllm.attention.layer import Attention
from vllm.attention.ops.merge_attn_states import merge_attn_states
from vllm.attention.utils.fa_utils import (flash_attn_supports_fp8,
                                           get_flash_attn_version)
from vllm.config import VllmConfig, get_layers_from_vllm_config
from vllm.distributed.kv_transfer.kv_connector.utils import (
    get_kv_connector_cache_layout)
from vllm.logger import init_logger
from vllm.platforms import current_platform
from vllm.utils import cdiv
from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
                                              CommonAttentionMetadata)
from vllm.v1.kv_cache_interface import AttentionSpec
from vllm.v1.worker.block_table import BlockTable

if TYPE_CHECKING:
    from vllm.v1.worker.gpu_model_runner import GPUModelRunner

if current_platform.is_cuda():
    from vllm.vllm_flash_attn import (flash_attn_varlen_func,
                                      get_scheduler_metadata)

logger = init_logger(__name__)


class FlashAttentionBackend(AttentionBackend):

    accept_output_buffer: bool = True

    @staticmethod
    def get_supported_head_sizes() -> list[int]:
        return [32, 64, 96, 128, 160, 192, 224, 256]

    @staticmethod
    def get_name() -> str:
        return "FLASH_ATTN_VLLM_V1"

    @staticmethod
    def get_impl_cls() -> type["FlashAttentionImpl"]:
        return FlashAttentionImpl

    @staticmethod
    def get_metadata_cls() -> type["AttentionMetadata"]:
        return FlashAttentionMetadata

    @staticmethod
    def get_builder_cls() -> type["FlashAttentionMetadataBuilder"]:
        return FlashAttentionMetadataBuilder

    @staticmethod
    def get_kv_cache_shape(
        num_blocks: int,
        block_size: int,
        num_kv_heads: int,
        head_size: int,
    ) -> tuple[int, ...]:
        if block_size % 16 != 0:
            raise ValueError("Block size must be a multiple of 16.")
        return (2, num_blocks, block_size, num_kv_heads, head_size)

    @staticmethod
    def get_kv_cache_stride_order() -> tuple[int, ...]:
        # NOTE When running disaggregated PD with NIXL, HND layout is used for
        # faster transfer. `stride_order` indicates the permutation that gets
        # us from `get_kv_cache_shape` to the actual memory layout we want.
        cache_layout = get_kv_connector_cache_layout()
        if cache_layout == "NHD":
            stride_order = (0, 1, 2, 3, 4)
        elif cache_layout == "HND":
            stride_order = (0, 1, 3, 2, 4)
        else:
            raise ValueError("Unknown cache layout format %s.", cache_layout)
        return stride_order


@dataclass
class FlashAttentionMetadata:
    # NOTE(sang): Definition of context_len, query_len, and seq_len.
    # |---------- N-1 iteration --------|
    # |---------------- N iteration ---------------------|
    # |- tokenA -|......................|-- newTokens ---|
    # |---------- context_len ----------|
    # |-------------------- seq_len ---------------------|
    #                                   |-- query_len ---|

    num_actual_tokens: int  # Number of tokens excluding padding.
    max_query_len: int
    query_start_loc: torch.Tensor
    max_seq_len: int
    seq_lens: torch.Tensor
    block_table: torch.Tensor
    slot_mapping: torch.Tensor

    # For cascade attention.
    use_cascade: bool
    common_prefix_len: int
    cu_prefix_query_lens: Optional[torch.Tensor]
    prefix_kv_lens: Optional[torch.Tensor]
    suffix_kv_lens: Optional[torch.Tensor]

    # Optional aot scheduling
    scheduler_metadata: Optional[torch.Tensor] = None
    prefix_scheduler_metadata: Optional[torch.Tensor] = None

    # for local attention
    @dataclass
    class LocalAttentionMetadata:
        local_query_start_loc: torch.Tensor
        local_seqused_k: torch.Tensor
        local_block_table: torch.Tensor
        local_max_query_len: int
        local_max_seq_len: int
        local_scheduler_metadata: Optional[torch.Tensor]

