Add files using upload-large-folder tool

.venv/lib/python3.11/site-packages/vllm/attention/__init__.py

@@ -0,0 +1,19 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from vllm.attention.backends.abstract import (AttentionBackend,
+                                              AttentionMetadata,
+                                              AttentionMetadataBuilder,
+                                              AttentionState, AttentionType)
+from vllm.attention.layer import Attention
+from vllm.attention.selector import get_attn_backend
+
+__all__ = [
+    "Attention",
+    "AttentionBackend",
+    "AttentionMetadata",
+    "AttentionType",
+    "AttentionMetadataBuilder",
+    "Attention",
+    "AttentionState",
+    "get_attn_backend",
+]

.venv/lib/python3.11/site-packages/vllm/attention/backends/abstract.py

@@ -0,0 +1,296 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from abc import ABC, abstractmethod
+from contextlib import contextmanager
+from dataclasses import dataclass, fields
+from typing import (TYPE_CHECKING, Any, Dict, Generic, List, Optional,
+                    Protocol, Set, Tuple, Type, TypeVar)
+
+import torch
+
+from vllm.multimodal import MultiModalPlaceholderMap
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner_base import (ModelRunnerBase,
+                                               ModelRunnerInputBase,
+                                               ModelRunnerInputBuilderBase)
+
+
+class AttentionType:
+    """
+    Attention type.
+    Use string to be compatible with `torch.compile`.
+    """
+    # Decoder attention between previous layer Q/K/V
+    DECODER = "decoder"
+    # Encoder attention between previous layer Q/K/V for encoder-decoder
+    ENCODER = "encoder"
+    # Encoder attention between previous layer Q/K/V
+    ENCODER_ONLY = "encoder_only"
+    # Attention between dec. Q and enc. K/V for encoder-decoder
+    ENCODER_DECODER = "encoder_decoder"
+
+
+class AttentionBackend(ABC):
+    """Abstract class for attention backends."""
+    # For some attention backends, we allocate an output tensor before
+    # calling the custom op. When piecewise cudagraph is enabled, this
+    # makes sure the output tensor is allocated inside the cudagraph.
+    accept_output_buffer: bool = False
+
+    @staticmethod
+    @abstractmethod
+    def get_name() -> str:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_impl_cls() -> Type["AttentionImpl"]:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_state_cls() -> Type["AttentionState"]:
+        raise NotImplementedError
+
+    @classmethod
+    def make_metadata(cls, *args, **kwargs) -> "AttentionMetadata":
+        return cls.get_metadata_cls()(*args, **kwargs)
+
+    @staticmethod
+    @abstractmethod
+    def get_builder_cls() -> Type["AttentionMetadataBuilder"]:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        raise NotImplementedError
+
+    def advance_step(self, model_input: "ModelRunnerInputBase",
+                     sampled_token_ids: Optional[torch.Tensor],
+                     block_size: int, num_seqs: int, num_queries: int) -> None:
+        raise NotImplementedError
+
+
+@dataclass
+class AttentionMetadata:
+    """Attention metadata for prefill and decode batched together."""
+    # Total number of prefill requests.
+    num_prefills: int
+    # Number of prefill tokens.
+    num_prefill_tokens: int
+    # Number of decode tokens. Note that it is equivalent to the number of
+    # decode requests.
+    num_decode_tokens: int
+    # (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
+
+    # The index maps that relate multi-modal embeddings to the corresponding
+    # placeholders.
+    #
+    # N.B. These aren't really related to attention and don't belong on this
+    # type -- this is just a temporary solution to make them available to
+    # `model_executable`.
+    multi_modal_placeholder_index_maps: Optional[Dict[
+        str, MultiModalPlaceholderMap.IndexMap]]
+
+    # Enable/disable KV scales calculation. This is so that we can disable the
+    # calculation until after prefill and cuda graph capture.
+    enable_kv_scales_calculation: bool
+
+    @property
+    @abstractmethod
+    def prefill_metadata(self) -> Optional["AttentionMetadata"]:
+        """Return the attention metadata that's required to run prefill
+        attention."""
+        pass
+
+    @property
+    @abstractmethod
+    def decode_metadata(self) -> Optional["AttentionMetadata"]:
+        """Return the attention metadata that's required to run decode
+        attention."""
+        pass
+
+    def asdict_zerocopy(self,
+                        skip_fields: Optional[Set[str]] = None
+                        ) -> Dict[str, Any]:
+        """Similar to dataclasses.asdict, but avoids deepcopying."""
+        if skip_fields is None:
+            skip_fields = set()
+        # Note that if we add dataclasses as fields, they will need
+        # similar handling.
+        return {
+            field.name: getattr(self, field.name)
+            for field in fields(self) if field.name not in skip_fields
+        }
+
+
+T = TypeVar("T", bound=AttentionMetadata)
+
+
+class AttentionState(ABC, Generic[T]):
+    """Holds attention backend-specific objects reused during the
+    lifetime of the model runner."""
+
+    @abstractmethod
+    def __init__(self, runner: "ModelRunnerBase"):
+        ...
+
+    @abstractmethod
+    @contextmanager
+    def graph_capture(self, max_batch_size: int):
+        """Context manager used when capturing CUDA graphs."""
+        yield
+
+    @abstractmethod
+    def graph_clone(self, batch_size: int) -> "AttentionState[T]":
+        """Clone attention state to save in CUDA graph metadata."""
+        ...
+
+    @abstractmethod
+    def graph_capture_get_metadata_for_batch(
+            self,
+            batch_size: int,
+            is_encoder_decoder_model: bool = False) -> T:
+        """Get attention metadata for CUDA graph capture of batch_size."""
+        ...
+
+    @abstractmethod
+    def get_graph_input_buffers(
+            self,
+            attn_metadata: T,
+            is_encoder_decoder_model: bool = False) -> Dict[str, Any]:
+        """Get attention-specific input buffers for CUDA graph capture."""
+        ...
+
+    @abstractmethod
+    def prepare_graph_input_buffers(
+            self,
+            input_buffers: Dict[str, Any],
+            attn_metadata: T,
+            is_encoder_decoder_model: bool = False) -> None:
+        """In-place modify input buffers dict for CUDA graph replay."""
+        ...
+
+    @abstractmethod
+    def begin_forward(self, model_input: "ModelRunnerInputBase") -> None:
+        """Prepare state for forward pass."""
+        ...
+
+
+class AttentionMetadataBuilder(ABC, Generic[T]):
+    """Abstract class for attention metadata builders."""
+
+    @abstractmethod
+    def __init__(self, input_builder: "ModelRunnerInputBuilderBase") -> None:
+        """Create the builder, remember some configuration and parameters."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def prepare(self) -> None:
+        """Prepare for one batch."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int) -> T:
+        """Build attention metadata with on-device tensors."""
+        raise NotImplementedError
+
+
+class AttentionLayer(Protocol):
+
+    _k_scale: torch.Tensor
+    _v_scale: torch.Tensor
+    _k_scale_float: float
+    _v_scale_float: float
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        ...
+
+
+class AttentionImpl(ABC, Generic[T]):
+
+    @abstractmethod
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: Optional[int] = None,
+        alibi_slopes: Optional[List[float]] = None,
+        sliding_window: Optional[int] = None,
+        kv_cache_dtype: str = "auto",
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+        attn_type: str = AttentionType.DECODER,
+    ) -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: T,
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        raise NotImplementedError
+
+
+class MLAAttentionImpl(AttentionImpl[T], Generic[T]):
+
+    @abstractmethod
+    def forward(
+        self,
+        layer: AttentionLayer,
+        hidden_states_or_cq: torch.Tensor,
+        kv_c_normed: torch.Tensor,
+        k_pe: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: T,
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        raise NotImplementedError

.venv/lib/python3.11/site-packages/vllm/attention/backends/blocksparse_attn.py

@@ -0,0 +1,457 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionLayer,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import (CommonAttentionState,
+                                           CommonMetadataBuilder)
+from vllm.attention.ops.blocksparse_attention.interface import (
+    LocalStridedBlockSparseAttn, get_head_sliding_step)
+from vllm.attention.ops.paged_attn import PagedAttention
+from vllm.distributed import (get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size)
+
+
+@dataclass
+class BlocksparseParams:
+    max_seqlen: int
+
+    # Num q heads per tensor-parallel rank/partition
+    num_heads: int  # per TP partition
+    # Num kv heads per tensor-parallel rank/partition
+    num_kv_heads: int
+
+    # block size used for blocksparse attention.
+    # This is the block_size used in `local_blocks`, `vert_stride`.
+    block_size: int
+
+    # Number of blocks for local attention, i.e., number of
+    # local attended tokens / `sparse_block_size`
+    local_blocks: int
+
+    # Attend to one block per every `vert_stride` blocks.
+    # Controlling the sparsity
+    vert_stride: int
+    """
+    If to use the same vertical stride offset for all heads, 
+    i.e., attend to the same block of tokens on all heads.
+    By default, it is False, i.e., attention on the non-local 
+    blocks depends on the `head_idx`, that is on
+    blocks satisfying 
+    `(block_idx + head_idx * head_sliding_step + 1) % vert_stride == 0`
+    where `head_sliding_step=max(1, int(vert_stride / num_total_heads))`,
+            `block_idx = position_id // sparse_block_size`.
+    See `..ops.blocksparse_attention.utils:get_sparse_attn_mask`
+    for more detail.
+    """
+    homo_head: bool = False
+
+    # If within a group, the kv offsets that each q attends is the same or no.
+    homo_head_group: bool = False
+
+    # Decided by homo_head and homo_head group
+    head_sliding_step: int = field(init=False)
+
+    # range of q heads to for a TP rank
+    active_head_range: Tuple = field(init=False)
+
+    def __post_init__(self):
+        assert self.block_size > 0
+        assert self.local_blocks >= 0
+        assert self.vert_stride >= 1
+        assert self.num_heads % self.num_kv_heads == 0
+
+        tp_size = get_tensor_model_parallel_world_size()
+        tp_rank = get_tensor_model_parallel_rank()
+        total_heads = tp_size * self.num_heads
+        total_kv_heads = tp_size * self.num_kv_heads
+
+        if self.homo_head:
+            self.head_sliding_step = 0
+        elif self.homo_head_group:
+            head_sliding_step = get_head_sliding_step(total_kv_heads,
+                                                      self.vert_stride)
+            # negative indicates sliding along kv heads, i.e., homo q group
+            self.head_sliding_step = -head_sliding_step
+        else:
+            self.head_sliding_step = get_head_sliding_step(
+                total_heads, self.vert_stride)
+
+        self.active_head_range = (
+            tp_rank * self.num_heads,
+            (tp_rank + 1) * self.num_heads,
+        )
+
+
+class BlocksparseFlashAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "BLOCK_SPARSE_FLASH_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> Type["BlocksparseFlashAttentionImpl"]:
+        return BlocksparseFlashAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return BlocksparseFlashAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["BlocksparseFlashAttentionMetadataBuilder"]:
+        return BlocksparseFlashAttentionMetadataBuilder
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: Dict[int, List[int]],
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class BlocksparseFlashAttentionMetadata(AttentionMetadata):
+    """A copy of Metadata for FlashAttentionBackend,
+    to avoid having to install flash_attn.
+
+    NOTE: Any python object stored here is not updated when it is
+    cuda-graph replayed. If you have values that need to be changed
+    dynamically, it should be stored in tensor. The tensor has to be
+    updated from `CUDAGraphRunner.forward` API.
+    """
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]]
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+
+    # NOTE(sang): Definition of context_len, query_len, and seq_len.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ----------------------|
+    #                                   |-- query_len ---|
+
+    # Maximum query length in the batch. None for decoding.
+    max_query_len: Optional[int]
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Maximum sequence length among decode batch. 0 if there are prefill
+    # requests only.
+    max_decode_seq_len: int
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    query_start_loc: Optional[torch.Tensor]
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor]
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor]
+
+    # (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]
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+    use_cuda_graph: bool
+
+    # Max number of query tokens for among request in the batch.
+    max_decode_query_len: Optional[int] = None
+
+    _cached_prefill_metadata: Optional[
+        "BlocksparseFlashAttentionMetadata"] = None
+    _cached_decode_metadata: Optional[
+        "BlocksparseFlashAttentionMetadata"] = None
+
+    @property
+    def prefill_metadata(
+            self) -> Optional["BlocksparseFlashAttentionMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            return self._cached_prefill_metadata
+
+        assert self.seq_lens is not None
+        assert self.seq_lens_tensor is not None
+        assert self.query_start_loc is not None
+        assert self.context_lens_tensor is not None
+        assert self.block_tables is not None
+        assert self.seq_start_loc is not None
+
+        self._cached_prefill_metadata = BlocksparseFlashAttentionMetadata(
+            num_prefills=self.num_prefills,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=0,
+            slot_mapping=self.slot_mapping[:self.num_prefill_tokens],
+            multi_modal_placeholder_index_maps=self.
+            multi_modal_placeholder_index_maps,
+            enable_kv_scales_calculation=self.enable_kv_scales_calculation,
+            seq_lens=self.seq_lens[:self.num_prefills],
+            seq_lens_tensor=self.seq_lens_tensor[:self.num_prefills],
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=self.max_prefill_seq_len,
+            max_decode_seq_len=0,
+            query_start_loc=self.query_start_loc[:self.num_prefills + 1],
+            seq_start_loc=self.seq_start_loc[:self.num_prefills + 1],
+            context_lens_tensor=self.context_lens_tensor[:self.num_prefills],
+            block_tables=self.block_tables[:self.num_prefills],
+            use_cuda_graph=False,
+        )
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> Optional["BlocksparseFlashAttentionMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            return self._cached_decode_metadata
+        assert self.block_tables is not None
+        assert self.seq_lens_tensor is not None
+
+        self._cached_decode_metadata = BlocksparseFlashAttentionMetadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=self.num_decode_tokens,
+            slot_mapping=self.slot_mapping[self.num_prefill_tokens:],
+            multi_modal_placeholder_index_maps=None,
+            enable_kv_scales_calculation=False,
+            seq_lens=None,
+            seq_lens_tensor=self.seq_lens_tensor[self.num_prefills:],
+            max_query_len=None,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.max_decode_seq_len,
+            query_start_loc=None,
+            seq_start_loc=None,
+            context_lens_tensor=None,
+            block_tables=self.block_tables[self.num_prefills:],
+            use_cuda_graph=self.use_cuda_graph,
+        )
+        return self._cached_decode_metadata
+
+
+class BlocksparseFlashAttentionMetadataBuilder(
+        CommonMetadataBuilder[BlocksparseFlashAttentionMetadata]):
+
+    _metadata_cls = BlocksparseFlashAttentionMetadata
+
+
+class BlocksparseFlashAttentionImpl(AttentionImpl):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prompt_tokens -------------->|
+    |<--prompt_0-->|<--prompt_1-->|...|<--prompt_N-1-->|
+
+    Otherwise, the layout is as follows:
+    |<------------------ num_generation_tokens (M) ----------------->|
+    |<--generation_0-->|..........|<--generation_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+
+    """
+
+    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: str = AttentionType.DECODER,
+    ) -> None:
+        assert blocksparse_params is not None
+        assert alibi_slopes is None, ValueError(
+            "Alibi not support for blocksparse flash attention.")
+        assert sliding_window is None, ValueError(
+            "sliding_window is invalid for blocksparse attention.")
+        assert logits_soft_cap is None, ValueError(
+            "logits_soft_cap is invalid for blocksparse attention.")
+
+        if "num_heads" not in blocksparse_params:
+            blocksparse_params["num_heads"] = num_heads
+        if "num_kv_heads" not in blocksparse_params:
+            blocksparse_params["num_kv_heads"] = num_kv_heads or num_heads
+        self.blocksparse_params = BlocksparseParams(**blocksparse_params)
+        self.kv_cache_dtype = kv_cache_dtype
+
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.alibi_slopes = alibi_slopes
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        self.local_blocks = self.blocksparse_params.local_blocks
+        self.vert_stride = self.blocksparse_params.vert_stride
+        self.sparse_block_size = self.blocksparse_params.block_size
+        self.head_sliding_step = self.blocksparse_params.head_sliding_step
+
+        suppored_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in suppored_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {suppored_head_sizes}.")
+
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
+
+        total_num_heads = num_heads * self.tp_size
+        self.bs_attn = LocalStridedBlockSparseAttn(
+            total_num_heads,
+            self.blocksparse_params.max_seqlen,
+            self.blocksparse_params.local_blocks,
+            self.blocksparse_params.vert_stride,
+            self.blocksparse_params.block_size,
+            homo_head=self.blocksparse_params.homo_head,
+            active_head_range=self.blocksparse_params.active_head_range,
+        )
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "BlocksparseFlashAttentionImpl")
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: BlocksparseFlashAttentionMetadata,
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Forward pass with FlashAttention and PagedAttention.
+
+        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]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        num_tokens, hidden_size = query.shape
+        # Reshape the query, key, and value tensors.
+        query = query.view(-1, self.num_heads, self.head_size)
+        key = key.view(-1, self.num_kv_heads, self.head_size)
+        value = value.view(-1, self.num_kv_heads, self.head_size)
+
+        if kv_cache.numel() > 0:
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            # Reshape the input keys and values and store them in the cache.
+            # If kv_cache is not provided, the new key and value tensors are
+            # not cached. This happens during the initial memory profiling run.
+
+            PagedAttention.write_to_paged_cache(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                attn_metadata.slot_mapping,
+                self.kv_cache_dtype,
+                layer._k_scale,
+                layer._v_scale,
+            )
+
+        if prefill_meta := attn_metadata.prefill_metadata:
+
+            # Prompt run.
+            # normal attention
+            # When block_tables are not filled, it means q and k are the
+            # prompt, and they have the same length.
+
+            assert kv_cache.numel() == 0 \
+                    or prefill_meta.block_tables is None \
+                    or prefill_meta.block_tables.numel() == 0, \
+                "Does not support prefix-enabled attention."
+
+            output = self.bs_attn(
+                q=query,
+                k=key,
+                v=value,
+                cu_seqlens_q=prefill_meta.seq_start_loc,
+                cu_seqlens_k=prefill_meta.seq_start_loc,
+                sm_scale=self.scale,
+            )
+
+        if decode_meta := attn_metadata.decode_metadata:
+            # Decoding run.
+            output = PagedAttention.forward_decode(
+                query,
+                key_cache,
+                value_cache,
+                decode_meta.block_tables,
+                decode_meta.seq_lens_tensor,
+                self.blocksparse_params.max_seqlen,
+                self.kv_cache_dtype,
+                self.num_kv_heads,
+                self.scale,
+                self.alibi_slopes,
+                layer._k_scale,
+                layer._v_scale,
+                tp_rank=self.tp_rank,
+                blocksparse_local_blocks=self.local_blocks,
+                blocksparse_vert_stride=self.vert_stride,
+                blocksparse_block_size=self.sparse_block_size,
+                blocksparse_head_sliding_step=self.head_sliding_step,
+            )
+
+        assert output is not None
+        # Reshape the output tensor.
+        return output.view(num_tokens, hidden_size)