    local_attn_metadata: Optional[LocalAttentionMetadata] = None


#
# Take in `query_start_loc_np` and `seq_lens_np` and break the sequences into
# local attention blocks, where each block is passed to the attention kernel
# as an independent local ("virtual") batch item.
#
# For example, if are performing a chunked prefill a batch of 3 sequences:
#   q_seqlens  = [4, 10, 5]
#   kv_seqlens = [6, 17, 9]
# Then normally for regular attention we would compute with an attention mask
#  for batch idx 0 (q_seqlens = 4, kv_seqlens = 6) like:
#   batch idx: 0 (q_seqlens = 4, kv_seqlens = 6)
#        k_toks >   0 1 2 3 4 5
#        q_toks v  _____________
#               0 | 1 1 1
#               1 | 1 1 1 1
#               2 | 1 1 1 1 1
#               3 | 1 1 1 1 1 1
#
# for local attention (with attn_chunk_size = 4) we would compute with an
#  attention mask like:
#   batch idx: 0  (q_seqlens = 4, kv_seqlens = 6, attn_chunk_size = 4)
#        k_toks >   0 1 2 3 4 5
#        q_toks v  _____________
#               0 | 1 1 1
#               1 | 1 1 1 1
#               2 |         1
#               3 |         1 1
#
# We can simulate this mask using standard flash-attention by breaking the
#  sequences into local ("virtual") batches, where each local batch item is a
#  local attention block, so in this case batch idx 0 would be broken up into:
#
#   local-batch idx: 0 (q_seqlens = 2, kv_seqlens = 4)  (batch 0)
#        k_toks >   0 1 2 3
#        q_toks v  _____________
#               0 | 1 1 1
#               1 | 1 1 1 1
#   local-batch idx: 1 (q_seqlens = 2, kv_seqlens = 2) (batch 0)
#        k_toks >   4 5
#        q_toks v  _____________
#               2 | 1
#               3 | 1 1
#
# e.g. if we have:
#   attn_chunk_size = 4
#   query_start_loc_np = [0, 4, 14, 19] (q_seqlens = [4, 10, 5])
# Then this function would return:
#                           __b0__  ______b1______  __b2__ < orig batch indices
#   q_seqlens_local    = [   2,  2,  1,  4,  4,  1,  4,  1]
#   cu_seqlens_q_local = [0, 4,  6, 10, 14, 18, 19, 23, 24]
#   seqlens_k_local    = [   4,  2,  4,  4,  4,  1,  4,  1]
#   block_table_local  : shape[local_virtual_batches, pages_per_local_batch]
def make_local_attention_virtual_batches(
    attn_chunk_size: int,
    query_start_loc_np: np.ndarray,
    seq_lens_np: np.ndarray,
    block_table: torch.Tensor,
    block_size: int = 0,
) -> tuple[np.ndarray, np.ndarray, np.ndarray, torch.Tensor]:
    q_seqlens = query_start_loc_np[1:] - query_start_loc_np[:-1]
    actual_batch_size = seq_lens_np.shape[0]

    # Handle if we are starting in the middle of a local attention block,
    #  we assume q_seqlens > 0 (for all elements), for each batch idx we compute
    #  the number of tokens that are not in the first local attention block and
    #  then we can simply use a cdiv for the rest.
    # For example if we have:
    #   attn_chunk_size = 4
    #   q_seqlens = [4, 10, 5]
    #   k_seqlens = [6, 17, 9]
    # Then we would get:
    #   new_tokens_in_first_block = [2, 1, 4]
    #   local_blocks = [2, 4, 2]
    q_tokens_in_first_block = np.minimum(
        attn_chunk_size - ((seq_lens_np - q_seqlens) % attn_chunk_size),
        q_seqlens).astype(np.int32)
    tokens_in_last_block = attn_chunk_size + (seq_lens_np % -attn_chunk_size)
    local_blocks = 1 + cdiv(q_seqlens - q_tokens_in_first_block,
                            attn_chunk_size)

    # Once we know the number of local blocks we can compute the request spans
    #  for each batch idx, we can figure out the number of "virtual" requests we
    #  have to make,
    # For the above example we would get:
    #   seqlens_q_local = [2, 2, 1, 4, 4, 1, 4, 1]
    #
    # First Get batched arange. (E.g., [2, 4, 2] -> [0, 1, 0, 1, 2, 3, 0, 1])
    #   (TODO: max a utility to share this code with _prepare_inputs)
    # arange step 1. [2, 4, 2] -> [2, 6, 8]
    cu_num_blocks = np.cumsum(local_blocks)
    virtual_batches = cu_num_blocks[-1]
    # arange step 2. [2, 6, 8] -> [0, 0, 2, 2, 2, 2, 6, 6]
    block_offsets = np.repeat(cu_num_blocks - local_blocks, local_blocks)
    # arange step 3. [0, 1, 0, 1, 2, 3, 0, 1]
    arange = np.arange(virtual_batches, dtype=np.int32) - block_offsets
    # also compute reverse arange (i.e. [1, 0, 3, 2, 1, 0, 1, 0])
    rarange = np.repeat(local_blocks, local_blocks) - arange - 1
    # Then we can compute the seqlens_q_local, handling the fact that the
    #  first and last blocks could be partial
    seqlens_q_local = \
        np.repeat(q_seqlens - q_tokens_in_first_block, local_blocks)
    # set the first block since this may be a partial block
    seqlens_q_local[arange == 0] = q_tokens_in_first_block
    # set the remaining blocks
    seqlens_q_local[arange > 0] = np.minimum(
        seqlens_q_local - attn_chunk_size * (arange - 1),
        attn_chunk_size)[arange > 0]

    # convert from q_seqlens to cu_seqlens_q
    cu_seqlens_q_local = np.pad(np.cumsum(seqlens_q_local), (1, 0))\
        .astype(np.int32)

    # compute the seqlens_k_local,
    #  basically a full local attention block for all but the last block in each
    #  batch
    # For our example this will be:
    #   seqlens_k_local = [4, 2, 4, 4, 4, 1, 4, 1]
    seqlens_k_local = np.full(cu_num_blocks[-1],
                              attn_chunk_size,
                              dtype=np.int32)
    seqlens_k_local[cu_num_blocks - 1] = tokens_in_last_block

    k_seqstarts_absolute = np.repeat(seq_lens_np, local_blocks) - \
        (rarange * attn_chunk_size + \
            np.repeat(tokens_in_last_block, local_blocks))
    # For the example the local attention blocks start at:
    #                           _b0_  _____b1_____  _b2_
    #   k_seqstarts_absolute = [0, 4, 4, 8, 12, 16, 4, 8]
    block_starts = k_seqstarts_absolute // block_size
    assert attn_chunk_size % block_size == 0, \
        f"attn_chunk_size {attn_chunk_size} is not " \
        f"divisible by block_size {block_size}"
    pages_per_local_batch = attn_chunk_size // block_size

    # Create a block_table for the local attention blocks
    # For out example if we have a block-table like (assuming block_size=2):
    #   block_table = [
    #     [ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9],  < batch 0
    #     [10, 11, 12, 13, 14, 15, 16, 17, 18, 19],  < batch 1
    #     [20, 21, 22, 23, 24, 25, 26, 27, 28, 29],  < batch 2
    #   ]
    # Then for the local batches we would want a block-table like
    #   block_table_local = [
    #     [  0,  1 ], < local-batch 0, (batch 0, starting from k[0])
    #     [  2,  3 ], < local-batch 1, (batch 0, starting from k[4])
    #     [ 12, 13 ], < local-batch 2, (batch 1, starting from k[4])
    #     [ 14, 15 ], < local-batch 3, (batch 1, starting from k[8])
    #     [ 16, 17 ], < local-batch 4, (batch 1, starting from k[12])
    #     [ 18, 19 ], < local-batch 5, (batch 1, starting from k[16])
    #     [ 22, 23 ], < local-batch 6, (batch 2, starting from k[4])
    #     [ 24, 25 ], < local-batch 7, (batch 2, starting from k[8])
    #   ]
    block_indices= np.broadcast_to(
        np.arange(pages_per_local_batch, dtype=np.int32),
        (virtual_batches, pages_per_local_batch)) \
            + np.expand_dims(block_starts, axis=1)
    block_indices = block_indices.flatten().clip(max=block_table.shape[1] - 1)
    batch_indices = np.repeat(np.arange(actual_batch_size, dtype=np.int32),
                              local_blocks * pages_per_local_batch)
    block_table_local = block_table[batch_indices, block_indices]\
        .view(virtual_batches, -1)

    return seqlens_q_local, cu_seqlens_q_local, seqlens_k_local, \
        block_table_local


def _get_sliding_window_configs(
        vllm_config: VllmConfig) -> set[Optional[tuple[int, int]]]:
    """Get the set of all sliding window configs used in the model."""
    sliding_window_configs: set[Optional[tuple[int, int]]] = set()
    layers = get_layers_from_vllm_config(vllm_config, Attention)
    for layer in layers.values():
        assert isinstance(layer.impl, FlashAttentionImpl)
        sliding_window_configs.add(layer.impl.sliding_window)
    return sliding_window_configs


class FlashAttentionMetadataBuilder(
        AttentionMetadataBuilder[FlashAttentionMetadata]):
    full_cudagraph_supported: ClassVar[bool] = get_flash_attn_version() == 3

    def __init__(self, runner: "GPUModelRunner", kv_cache_spec: AttentionSpec,
                 block_table: BlockTable):
        model_config = runner.model_config
        compilation_config = runner.vllm_config.compilation_config