.venv/lib/python3.11/site-packages/vllm/attention/backends/flash_attn.py

@@ -0,0 +1,942 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Attention layer with FlashAttention."""
+from collections import defaultdict
+from dataclasses import dataclass
+from itertools import accumulate
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionLayer,
+                                              AttentionMetadata,
+                                              AttentionMetadataBuilder,
+                                              AttentionType)
+from vllm.attention.backends.utils import (
+    PAD_SLOT_ID, CommonAttentionState, compute_slot_mapping,
+    compute_slot_mapping_start_idx, get_num_prefill_decode_query_kv_tokens,
+    get_seq_len_block_table_args, is_all_cross_attn_metadata_set,
+    is_all_encoder_attn_metadata_set, is_block_tables_empty)
+from vllm.envs import VLLM_FLASH_ATTN_VERSION
+from vllm.logger import init_logger
+from vllm.multimodal import MultiModalPlaceholderMap
+from vllm.platforms import current_platform
+from vllm.utils import async_tensor_h2d, make_tensor_with_pad
+from vllm.vllm_flash_attn import (fa_version_unsupported_reason,
+                                  flash_attn_varlen_func,
+                                  flash_attn_with_kvcache,
+                                  is_fa_version_supported)
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import (ModelInputForGPUBuilder,
+                                          ModelInputForGPUWithSamplingMetadata)
+
+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"
+
+    @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_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @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 swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        src_key_cache = src_kv_cache[0]
+        dst_key_cache = dst_kv_cache[0]
+        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]
+        ops.swap_blocks(src_value_cache, dst_value_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        key_caches = [kv_cache[0] for kv_cache in kv_caches]
+        value_caches = [kv_cache[1] for kv_cache in kv_caches]
+
+        ops.copy_blocks(key_caches, value_caches, src_to_dists)
+
+
+@dataclass
+class FlashAttentionMetadata(AttentionMetadata):
+    """Metadata for FlashAttentionBackend.
+
+    NOTE: Any python object stored here is not updated when it is
+    cuda-graph replayed. If you have values that need to be changed
+    dynamically, it should be stored in tensor. The tensor has to be
+    updated from `CUDAGraphRunner.forward` API.
+    """
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]]
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+
+    # NOTE(sang): Definition of context_len, query_len, and seq_len.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ---------------------|
+    #                                   |-- query_len ---|
+
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Maximum sequence length among decode batch. 0 if there are prefill
+    # requests only.
+    max_decode_seq_len: int
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor]
+
+    # (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]
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+
+    use_cuda_graph: bool
+
+    # Maximum query length in the batch.
+    max_query_len: Optional[int] = None
+
+    # Max number of query tokens among request in the batch.
+    max_decode_query_len: Optional[int] = None
+
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    query_start_loc: Optional[torch.Tensor] = None
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor] = None
+
+    _cached_prefill_metadata: Optional["FlashAttentionMetadata"] = None
+    _cached_decode_metadata: Optional["FlashAttentionMetadata"] = None
+
+    # Begin encoder attn & enc/dec cross-attn fields...
+
+    # Encoder sequence lengths representation
+    encoder_seq_lens: Optional[List[int]] = None
+    encoder_seq_lens_tensor: Optional[torch.Tensor] = None
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    encoder_seq_start_loc: Optional[torch.Tensor] = None
+    # Maximum sequence length among encoder sequences
+    max_encoder_seq_len: Optional[int] = None
+    # Number of tokens input to encoder
+    num_encoder_tokens: Optional[int] = None
+
+    # Cross-attention memory-mapping data structures: slot mapping
+    # and block tables
+    cross_slot_mapping: Optional[torch.Tensor] = None
+    cross_block_tables: Optional[torch.Tensor] = None
+
+    @property
+    def is_all_encoder_attn_metadata_set(self):
+        '''
+        All attention metadata required for encoder attention is set.
+        '''
+        return is_all_encoder_attn_metadata_set(self)
+
+    @property
+    def is_all_cross_attn_metadata_set(self):
+        '''
+        All attention metadata required for enc/dec cross-attention is set.
+
+        Superset of encoder attention required metadata.
+        '''
+        return is_all_cross_attn_metadata_set(self)
+
+    @property
+    def prefill_metadata(self) -> Optional["FlashAttentionMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            return self._cached_prefill_metadata
+
+        assert ((self.seq_lens is not None)
+                or (self.encoder_seq_lens is not None))
+        assert ((self.seq_lens_tensor is not None)
+                or (self.encoder_seq_lens_tensor is not None))
+
+        # Compute some attn_metadata fields which default to None
+        query_start_loc = (None if self.query_start_loc is None else
+                           self.query_start_loc[:self.num_prefills + 1])
+        slot_mapping = (None if self.slot_mapping is None else
+                        self.slot_mapping[:self.num_prefill_tokens])
+        seq_lens = (None if self.seq_lens is None else
+                    self.seq_lens[:self.num_prefills])
+        seq_lens_tensor = (None if self.seq_lens_tensor is None else
+                           self.seq_lens_tensor[:self.num_prefills])
+        seq_start_loc = (None if self.seq_start_loc is None else
+                         self.seq_start_loc[:self.num_prefills + 1])
+        context_lens_tensor = (None if self.context_lens_tensor is None else
+                               self.context_lens_tensor[:self.num_prefills])
+        block_tables = (None if self.block_tables is None else
+                        self.block_tables[:self.num_prefills])
+
+        self._cached_prefill_metadata = FlashAttentionMetadata(
+            num_prefills=self.num_prefills,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=0,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=self.
+            multi_modal_placeholder_index_maps,
+            enable_kv_scales_calculation=self.enable_kv_scales_calculation,
+            seq_lens=seq_lens,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=self.max_prefill_seq_len,
+            max_decode_query_len=0,
+            max_decode_seq_len=0,
+            query_start_loc=query_start_loc,
+            seq_start_loc=seq_start_loc,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=False,
+            # Begin encoder & cross attn fields below...
+            encoder_seq_lens=self.encoder_seq_lens,
+            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
+            encoder_seq_start_loc=self.encoder_seq_start_loc,
+            max_encoder_seq_len=self.max_encoder_seq_len,
+            cross_slot_mapping=self.cross_slot_mapping,
+            cross_block_tables=self.cross_block_tables)
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> Optional["FlashAttentionMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            return self._cached_decode_metadata
+        assert ((self.seq_lens_tensor is not None)
+                or (self.encoder_seq_lens_tensor is not None))
+
+        # Compute some attn_metadata fields which default to None
+        slot_mapping = (None if self.slot_mapping is None else
+                        self.slot_mapping[self.num_prefill_tokens:])
+        seq_lens_tensor = (None if self.seq_lens_tensor is None else
+                           self.seq_lens_tensor[self.num_prefills:])
+        block_tables = (None if self.block_tables is None else
+                        self.block_tables[self.num_prefills:])
+
+        self._cached_decode_metadata = FlashAttentionMetadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=self.num_decode_tokens,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=None,
+            enable_kv_scales_calculation=True,
+            seq_lens=None,
+            seq_lens_tensor=seq_lens_tensor,
+            max_decode_query_len=self.max_decode_query_len,
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.max_decode_seq_len,
+            # Batch may be composed of prefill|decodes, adjust query start
+            # indices to refer to the start of decodes. E.g.
+            # in tokens:[3 prefills|6 decodes], query_start_loc=[3,9] => [0,6].
+            query_start_loc=(self.query_start_loc[self.num_prefills:] -
+                             self.query_start_loc[self.num_prefills])
+            if self.query_start_loc is not None else None,
+            seq_start_loc=self.seq_start_loc[self.num_prefills:]
+            if self.seq_start_loc is not None else None,
+            context_lens_tensor=None,
+            block_tables=block_tables,
+            use_cuda_graph=self.use_cuda_graph,
+            # Begin encoder & cross attn fields below...
+            encoder_seq_lens=self.encoder_seq_lens,
+            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
+            encoder_seq_start_loc=self.encoder_seq_start_loc,
+            max_encoder_seq_len=self.max_encoder_seq_len,
+            cross_slot_mapping=self.cross_slot_mapping,
+            cross_block_tables=self.cross_block_tables)
+        return self._cached_decode_metadata
+
+    def advance_step(self,
+                     model_input: "ModelInputForGPUWithSamplingMetadata",
+                     sampled_token_ids: Optional[torch.Tensor],
+                     block_size: int,
+                     num_seqs: int,
+                     num_queries: int,
+                     turn_prefills_into_decodes: bool = False):
+        """
+        Update metadata in-place to advance one decode step.
+        """
+        # When using cudagraph, the num_seqs is padded to the next captured
+        # batch sized, but num_queries tracks the actual number of requests in
+        # the batch. For --enforce-eager mode, num_seqs == num_queries
+        if num_seqs != num_queries:
+            assert num_seqs > num_queries
+            assert self.use_cuda_graph
+
+        if turn_prefills_into_decodes:
+            # When Mutli-Step is enabled with Chunked-Prefill, prefills and
+            # decodes are scheduled together. In the first step, all the
+            # prefills turn into decodes. This update reflects that
+            # conversion.
+            assert self.num_decode_tokens + self.num_prefills == num_seqs
+            self.num_decode_tokens += self.num_prefills
+            self.num_prefills = 0
+            self.num_prefill_tokens = 0
+            self.max_prefill_seq_len = 0
+            self.max_query_len = 1
+
+            self.slot_mapping = self.slot_mapping[:num_seqs]
+        else:
+            assert self.seq_lens is not None
+            assert self.max_decode_seq_len == max(self.seq_lens)
+
+        assert self.num_prefills == 0
+        assert self.num_prefill_tokens == 0
+        assert self.num_decode_tokens == num_seqs
+        assert self.slot_mapping.shape == (num_seqs, )
+
+        assert self.seq_lens is not None
+        assert len(self.seq_lens) == num_seqs
+        assert self.seq_lens_tensor is not None
+        assert self.seq_lens_tensor.shape == (num_seqs, )
+        assert self.max_query_len == 1
+        assert self.max_prefill_seq_len == 0
+
+        assert self.query_start_loc is not None
+        assert self.query_start_loc.shape == (num_queries + 1, )
+        assert self.seq_start_loc is not None
+        assert self.seq_start_loc.shape == (num_seqs + 1, )
+
+        assert self.context_lens_tensor is not None
+        assert self.context_lens_tensor.shape == (num_queries, )
+
+        assert self.block_tables is not None
+        assert self.block_tables.shape[0] == num_seqs
+
+        # Update query lengths. Note that we update only queries and not seqs,
+        # since tensors may be padded due to captured cuda graph batch size
+        for i in range(num_queries):
+            self.seq_lens[i] += 1
+        self.max_decode_seq_len = max(self.seq_lens)
+
+        ops.advance_step_flashattn(num_seqs=num_seqs,
+                                   num_queries=num_queries,
+                                   block_size=block_size,
+                                   input_tokens=model_input.input_tokens,
+                                   sampled_token_ids=sampled_token_ids,
+                                   input_positions=model_input.input_positions,
+                                   seq_lens=self.seq_lens_tensor,
+                                   slot_mapping=self.slot_mapping,
+                                   block_tables=self.block_tables)
+
+
+class FlashAttentionMetadataBuilder(
+        AttentionMetadataBuilder[FlashAttentionMetadata]):
+
+    def __init__(self, input_builder: "ModelInputForGPUBuilder"):
+        self.input_builder = input_builder
+        self.runner = input_builder.runner
+        self.sliding_window = input_builder.sliding_window
+        self.block_size = input_builder.block_size
+
+    def prepare(self):
+        self.slot_mapping: List[int] = []
+        self.prefill_seq_lens: List[int] = []
+        self.context_lens: List[int] = []
+        self.block_tables: List[List[int]] = []
+        self.curr_seq_lens: List[int] = []
+        self.multimodal_placeholder_maps: Dict[
+            str,
+            MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
+        self.num_prefills = 0
+        self.num_prefill_tokens = 0
+        self.num_decode_tokens = 0
+        self.has_prefix_cache_hit = False
+
+    def _add_seq_group(
+            self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
+            chunked_prefill_enabled: bool, prefix_cache_hit: bool):
+        """Add a sequence group to the metadata. Specifically update/append
+        1. context length.
+        2. block table.
+        3. slot mapping.
+        """
+        is_prompt = inter_data.is_prompt
+        block_tables = inter_data.block_tables
+
+        for (seq_id, token_len, seq_len, curr_seq_len, query_len, context_len,
+             curr_sliding_window_block) in zip(
+                 inter_data.seq_ids, [len(t) for t in inter_data.input_tokens],
+                 inter_data.orig_seq_lens, inter_data.seq_lens,
+                 inter_data.query_lens, inter_data.context_lens,
+                 inter_data.curr_sliding_window_blocks):
+            self.context_lens.append(context_len)
+
+            if is_prompt:
+                mm_maps = inter_data.multi_modal_placeholder_maps
+                if mm_maps:
+                    for modality, placeholders in mm_maps.items():
+                        self.multimodal_placeholder_maps[modality].extend(
+                            placeholders)
+
+                self.num_prefills += 1
+                self.num_prefill_tokens += token_len
+                self.prefill_seq_lens.append(seq_len)
+            else:
+                self.num_decode_tokens += query_len
+                self.curr_seq_lens.append(curr_seq_len)
+
+            # Compute block table.
+            # TODO(sang): Combine chunked prefill and prefix caching by
+            # only allowing multiple of block_size chunk size.
+            # NOTE: This only works for oooooooxxx style attention.
+            block_table = []
+            if prefix_cache_hit:
+                # NOTE(woosuk): For flash-attn, the block table should
+                # include the entries for the incoming prefill tokens.
+                block_table = block_tables[seq_id]
+            elif ((chunked_prefill_enabled or not is_prompt)
+                  and block_tables is not None):
+                if curr_sliding_window_block == 0:
+                    block_table = block_tables[seq_id]
+                else:
+                    block_table = block_tables[seq_id][
+                        -curr_sliding_window_block:]
+            self.block_tables.append(block_table)
+
+            # Compute slot mapping.
+            is_profile_run = is_block_tables_empty(block_tables)
+            start_idx = compute_slot_mapping_start_idx(is_prompt, query_len,
+                                                       context_len,
+                                                       self.sliding_window)
+            compute_slot_mapping(is_profile_run, self.slot_mapping, seq_id,
+                                 seq_len, context_len, start_idx,
+                                 self.block_size, inter_data.block_tables)
+
+    def _get_graph_runner_block_tables(
+            self, num_seqs: int,
+            block_tables: List[List[int]]) -> torch.Tensor:
+        # The shape of graph_block_tables is
+        # [max batch size, max context len // block size].
+        max_batch_size, max_blocks = self.runner.graph_block_tables.shape
+        assert max_batch_size >= num_seqs
+
+        graph_block_tables = self.runner.graph_block_tables[:num_seqs]
+        for i, block_table in enumerate(block_tables):
+            if block_table:
+                num_blocks = len(block_table)
+                if num_blocks <= max_blocks:
+                    graph_block_tables[i, :num_blocks] = block_table
+                else:
+                    # It may be possible to have more blocks allocated due
+                    # to lookahead slots of multi-step, however, they are
+                    # not used anyway, so can be safely ignored.
+                    graph_block_tables[
+                        i, :max_blocks] = block_table[:max_blocks]
+
+        return torch.from_numpy(graph_block_tables).to(
+            device=self.runner.device, non_blocking=True)
+