        self.runner = runner
        self.num_heads_q = model_config.get_num_attention_heads(
            runner.parallel_config)
        self.num_heads_kv = model_config.get_num_kv_heads(
            runner.parallel_config)
        self.headdim = model_config.get_head_size()
        self.block_size = kv_cache_spec.block_size
        self.kv_cache_spec = kv_cache_spec
        self.block_table = block_table

        self.aot_schedule = (get_flash_attn_version() == 3)
        self.use_full_cuda_graph = compilation_config.full_cuda_graph
        if self.use_full_cuda_graph and not self.aot_schedule:
            raise ValueError("Full CUDA graph mode requires AOT scheduling, "
                             "which requires FlashAttention 3.")
        self.scheduler_metadata = torch.zeros(self.runner.max_num_reqs + 1,
                                              dtype=torch.int32,
                                              device=self.runner.device)

        # Sliding window size to be used with the AOT scheduler will be
        # populated on first build() call.
        self.aot_sliding_window: Optional[tuple[int, int]] = None

    def build(
        self, common_prefix_len: int,
        common_attn_metadata: CommonAttentionMetadata
    ) -> FlashAttentionMetadata:
        num_reqs = common_attn_metadata.num_reqs
        num_actual_tokens = common_attn_metadata.num_actual_tokens
        max_query_len = common_attn_metadata.max_query_len

        max_seq_len = int(self.runner.seq_lens_np[:num_reqs].max())
        query_start_loc = common_attn_metadata.query_start_loc
        seq_lens = common_attn_metadata.seq_lens
        block_table = self.block_table
        block_table_tensor = block_table.get_device_tensor()[:num_reqs]

        block_table.slot_mapping[:num_actual_tokens].copy_(
            block_table.slot_mapping_cpu[:num_actual_tokens],
            non_blocking=True)
        # Fill unused with -1. Needed for reshape_and_cache in full cuda graph
        # mode.
        block_table.slot_mapping[num_actual_tokens:].fill_(-1)

        slot_mapping = block_table.slot_mapping[:num_actual_tokens]

        if self.aot_sliding_window is None:
            self.aot_sliding_window = (-1, -1)
            # For the AOT scheduler we need the sliding window value to be
            # constant for all layers to. We have to populate this on the first
            # build() call so the layers are constructed (cannot populate)
            # in __init__.
            if self.aot_schedule:
                sliding_window_configs = _get_sliding_window_configs(
                    self.runner.vllm_config)
                if len(sliding_window_configs) == 1:
                    sliding_window_config = sliding_window_configs.pop()
                    if sliding_window_config is not None:
                        self.aot_sliding_window = sliding_window_config
                elif len(sliding_window_configs) > 1:
                    self.aot_schedule = False

        def schedule(batch_size, cu_query_lens, max_query_len, seqlens,
                     max_seq_len, causal):
            if self.aot_schedule:
                return get_scheduler_metadata(
                    batch_size=batch_size,
                    max_seqlen_q=max_query_len,
                    max_seqlen_k=max_seq_len,
                    cache_seqlens=seqlens,
                    num_heads_q=self.num_heads_q,
                    num_heads_kv=self.num_heads_kv,
                    headdim=self.headdim,
                    page_size=self.block_size,
                    cu_seqlens_q=cu_query_lens,
                    causal=causal,
                    window_size=self.aot_sliding_window,
                )
            return None

        # for local attention
        local_attn_metadata = None
        if self.runner.attention_chunk_size is not None:
            seqlens_q_local_np, virt_q_cu_seqlens_np, virt_k_seqlens_np, \
                virt_block_table_tensor = make_local_attention_virtual_batches(
                    self.runner.attention_chunk_size,
                    self.runner.query_start_loc_np[:num_reqs + 1],
                    self.runner.seq_lens_np[:num_reqs],
                    block_table_tensor,
                    self.block_size,
                )
            local_query_start_loc = torch.from_numpy(virt_q_cu_seqlens_np).to(
                self.runner.device, non_blocking=True)
            local_seqused_k = torch.from_numpy(virt_k_seqlens_np).to(
                self.runner.device, non_blocking=True)
            local_max_query_len = seqlens_q_local_np.max()
            local_max_seq_len = virt_k_seqlens_np.max()
            local_scheduler_metadata = schedule(
                batch_size=local_query_start_loc.shape[0] - 1,
                cu_query_lens=local_query_start_loc,
                max_query_len=local_max_query_len,
                seqlens=local_seqused_k,
                max_seq_len=local_max_seq_len,
                causal=True)