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int):
+        """Build attention metadata with on-device tensors.
+
+        Args:
+            seq_lens: The maybe padded sequence lengths of the input sequences.
+            query_lens: The query lengths of the input sequences.
+            cuda_graph_pad_size: The padding size for cuda graph.
+                                 -1 if cuda graph is not used.
+            batch_size: The maybe padded batch size.
+        """
+        prefix_cache_hit = any([
+            inter_data.prefix_cache_hit
+            for inter_data in self.input_builder.inter_data_list
+        ])
+        for inter_data in self.input_builder.inter_data_list:
+            self._add_seq_group(inter_data,
+                                self.input_builder.chunked_prefill_enabled,
+                                prefix_cache_hit)
+
+        device = self.runner.device
+        use_captured_graph = cuda_graph_pad_size != -1
+
+        max_query_len = max(query_lens)
+        decode_query_lens = query_lens[self.num_prefills:]
+        if len(decode_query_lens) > 0:
+            max_decode_query_len = max(decode_query_lens)
+        else:
+            max_decode_query_len = 1
+        max_prefill_seq_len = max(self.prefill_seq_lens, default=0)
+        max_decode_seq_len = max(self.curr_seq_lens, default=0)
+        num_decode_tokens = self.num_decode_tokens
+        query_start_loc = list(accumulate(query_lens, initial=0))
+        seq_start_loc = list(accumulate(seq_lens, initial=0))
+
+        num_seqs = len(seq_lens)
+        if use_captured_graph:
+            self.slot_mapping.extend([PAD_SLOT_ID] * cuda_graph_pad_size)
+            self.block_tables.extend([] * cuda_graph_pad_size)
+            num_decode_tokens = batch_size - self.num_prefill_tokens
+            block_tables = self._get_graph_runner_block_tables(
+                num_seqs, self.block_tables)
+        else:
+            block_tables = make_tensor_with_pad(
+                self.block_tables,
+                pad=0,
+                dtype=torch.int,
+                device=device,
+            )
+        assert max_query_len > 0, ("query_lens: {}".format(query_lens))
+
+        assert device is not None
+        context_lens_tensor = async_tensor_h2d(self.context_lens, torch.int,
+                                               device, self.runner.pin_memory)
+        seq_lens_tensor = async_tensor_h2d(seq_lens, torch.int, device,
+                                           self.runner.pin_memory)
+        slot_mapping_tensor = async_tensor_h2d(self.slot_mapping, torch.long,
+                                               device, self.runner.pin_memory)
+        query_start_loc_tensor = async_tensor_h2d(query_start_loc, torch.int32,
+                                                  device,
+                                                  self.runner.pin_memory)
+        seq_start_loc_tensor = async_tensor_h2d(seq_start_loc, torch.int32,
+                                                device, self.runner.pin_memory)
+        placeholder_index_maps = {
+            modality: placeholder_map.index_map()
+            for modality, placeholder_map in
+            self.multimodal_placeholder_maps.items()
+        }
+
+        return FlashAttentionMetadata(
+            num_prefills=self.num_prefills,
+            slot_mapping=slot_mapping_tensor,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=num_decode_tokens,
+            seq_lens=seq_lens,
+            multi_modal_placeholder_index_maps=placeholder_index_maps,
+            enable_kv_scales_calculation=True,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=max_query_len,
+            max_decode_query_len=max_decode_query_len,
+            max_prefill_seq_len=max_prefill_seq_len,
+            max_decode_seq_len=max_decode_seq_len,
+            query_start_loc=query_start_loc_tensor,
+            seq_start_loc=seq_start_loc_tensor,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=use_captured_graph,
+        )
+
+
+class FlashAttentionImpl(AttentionImpl):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prefill_tokens ----------------->|	
+    |<--prefill_0-->|<--prefill_1-->|...|<--prefill_N-1--->|
+
+    Otherwise, the layout is as follows:	
+    |<----------------- num_decode_tokens ------------------>|	
+    |<--decode_0-->|..........|<--decode_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+
+    If chunked prefill is enabled, prefill tokens and decode tokens can be
+    batched together in a flattened 1D query.
+
+    |<----- num_prefill_tokens ---->|<------- num_decode_tokens --------->|
+    |<-prefill_0->|...|<-prefill_N-1->|<--decode_0-->|...|<--decode_M-1-->|
+
+    Currently, cuda graph is disabled for chunked prefill, meaning there's no
+    padding between prefill and decode tokens.
+    """
+
+    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: str = AttentionType.DECODER,
+    ) -> 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
+        self.sliding_window = ((sliding_window - 1,
+                                0) if sliding_window is not None else (-1, -1))
+        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
+
+        assert self.num_heads % self.num_kv_heads == 0
+        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}.")
+        self.attn_type = attn_type
+
+        # if hopper default to FA3, otherwise stick to FA2 for now
+        # TODO(lucas): profile FA3 on ampere to see if it makes sense to
+        #  use FA3 as default for both
+        if current_platform.get_device_capability()[0] >= 9:
+            self.fa_version = 3 if is_fa_version_supported(3) else 2
+        else:
+            self.fa_version = 2
+
+        if VLLM_FLASH_ATTN_VERSION is not None:
+            assert VLLM_FLASH_ATTN_VERSION in [2, 3]
+            self.fa_version = VLLM_FLASH_ATTN_VERSION
+
+        if not is_fa_version_supported(self.fa_version):
+            logger.error("Cannot use FA version %d is not supported due to %s",
+                         self.fa_version,
+                         fa_version_unsupported_reason(self.fa_version))
+
+        assert is_fa_version_supported(self.fa_version)
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: FlashAttentionMetadata,
+        output: 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]
+            output: shape = [num_tokens, num_heads, head_size]
+            kv_cache = [2, num_blocks, block_size, num_kv_heads, head_size]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+        NOTE: It in-place updates the output tensor.
+        """
+        # NOTE(woosuk): FlashAttention does not support FP8 KV cache.
+        assert layer._k_scale_float == 1.0 and layer._v_scale_float == 1.0, (
+            "key/v_scale is not supported in FlashAttention.")
+
+        assert output is not None, "Output tensor must be provided."
+
+        attn_type = self.attn_type
+        if (attn_type == AttentionType.ENCODER
+                and (not attn_metadata.is_all_encoder_attn_metadata_set)):
+            raise AttributeError("Encoder attention requires setting "
+                                 "encoder metadata attributes.")
+        elif (attn_type == AttentionType.ENCODER_DECODER
+              and (not attn_metadata.is_all_cross_attn_metadata_set)):
+            raise AttributeError("Encoder/decoder cross-attention "
+                                 "requires setting cross-attention "
+                                 "metadata attributes.")
+
+        kv_cache_dtype: str = self.kv_cache_dtype
+        softmax_scale: float = self.scale
+        window_size = self.sliding_window
+        alibi_slopes: Optional[torch.Tensor] = self.alibi_slopes
+        logits_soft_cap: Optional[float] = self.logits_soft_cap
+
+        if kv_cache.numel() > 0:
+            key_cache = kv_cache[0]
+            value_cache = kv_cache[1]
+            # We skip updating the KV cache under two conditions:
+            #  a. When the Attention Type is ENCODER. In this phase, we compute
+            #     only the encoder attention without updating the cache.
+            #  b. When both Key and Value are None. This occurs during
+            #     cross-attention computation in the decoding phase, where the
+            #     KV cache is already populated with the cross-attention
+            #     tensor. Thus, we skip cache updates during this time.
+            if (attn_type != AttentionType.ENCODER) and (key is not None) and (
+                    value is not None):
+                if attn_type == AttentionType.ENCODER_DECODER:
+                    # Update cross-attention KV cache (prefill-only)
+                    updated_slot_mapping = attn_metadata.cross_slot_mapping
+                else:
+                    # Update self-attention KV cache (prefill/decode)
+                    updated_slot_mapping = attn_metadata.slot_mapping
+
+                # Reshape the input keys and values and store them in the cache.
+                # If kv_cache is not provided, the new key and value tensors are
+                # not cached. This happens during the initial memory
+                # profiling run.
+                torch.ops._C_cache_ops.reshape_and_cache_flash(
+                    key,
+                    value,
+                    kv_cache[0],
+                    kv_cache[1],
+                    updated_slot_mapping.flatten(),  # type: ignore[union-attr]
+                    kv_cache_dtype,
+                    layer._k_scale,
+                    layer._v_scale,
+                )
+
+        (num_prefill_query_tokens, num_prefill_kv_tokens,
+        num_decode_query_tokens) = \
+            get_num_prefill_decode_query_kv_tokens(attn_metadata, attn_type)
+        decode_query = query[num_prefill_query_tokens:]
+        decode_output = output[num_prefill_query_tokens:]
+        # QKV for prefill.
+        query = query[:num_prefill_query_tokens]
+        prefill_output = output[:num_prefill_query_tokens]
+        assert query.shape[0] == num_prefill_query_tokens
+        assert decode_query.shape[0] == num_decode_query_tokens
+
+        if prefill_meta := attn_metadata.prefill_metadata:
+            # Prompt run.
+            if (kv_cache.numel() == 0 or prefill_meta.block_tables is None
+                    or prefill_meta.block_tables.numel() == 0):
+                # normal attention
+                # When block_tables are not filled, it means q and k are the
+                # prompt, and they have the same length.
+                q_seq_start_loc, q_seq_len, k_seq_start_loc, k_seq_len = \
+                    _get_query_key_seq_metadata(prefill_meta, True, attn_type)
+
+                key = key[:num_prefill_kv_tokens]
+                value = value[:num_prefill_kv_tokens]
+
+                flash_attn_varlen_func(
+                    q=query,
+                    k=key,
+                    v=value,
+                    cu_seqlens_q=q_seq_start_loc,
+                    cu_seqlens_k=k_seq_start_loc,
+                    max_seqlen_q=q_seq_len,
+                    max_seqlen_k=k_seq_len,
+                    softmax_scale=softmax_scale,
+                    causal=_get_causal_option(attn_type),
+                    window_size=window_size,
+                    alibi_slopes=alibi_slopes,
+                    softcap=logits_soft_cap,
+                    out=prefill_output,
+                    fa_version=self.fa_version,
+                )
+            else:
+                # prefix-enabled attention
+                assert attn_type == AttentionType.DECODER, (
+                    "Only decoder-only models support prefix caching")
+                assert prefill_meta.seq_lens is not None
+                max_seq_len = max(prefill_meta.seq_lens)
+                flash_attn_varlen_func(  # noqa
+                    q=query,
+                    k=key_cache,
+                    v=value_cache,
+                    cu_seqlens_q=prefill_meta.query_start_loc,
+                    max_seqlen_q=prefill_meta.max_query_len,
+                    seqused_k=prefill_meta.seq_lens_tensor,
+                    max_seqlen_k=max_seq_len,
+                    softmax_scale=softmax_scale,
+                    causal=True,
+                    window_size=window_size,
+                    alibi_slopes=alibi_slopes,
+                    block_table=prefill_meta.block_tables,
+                    softcap=logits_soft_cap,
+                    out=prefill_output,
+                    fa_version=self.fa_version,
+                )
+
+        if decode_meta := attn_metadata.decode_metadata:
+            # Decoding run.
+            # Use flash_attn_varlen_func kernel for speculative decoding
+            # because different queries might have different lengths.
+
+            assert decode_meta.max_decode_query_len is not None
+            # use only for actual varlen decoding
+            if decode_meta.max_decode_query_len > 1:
+                assert attn_type == AttentionType.DECODER, (
+                    "Only decoder-only models support max_decode_query_len > 1"
+                )
+                flash_attn_varlen_func(
+                    q=decode_query,
+                    k=key_cache,
+                    v=value_cache,
+                    cu_seqlens_q=decode_meta.query_start_loc,
+                    max_seqlen_q=decode_meta.max_decode_query_len,
+                    seqused_k=decode_meta.seq_lens_tensor,
+                    max_seqlen_k=decode_meta.max_decode_seq_len,
+                    softmax_scale=softmax_scale,
+                    causal=True,
+                    window_size=window_size,
+                    alibi_slopes=alibi_slopes,
+                    softcap=logits_soft_cap,
+                    block_table=decode_meta.block_tables,
+                    out=decode_output,
+                    fa_version=self.fa_version,
+                )
+            else:
+                # Use flash_attn_with_kvcache for normal decoding.
+                (
+                    seq_lens_arg,
+                    _,
+                    block_tables_arg,
+                ) = get_seq_len_block_table_args(decode_meta, False, attn_type)
+                flash_attn_with_kvcache(
+                    q=decode_query.unsqueeze(1),
+                    k_cache=key_cache,
+                    v_cache=value_cache,
+                    block_table=block_tables_arg,
+                    cache_seqlens=seq_lens_arg,
+                    softmax_scale=softmax_scale,
+                    causal=True,
+                    window_size=window_size,
+                    alibi_slopes=alibi_slopes,
+                    softcap=logits_soft_cap,
+                    out=decode_output.unsqueeze(1),
+                    fa_version=self.fa_version,
+                )
+        return output
+
+
+def _get_query_key_seq_metadata(
+    attn_metadata,
+    is_prompt: bool,
+    attn_type: str,
+) -> tuple:
+    """
+    Returns sequence metadata for key and query based on the specified 
+    attention type and whether input is a prompt.
+
+    This function computes the starting locations and maximum sequence lengths 
+    for key and query sequences for different attention types.
+
+    Args:
+        attn_metadata: The attention metadata object
+        is_prompt (bool): A flag indicating if the input is a prompt
+        attn_type (AttentionType): The type of attention being used.
+
+    Returns:
+        tuple: A tuple containing four integers:
+            - Starting location for the query sequence.
+            - Maximum sequence length for the query sequence.
+            - Starting location for the key sequence.
+            - Maximum sequence length for the key sequence.
+
+    Raises:
+        AttributeError: If an invalid attention type is provided.
+    """
+    if attn_type == AttentionType.DECODER:
+        # Decoder self-attention
+        # Choose max_seq_len based on whether we are in prompt_run
+        if is_prompt:
+            max_seq_len = attn_metadata.max_prefill_seq_len
+        else:
+            max_seq_len = attn_metadata.max_decode_seq_len
+        return (attn_metadata.seq_start_loc, max_seq_len,
+                attn_metadata.seq_start_loc, max_seq_len)
+
+    elif attn_type == AttentionType.ENCODER_DECODER:
+        # This is cross attention between the where the key
+        # is the precomputed encoder attention and query
+        # is the input sequence.
+        # Choose query max length based on whether it is prompt
+        # or not.
+        if is_prompt:
+            max_seq_len = attn_metadata.max_prefill_seq_len
+        else:
+            max_seq_len = attn_metadata.max_decode_seq_len
+        return (attn_metadata.seq_start_loc, max_seq_len,
+                attn_metadata.encoder_seq_start_loc,
+                attn_metadata.max_encoder_seq_len)
+    elif attn_type == AttentionType.ENCODER:
+        # For encoder attention both the query and the key are same i.e the
+        # encoder sequence.
+        return (attn_metadata.encoder_seq_start_loc,
+                attn_metadata.max_encoder_seq_len,
+                attn_metadata.encoder_seq_start_loc,
+                attn_metadata.max_encoder_seq_len)
+    elif attn_type == AttentionType.ENCODER_ONLY:
+        assert is_prompt, "Should not have decode for encoder only model."
+        return (attn_metadata.seq_start_loc, attn_metadata.max_prefill_seq_len,
+                attn_metadata.seq_start_loc, attn_metadata.max_prefill_seq_len)
+    else:
+        raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+
+def _get_causal_option(attn_type: str) -> bool:
+    """
+    Determine whether the given attention type is suitable for causal 
+    attention mechanisms.
+
+    Args:
+        attn_type (AttentionType): The type of attention being evaluated
+
+    Returns:
+        bool: Returns `True` if the attention type is suitable for causal 
+        attention (i.e., not encoder, encoder-only, or encoder-decoder), 
+        otherwise returns `False`.
+    """
+    return not (attn_type == AttentionType.ENCODER
+                or attn_type == AttentionType.ENCODER_ONLY
+                or attn_type == AttentionType.ENCODER_DECODER)

.venv/lib/python3.11/site-packages/vllm/attention/backends/flashinfer.py

@@ -0,0 +1,1066 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import dataclasses
+from collections import defaultdict
+from contextlib import contextmanager
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Set, Tuple, Type
+
+from vllm.multimodal import MultiModalPlaceholderMap
+
+try:
+    from flashinfer import BatchDecodeWithPagedKVCacheWrapper
+    from flashinfer.decode import CUDAGraphBatchDecodeWithPagedKVCacheWrapper
+    from flashinfer.prefill import BatchPrefillWithPagedKVCacheWrapper
+
+    from vllm.vllm_flash_attn import flash_attn_varlen_func
+    FLASHINFER_WORKSPACE_BUFFER_SIZE = 256 * 1024 * 1024
+except ImportError:
+    # Avoid turning these types into variables during type checking
+    if not TYPE_CHECKING:
+        BatchDecodeWithPagedKVCacheWrapper = None
+        CUDAGraphBatchDecodeWithPagedKVCacheWrapper = None
+        BatchPrefillWithPagedKVCacheWrapper = None
+    FLASHINFER_WORKSPACE_BUFFER_SIZE = 0
+
+import torch
+
+import vllm.envs as envs
+from vllm import _custom_ops as ops
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionLayer,
+                                              AttentionMetadata,
+                                              AttentionMetadataBuilder,
+                                              AttentionState, AttentionType)
+from vllm.attention.backends.utils import (PAD_SLOT_ID, compute_slot_mapping,
+                                           compute_slot_mapping_start_idx,
+                                           is_block_tables_empty)
+from vllm.attention.layer import Attention
+from vllm.attention.ops.paged_attn import PagedAttention
+from vllm.config import VllmConfig, get_current_vllm_config
+from vllm.utils import (async_tensor_h2d, get_kv_cache_torch_dtype,
+                        make_tensor_with_pad)
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import (ModelInputForGPUBuilder,
+                                          ModelInputForGPUWithSamplingMetadata)
+
+
+class FlashInferBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "FLASHINFER"
+
+    @staticmethod
+    def get_impl_cls() -> Type["FlashInferImpl"]:
+        return FlashInferImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return FlashInferMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["FlashInferMetadataBuilder"]:
+        return FlashInferMetadataBuilder
+
+    @staticmethod
+    def get_state_cls() -> Type["FlashInferState"]:
+        return FlashInferState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return (num_blocks, 2, block_size, num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+    @staticmethod
+    def get_supported_head_sizes() -> List[int]:
+        return [64, 128, 256]
+
+    @staticmethod
+    def get_fp8_dtype_for_flashinfer(kv_cache_dtype: str) -> torch.dtype:
+        if kv_cache_dtype in ("fp8", "fp8_e4m3"):
+            return torch.float8_e4m3fn
+        elif kv_cache_dtype == "fp8_e5m2":
+            return torch.float8_e5m2
+        else:
+            raise ValueError(f"Unrecognized FP8 dtype: {kv_cache_dtype}")
+
+
+@dataclass
+class PerLayerParameters:
+    """
+    Currently, FlashInfer backend only support models in which all layers share
+    the same values for the following hyperparameters.
+    """
+
+    window_left: int
+    logits_soft_cap: Optional[float]
+    sm_scale: float
+
+
+def get_per_layer_parameters(
+        vllm_config: VllmConfig) -> Dict[str, PerLayerParameters]:
+    """
+    Scan all attention layers and determine some hyperparameters
+    to use during `plan`.
+    """
+
+    layers = vllm_config.compilation_config.static_forward_context
+    per_layer_params: Dict[str, PerLayerParameters] = {}
+
+    for key, layer in layers.items():
+        assert isinstance(layer, Attention)
+
+        impl = layer.impl
+        assert isinstance(impl, FlashInferImpl)
+
+        # Infer hyperparameters from the attention layer
+        window_size = impl.sliding_window
+        window_left = window_size[0] if window_size is not None else -1
+        logits_soft_cap = impl.logits_soft_cap
+        sm_scale = impl.scale
+
+        per_layer_params[key] = PerLayerParameters(window_left,
+                                                   logits_soft_cap, sm_scale)
+
+    return per_layer_params
+
+
+def infer_global_hyperparameters(
+        per_layer_params: Dict[str, PerLayerParameters]) -> PerLayerParameters:
+    """
+    Currently, FlashInfer backend only support models in which all layers share
+    the same values for the following hyperparameters:
+    - `window_left`
+    - `logits_soft_cap`
+    - `sm_scale`
+
+    So this function asserts that all layers share the same values for these
+    hyperparameters and returns the global values.
+    """
+
+    assert len(per_layer_params) > 0, "No attention layers found in the model."
+
+    param_sets = list(per_layer_params.values())
+    global_params = param_sets[0]
+    for params in param_sets:
+        assert params == global_params, (
+            "FlashInfer backend currently only supports models in which all "
+            "layers share the same values for the following hyperparameters: "
+            "`window_left`, `logits_soft_cap`, `sm_scale`.")
+
+    return global_params
+
+
+class FlashInferState(AttentionState):
+
+    def __init__(self, runner):
+        self.runner = runner
+        self._is_graph_capturing = False
+        self._workspace_buffer = None
+        self._decode_wrapper = None
+        self._prefill_wrapper = None
+
+        # Global hyperparameters shared by all attention layers
+        self.global_hyperparameters: Optional[PerLayerParameters] = None
+
+        self.vllm_config = get_current_vllm_config()
+
+    def _get_workspace_buffer(self):
+        if self._workspace_buffer is None:
+            self._workspace_buffer = torch.empty(
+                FLASHINFER_WORKSPACE_BUFFER_SIZE,
+                dtype=torch.uint8,
+                device=self.runner.device)
+        return self._workspace_buffer
+
+    def _get_prefill_wrapper(self):
+        if self._prefill_wrapper is None:
+            self._prefill_wrapper = BatchPrefillWithPagedKVCacheWrapper(
+                self._get_workspace_buffer(), "NHD")
+        return self._prefill_wrapper
+
+    def _get_decode_wrapper(self):
+        if self._decode_wrapper is None:
+            num_qo_heads = (self.runner.model_config.get_num_attention_heads(
+                self.runner.parallel_config))
+            num_kv_heads = self.runner.model_config.get_num_kv_heads(
+                self.runner.parallel_config)
+            use_tensor_cores = envs.VLLM_FLASHINFER_FORCE_TENSOR_CORES or (
+                num_qo_heads // num_kv_heads > 4)
+            self._decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
+                self._get_workspace_buffer(),
+                "NHD",
+                use_tensor_cores=use_tensor_cores)
+        return self._decode_wrapper
+
+    @contextmanager
+    def graph_capture(self, max_batch_size: int):
+        self._is_graph_capturing = True
+        self._graph_decode_wrapper = None
+        self._graph_slot_mapping = torch.full((max_batch_size, ),
+                                              PAD_SLOT_ID,
+                                              dtype=torch.long,
+                                              device=self.runner.device)
+        self._graph_seq_lens = torch.ones(max_batch_size,
+                                          dtype=torch.int32,
+                                          device=self.runner.device)
+        self._graph_block_tables = torch.from_numpy(
+            self.runner.graph_block_tables).to(device=self.runner.device)
+        self._graph_decode_workspace_buffer = self._get_workspace_buffer()
+        self._graph_indices_buffer = torch.empty(
+            max_batch_size * self.runner.cache_config.num_gpu_blocks,
+            dtype=torch.int32,
+            device=self.runner.device)
+        self._graph_indptr_buffer = torch.empty(max_batch_size + 1,
+                                                dtype=torch.int32,
+                                                device=self.runner.device)
+        self._graph_last_page_len_buffer = torch.empty(
+            max_batch_size, dtype=torch.int32, device=self.runner.device)
+        yield
+        self._is_graph_capturing = False
+        del self._graph_slot_mapping
+        del self._graph_seq_lens
+        del self._graph_block_tables
+        del self._graph_decode_workspace_buffer
+        del self._graph_indices_buffer
+        del self._graph_indptr_buffer
+        del self._graph_last_page_len_buffer
+        del self._graph_decode_wrapper
+
+    def graph_clone(self, batch_size: int):
+        assert self._is_graph_capturing
+        state = self.__class__(self.runner)
+        state._workspace_buffer = self._graph_decode_workspace_buffer
+        state._decode_wrapper = self._graph_decode_wrapper
+        state._prefill_wrapper = self._get_prefill_wrapper()
+        return state
+
+    def graph_capture_get_metadata_for_batch(
+            self, batch_size: int, is_encoder_decoder_model: bool = False):
+        assert self._is_graph_capturing
+        _indptr_buffer = self._graph_indptr_buffer[:batch_size + 1]
+        _last_page_len_buffer = self._graph_last_page_len_buffer[:batch_size]
+
+        num_qo_heads = (self.runner.model_config.get_num_attention_heads(
+            self.runner.parallel_config))
+        num_kv_heads = self.runner.model_config.get_num_kv_heads(
+            self.runner.parallel_config)
+        use_tensor_cores = envs.VLLM_FLASHINFER_FORCE_TENSOR_CORES or (
+            num_qo_heads // num_kv_heads > 4)
+        self._graph_decode_wrapper = \
+            CUDAGraphBatchDecodeWithPagedKVCacheWrapper(
+            self._graph_decode_workspace_buffer, _indptr_buffer,
+            self._graph_indices_buffer, _last_page_len_buffer, "NHD",
+            use_tensor_cores)
+        if self.runner.kv_cache_dtype.startswith("fp8"):
+            kv_cache_dtype = FlashInferBackend.get_fp8_dtype_for_flashinfer(
+                self.runner.kv_cache_dtype)
+        else:
+            kv_cache_dtype = get_kv_cache_torch_dtype(
+                self.runner.kv_cache_dtype, self.runner.model_config.dtype)
+
+        paged_kv_indptr_tensor_host = torch.arange(0,
+                                                   batch_size + 1,
+                                                   dtype=torch.int32)
+        paged_kv_indices_tensor_host = torch.arange(0,
+                                                    batch_size,
+                                                    dtype=torch.int32)
+        paged_kv_last_page_len_tensor_host = torch.full((batch_size, ),
+                                                        self.runner.block_size,
+                                                        dtype=torch.int32)
+        query_start_loc_host = torch.arange(0,
+                                            batch_size + 1,
+                                            dtype=torch.int32)
+
+        global_params = infer_global_hyperparameters(
+            get_per_layer_parameters(self.vllm_config))
+
+        attn_metadata = self.runner.attn_backend.make_metadata(
+            num_prefills=0,
+            slot_mapping=self._graph_slot_mapping[:batch_size],
+            multi_modal_placeholder_index_maps=None,
+            enable_kv_scales_calculation=False,
+            num_prefill_tokens=0,
+            num_decode_tokens=batch_size,
+            max_prefill_seq_len=0,
+            block_tables=self._graph_block_tables,
+            paged_kv_indptr=paged_kv_indptr_tensor_host,
+            paged_kv_indices=paged_kv_indices_tensor_host,
+            paged_kv_last_page_len=paged_kv_last_page_len_tensor_host,
+            num_qo_heads=num_qo_heads,
+            num_kv_heads=num_kv_heads,
+            head_dim=self.runner.model_config.get_head_size(),
+            page_size=self.runner.block_size,
+            seq_start_loc=None,
+            query_start_loc=query_start_loc_host,
+            device=self.runner.device,
+            data_type=kv_cache_dtype,
+            q_data_type=self.runner.model_config.dtype,
+            use_cuda_graph=True,
+            decode_wrapper=self._graph_decode_wrapper,
+            prefill_wrapper=None,
+            **dataclasses.asdict(global_params),
+        )
+        attn_metadata.begin_forward()
+        return attn_metadata
+
+    def get_graph_input_buffers(self,
+                                attn_metadata,
+                                is_encoder_decoder_model: bool = False):
+        return {
+            "slot_mapping": attn_metadata.slot_mapping,
+        }
+
+    def prepare_graph_input_buffers(self,
+                                    input_buffers,
+                                    attn_metadata,
+                                    is_encoder_decoder_model: bool = False):
+        return
+
+    def begin_forward(self, model_input):
+        assert not self._is_graph_capturing
+        state = self
+        use_cuda_graph = model_input.attn_metadata.use_cuda_graph
+        is_decode = model_input.attn_metadata.num_prefills == 0
+        # In case of multistep chunked-prefill, there might be prefill requests
+        # scheduled while CUDA graph mode is enabled. We don't run graph in that
+        # case.
+        if use_cuda_graph and is_decode:
+            batch_size = model_input.input_tokens.shape[0]
+            state = (self.runner.graph_runners[model_input.virtual_engine]
+                     [batch_size].attn_state)
+        model_input.attn_metadata.prefill_wrapper = state._get_prefill_wrapper(
+        )
+        model_input.attn_metadata.decode_wrapper = state._get_decode_wrapper()
+        model_input.attn_metadata.begin_forward()
+
+
+@dataclass
+class FlashInferMetadata(AttentionMetadata):
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Number of query tokens for each request in the batch.
+    # Currently, we require that all requests have the same number of query
+    # tokens during the decoding phase. When speculavie decoding is enabled,
+    # decode_query_len might be greater than 1. In all other cases, it is 1.
+    decode_query_len: Optional[int] = 1
+
+    use_cuda_graph: bool = True
+
+    prefill_wrapper: Optional[BatchPrefillWithPagedKVCacheWrapper] = None
+    decode_wrapper: Optional[BatchDecodeWithPagedKVCacheWrapper] = None
+
+    # Metadata for the prefill stage
+    seq_start_loc: Optional[torch.Tensor] = None
+    query_start_loc: Optional[torch.Tensor] = None
+    block_tables: Optional[torch.Tensor] = None
+
+    # used for GPU in-place advance_step
+    seq_lens_tensor: Optional[torch.Tensor] = None
+    block_table_bound: Optional[torch.Tensor] = None
+
+    # An example for paged_kv_indices, paged_kv_indptr:
+    # request 1, page indices [0, 5, 8]
+    # request 2, page indices [1, 6, 7]
+    # request 3, page indices [3, 4]
+    # paged_kv_indices is a concatenation of page indices of all requests:
+    # [0, 5, 8, 1, 6, 7, 3, 4]
+    # paged_kv_indptr is used to index into paged_kv_indices:
+    # [0, 3, 6, 8]
+    # The indptr of the paged kv cache, shape: [batch_size + 1]
+    paged_kv_indptr: Optional[torch.Tensor] = None
+    # The page indices of the paged kv cache
+    paged_kv_indices: Optional[torch.Tensor] = None
+    # The number of entries in the last page of each request in
+    # the paged kv cache, shape: [batch_size]
+    paged_kv_last_page_len: Optional[torch.Tensor] = None
+    # The number of query/output heads
+    num_qo_heads: Optional[int] = None
+    # The number of key/value heads
+    num_kv_heads: Optional[int] = None
+    # The dimension of the attention heads
+    head_dim: Optional[int] = None
+    # Block size of vllm
+    page_size: Optional[int] = None
+    # The data type of the paged kv cache
+    data_type: torch.dtype = None
+    # The data type of the query
+    q_data_type: torch.dtype = None
+    # FlashInfer 0.2 encourages passing host tensors
+    device: torch.device = torch.device("cpu")
+    is_profile_run: bool = False
+
+    # The FlashInfer backend currently supports only models in which all layers
+    # share the same following hyperparameters:
+
+    # The left (inclusive) window size for the attention window, when
+    # set to `-1`, the window size will be set to the full length of
+    # the sequence. Defaults to `-1`.
+    window_left: int = -1
+    # The attention logits soft capping value (used in Gemini, Grok and
+    # Gemma-2, etc.), if not provided, will be set to `0`. If greater
+    # than 0, the logits will be capped according to formula:
+    # $$\texttt{logits\_soft\_cap} \times
+    # \mathrm{tanh}(x / \texttt{logits\_soft\_cap})$$,
+    # where $x$ is the input logits.
+    logits_soft_cap: Optional[float] = None
+    # The scale used in softmax, if not provided, will be set to
+    # `1.0 / sqrt(head_dim)`.
+    sm_scale: Optional[float] = None
+
+    def __post_init__(self):
+        # Refer to
+        # https://github.com/flashinfer-ai/flashinfer/blob/3d55c71a62052c590c130897d3a3db49b14fcc34/include/flashinfer/utils.cuh#L157
+        supported_head_sizes = FlashInferBackend.get_supported_head_sizes()
+        if self.head_dim is not None and self.head_dim \
+                not in supported_head_sizes:
+            raise ValueError(
+                f"Only {supported_head_sizes} are supported for head_dim,",
+                f"received {self.head_dim}.")
+
+    def begin_forward(self):
+        if self.num_prefill_tokens > 0:
+            if self.paged_kv_indices is None:
+                return
+
+            assert self.prefill_wrapper is not None
+            assert self.query_start_loc is not None
+            assert self.paged_kv_indices is not None
+            assert self.paged_kv_indptr is not None
+            assert self.paged_kv_last_page_len is not None
+            assert self.block_table_bound is not None
+            assert self.seq_lens_tensor is not None
+            self.query_start_loc = self.query_start_loc[:self.num_prefills + 1]
+            batch_size = self.query_start_loc.shape[0] - 1
+            assert batch_size >= 0
+            # We will use flash attention for profiling to
+            # determine the number of blocks. Therefore,
+            # we don't need to prepare the input for flashinfer for profile run.
+            if not self.is_profile_run:
+                self.paged_kv_indptr = self.paged_kv_indptr.to(self.device)
+                self.paged_kv_last_page_len = self.paged_kv_last_page_len.to(
+                    self.device)
+                self.block_table_bound = self.block_table_bound.to(self.device)
+                self.seq_lens_tensor = self.seq_lens_tensor.to(self.device)
+                self.paged_kv_indices = self.paged_kv_indices.to(self.device)
+                self.prefill_wrapper.plan(
+                    self.query_start_loc,
+                    self.paged_kv_indptr[:self.num_prefills + 1],
+                    self.paged_kv_indices,
+                    self.paged_kv_last_page_len[:self.num_prefills],
+                    self.num_qo_heads,
+                    self.num_kv_heads,
+                    self.head_dim,
+                    self.page_size,
+                    causal=True,
+                    sm_scale=self.sm_scale,
+                    window_left=self.window_left,
+                    logits_soft_cap=self.logits_soft_cap,
+                    q_data_type=self.q_data_type,
+                    kv_data_type=self.data_type)
+        if self.num_decode_tokens > 0:
+            assert self.paged_kv_indices is not None
+            assert self.paged_kv_indptr is not None
+            assert self.paged_kv_last_page_len is not None
+            self.paged_kv_indices = self.paged_kv_indices.to(self.device)
+            self.paged_kv_indptr = self.paged_kv_indptr.to(self.device)
+            self.paged_kv_last_page_len = self.paged_kv_last_page_len.to(
+                self.device)
+            # handle model warmup path
+            if self.block_table_bound is not None:
+                self.block_table_bound = self.block_table_bound.to(self.device)
+            if self.seq_lens_tensor is not None:
+                self.seq_lens_tensor = self.seq_lens_tensor.to(self.device)
+
+            assert self.decode_wrapper is not None
+            self.decode_wrapper.plan(
+                self.paged_kv_indptr[self.num_prefills:],
+                self.paged_kv_indices,
+                self.paged_kv_last_page_len[self.num_prefills:],
+                self.num_qo_heads,
+                self.num_kv_heads,
+                self.head_dim,
+                self.page_size,
+                # Disable flashinfer's pos encoding and use vllm's rope.
+                pos_encoding_mode="NONE",
+                window_left=self.window_left,
+                logits_soft_cap=self.logits_soft_cap,
+                sm_scale=self.sm_scale,
+                # kv-cache data type.
+                kv_data_type=self.data_type,
+                # query data type.
+                q_data_type=self.q_data_type)
+
+    def asdict_zerocopy(self,
+                        skip_fields: Optional[Set[str]] = None
+                        ) -> Dict[str, Any]:
+        if skip_fields is None:
+            skip_fields = set()
+        # We need to skip the prefill/decode_wrapper field since it cannot be
+        # broadcasted with nccl when TP is enabled.
+        skip_fields.add('prefill_wrapper')
+        skip_fields.add('decode_wrapper')
+        return super().asdict_zerocopy(skip_fields)
+
+    @property
+    def prefill_metadata(self) -> Optional["FlashInferMetadata"]:
+        if self.num_prefills == 0:
+            return None
+        return self
+
+    @property
+    def decode_metadata(self) -> Optional["FlashInferMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+        return self
+
+    def advance_step(self,
+                     model_input: "ModelInputForGPUWithSamplingMetadata",
+                     sampled_token_ids: Optional[torch.Tensor],
+                     block_size: int,
+                     num_seqs: int,
+                     num_queries: int,
+                     turn_prefills_into_decodes: bool = False):
+        """
+        Update metadata in-place to advance one decode step.
+        """
+
+        if turn_prefills_into_decodes:
+            # When Multi-Step is enabled with Chunked-Prefill, prefills and