            local_attn_metadata = FlashAttentionMetadata.LocalAttentionMetadata(
                local_query_start_loc=local_query_start_loc,
                local_seqused_k=local_seqused_k,
                local_block_table=virt_block_table_tensor,
                local_max_query_len=local_max_query_len,
                local_max_seq_len=local_max_seq_len,
                local_scheduler_metadata=local_scheduler_metadata,
            )

        use_cascade = common_prefix_len > 0

        if use_cascade:
            cu_prefix_query_lens = torch.tensor([0, num_actual_tokens],
                                                dtype=torch.int32,
                                                device=self.runner.device)
            prefix_kv_lens = torch.tensor([common_prefix_len],
                                          dtype=torch.int32,
                                          device=self.runner.device)
            suffix_kv_lens = (self.runner.seq_lens_np[:num_reqs] -
                              common_prefix_len)
            suffix_kv_lens = torch.from_numpy(suffix_kv_lens).to(
                self.runner.device)
            prefix_scheduler_metadata = schedule(
                batch_size=1,
                cu_query_lens=cu_prefix_query_lens,
                max_query_len=num_actual_tokens,
                seqlens=prefix_kv_lens,
                max_seq_len=common_prefix_len,
                causal=False)
            scheduler_metadata = schedule(batch_size=num_reqs,
                                          cu_query_lens=query_start_loc,
                                          max_query_len=max_query_len,
                                          seqlens=suffix_kv_lens,
                                          max_seq_len=max_seq_len -
                                          common_prefix_len,
                                          causal=True)
        else:
            cu_prefix_query_lens = None
            prefix_kv_lens = None
            suffix_kv_lens = None
            prefix_scheduler_metadata = None
            scheduler_metadata = schedule(batch_size=num_reqs,
                                          cu_query_lens=query_start_loc,
                                          max_query_len=max_query_len,
                                          seqlens=seq_lens,
                                          max_seq_len=max_seq_len,
                                          causal=True)

        if self.use_full_cuda_graph:
            assert scheduler_metadata is not None
            n = scheduler_metadata.shape[0]
            self.scheduler_metadata[:n].copy_(scheduler_metadata,
                                              non_blocking=True)
            # NOTE(woosuk): We should zero out the rest of the scheduler
            # metadata to guarantee the correctness. Otherwise, some thread
            # blocks may use the invalid scheduler metadata and overwrite the
            # output buffer.
            self.scheduler_metadata[n:] = 0
            scheduler_metadata = self.scheduler_metadata[:n]

        attn_metadata = FlashAttentionMetadata(
            num_actual_tokens=num_actual_tokens,
            max_query_len=max_query_len,
            query_start_loc=query_start_loc,
            max_seq_len=max_seq_len,
            seq_lens=seq_lens,
            block_table=block_table_tensor,
            slot_mapping=slot_mapping,
            use_cascade=use_cascade,
            common_prefix_len=common_prefix_len,
            scheduler_metadata=scheduler_metadata,
            cu_prefix_query_lens=cu_prefix_query_lens,
            prefix_kv_lens=prefix_kv_lens,
            suffix_kv_lens=suffix_kv_lens,
            local_attn_metadata=local_attn_metadata,
            prefix_scheduler_metadata=prefix_scheduler_metadata,
        )
        return attn_metadata

    def can_run_in_cudagraph(
            self, common_attn_metadata: CommonAttentionMetadata) -> bool:
        # Full CUDA Graph always supported (FA2 support checked separately)
        return True

    def use_cascade_attention(self, *args, **kwargs) -> bool:
        return use_cascade_attention(*args, **kwargs)


class FlashAttentionImpl(AttentionImpl):

    def __init__(
        self,
        num_heads: int,
        head_size: int,
        scale: float,
        num_kv_heads: int,
        alibi_slopes: Optional[list[float]],
        sliding_window: Optional[int],
        kv_cache_dtype: str,
        blocksparse_params: Optional[dict[str, Any]] = None,
        logits_soft_cap: Optional[float] = None,
        attn_type: AttentionType = AttentionType.DECODER,
        kv_sharing_target_layer_name: Optional[str] = None,
        use_irope: bool = False,
    ) -> None:
        if blocksparse_params is not None:
            raise ValueError(
                "FlashAttention does not support block-sparse attention.")
        self.num_heads = num_heads
        self.head_size = head_size
        self.scale = float(scale)
        self.num_kv_heads = num_kv_heads
        if alibi_slopes is not None:
            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
        self.alibi_slopes = alibi_slopes
        if sliding_window is None:
            self.sliding_window = (-1, -1)
        else:
            self.sliding_window = (sliding_window - 1, 0)
        self.kv_cache_dtype = kv_cache_dtype
        if logits_soft_cap is None:
            # In flash-attn, setting logits_soft_cap as 0 means no soft cap.
            logits_soft_cap = 0
        self.logits_soft_cap = logits_soft_cap
        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name

        self.num_queries_per_kv = self.num_heads // self.num_kv_heads

        support_head_sizes = FlashAttentionBackend.get_supported_head_sizes()
        if head_size not in support_head_sizes:
            raise ValueError(
                f"Head size {head_size} is not supported by FlashAttention. "
                f"Supported head sizes are: {support_head_sizes}. "
                "Set VLLM_USE_V1=0 to use another attention backend.")

        if attn_type != AttentionType.DECODER:
            raise NotImplementedError("Encoder self-attention and "
                                      "encoder/decoder cross-attention "
                                      "are not implemented for "
                                      "FlashAttentionImpl")
        self.use_irope = use_irope
        self.vllm_flash_attn_version = get_flash_attn_version()
        if is_quantized_kv_cache(self.kv_cache_dtype) \
            and not flash_attn_supports_fp8():
            raise NotImplementedError(
                "FlashAttention does not support fp8 kv-cache on this device.")

    def forward(
        self,
        layer: torch.nn.Module,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: FlashAttentionMetadata,
        output: Optional[torch.Tensor] = None,
        output_scale: Optional[torch.Tensor] = None,
    ) -> torch.Tensor:
        """Forward pass with FlashAttention.

        Args:
            query: shape = [num_tokens, num_heads, head_size]
            key: shape = [num_tokens, num_kv_heads, head_size]
            value: shape = [num_tokens, num_kv_heads, head_size]
            kv_cache = [2, num_blocks, block_size, num_kv_heads, head_size]
            attn_metadata: Metadata for attention.
        Returns:
            shape = [num_tokens, num_heads * head_size]
        NOTE: FP8 quantization, flash-attn expect the size of
              {q,k,v}_descale to be (num_sequences, num_kv_heads).
              We use torch's .expand() to avoid duplicating values
        """
        assert output is not None, "Output tensor must be provided."

        if output_scale is not None:
            raise NotImplementedError(
                "fused output quantization is not yet supported"
                " for FlashAttentionImpl")

        if attn_metadata is None:
            # Profiling run.
            return output

        # IMPORTANT!
        # NOTE(woosuk): With piece-wise CUDA graphs, this method is executed in
        # eager-mode PyTorch. Thus, we need to be careful about any CPU overhead
        # in this method. For example, `view` and `slice` (or `[:n]`) operations
        # are surprisingly slow even in the case they do not invoke any GPU ops.
        # Minimize the PyTorch ops in this method as much as possible.
        # Whenever making a change in this method, please benchmark the
        # performance to make sure it does not introduce any overhead.

        num_actual_tokens = attn_metadata.num_actual_tokens
        key_cache, value_cache = kv_cache.unbind(0)

        if self.kv_sharing_target_layer_name is None:
            # Reshape the input keys and values and store them in the cache.
            # Skip this if sharing KV cache with an earlier attention layer.
            # NOTE(woosuk): Here, key and value are padded while slot_mapping is
            # not padded. However, we don't need to do key[:num_actual_tokens]
            # and value[:num_actual_tokens] because the reshape_and_cache_flash
            # op uses the slot_mapping's shape to determine the number of
            # actual tokens.
            torch.ops._C_cache_ops.reshape_and_cache_flash(
                key,
                value,
                key_cache,
                value_cache,
                attn_metadata.slot_mapping,
                self.kv_cache_dtype,
                layer._k_scale,
                layer._v_scale,
            )

        if self.kv_cache_dtype.startswith("fp8"):
            key_cache = key_cache.view(torch.float8_e4m3fn)
            value_cache = value_cache.view(torch.float8_e4m3fn)
            num_tokens, num_heads, head_size = query.shape
            query, _ = ops.scaled_fp8_quant(
                query.reshape(
                    (num_tokens, num_heads * head_size)).contiguous(),
                layer._q_scale)
            query = query.reshape((num_tokens, num_heads, head_size))