+            # decodes are scheduled together. In the first step, all the
+            # prefills turn into decodes. This update reflects that
+            # conversion.
+            assert self.num_decode_tokens + self.num_prefills == num_seqs
+            # Flashinfer doesn't support speculative decoding + chunked-prefill
+            # + multi-step scheduling yet.
+            assert self.decode_query_len == 1
+            self.num_decode_tokens += self.num_prefills
+            self.num_prefills = 0
+            self.num_prefill_tokens = 0
+            self.max_prefill_seq_len = 0
+            self.max_query_len = 1
+
+            self.slot_mapping = self.slot_mapping[:num_seqs]
+        else:
+            assert self.seq_lens_tensor is not None
+
+        assert num_seqs > 0
+        assert num_queries > 0
+        assert model_input.attn_metadata is not None
+        assert sampled_token_ids is not None
+
+        # When using cudagraph, the num_seqs is padded to the next captured
+        # batch sized, but num_queries tracks the actual number of requests in
+        # the batch. For --enforce-eager mode, num_seqs == num_queries
+        if num_seqs != num_queries:
+            assert num_seqs > num_queries
+            assert self.use_cuda_graph
+
+        model_input.input_tokens[:num_queries] = sampled_token_ids.flatten()
+
+        # Update GPU tensors
+        ops.advance_step_flashinfer(
+            num_seqs=num_seqs,
+            num_queries=num_queries,
+            block_size=block_size,
+            input_tokens=model_input.input_tokens,
+            sampled_token_ids=model_input.input_tokens,
+            input_positions=model_input.input_positions,
+            seq_lens=self.seq_lens_tensor,
+            slot_mapping=self.slot_mapping,
+            block_tables=self.block_tables,
+            paged_kv_indices=self.paged_kv_indices,
+            paged_kv_indptr=self.paged_kv_indptr,
+            paged_kv_last_page_len=self.paged_kv_last_page_len,
+            block_table_bound=self.block_table_bound)
+
+
+class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
+
+    def __init__(self, input_builder: "ModelInputForGPUBuilder"):
+
+        self.input_builder = input_builder
+        self.runner = input_builder.runner
+
+        self.sliding_window = input_builder.sliding_window
+        self.block_size = input_builder.block_size
+
+        # Global hyperparameters shared by all attention layers
+        self.global_hyperparameters: Optional[PerLayerParameters] = None
+
+        self.vllm_config = get_current_vllm_config()
+
+    def prepare(self):
+        self.slot_mapping: List[int] = []
+        self.prefill_seq_lens: List[int] = []
+        self.context_lens: List[int] = []
+        self.block_tables: List[List[int]] = []
+        self.curr_seq_lens: List[int] = []
+        self.multimodal_placeholder_maps: Dict[
+            str,
+            MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
+        self.num_prefills = 0
+        self.num_prefill_tokens = 0
+        self.num_decode_tokens = 0
+
+        # Please follow https://docs.flashinfer.ai/tutorials/kv_layout.html#page-layout
+        # for the precise definition of the following fields.
+        # An example:
+        # request 1, page indices [0, 5, 8]
+        # request 2, page indices [1, 6, 7]
+        # request 3, page indices [3, 4]
+        # paged_kv_indices is a concatenation of page indices of all requests:
+        # [0, 5, 8, 1, 6, 7, 3, 4]
+        # paged_kv_indptr is used to index into paged_kv_indices:
+        # [0, 3, 6, 8]
+        self.paged_kv_indices: List[int] = []
+        # 0 at the beginning of paged_kv_indptr indicates the start of the
+        # first request’s page indices in the paged_kv_indices list.
+        self.paged_kv_indptr: List[int] = [0]
+        # paged_kv_last_page_len is the length of the last page of each request
+        self.paged_kv_last_page_len: List[int] = []
+        self.total_blocks = 0
+        self.is_profile_run: bool = False
+
+        if self.global_hyperparameters is None:
+            # Infer global hyperparameters, since currently we only support
+            # models in which all layers share the same values for the
+            # following hyperparameters:
+            # - `window_left`
+            # - `logits_soft_cap`
+            # - `sm_scale`
+            inferred_params = infer_global_hyperparameters(
+                get_per_layer_parameters(self.vllm_config))
+            self.global_hyperparameters = inferred_params
+            self.window_left = inferred_params.window_left
+            self.logits_soft_cap = inferred_params.logits_soft_cap
+            self.sm_scale = inferred_params.sm_scale
+
+    def _add_seq_group(
+            self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
+            chunked_prefill_enabled: bool):
+        """Add a sequence group to the metadata. Specifically update/append
+        1. context length.
+        2. block table.
+        3. slot mapping.
+        """
+        is_prompt = inter_data.is_prompt
+        block_tables = inter_data.block_tables
+        computed_block_nums = inter_data.computed_block_nums
+
+        for (seq_id, token_len, seq_len, curr_seq_len, query_len, context_len,
+             curr_sliding_window_block) in zip(
+                 inter_data.seq_ids, [len(t) for t in inter_data.input_tokens],
+                 inter_data.orig_seq_lens, inter_data.seq_lens,
+                 inter_data.query_lens, inter_data.context_lens,
+                 inter_data.curr_sliding_window_blocks):
+            self.context_lens.append(context_len)
+            if is_prompt:
+                mm_maps = inter_data.multi_modal_placeholder_maps
+                if mm_maps:
+                    for modality, placeholders in mm_maps.items():
+                        self.multimodal_placeholder_maps[modality].extend(
+                            placeholders)
+                self.num_prefills += 1
+                self.num_prefill_tokens += token_len
+                self.prefill_seq_lens.append(seq_len)
+            else:
+                assert query_len == 1, (
+                    "seq_len: {}, context_len: {}, query_len: {}".format(
+                        seq_len, context_len, query_len))
+                self.num_decode_tokens += query_len
+                self.curr_seq_lens.append(curr_seq_len)
+
+            # Compute block table.
+            # TODO(sang): Combine chunked prefill and prefix caching by
+            # only allowing multiple of block_size chunk size.
+            # NOTE: This only works for oooooooxxx style attention.
+            block_table = []
+            if inter_data.prefix_cache_hit:
+                block_table = computed_block_nums
+            elif ((chunked_prefill_enabled or not is_prompt)
+                  and block_tables is not None):
+                block_table = block_tables[seq_id][-curr_sliding_window_block:]
+            self.block_tables.append(block_table)
+
+            is_profile_run = is_block_tables_empty(block_tables)
+
+            # Compute slot mapping.
+            start_idx = compute_slot_mapping_start_idx(is_prompt, query_len,
+                                                       context_len,
+                                                       self.sliding_window)
+            compute_slot_mapping(is_profile_run, self.slot_mapping, seq_id,
+                                 seq_len, context_len, start_idx,
+                                 self.block_size, inter_data.block_tables)
+
+            # It is not necessary to add paged_kv_indices, paged_kv_indptr,
+            # and paged_kv_last_page_len for profile run because we will
+            # create dummy inputs.
+            if is_profile_run:
+                self.is_profile_run = is_profile_run
+                return
+
+            block_table = block_tables[seq_id]
+            self._update_paged_kv_tensors(block_table, seq_len)
+
+    def _update_paged_kv_tensors(self, block_table: List[int], seq_len: int):
+        # Get the number of valid blocks based on sequence length.
+        # If seq_len = 16, block_size = 16,
+        # block_table_bound is 1 with 1 valid block.
+        # If seq_len = 15, block_size = 16,
+        # block_table_bound is 0 + 1 with 1 valid block.
+        self.total_blocks += len(block_table)
+        block_table_bound = seq_len // self.block_size + 1 \
+                            if seq_len % self.block_size != 0 \
+                            else seq_len // self.block_size
+        self.paged_kv_indices.extend(block_table[:block_table_bound])
+        self.paged_kv_indptr.append(self.paged_kv_indptr[-1] +
+                                    block_table_bound)
+
+        last_page_len = seq_len % self.block_size
+        if last_page_len == 0:
+            last_page_len = self.block_size
+        self.paged_kv_last_page_len.append(last_page_len)
+
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int):
+        """Build attention metadata with on-device tensors.
+
+        Args:
+            seq_lens: The maybe padded sequence lengths of the input sequences.
+            query_lens: The query lengths of the input sequences.
+            cuda_graph_pad_size: The padding size for cuda graph.
+                                 -1 if cuda graph is not used.
+            batch_size: The maybe padded batch size.
+        """
+        for inter_data in self.input_builder.inter_data_list:
+            self._add_seq_group(inter_data,
+                                self.input_builder.chunked_prefill_enabled)
+
+        device = self.runner.device
+        use_captured_graph = cuda_graph_pad_size != -1
+
+        max_prefill_seq_len = max(self.prefill_seq_lens, default=0)
+        num_decode_tokens = self.num_decode_tokens
+        decode_query_len = max(query_lens[self.num_prefills:], default=1)
+
+        if use_captured_graph:
+            self.slot_mapping.extend([PAD_SLOT_ID] * cuda_graph_pad_size)
+            self.block_tables.extend([] * cuda_graph_pad_size)
+            num_decode_tokens = batch_size - self.num_prefill_tokens
+
+            # The shape of graph_block_tables is
+            # [max batch size, max context len // block size].
+            input_block_tables = self.runner.graph_block_tables[:batch_size]
+            max_blocks = input_block_tables.shape[1]
+            for i, block_table in enumerate(self.block_tables):
+                if block_table:
+                    num_blocks = len(block_table)
+                    if num_blocks <= max_blocks:
+                        input_block_tables[i, :num_blocks] = block_table
+                    else:
+                        # It may be possible to have more blocks allocated due
+                        # to lookahead slots of multi-step, however, they are
+                        # not used anyway, so can be safely ignored.
+                        input_block_tables[
+                            i, :max_blocks] = block_table[:max_blocks]
+
+            block_tables = torch.from_numpy(input_block_tables).to(
+                device, non_blocking=True)
+
+            last_paged_kv_indptr = self.paged_kv_indptr[-1]
+            self.paged_kv_indptr.extend([last_paged_kv_indptr] *
+                                        cuda_graph_pad_size)
+            self.paged_kv_last_page_len.extend([0] * cuda_graph_pad_size)
+        else:
+            block_tables = make_tensor_with_pad(
+                self.block_tables,
+                pad=0,
+                dtype=torch.int,
+                device=device,
+            )
+
+        assert device is not None
+        seq_lens_tensor = async_tensor_h2d(seq_lens, torch.int, device,
+                                           self.runner.pin_memory)
+        query_lens_tensor = async_tensor_h2d(query_lens, torch.long, device,
+                                             self.runner.pin_memory)
+        slot_mapping_tensor = async_tensor_h2d(self.slot_mapping, torch.long,
+                                               device, self.runner.pin_memory)
+        query_start_loc = torch.zeros(query_lens_tensor.shape[0] + 1,
+                                      dtype=torch.int32,
+                                      device=device)
+        seq_start_loc = torch.zeros(seq_lens_tensor.shape[0] + 1,
+                                    dtype=torch.int32,
+                                    device=device)
+        placeholder_index_maps = {
+            modality: placeholder_map.index_map()
+            for modality, placeholder_map in
+            self.multimodal_placeholder_maps.items()
+        }
+        torch.cumsum(seq_lens_tensor,
+                     dim=0,
+                     dtype=seq_start_loc.dtype,
+                     out=seq_start_loc[1:])
+        torch.cumsum(query_lens_tensor,
+                     dim=0,
+                     dtype=query_start_loc.dtype,
+                     out=query_start_loc[1:])
+
+        if len(self.paged_kv_indptr) > 0:
+            # extend to the maximum number of blocks as returned by the
+            # scheduler
+            self.paged_kv_indices.extend(
+                [0] * (self.total_blocks - len(self.paged_kv_indices)))
+            paged_kv_indices_tensor = torch.tensor(self.paged_kv_indices,
+                                                   device="cpu",
+                                                   dtype=torch.int)
+            paged_kv_indptr_tensor = torch.tensor(self.paged_kv_indptr,
+                                                  device="cpu",
+                                                  dtype=torch.int)
+            paged_kv_last_page_len_tensor = torch.tensor(
+                self.paged_kv_last_page_len, device="cpu", dtype=torch.int)
+            block_table_bound_tensor = torch.zeros(len(self.paged_kv_indptr) -
+                                                   1,
+                                                   device="cpu",
+                                                   dtype=torch.int)
+        else:
+            paged_kv_indices_tensor = None
+            paged_kv_indptr_tensor = None
+            paged_kv_last_page_len_tensor = None
+            block_table_bound_tensor = None
+
+        if self.runner.kv_cache_dtype.startswith("fp8"):
+            kv_cache_dtype = FlashInferBackend.get_fp8_dtype_for_flashinfer(
+                self.runner.kv_cache_dtype)
+        else:
+            kv_cache_dtype = get_kv_cache_torch_dtype(
+                self.runner.kv_cache_dtype, self.runner.model_config.dtype)
+
+        return FlashInferMetadata(
+            decode_query_len=decode_query_len,
+            num_prefills=self.num_prefills,
+            slot_mapping=slot_mapping_tensor,
+            multi_modal_placeholder_index_maps=placeholder_index_maps,
+            enable_kv_scales_calculation=False,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=num_decode_tokens,
+            max_prefill_seq_len=max_prefill_seq_len,
+            block_tables=block_tables,
+            paged_kv_indptr=paged_kv_indptr_tensor,
+            paged_kv_indices=paged_kv_indices_tensor,
+            paged_kv_last_page_len=paged_kv_last_page_len_tensor,
+            block_table_bound=block_table_bound_tensor,
+            seq_lens_tensor=seq_lens_tensor,
+            num_qo_heads=self.runner.model_config.get_num_attention_heads(
+                self.runner.parallel_config),
+            num_kv_heads=self.runner.model_config.get_num_kv_heads(
+                self.runner.parallel_config),
+            head_dim=self.runner.model_config.get_head_size(),
+            page_size=self.block_size,
+            seq_start_loc=seq_start_loc,
+            query_start_loc=query_start_loc,
+            device=device,
+            data_type=kv_cache_dtype,
+            q_data_type=self.runner.model_config.dtype,
+            use_cuda_graph=use_captured_graph,
+            is_profile_run=self.is_profile_run,
+            window_left=self.window_left,
+            logits_soft_cap=self.logits_soft_cap,
+            sm_scale=self.sm_scale,
+        )
+
+
+class FlashInferImpl(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: str = AttentionType.DECODER,
+    ) -> None:
+        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
+        self.sliding_window = ((sliding_window - 1,
+                                0) if sliding_window is not None else (-1, -1))
+        self.kv_cache_dtype = kv_cache_dtype
+        self.logits_soft_cap = logits_soft_cap
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "FlashInferImpl")
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: FlashInferMetadata,
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+
+        # TODO: directly write to output tensor
+        num_heads: int = self.num_heads
+        head_size: int = self.head_size
+        num_kv_heads: int = self.num_kv_heads
+        kv_cache_dtype: str = self.kv_cache_dtype
+        softmax_scale: float = self.scale
+        window_size = self.sliding_window
+        alibi_slopes = self.alibi_slopes
+        logits_soft_cap = self.logits_soft_cap
+
+        num_tokens, hidden_size = query.shape
+        query = query.view(-1, num_heads, head_size)
+        key = key.view(-1, num_kv_heads, head_size)
+        value = value.view(-1, num_kv_heads, head_size)
+
+        if kv_cache.numel() > 0:
+            # Use the same reshape and cache kernel as flash attention.
+            ops.reshape_and_cache_flash(
+                key,
+                value,
+                kv_cache[:, 0],
+                kv_cache[:, 1],
+                attn_metadata.slot_mapping.flatten(),
+                kv_cache_dtype,
+                layer._k_scale,
+                layer._v_scale,
+            )
+            # The FlashInfer api requires data to be in fp8_e4m3 or fp8_e5m2
+            # to process the cache when the kv_cache_dtype is fp8
+            if kv_cache_dtype.startswith("fp8"):
+                torch_dtype = FlashInferBackend.get_fp8_dtype_for_flashinfer(
+                    kv_cache_dtype)
+                kv_cache = kv_cache.view(torch_dtype)
+
+        num_prefill_tokens = attn_metadata.num_prefill_tokens
+        num_decode_tokens = attn_metadata.num_decode_tokens
+        assert key.shape[0] == num_prefill_tokens + num_decode_tokens, \
+                    f"key : {key.shape} : #prefill tokens {num_prefill_tokens} : #decode tokens {num_decode_tokens}" # noqa
+        assert value.shape[0] == num_prefill_tokens + num_decode_tokens, \
+                    f"value : {value.shape} : #prefill toks {num_prefill_tokens} : #decode toks {num_decode_tokens}" # noqa
+        query = query.contiguous(
+        )  # Flashinfer requires query to be contiguous
+        # Query for decode. KV is not needed because it is already cached.
+        # QKV for prefill.
+        decode_query = query[num_prefill_tokens:]
+        query = query[:num_prefill_tokens]
+
+        key = key[:num_prefill_tokens]
+        value = value[:num_prefill_tokens]
+
+        assert query.shape[0] == num_prefill_tokens
+        assert decode_query.shape[0] == num_decode_tokens
+
+        window_left = window_size[0] if window_size is not None else -1
+
+        prefill_output: Optional[torch.Tensor] = None
+        decode_output: Optional[torch.Tensor] = None
+        if prefill_meta := attn_metadata.prefill_metadata:
+            # We will use flash attention for prefill
+            # when kv_cache is not provided.
+            # This happens when vllm runs the profiling to
+            # determine the number of blocks.
+            if kv_cache.numel() == 0:
+                prefill_output = flash_attn_varlen_func(
+                    q=query,
+                    k=key,
+                    v=value,
+                    cu_seqlens_q=prefill_meta.seq_start_loc,
+                    cu_seqlens_k=prefill_meta.seq_start_loc,
+                    max_seqlen_q=prefill_meta.max_prefill_seq_len,
+                    max_seqlen_k=prefill_meta.max_prefill_seq_len,
+                    softmax_scale=softmax_scale,
+                    causal=True,
+                    window_size=window_size,
+                    alibi_slopes=alibi_slopes,
+                )
+            else:
+                assert prefill_meta is not None
+                assert prefill_meta.prefill_wrapper is not None
+
+                assert prefill_meta.prefill_wrapper._causal
+                assert prefill_meta.prefill_wrapper._window_left == window_left
+                assert prefill_meta.prefill_wrapper._logits_soft_cap == (
+                    logits_soft_cap or 0.0)
+                assert prefill_meta.prefill_wrapper._sm_scale == softmax_scale
+
+                prefill_output = prefill_meta.prefill_wrapper.run(
+                    query,
+                    kv_cache,
+                    k_scale=layer._k_scale_float,
+                    v_scale=layer._v_scale_float,
+                )
+        if decode_meta := attn_metadata.decode_metadata:
+            assert decode_meta is not None
+            assert decode_meta.decode_wrapper is not None
+
+            assert decode_meta.decode_wrapper._window_left == window_left
+            assert decode_meta.decode_wrapper._logits_soft_cap == (
+                logits_soft_cap or 0.0)
+            assert decode_meta.decode_wrapper._sm_scale == softmax_scale
+
+            decode_output = decode_meta.decode_wrapper.run(
+                decode_query,
+                kv_cache,
+                k_scale=layer._k_scale_float,
+                v_scale=layer._v_scale_float,
+            )
+
+        if prefill_output is None and decode_output is not None:
+            # Decode only batch.
+            output, num_tokens = decode_output, num_decode_tokens
+        elif decode_output is None and prefill_output is not None:
+            # Prefill only batch.
+            output, num_tokens = prefill_output, num_prefill_tokens
+        else:
+            # Chunked prefill batch does not work with speculative decoding in
+            # FlashInfer backend, so the query length for decode should be 1.
+            assert prefill_output is not None
+            assert decode_output is not None
+            assert decode_meta is not None
+            assert decode_meta.decode_query_len == 1
+            decode_output = decode_output.squeeze(1)
+            output = torch.cat([prefill_output, decode_output], dim=0)
+        return output.view(num_tokens, hidden_size)