        # Compute attention and update output up to `num_actual_tokens`.
        use_local_attn = \
            (self.use_irope and attn_metadata.local_attn_metadata is not None)

        if not attn_metadata.use_cascade or use_local_attn:
            if use_local_attn:
                assert attn_metadata.local_attn_metadata is not None
                local_metadata = attn_metadata.local_attn_metadata
                cu_seqlens_q = local_metadata.local_query_start_loc
                seqused_k = local_metadata.local_seqused_k
                max_seqlen_q = local_metadata.local_max_query_len
                max_seqlen_k = local_metadata.local_max_seq_len
                block_table = local_metadata.local_block_table
                scheduler_metadata = local_metadata.local_scheduler_metadata
            else:
                cu_seqlens_q = attn_metadata.query_start_loc
                seqused_k = attn_metadata.seq_lens
                max_seqlen_q = attn_metadata.max_query_len
                max_seqlen_k = attn_metadata.max_seq_len
                block_table = attn_metadata.block_table
                scheduler_metadata = attn_metadata.scheduler_metadata

            descale_shape = (cu_seqlens_q.shape[0] - 1, key.shape[1])

            flash_attn_varlen_func(
                q=query[:num_actual_tokens],
                k=key_cache,
                v=value_cache,
                out=output[:num_actual_tokens],
                cu_seqlens_q=cu_seqlens_q,
                max_seqlen_q=max_seqlen_q,
                seqused_k=seqused_k,
                max_seqlen_k=max_seqlen_k,
                softmax_scale=self.scale,
                causal=True,
                alibi_slopes=self.alibi_slopes,
                window_size=self.sliding_window,
                block_table=block_table,
                softcap=self.logits_soft_cap,
                scheduler_metadata=scheduler_metadata,
                fa_version=self.vllm_flash_attn_version,
                q_descale=layer._q_scale.expand(descale_shape),
                k_descale=layer._k_scale.expand(descale_shape),
                v_descale=layer._v_scale.expand(descale_shape),
            )
            return output

        assert not use_local_attn, (
            "Cascade attention does not support local attention.")
        # Cascade attention (rare case).
        cascade_attention(
            output[:num_actual_tokens],
            query[:num_actual_tokens],
            key_cache,
            value_cache,
            cu_query_lens=attn_metadata.query_start_loc,
            max_query_len=attn_metadata.max_query_len,
            cu_prefix_query_lens=attn_metadata.cu_prefix_query_lens,
            prefix_kv_lens=attn_metadata.prefix_kv_lens,
            suffix_kv_lens=attn_metadata.suffix_kv_lens,
            max_kv_len=attn_metadata.max_seq_len,
            softmax_scale=self.scale,
            alibi_slopes=self.alibi_slopes,
            sliding_window=self.sliding_window,
            logits_soft_cap=self.logits_soft_cap,
            block_table=attn_metadata.block_table,
            common_prefix_len=attn_metadata.common_prefix_len,
            fa_version=self.vllm_flash_attn_version,
            prefix_scheduler_metadata=attn_metadata.prefix_scheduler_metadata,
            suffix_scheduler_metadata=attn_metadata.scheduler_metadata,
            q_descale=layer._q_scale,
            k_descale=layer._k_scale,
            v_descale=layer._v_scale,
        )
        return output


def use_cascade_attention(
    common_prefix_len: int,
    query_lens: np.ndarray,
    num_query_heads: int,
    num_kv_heads: int,
    use_alibi: bool,
    use_sliding_window: bool,
    num_sms: int,
) -> bool:
    """Decide whether to use cascade attention.

    This function 1) checks whether cascade attention is supported with the
    given configuration, and 2) heuristically decides whether using cascade
    attention can improve performance.
    """
    # Too short common prefix. Probably not worth using cascade attention.
    # We use an arbitrary threshold of 256 tokens. TODO: Tune this threshold.
    # NOTE(woosuk): This is the common case. We should return False as soon as
    # possible to avoid any unnecessary computation.
    if common_prefix_len < 256:
        return False
    # Cascade attention is currently not supported with these variants.
    if use_alibi or use_sliding_window:
        return False
    # Too few queries. Probably not worth using cascade attention.
    # We use an arbitrary threshold of 8 queries. TODO: Tune this threshold.
    num_reqs = len(query_lens)
    if num_reqs < 8:
        return False

    # Heuristics to decide whether using cascade attention is beneficial.
    # 1. When FlashDecoding is not used for normal attention, cascade attention
    #    is likely to be faster since it saves memory bandwidth.
    num_queries_per_kv = num_query_heads // num_kv_heads
    # The criteria for using FlashDecoding can be found in the following link:
    # https://github.com/vllm-project/flash-attention/blob/96266b1111111f3d11aabefaf3bacbab6a89d03c/csrc/flash_attn/flash_api.cpp#L535
    use_flash_decoding = (num_queries_per_kv > 1 and not use_sliding_window
                          and not use_alibi and np.all(query_lens == 1))
    if not use_flash_decoding:
        # Use cascade attention.
        return True

    # 2. When FlashDecoding is used for normal attention, it is not clear
    #    whether cascade attention is beneficial, because FlashDecoding can
    #    launch more CTAs than cascade attention.
    #    We use a simple performance model to compare the two methods.
    #    NOTE(woosuk): The performance model is very rough and may not be
    #    accurate.
    num_tokens = num_reqs
    # NOTE(woosuk): These are default tile sizes. flash-attn might use
    # different tile sizes (e.g., 64 or 256) depending on the configuration.
    q_tile_size = 128
    kv_tile_size = 128
    num_prefix_tiles = cdiv(common_prefix_len, kv_tile_size)

    cascade_ctas = num_query_heads * cdiv(num_tokens, q_tile_size)
    cascade_waves = cdiv(cascade_ctas, num_sms)
    cascade_time = cascade_waves * num_prefix_tiles

    flash_decoding_ctas = (num_reqs * num_kv_heads *
                           cdiv(num_queries_per_kv, q_tile_size))
    flash_decoding_ctas *= num_prefix_tiles
    flash_decoding_time = cdiv(flash_decoding_ctas, num_sms)

    # Use cascade attention if it is faster than FlashDecoding.
    return cascade_time < flash_decoding_time


def cascade_attention(
    output: torch.Tensor,
    query: torch.Tensor,
    key_cache: torch.Tensor,
    value_cache: torch.Tensor,
    cu_query_lens: torch.Tensor,
    max_query_len: int,
    cu_prefix_query_lens: torch.Tensor,
    prefix_kv_lens: torch.Tensor,
    suffix_kv_lens: torch.Tensor,
    max_kv_len: int,
    softmax_scale: float,
    alibi_slopes: Optional[torch.Tensor],
    sliding_window: tuple[int, int],
    logits_soft_cap: float,
    block_table: torch.Tensor,
    common_prefix_len: int,
    fa_version: int,
    prefix_scheduler_metadata: Optional[torch.Tensor] = None,
    suffix_scheduler_metadata: Optional[torch.Tensor] = None,
    q_descale: Optional[torch.Tensor] = None,
    k_descale: Optional[torch.Tensor] = None,
    v_descale: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    assert alibi_slopes is None, ("Cascade attention does not support ALiBi.")
    # TODO: Support sliding window.
    assert sliding_window == (-1, -1), (
        "Cascade attention does not support sliding window.")

    num_tokens = query.shape[0]
    block_size = key_cache.shape[-3]
    assert common_prefix_len % block_size == 0
    num_common_kv_blocks = common_prefix_len // block_size
    assert num_common_kv_blocks > 0
    descale_shape = (cu_prefix_query_lens.shape[0] - 1, key_cache.shape[-2])

    # Process shared prefix.
    prefix_output, prefix_lse = flash_attn_varlen_func(
        q=query,
        k=key_cache,
        v=value_cache,
        cu_seqlens_q=cu_prefix_query_lens,
        seqused_k=prefix_kv_lens,
        max_seqlen_q=num_tokens,
        max_seqlen_k=common_prefix_len,
        softmax_scale=softmax_scale,
        causal=False,
        window_size=sliding_window,
        block_table=block_table[:1],
        softcap=logits_soft_cap,
        return_softmax_lse=True,
        scheduler_metadata=prefix_scheduler_metadata,
        fa_version=fa_version,
        q_descale=q_descale.expand(descale_shape)
        if q_descale is not None else None,
        k_descale=k_descale.expand(descale_shape)
        if k_descale is not None else None,
        v_descale=v_descale.expand(descale_shape)
        if v_descale is not None else None,
    )

    descale_shape = (cu_query_lens.shape[0] - 1, key_cache.shape[-2])

    # Process suffix per query.
    suffix_output, suffix_lse = flash_attn_varlen_func(
        q=query,
        k=key_cache,
        v=value_cache,
        cu_seqlens_q=cu_query_lens,
        seqused_k=suffix_kv_lens,
        max_seqlen_q=max_query_len,
        max_seqlen_k=max_kv_len - common_prefix_len,
        softmax_scale=softmax_scale,
        causal=True,
        window_size=sliding_window,
        block_table=block_table[:, num_common_kv_blocks:],
        softcap=logits_soft_cap,
        return_softmax_lse=True,
        scheduler_metadata=suffix_scheduler_metadata,
        fa_version=fa_version,
        q_descale=q_descale.expand(descale_shape)
        if q_descale is not None else None,
        k_descale=k_descale.expand(descale_shape)
        if k_descale is not None else None,
        v_descale=v_descale.expand(descale_shape)
        if v_descale is not None else None,
    )

    # Merge prefix and suffix outputs, and store the result in output.
    merge_attn_states(output, prefix_output, prefix_lse, suffix_output,
                      suffix_lse)