.venv/lib/python3.11/site-packages/vllm/attention/backends/hpu_attn.py

@@ -0,0 +1,293 @@
+# SPDX-License-Identifier: Apache-2.0
+
+###############################################################################
+# Copyright (C) 2024 Habana Labs, Ltd. an Intel Company
+###############################################################################
+
+import os
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+import vllm_hpu_extension.ops as ops
+from vllm_hpu_extension.utils import (Matmul, ModuleFusedSDPA, Softmax,
+                                      VLLMKVCache)
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionLayer,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import CommonAttentionState
+from vllm.attention.ops.hpu_paged_attn import (HPUPagedAttention,
+                                               HPUPagedAttentionMetadata)
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class HPUAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "HPU_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> Type["HPUAttentionImpl"]:
+        return HPUAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return HPUAttentionMetadata
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return HPUPagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                    num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        HPUPagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: Dict[int, List[int]],
+    ) -> None:
+        HPUPagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class HPUAttentionMetadata(HPUPagedAttentionMetadata, AttentionMetadata):
+    """Metadata for HPUAttentionbackend."""
+    # Currently, input sequences can only contain all prompts
+    # or all decoding. True if all sequences are prompts.
+    is_prompt: bool
+    attn_bias: Optional[torch.Tensor]
+    seq_lens_tensor: Optional[torch.Tensor]
+
+
+class HPUAttentionImpl(AttentionImpl, torch.nn.Module):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prefill_tokens ----------------->|
+    |<--prefill_0-->|<--prefill_1-->|...|<--prefill_N-1--->|
+
+    Otherwise, the layout is as follows:
+    |<----------------- num_decode_tokens ------------------>|
+    |<--decode_0-->|..........|<--decode_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+    """
+
+    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,
+        max_seq_len: int = 4096,
+        attn_type: str = AttentionType.DECODER,
+    ) -> None:
+        super(AttentionImpl, self).__init__()
+        self.kv_cache_dtype = kv_cache_dtype
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.matmul_qk = Matmul()
+        self.softmax = Softmax()
+        self.matmul_av = Matmul()
+        self.batch2block_matmul = Matmul()
+        self.block2batch_matmul = Matmul()
+        # NOTE(kzawora): Contiguous PA is off until model runner supports it
+        self.k_cache = VLLMKVCache()
+        self.k_cache.use_contiguous_pa = False
+        self.v_cache = VLLMKVCache()
+        self.v_cache.use_contiguous_pa = False
+        # NOTE(kzawora): Pipelined PA is off until model runner supports it
+        ops.pa_impl = ops.pa
+
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.sliding_window = sliding_window
+        self.alibi_slopes = alibi_slopes
+        if alibi_slopes is not None:
+            alibi_slopes_tensor = torch.tensor(alibi_slopes,
+                                               dtype=torch.bfloat16)
+            self.alibi_slopes = alibi_slopes_tensor
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        self.prefill_usefusedsdpa = os.getenv('VLLM_PROMPT_USE_FUSEDSDPA',
+                                              '0').lower() in ['1', 'true']
+        self.fused_scaled_dot_product_attention = None
+        if self.prefill_usefusedsdpa:
+            assert alibi_slopes is None, \
+                'Prefill with FusedSDPA not supported with alibi slopes!'
+            try:
+                from habana_frameworks.torch.hpex.kernels import FusedSDPA
+                self.fused_scaled_dot_product_attention = ModuleFusedSDPA(
+                    FusedSDPA)
+            except ImportError:
+                logger().warning("Could not import HPU FusedSDPA kernel. "
+                                 "vLLM will use native implementation.")
+
+        suppored_head_sizes = HPUPagedAttention.get_supported_head_sizes()
+        if head_size not in suppored_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {suppored_head_sizes}.")
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "HPUAttentionImpl")
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: HPUAttentionMetadata,
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Forward pass with xFormers and PagedAttention.
+
+        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]
+        """
+        batch_size, seq_len, hidden_size = query.shape
+        _, seq_len_kv, _ = key.shape
+
+        query = query.view(-1, self.num_heads, self.head_size)
+        key = key.view(-1, self.num_kv_heads, self.head_size)
+        value = value.view(-1, self.num_kv_heads, self.head_size)
+        block_indices = attn_metadata.block_indices
+        block_offsets = attn_metadata.block_offsets
+        if attn_metadata.is_prompt:
+            key = key.unflatten(0, (block_indices.size(0), -1))
+            value = value.unflatten(0, (block_indices.size(0), -1))
+        if kv_cache is not None:
+            key_cache, value_cache = HPUPagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            # Reshape the input keys and values and store them in the cache.
+            # If kv_cache is not provided, the new key and value tensors are
+            # not cached. This happens during the initial memory profiling run.
+            key_cache = self.k_cache(key, key_cache, block_indices,
+                                     block_offsets)
+            value_cache = self.v_cache(value, value_cache, block_indices,
+                                       block_offsets)
+
+        if attn_metadata.is_prompt:
+            # Prompt run.
+            if not self.prefill_usefusedsdpa:
+                # TODO: move this outside of model
+                assert attn_metadata.attn_bias is not None, \
+                        'attn_bias must be set before calling model.forward!'
+                attn_bias = attn_metadata.attn_bias
+                if self.alibi_slopes is not None:
+                    position_bias = _make_alibi_bias(self.alibi_slopes,
+                                                     self.num_kv_heads,
+                                                     attn_bias.dtype,
+                                                     attn_bias.shape[-1])
+                    attn_bias = attn_bias.tile((1, self.num_kv_heads, 1, 1))
+                    attn_bias.add_(position_bias)
+            else:
+                attn_bias = None
+
+            query_shape = (batch_size, seq_len, self.num_heads, self.head_size)
+            kv_shape = (batch_size, seq_len_kv, self.num_kv_heads,
+                        self.head_size)
+            out = ops.prompt_attention(
+                query.view(query_shape),
+                key.view(kv_shape),
+                value.view(kv_shape),
+                attn_bias=attn_bias,
+                p=0.0,
+                scale=self.scale,
+                matmul_qk_op=self.matmul_qk,
+                softmax_op=self.softmax,
+                matmul_av_op=self.matmul_av,
+                fsdpa_op=self.fused_scaled_dot_product_attention,
+            )
+            output = out.reshape(batch_size, seq_len, hidden_size)
+        else:
+            # Decoding run.
+            output = HPUPagedAttention.forward_decode(
+                query=query,
+                key_cache=key_cache,
+                value_cache=value_cache,
+                block_list=attn_metadata.block_list,
+                block_mapping=attn_metadata.block_mapping,
+                block_bias=attn_metadata.attn_bias,
+                block_scales=attn_metadata.block_scales,
+                block_groups=None,
+                scale=self.scale,
+                matmul_qk_op=self.matmul_qk,
+                matmul_av_op=self.matmul_av,
+                batch2block_matmul_op=self.batch2block_matmul,
+                block2batch_matmul_op=self.block2batch_matmul,
+                keys_fetch_func=self.k_cache.fetch_from_cache,
+                values_fetch_func=self.v_cache.fetch_from_cache)
+        # Reshape the output tensor.
+        return output.view(batch_size, seq_len, hidden_size)
+
+
+def _make_alibi_bias(
+    alibi_slopes: torch.Tensor,
+    num_kv_heads: int,
+    dtype: torch.dtype,
+    seq_len: int,
+) -> torch.Tensor:
+    bias = torch.arange(seq_len, dtype=dtype)
+    # NOTE(zhuohan): HF uses
+    #     `bias = bias[None, :].repeat(seq_len, 1)`
+    # here. We find that both biases give the same results, but
+    # the bias below more accurately follows the original ALiBi
+    # paper.
+    # Calculate a matrix where each element represents ith element- jth
+    # element.
+    bias = bias[None, :] - bias[:, None]
+
+    padded_len = (seq_len + 7) // 8 * 8
+    num_heads = alibi_slopes.shape[0]
+    bias = torch.empty(
+        1,  # batch size
+        num_heads,
+        seq_len,
+        padded_len,
+        device=alibi_slopes.device,
+        dtype=dtype,
+    )[:, :, :, :seq_len].copy_(bias)
+    bias.mul_(alibi_slopes[:, None, None])
+    if num_heads != num_kv_heads:
+        bias = bias.unflatten(1, (num_kv_heads, num_heads // num_kv_heads))
+    return bias

.venv/lib/python3.11/site-packages/vllm/attention/backends/ipex_attn.py

@@ -0,0 +1,387 @@
+# SPDX-License-Identifier: Apache-2.0
+""" Attention layer with torch scaled_dot_product_attention
+    and PagedAttention."""
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+
+from vllm._ipex_ops import ipex_ops
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionLayer,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import CommonAttentionState
+from vllm.attention.ops.paged_attn import (PagedAttention,
+                                           PagedAttentionMetadata)
+
+_PARTITION_SIZE = 512
+
+
+class IpexAttnBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "IPEX"
+
+    @staticmethod
+    def get_impl_cls() -> Type["IpexAttnBackendImpl"]:
+        return IpexAttnBackendImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["IpexAttnMetadata"]:
+        return IpexAttnMetadata
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        from vllm._ipex_ops import ipex_ops as ops
+        ops.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        from vllm._ipex_ops import ipex_ops as ops
+        key_caches = [kv_cache[0] for kv_cache in kv_caches]
+        value_caches = [kv_cache[1] for kv_cache in kv_caches]
+        ops.copy_blocks(key_caches, value_caches, src_to_dists)
+
+
+@dataclass
+class IpexAttnMetadata(AttentionMetadata, PagedAttentionMetadata):
+    """Metadata for IpexAttnBackend.
+    """
+    # Currently, input sequences can only contain all prompts
+    # or all decoding. True if all sequences are prompts.
+    is_prompt: bool
+    slot_mapping: torch.Tensor
+    seq_lens: Optional[List[int]]
+    seqlen_q: Optional[torch.Tensor]
+    max_seqlen: Optional[int]
+
+    def __post_init__(self):
+        # Set during the execution of the first attention op.
+        # It is a list because it is needed to set per prompt
+        # when alibi slopes is used. It is because of the limitation
+        # from xformer API.
+        # will not appear in the __repr__ and __init__
+        self.attn_bias: Optional[List[torch.Tensor]] = None
+
+    @property
+    def prefill_metadata(self) -> Optional["IpexAttnMetadata"]:
+        # Currently chunked prefill is not supported
+        if self.num_decode_tokens == 0:
+            assert self.num_prefills > 0
+            return self
+
+        return None
+
+    @property
+    def decode_metadata(self) -> Optional["IpexAttnMetadata"]:
+        # Currently chunked prefill is not supported
+        if self.num_prefills > 0:
+            assert self.num_decode_tokens == 0
+            return None
+
+        return self
+
+
+class IpexAttnBackendImpl(AttentionImpl[IpexAttnMetadata]):
+
+    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: str = AttentionType.DECODER,
+    ) -> None:
+        if blocksparse_params is not None:
+            raise ValueError(
+                "IPEX backend 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
+        self.sliding_window = sliding_window
+        self.kv_cache_dtype = kv_cache_dtype
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+        self.need_mask = (self.alibi_slopes is not None
+                          or self.sliding_window is not None)
+        if logits_soft_cap is None:
+            logits_soft_cap = 0
+        self.logits_soft_cap = logits_soft_cap
+
+        supported_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in supported_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {supported_head_sizes}.")
+        if kv_cache_dtype != "auto":
+            raise NotImplementedError(
+                "IPEX backend does not support FP8 KV cache. "
+                "Please use xFormers backend instead.")
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "IpexAttnBackendImpl")
+
+    def split_kv_cache(
+        self,
+        kv_cache: torch.Tensor,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        x = 1
+        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
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: IpexAttnMetadata,  # type: ignore
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Forward pass with IPEX varlen_attention and PagedAttention.
+
+        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]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        assert layer._k_scale_float == 1.0 and layer._v_scale_float == 1.0
+        num_tokens, hidden_size = query.shape
+        # Reshape the query, key, and value tensors.
+        query = query.view(-1, self.num_heads, self.head_size)
+        key = key.view(-1, self.num_kv_heads, self.head_size)
+        value = value.view(-1, self.num_kv_heads, self.head_size)
+
+        if kv_cache.numel() > 0:
+            key_cache, value_cache = self.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+            ipex_ops.reshape_and_cache(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                attn_metadata.slot_mapping.flatten(),
+                self.kv_cache_dtype,
+                layer._k_scale,
+                layer._v_scale,
+            )
+
+        if attn_metadata.is_prompt:
+            assert attn_metadata.seq_lens is not None
+            if (kv_cache.numel() == 0
+                    or attn_metadata.block_tables.numel() == 0):
+                if self.num_kv_heads != self.num_heads:
+                    key = key.repeat_interleave(self.num_queries_per_kv, dim=1)
+                    value = value.repeat_interleave(self.num_queries_per_kv,
+                                                    dim=1)
+
+                if attn_metadata.attn_bias is None:
+                    if self.alibi_slopes is not None:
+                        att_masks = _make_alibi_bias(
+                            self.alibi_slopes, query.dtype,
+                            attn_metadata.seq_lens)  # type: ignore
+                    elif self.sliding_window is not None:
+                        att_masks = _make_sliding_window_bias(
+                            attn_metadata.seq_lens, self.sliding_window,
+                            query.dtype)  # type: ignore
+                    else:
+                        att_masks = _make_sliding_window_bias(
+                            attn_metadata.seq_lens, None, dtype=query.dtype)
+                    attn_metadata.attn_bias = att_masks
+
+                output = torch.empty(
+                    (num_tokens, self.num_heads, self.head_size),
+                    dtype=query.dtype,
+                    device=query.device)
+                ipex_ops.varlen_attention(
+                    query,
+                    key,
+                    value,
+                    output,
+                    attn_metadata.seqlen_q,
+                    attn_metadata.seqlen_q,
+                    attn_metadata.max_seqlen,
+                    attn_metadata.max_seqlen,
+                    pdropout=0.0,
+                    softmax_scale=self.scale,
+                    zero_tensors=False,
+                    is_causal=True,
+                    return_softmax=False,
+                    gen_=None,
+                    logits_soft_cap=self.logits_soft_cap,
+                )
+            else:
+                # prefix-enabled attention
+                raise RuntimeError(
+                    "IPEX backend doesn't support prefix decoding.")
+
+        else:
+            # Decoding run.
+            max_seq_len = attn_metadata.max_decode_seq_len
+            output = torch.empty_like(query)
+            block_size = value_cache.shape[3]
+            num_seqs, num_heads, head_size = query.shape
+            max_num_partitions = ((max_seq_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_seq_len <= 8192 and
+                      (max_num_partitions == 1 or num_seqs * num_heads > 512))
+            if use_v1:
+                # Run PagedAttention V1.
+                ipex_ops.paged_attention_v1(
+                    output,
+                    query,
+                    key_cache,
+                    value_cache,
+                    self.num_kv_heads,
+                    self.scale,
+                    attn_metadata.block_tables,
+                    attn_metadata.seq_lens_tensor,
+                    block_size,
+                    max_seq_len,
+                    self.alibi_slopes,
+                    self.kv_cache_dtype,
+                    layer._k_scale,
+                    layer._v_scale,
+                )
+            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)
+                ipex_ops.paged_attention_v2(
+                    output,
+                    exp_sums,
+                    max_logits,
+                    tmp_output,
+                    query,
+                    key_cache,
+                    value_cache,
+                    self.num_kv_heads,
+                    self.scale,
+                    attn_metadata.block_tables,
+                    attn_metadata.seq_lens_tensor,
+                    block_size,
+                    max_seq_len,
+                    self.alibi_slopes,
+                    self.kv_cache_dtype,
+                    layer._k_scale,
+                    layer._v_scale,
+                )
+
+            # Reshape the output tensor.
+        return output.view(-1, self.num_heads * self.head_size)
+
+
+def _make_alibi_bias(
+    alibi_slopes: torch.Tensor,
+    dtype: torch.dtype,
+    seq_lens: List[int],
+) -> List[torch.Tensor]:
+    attn_biases = []
+    for seq_len in seq_lens:
+        bias = torch.arange(seq_len, dtype=dtype, device=alibi_slopes.device)
+        # NOTE(zhuohan): HF uses
+        #     `bias = bias[None, :].repeat(seq_len, 1)`
+        # here. We find that both biases give the same results, but
+        # the bias below more accurately follows the original ALiBi
+        # paper.
+        bias = bias[None, :] - bias[:, None]
+
+        num_heads = alibi_slopes.shape[0]
+        bias = bias[None, :].repeat((num_heads, 1, 1))
+        bias.mul_(alibi_slopes[:, None, None])
+        inf_mask = torch.empty(
+            (1, seq_len, seq_len),
+            dtype=bias.dtype,
+            device=alibi_slopes.device).fill_(-torch.inf).triu_(diagonal=1)
+        attn_biases.append((bias + inf_mask).to(dtype))
+
+    return attn_biases
+
+
+def _make_sliding_window_bias(
+    seq_lens: List[int],
+    window_size: Optional[int],
+    dtype: torch.dtype,
+) -> List[torch.Tensor]:
+    attn_biases = []
+    for seq_len in seq_lens:
+        tensor = torch.full(
+            (1, seq_len, seq_len),
+            dtype=dtype,
+            fill_value=1,
+        )
+        shift = 0
+        mask = torch.tril(tensor, diagonal=shift).to(dtype)  # type: ignore
+        if window_size is not None:
+            mask = torch.triu(mask, diagonal=shift - window_size + 1)
+        mask = torch.log(mask)
+        attn_biases.append(mask.to(dtype))
+
+    return attn_biases

.venv/lib/python3.11/site-packages/vllm/attention/backends/mla/utils.py

@@ -0,0 +1,541 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from abc import abstractmethod
+from dataclasses import dataclass
+from typing import Any, Dict, Generic, List, Optional, Tuple
+
+import torch
+from compressed_tensors.quantization import QuantizationStrategy
+
+from vllm import _custom_ops as ops
+from vllm import envs
+from vllm.attention.backends.abstract import (AttentionLayer,
+                                              AttentionMetadata,
+                                              MLAAttentionImpl, T)
+from vllm.distributed import (get_tensor_model_parallel_world_size,
+                              tensor_model_parallel_all_reduce)
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               LinearBase, RowParallelLinear,
+                                               UnquantizedLinearMethod)
+from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import (  # noqa: E501
+    CompressedTensorsLinearMethod)
+from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsW8A8Fp8)
+from vllm.model_executor.layers.quantization.fp8 import Fp8LinearMethod
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    apply_fp8_linear_generic, current_platform_fp8_dtype, is_fp8)
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    scaled_dequantize, scaled_quantize)
+from vllm.model_executor.layers.rotary_embedding import (
+    DeepseekScalingRotaryEmbedding, RotaryEmbedding)
+
+try:
+    from vllm.vllm_flash_attn import flash_attn_varlen_func
+except ImportError:
+    from flash_attn import flash_attn_varlen_func
+
+
+@dataclass
+class MLACommonMetadata(AttentionMetadata):
+    # Input positions for rotrary embeddings since for MLA the rotary
+    # position embeddings are applied inside the attention backend
+    input_positions: torch.Tensor
+
+
+class MLACommonImpl(MLAAttentionImpl[T], Generic[T]):
+    """
+    Common class for implementing repeated parts
+
+    Main reference: DeepseekV2 paper, and FlashInfer Implementation
+    (https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551).
+
+    Deepseek's MLA attention works the following way:
+    * Use a single latent vector to represent the entire KV cache.
+    * The attention "simulates" a multi-head attention, while the compute is
+      similar to multi-query attention.
+    * The dataflow is as follows,
+
+        * B: batch/sequence length
+        * H: hidden size
+        * N: number of attention heads
+        * Lq: latent dimension for Q
+        * Lkv: latent dimension for K/V
+        * P: nope dimension, P+R is the actual head_dim in common attention.
+        * R: rope dimension, this slide of the head_dim goes through rope.
+        * V: V head dim.
+        * kv_c: latent/compressed KV
+        * q_c: latent/compressed Q
+
+        #
+        # Outside the MLA attention backend
+        #
+
+        1. The hidden states (B, H) are projected down into cq (B, Lq) and
+           kv_c_k_pe (B, Lkv+R).
+        2. The kv_c_k_pe is split into kv_c (B, Lkv) and k_pe (B, R). cq
+           and kv_c are normalized.
+
+        #
+        # Inside the MLA attention backend
+        #
+
+        * if prefill:
+
+        3. The q_c is then projected up into the multi-head version.
+           * q_c goes from (B, Lq) to (B, N, (P+R)), which is split into q_nope
+             (B, N, P) and q_pe (B, N, R).
+        4. q_pe, k_pe are then passed through rotary embeddings.
+        5. kv_c and k_pe are concatenated and inserted into the cache
+        6. The kv_c is then projected up into the multi-head version.
+           * kv_c goes from (B, Lkv) to (B, N, (P+V)) which has the nope
+             dimensions for K and V, which is split into k_nope (B, N, P)
+             and v (B, N, V).
+        7. q (B, N, (P+R)) and k (B, N, (P+R)) matrices are assembled from
+           q_nope, q_pe, k_nope, k_pe.
+        8. Attention is computued with q, k, v.
+        9. The attention computation returns (B, N, V), which is projected back
+           to (B, H) using out projection.
+
+        * if decode:
+
+        3. Here's the change, we do not perform up the full up projection for
+           q_c, and there is no up projection at all for kv_c. This is
+           achieved by the technique of "weight absorption". The paper says
+           "Fortunately, due to the associative law of matrix multiplication,
+           we can absorb WUK into WUQ, and WUV into WO"
+           * The q up projection turns (B, Lq) into (B, N, (P+R)), we split it
+             into W_UQ (Lq, N, P) and W_QR (Lq, N, R).
+           * The kv_c up projection turns (B, Lkv) into (B, N, (P+V)), we split
+             it into W_UK (Lkv, N, P) and W_UV (Lkv, N, V).
+           * The out projection shape W_O (N*V, H) turns (B, N, V) into (B, H).
+           * We can precompute the product of W_UQ and W_UK into
+             W_UQ_UK (Lq, N, Lkv), which is possible due to QK^T operation in
+             attention.
+           * We can precompute the product of W_UV and W_O into
+             W_UV_O (N, Lkv, H), which is possible due to V@O as the
+             "epilogue" of attention
+        4. We still need to compute q_pe (B, N, R) by applying W_QR to q_latent.
+        5. q_pe, k_pe are then passed through rotary embeddings.
+        6. kv_c and k_pe are concatenated and inserted into the cache
+        7. By applying W_UQ_UK to q_latent, we have the new q_nope of shape
+           (B, N, Lkv).
+        8. q (B, N, (Lkv+R)), k (B, (Lkv+R)) are assembled from q_nope, q_pe,
+           kv_a, k_pe. v (B, Lkv) is exactly the same vector as kv_a.
+        9. The attention is computed with q, k, v. Note that we just performed
+           a MQA attention with (LKv+R) as our head dim.
+        10. The KV cache is updated using the new entries k (B, N, (Lkv+R)),
+           which included the v and rope values.
+        11. The attention computation returns (B, N, Lkv), which is projected
+           back to (B, H) using W_UV_O.
+
+    From @tsu-bin's calculation, we only want to use the absorption technique
+    for decode. The prefill algorithm should still use the up-projected MHA
+    for less flops and memory usage.
+
+    """
+
+    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]],
+        logits_soft_cap: Optional[float],
+        attn_type: str,
+        # MLA Specific Arguments
+        q_lora_rank: Optional[int],
+        kv_lora_rank: int,
+        qk_nope_head_dim: int,
+        qk_rope_head_dim: int,
+        qk_head_dim: int,
+        v_head_dim: int,
+        rotary_emb: RotaryEmbedding,
+        # q_proj should be q_b_proj if q_lora_rank is not None, but from an
+        # attention backend perspective we rely on the layer to pass in the
+        # correct matrix
+        q_proj: ColumnParallelLinear,
+        kv_b_proj: ColumnParallelLinear,
+        o_proj: RowParallelLinear,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        self.kv_cache_dtype = kv_cache_dtype
+
+        self.q_lora_rank = q_lora_rank
+        self.kv_lora_rank = kv_lora_rank
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.qk_head_dim = qk_head_dim
+        self.v_head_dim = v_head_dim
+
+        self.rotary_emb = rotary_emb
+        self.use_yarn_rope = isinstance(rotary_emb,
+                                        DeepseekScalingRotaryEmbedding)
+        self.q_proj = q_proj
+        self.kv_b_proj = kv_b_proj
+        self.o_proj = o_proj
+
+    def _v_up_proj_and_o_proj(self, x):
+        if envs.VLLM_MLA_PERFORM_MATRIX_ABSORPTION:
+            if is_fp8(self.W_UV_O):
+                output_parallel = apply_fp8_linear_generic(
+                    x.flatten(start_dim=1), self.W_UV_O, self.W_UV_O_scales,
+                    self.reqaunt_input_group_shape,
+                    self.reqaunt_weight_group_shape)
+            else:
+                output_parallel = torch.matmul(x.flatten(start_dim=1),
+                                               self.W_UV_O)
+            if self.tp_size > 1:
+                output = tensor_model_parallel_all_reduce(output_parallel)
+            else:
+                output = output_parallel
+            return output
+        else:
+            x = torch.einsum("bnl,lnv->bnv", x, self.W_UV)
+            return self.o_proj(x.reshape(-1,
+                                         self.num_heads * self.v_head_dim))[0]
+
+    def _q_proj_and_k_up_proj(self, x):
+        if envs.VLLM_MLA_PERFORM_MATRIX_ABSORPTION:
+            if is_fp8(self.W_Q_UK):
+                return apply_fp8_linear_generic(
+                    x, self.W_Q_UK, self.W_Q_UK_scales,
+                    self.reqaunt_input_group_shape,
+                    self.reqaunt_weight_group_shape).view(
+                        -1, self.num_heads, self.kv_lora_rank)
+            return torch.matmul(x, self.W_Q_UK)\
+                .view(-1, self.num_heads, self.kv_lora_rank)
+        else:
+            x = torch.matmul(x, self.W_Q)\
+                .view(-1, self.num_heads, self.qk_nope_head_dim)
+            return torch.einsum("bnp,lnp->bnl", x, self.W_UK)\
+                .view(-1, self.num_heads, self.kv_lora_rank)
+
+    def process_weights_after_loading(self, act_dtype: torch.dtype):
+
+        def is_layer_fp8(layer: LinearBase) -> bool:
+            return isinstance(layer.quant_method, Fp8LinearMethod) or\
+                (isinstance(layer.quant_method, CompressedTensorsLinearMethod)\
+                and isinstance(layer.scheme, CompressedTensorsW8A8Fp8))
+
+        def quantization_scheme_supported(layer: LinearBase) -> bool:
+            return isinstance(layer.quant_method, UnquantizedLinearMethod) or \
+                is_layer_fp8(layer)
+
+        # TODO(lucas) This is very gross, we need a more wide scale refactor of
+        # all the FP8 code with a more standard way of
+        # defining schemes/group-shapes, we should also potentially force
+        # quant_methods to support a decompress function
+        #
+        # returns input_group_shape, weight_group_shape
+        def get_scale_group_shapes_for_fp8(layer: LinearBase) -> \
+            Tuple[Tuple[int, int], Tuple[int, int]]:
+            if isinstance(layer.quant_method, Fp8LinearMethod):
+                if layer.quant_method.block_quant is not None:
+                    weight_block_size = \
+                        layer.quant_method.quant_config.weight_block_size
+                    # per-token-group (1, X), block-quantized (X, Y)
+                    return (1, weight_block_size[-1]), weight_block_size
+                else:
+                    return (-1, -1), (-1, -1)  # per-tensor, per-tensor
+            elif isinstance(layer.quant_method, CompressedTensorsLinearMethod)\
+                and isinstance(layer.scheme, CompressedTensorsW8A8Fp8):
+                # this is hacky but we always assume the for
+                # CompressedTensorsW8A8Fp8 the input is dynamic per-token
+                # we ignore if it is static-per-tensor since we are going to
+                # requantize after later anyways
+                strategy = layer.scheme.strategy
+                if strategy == QuantizationStrategy.TENSOR:
+                    return (1, -1), (-1, -1)  # per-token, per-tensor
+                elif strategy == QuantizationStrategy.CHANNEL:
+                    return (1, -1), (-1, 1)  # per-token, per-channel
+                else:
+                    raise NotImplementedError(
+                        f"QuantizationStrategy.{strategy} is not supported for "
+                        "fp8 MLA, please run with VLLM_MLA_DISABLE=1")
+            else:
+                raise NotImplementedError(
+                    "Can't determine scale group shapes for "
+                    f"{layer.quant_method}, please run with VLLM_MLA_DISABLE=1"
+                )
+
+        def get_scales(layer: LinearBase) -> torch.Tensor:
+            if hasattr(layer, "weight_scale_inv"):
+                return layer.weight_scale_inv
+            return layer.weight_scale
+
+        def get_and_maybe_dequant_weights(layer: LinearBase):
+            if is_layer_fp8(layer):
+                if isinstance(layer.quant_method, \
+                    CompressedTensorsLinearMethod) and \
+                    isinstance(layer.scheme, CompressedTensorsW8A8Fp8):
+                    # NOTE(lucas): note sure why but `CompressedTensorsW8A8Fp8`
+                    # seems to store weights as (input, output) instead of
+                    # (output, input) so we need to transpose
+                    weight = layer.weight.T  # standardize to (output, input)
+                else:
+                    weight = layer.weight
+                _, weight_scale_group_shape = \
+                    get_scale_group_shapes_for_fp8(layer)
+                scales = get_scales(layer)
+
+                return scaled_dequantize(weight, scales,
+                                         weight_scale_group_shape)
+            else:
+                return layer.weight
+
+        if not (quantization_scheme_supported(self.kv_b_proj) and\
+            quantization_scheme_supported(self.q_proj) and\
+                quantization_scheme_supported(self.o_proj)):
+            raise NotImplementedError(
+                "Only FP8 and UnquantizedLinearMethod are supported for MLA"
+                ", please run with VLLM_MLA_DISABLE=1")
+
+        weight_dtype = self.kv_b_proj.weight.dtype
+        assert self.o_proj.weight.dtype == weight_dtype
+        assert self.q_proj.weight.dtype == weight_dtype
+
+        kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj).T
+        assert kv_b_proj_weight.shape == (
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim)), (
+                f"{kv_b_proj_weight.shape=}, "
+                f"{self.kv_lora_rank=}, "
+                f"{self.num_heads=}, "
+                f"{self.qk_nope_head_dim=}, "
+                f"{self.v_head_dim=}")
+        kv_b_proj_weight = kv_b_proj_weight.view(
+            self.kv_lora_rank,
+            self.num_heads,
+            self.qk_nope_head_dim + self.v_head_dim,
+        )
+
+        W_UK, W_UV = kv_b_proj_weight.split(
+            [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        q_proj_weight = get_and_maybe_dequant_weights(self.q_proj).T\
+                .view(-1, self.num_heads, self.qk_head_dim)
+
+        # can be W_Q or W_UQ depending q_lora_rank, the former if
+        # q_lora_rank is None, the latter otherwise. From the Attention backend
+        # perspective though we call these both W_Q and rely on the layer
+        # to pass in the correct matrix
+        W_Q = q_proj_weight[..., :self.qk_nope_head_dim]
+        self.W_QR = q_proj_weight[..., self.qk_nope_head_dim:]\
+            .flatten(start_dim=1).contiguous()
+
+        # W_QR is small so for simplicity we dont bother requantizing it
+        self.W_QR = self.W_QR.to(act_dtype)
+
+        if envs.VLLM_MLA_PERFORM_MATRIX_ABSORPTION:
+            requantization_enabled = not envs.VLLM_MLA_DISABLE_REQUANTIZATION
+            if is_fp8(weight_dtype) and requantization_enabled:
+                # This assumes it wise to requantize using the same group shapes
+                # (i.e. strategy, per-tensor, per-channel, block etc.) that the
+                # weights were originally quantized
+                requant_input_group_shape, requant_weight_group_shape = \
+                    get_scale_group_shapes_for_fp8(self.q_proj)
+                assert (requant_input_group_shape, requant_weight_group_shape)\
+                    == get_scale_group_shapes_for_fp8(self.kv_b_proj)
+                assert (requant_input_group_shape, requant_weight_group_shape)\
+                    == get_scale_group_shapes_for_fp8(self.o_proj)
+                self.reqaunt_input_group_shape = requant_input_group_shape
+                self.reqaunt_weight_group_shape = requant_weight_group_shape
+
+            #
+            # Perform matrix-absorption following
+            #     https://github.com/flashinfer-ai/flashinfer/pull/551
+            # for decode, as a result we end up with absorbed weights for decode
+            # and another copy of raw weights for prefill.
+            #
+            self.W_UK, self.W_UV = kv_b_proj_weight.split(
+                [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+            # We absorb `W_UK` into `W_Q` resulting in either W_Q_UK or W_UQ_UK
+            # depending q_lora_rank, the former if q_lora_rank is None, the
+            # latter otherwise
+            # basically if q_lora_rank is none we are absorbing into q_proj
+            # instead of UQ
+            W_Q_UK = torch.einsum("qnd,lnd -> qnl", W_Q, W_UK)\
+                .flatten(start_dim=1).contiguous()
+
+            if is_fp8(weight_dtype) and requantization_enabled:
+                W_Q_UK, W_Q_UK_scales = scaled_quantize(
+                    W_Q_UK,
+                    self.reqaunt_weight_group_shape,
+                    quant_dtype=current_platform_fp8_dtype)
+                # For FP8 save the transpose so we can use
+                # `apply_w8a8_block_fp8_linear` directly
+                self.W_Q_UK = W_Q_UK.T.contiguous()
+                self.W_Q_UK_scales = W_Q_UK_scales.T.contiguous()
+            else:
+                self.W_Q_UK = W_Q_UK.to(act_dtype)
+
+            W_O = get_and_maybe_dequant_weights(self.o_proj)\
+                .view(-1, self.num_heads, self.v_head_dim)
+            W_UV_O = torch.einsum("lnd,hnd -> nlh", W_UV, W_O)\
+                .flatten(start_dim=0, end_dim=1).contiguous()
+
+            if is_fp8(weight_dtype) and requantization_enabled:
+                W_UV_O, W_UV_O_scales = scaled_quantize(
+                    W_UV_O,
+                    self.reqaunt_weight_group_shape,
+                    quant_dtype=current_platform_fp8_dtype)
+                # For FP8 save the transpose so we can use
+                # `apply_w8a8_block_fp8_linear` directly
+                self.W_UV_O = W_UV_O.T.contiguous()
+                self.W_UV_O_scales = W_UV_O_scales.T.contiguous()
+            else:
+                self.W_UV_O = W_UV_O.to(act_dtype)
+
+            self.tp_size = get_tensor_model_parallel_world_size()
+        else:
+            if is_fp8(weight_dtype):
+                raise NotImplementedError(
+                    "Currently fp8 requires matrix absorption")
+
+            self.W_UV = W_UV
+            self.W_UK = W_UK
+            self.W_Q = W_Q.flatten(start_dim=1)
+
+    @abstractmethod
+    def _forward_prefill(
+        self,
+        q: torch.Tensor,
+        kv_c_normed: torch.Tensor,
+        k_pe: torch.Tensor,
+        attn_metadata: T,
+    ) -> torch.Tensor:
+        raise NotImplementedError
+
+    @abstractmethod
+    def _forward_decode(
+        self,
+        q_nope: torch.Tensor,
+        q_pe: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: T,
+    ) -> torch.Tensor:
+        raise NotImplementedError
+
+    def apply_pure_rope(
+        self,
+        input_positions: torch.Tensor,
+        q_pe: torch.Tensor,
+        k_pe: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        seq_len = input_positions.size(0)
+        ori_q_pe_shape, ori_k_pe_shape = q_pe.shape, k_pe.shape
+
+        q_pe, k_pe = self.rotary_emb(
+            input_positions,
+            q_pe.reshape(seq_len, -1),
+            k_pe.reshape(seq_len, -1),
+        )
+        q_pe, k_pe = q_pe.view(ori_q_pe_shape), k_pe.view(ori_k_pe_shape)
+
+        return q_pe, k_pe
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        hidden_states_or_q_c: torch.Tensor,  # query in unified attn
+        k_c_normed: torch.Tensor,  # key in unified attn
+        k_pe: torch.Tensor,  # value in unified attn
+        kv_cache: torch.Tensor,
+        attn_metadata: T,
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if output is not None:
+            raise NotImplementedError(
+                "output is not yet supported for MLAImplBase")
+
+        is_decode = attn_metadata.decode_metadata is not None
+        is_prefill = attn_metadata.prefill_metadata is not None
+
+        if (is_decode and is_prefill):
+            raise NotImplementedError(
+                "chunked prefill is not supported for MLAImplBase")
+
+        # Restore head dim (for rotary embedding)
+        k_pe = k_pe.unsqueeze(1)
+        assert hasattr(attn_metadata, "input_positions")
+        rope_fn = (self.rotary_emb
+                   if self.use_yarn_rope else self.apply_pure_rope)
+
+        if is_decode:
+            q_nope = self._q_proj_and_k_up_proj(hidden_states_or_q_c)
+            q_pe = torch.matmul(hidden_states_or_q_c, self.W_QR)\
+                .view(-1, self.num_heads, self.qk_rope_head_dim)
+            q_pe, k_pe = rope_fn(attn_metadata.input_positions, q_pe, k_pe)
+        else:
+            assert is_prefill
+            q = self.q_proj(hidden_states_or_q_c)[0]\
+                .view(-1, self.num_heads, self.qk_head_dim)
+
+            # TODO(lucas): there must be a nicer way to write this line
+            q[..., self.qk_nope_head_dim:], k_pe = \
+                rope_fn(
+                    attn_metadata.input_positions,
+                    q[..., self.qk_nope_head_dim:], k_pe)
+
+        # write the latent and rope to kv cache
+        if kv_cache.numel() > 0:
+            ops.concat_and_cache_mla(
+                k_c_normed,
+                k_pe.squeeze(1),
+                kv_cache,
+                attn_metadata.slot_mapping.flatten(),
+                kv_cache_dtype=self.kv_cache_dtype,
+                scale=layer._k_scale,
+            )
+
+        if attn_metadata.prefill_metadata is not None:
+            return self._forward_prefill(q, k_c_normed, k_pe, attn_metadata)
+
+        if attn_metadata.decode_metadata is not None:
+            return self._forward_decode(q_nope, q_pe, kv_cache, attn_metadata)
+
+    # Optional common flash-attn based prefill
+    def _forward_prefill_flash(
+        self,
+        q: torch.Tensor,
+        k_c_normed: torch.Tensor,
+        k_pe: torch.Tensor,
+        seq_start_loc: torch.Tensor,
+        max_prefill_seq_len: int,
+    ) -> torch.Tensor:
+
+        kv_nope = self.kv_b_proj(k_c_normed)[0]\
+            .view(-1, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
+        k_nope, v = kv_nope\
+            .split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        k = torch.cat((k_nope, k_pe.expand((*k_nope.shape[:-1], -1))), dim=-1)
+
+        # For MLA the v head dim is smaller than qk head dim so we pad out
+        # v with 0s to match the qk head dim
+        v_padded = torch.nn.functional.pad(v, [0, q.shape[-1] - v.shape[-1]],
+                                           value=0)
+
+        attn_output = flash_attn_varlen_func(
+            q=q,
+            k=k,
+            v=v_padded,
+            cu_seqlens_q=seq_start_loc,
+            cu_seqlens_k=seq_start_loc,
+            max_seqlen_q=max_prefill_seq_len,
+            max_seqlen_k=max_prefill_seq_len,
+            softmax_scale=self.scale,
+            causal=True,
+        )
+        attn_output = attn_output\
+            .view(-1, self.num_heads, q.shape[-1])[..., :v.shape[-1]]\
+                .reshape(-1, self.num_heads * v.shape[-1])
+
+        return self.o_proj(attn_output)[0]

.venv/lib/python3.11/site-packages/vllm/attention/backends/openvino.py

@@ -0,0 +1,146 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Type
+
+import openvino as ov
+import torch
+
+from vllm.attention.backends.abstract import (AttentionBackend,
+                                              AttentionMetadata)
+from vllm.attention.backends.utils import CommonAttentionState
+from vllm.multimodal import MultiModalPlaceholderMap
+
+
+def copy_cache_block(src_tensor: ov.Tensor, dst_tensor: ov.Tensor,
+                     src_offset: int, dst_offset: int) -> None:
+
+    def create_roi_tensor(
+        tensor: ov.Tensor,
+        block_number: int,
+    ) -> ov.Tensor:
+        roi_begin = ov.runtime.Coordinate([0, 0, 0, 0])
+        roi_end = ov.runtime.Coordinate(tensor.get_shape())
+
+        roi_begin[0] = block_number
+        roi_end[0] = block_number + 1
+
+        if isinstance(tensor, ov.Tensor):
+            return ov.Tensor(tensor, roi_begin, roi_end)
+        else:
+            return ov.RemoteTensor(tensor, roi_begin, roi_end)
+
+    src_roi_tensor = \
+        create_roi_tensor(src_tensor, src_offset)
+    dst_roi_tensor = \
+        create_roi_tensor(dst_tensor, dst_offset)
+    src_roi_tensor.copy_to(dst_roi_tensor)
+
+
+class OpenVINOAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "OPENVINO"
+
+    @staticmethod
+    def get_impl_cls():
+        # OpenVINO implements PagedAttention as part of the Optimum
+        # exported model
+        raise NotImplementedError
+
+    @staticmethod
+    def make_metadata(*args, **kwargs) -> "AttentionMetadata":
+        raise NotImplementedError
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def make_openvino_metadata(*args, **kwargs) -> "OpenVINOAttentionMetadata":
+        return OpenVINOAttentionMetadata(*args, **kwargs)
+
+    @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, num_kv_heads, block_size, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_tensor: ov.Tensor,
+        dst_tensor: ov.Tensor,
+        src_to_dists: List[Tuple[int, int]],
+    ) -> None:
+        for src, dst in src_to_dists:
+            copy_cache_block(src_tensor, dst_tensor, src, dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[Tuple[ov.Tensor, ov.Tensor]],
+        src_to_dists: List[Tuple[int, int]],
+    ) -> None:
+        for src, dst in src_to_dists:
+            for key_cache, value_cache in kv_caches:
+                copy_cache_block(key_cache, key_cache, src, dst)
+                copy_cache_block(value_cache, value_cache, src, dst)
+
+
+@dataclass
+class OpenVINOAttentionMetadata:
+    """Metadata for OpenVINOAttentionBackend.
+
+    Basic terms used below:
+    - batch_size_in_sequences - total number of sequences to execute​
+    - prompt_lens – per sequence size number of scheduled tokens​
+    - batch_size_in_tokens = sum(prompt_lens)​
+    - max_context_len = max(context_lens)​
+    - max_num_blocks = div_up(max_context_len / BLOCK_SIZE)​
+    - num_blocks – total number of blocks in block_indices​
+    """
+
+    # Describes past KV cache size for each sequence within a batch
+    # Shape: [batch_size_in_sequences]
+    # Type: i32​
+    past_lens: torch.Tensor
+
+    # Describes start indices of input / speculative tokens from
+    # current sequences within a batch sequence​
+    # Shape: [batch_size_in_sequences + 1]​
+    # Type: i32
+    subsequence_begins: torch.Tensor
+
+    # Describes block tables for each sequence within a batch​ -
+    # indices along 0th dimension in key_cache and value_cache inputs​
+    # Shape: [num_blocks]
+    # Type: i32​
+    block_indices: torch.Tensor
+
+    # Describes block tables for each sequence within a batch​ -
+    # for i-th element, it is an index in block_indices with the
+    # first block belonging to i-th sequence​
+    # Shape: [batch_size_in_sequences + 1]
+    # Type: i32​
+    block_indices_begins: torch.Tensor
+
+    # Describes max context length
+    # Shape: scalar
+    # Type: i32
+    max_context_len: torch.Tensor
+
+    # The index maps that relate multi-modal embeddings to the corresponding
+    # placeholders.
+    #
+    # N.B. These aren't really related to attention and don't belong on this
+    # type -- this is just a temporary solution to make them available to
+    # `model_executable`.
+    multi_modal_placeholder_index_maps: Optional[Dict[
+        str, MultiModalPlaceholderMap.IndexMap]]
+
+    # Enable/disable KV scales calculation. This is so that we can disable the
+    # calculation until after prefill and cuda graph capture.
+    enable_kv_scales_calculation: bool

.venv/lib/python3.11/site-packages/vllm/attention/backends/pallas.py

@@ -0,0 +1,337 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+import torch_xla.experimental.custom_kernel  # Required to register custom ops.
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionLayer,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import CommonAttentionState
+
+
+class PallasAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "PALLAS"
+
+    @staticmethod
+    def get_impl_cls() -> Type["PallasAttentionBackendImpl"]:
+        return PallasAttentionBackendImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["PallasMetadata"]:
+        return PallasMetadata
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return (num_kv_heads, num_blocks, block_size, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        raise RuntimeError("swap_blocks is not used for the TPU backend.")
+
+    @torch.compile(backend="openxla")
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[Tuple[torch.Tensor, torch.Tensor]],
+        src_to_dists: Tuple[torch.Tensor, torch.Tensor],
+    ) -> None:
+        src_indices, dst_indices = src_to_dists
+        for k_cache, v_cache in kv_caches:
+            torch.ops.xla.dynamo_set_buffer_donor_(k_cache, True)
+            k_cache[:, dst_indices] = k_cache[:, src_indices]
+            torch.ops.xla.dynamo_set_buffer_donor_(v_cache, True)
+            v_cache[:, dst_indices] = v_cache[:, src_indices]
+
+
+@dataclass
+class PallasMetadata(AttentionMetadata):
+
+    # Currently, input sequences can only contain all prefills
+    # or all decoding.
+    block_tables: Optional[torch.Tensor] = None
+    context_lens: Optional[torch.Tensor] = None
+    effective_query_lens: Optional[torch.Tensor] = None
+
+    @property
+    def prefill_metadata(self) -> Optional["PallasMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        assert self.num_decode_tokens == 0
+        return self
+
+    @property
+    def decode_metadata(self) -> Optional["PallasMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        assert self.num_prefills == 0
+        assert self.num_prefill_tokens == 0
+        assert self.block_tables is not None
+        assert self.context_lens is not None
+        return self
+
+
+class PallasAttentionBackendImpl(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: str = AttentionType.DECODER,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+        self.logits_soft_cap = logits_soft_cap
+        if head_size % 128 != 0:
+            raise NotImplementedError("Head size must be a multiple of 128.")
+        if alibi_slopes is not None:
+            raise NotImplementedError("Alibi slopes is not supported.")
+        if sliding_window is not None:
+            raise NotImplementedError("Sliding window is not supported.")
+        if kv_cache_dtype != "auto":
+            raise NotImplementedError("FP8 KV cache dtype is not supported.")
+        if blocksparse_params is not None:
+            raise NotImplementedError("Blocksparse is not supported.")
+
+        if torch_xla.tpu.version() < 4:
+            raise NotImplementedError("TPU version must be 4 or higher.")
+
+        self.megacore_mode = None
+        tpu_env = torch_xla.tpu.get_tpu_env()
+        tpu_type = (tpu_env.get("ACCELERATOR_TYPE", None)
+                    or tpu_env.get("TYPE", None)
+                    or tpu_env.get("TPU_ACCELERATOR_TYPE", None))
+        assert tpu_type is not None
+        tpu_type = tpu_type.lower()
+
+        if (("lite" not in tpu_type) and ("v6" not in tpu_type)):
+            if self.num_kv_heads % 2 == 0:
+                self.megacore_mode = "kv_head"
+            else:
+                # NOTE(woosuk): If the batch size is not a multiple of 2, the
+                # megacore mode will be None.
+                self.megacore_mode = "batch"
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "PallasAttentionBackendImpl")
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: Tuple[torch.Tensor, torch.Tensor],
+        attn_metadata: PallasMetadata,
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Forward pass with Pallas attention.
+
+        Args:
+            query: shape = [batch_size, seq_len, num_heads * head_size]
+            key: shape = [batch_size, seq_len, num_kv_heads * head_size]
+            value: shape = [batch_size, seq_len, num_kv_heads * head_size]
+            kv_cache[0] = [num_kv_heads, num_blocks, block_size, head_size]
+            kv_cache[1] = [num_kv_heads, num_blocks, block_size, head_size]
+                NOTE: kv_cache[0] and kv_cache[1] will be an empty tensor 
+                with shape [0] for profiling run.
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [batch_size, seq_len, num_heads * head_size]
+        """
+        assert layer._k_scale_float == 1.0 and layer._v_scale_float == 1.0
+        batch_size, seq_len, hidden_size = query.shape
+        query = query.view(batch_size, seq_len, self.num_heads, self.head_size)
+        key = key.view(batch_size, seq_len, self.num_kv_heads, self.head_size)
+        value = value.view(batch_size, seq_len, self.num_kv_heads,
+                           self.head_size)
+
+        if kv_cache[0].numel() > 0:
+            slot_mapping = attn_metadata.slot_mapping
+            key_cache, value_cache = kv_cache
+            write_to_kv_cache(key, value, key_cache, value_cache, slot_mapping)
+
+        query = query * self.scale
+        if attn_metadata.num_prefills > 0:
+            if attn_metadata.block_tables is None:
+                # Prefill without paged KV cache.
+                assert seq_len % 16 == 0, (
+                    "Pallas FlashAttention kernel requires seq_len to be a "
+                    f"multiple of 16 but got {seq_len}")
+
+                # Handle GQA/MQA.
+                if self.num_kv_heads != self.num_heads:
+                    key = key.repeat_interleave(self.num_queries_per_kv,
+                                                dim=-2)
+                    key = key.view(batch_size, seq_len, self.num_heads,
+                                   self.head_size)
+                    value = value.repeat_interleave(self.num_queries_per_kv,
+                                                    dim=-2)
+                    value = value.view(batch_size, seq_len, self.num_heads,
+                                       self.head_size)
+                # FlashAttention kernel requires the input shape to be
+                # [batch_size, num_heads, seq_len, d_model]
+                # while the input is [batch_size, seq_len, num_heads, d_model].
+                # Permute the input to match the required format.
+                output = torch.ops.xla.flash_attention(
+                    query.permute(0, 2, 1, 3),
+                    key.permute(0, 2, 1, 3),
+                    value.permute(0, 2, 1, 3),
+                    True,
+                )
+                output = output.permute(0, 2, 1, 3)
+            else:
+                # Prefill with paged KV cache.
+                # TODO(woosuk): Tune the below knobs.
+                num_kv_pages_per_compute_block = 16
+                num_queries_per_compute_block = 16
+                assert seq_len % num_queries_per_compute_block == 0
+                output = torch.ops.xla.multi_queries_paged_attention(
+                    query,
+                    key_cache,
+                    value_cache,
+                    attn_metadata.context_lens,
+                    attn_metadata.block_tables,
+                    attn_metadata.effective_query_lens,
+                    num_kv_pages_per_compute_block,
+                    num_queries_per_compute_block,
+                    use_kernel=True,
+                    attn_logits_soft_cap=self.logits_soft_cap,
+                )
+        else:
+            # Decoding run.
+            assert kv_cache[0].numel() > 0
+            query = query.squeeze(dim=1)
+            pages_per_compute_block = 16  # TODO(woosuk): Tune this value.
+
+            assert attn_metadata.block_tables is not None
+            assert attn_metadata.context_lens is not None
+            # NOTE(woosuk): The PagedAttention Pallas kernel stores the entire
+            # block table in SMEM. Therefore, if the block table is too large,
+            # the kernel compilation will fail. To avoid this, we split the
+            # batch dimension into smaller chunks and run the kernel multiple
+            # times.
+            MAX_SMEM_USAGE = 512 * 1024
+            size_per_seq = 4 * attn_metadata.block_tables.shape[1]
+            max_num_seq = MAX_SMEM_USAGE // size_per_seq
+
+            if batch_size <= max_num_seq:
+                output = paged_attention(
+                    query,
+                    key_cache,
+                    value_cache,
+                    attn_metadata.context_lens,
+                    attn_metadata.block_tables,
+                    pages_per_compute_block,
+                    self.megacore_mode,
+                    attn_logits_soft_cap=self.logits_soft_cap,
+                )
+            else:
+                chunk_size = max_num_seq
+                # Make sure the chunk size is a multiple of 2.
+                chunk_size = chunk_size // 2 * 2
+                num_chunks = (batch_size + chunk_size - 1) // chunk_size
+
+                output = torch.empty_like(query)
+                for chunk_idx in range(num_chunks):
+                    chunk_start = chunk_idx * chunk_size
+                    chunk_end = chunk_start + chunk_size
+                    # NOTE(woosuk): We skip this line because it causes Dynamo
+                    # compilation error. Instead, we rely on the slice operation
+                    # to handle the out-of-bound case.
+                    # chunk_end = min(chunk_end, batch_size)
+                    chunk_output = paged_attention(
+                        query[chunk_start:chunk_end],
+                        key_cache,
+                        value_cache,
+                        attn_metadata.context_lens[chunk_start:chunk_end],
+                        attn_metadata.block_tables[chunk_start:chunk_end],
+                        pages_per_compute_block,
+                        self.megacore_mode,
+                        attn_logits_soft_cap=self.logits_soft_cap,
+                    )
+                    output[chunk_start:chunk_end] = chunk_output
+
+        # Reshape the output tensor.
+        return output.reshape(batch_size, seq_len, hidden_size)
+
+
+def write_to_kv_cache(
+    key: torch.Tensor,
+    value: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    slot_mapping: torch.Tensor,
+) -> None:
+    torch.ops.xla.dynamo_set_buffer_donor_(key_cache, True)
+    torch.ops.xla.dynamo_set_buffer_donor_(value_cache, True)
+
+    key = key.flatten(0, 2)
+    value = value.flatten(0, 2)
+    key_cache = key_cache.flatten(0, 2)
+    value_cache = value_cache.flatten(0, 2)
+    key_cache.index_copy_(0, slot_mapping, key)
+    value_cache.index_copy_(0, slot_mapping, value)
+
+
+def paged_attention(
+    query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    context_lens: torch.Tensor,
+    block_tables: torch.Tensor,
+    pages_per_compute_block: int,
+    megacore_mode: Optional[str],
+    *,
+    attn_logits_soft_cap: Optional[float],
+) -> torch.Tensor:
+    batch_size = query.shape[0]
+    if megacore_mode == "batch" and batch_size % 2 != 0:
+        megacore_mode = None
+    else:
+        megacore_mode = megacore_mode
+
+    return torch.ops.xla.paged_attention(
+        query,
+        key_cache,
+        value_cache,
+        context_lens,
+        block_tables,
+        pages_per_compute_block,
+        megacore_mode=megacore_mode,
+        attn_logits_soft_cap=attn_logits_soft_cap,
+    )

.venv/lib/python3.11/site-packages/vllm/attention/backends/placeholder_attn.py

@@ -0,0 +1,410 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from collections import defaultdict
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Type
+
+import torch
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata,
+                                              AttentionMetadataBuilder)
+from vllm.attention.backends.utils import CommonAttentionState
+from vllm.multimodal import MultiModalPlaceholderMap
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import (ModelInputForGPUBuilder,
+                                          ModelInputForGPUWithSamplingMetadata)
+
+# Placeholder attention backend for models like Mamba and pooling models that
+# lack attention.
+
+
+class PlaceholderAttentionBackend(AttentionBackend):
+    """Placeholder backend for when no attention is needed."""
+
+    @staticmethod
+    def get_name() -> str:
+        return "NO_ATTENTION"
+
+    @staticmethod
+    def get_impl_cls() -> Type["PlaceholderAttentionImpl"]:
+        return PlaceholderAttentionImpl
+
+    @staticmethod
+    def get_builder_cls() -> Type["PlaceholderAttentionMetadataBuilder"]:
+        return PlaceholderAttentionMetadataBuilder
+
+    @staticmethod
+    def get_metadata_cls() -> Type["PlaceholderAttentionMetadata"]:
+        return PlaceholderAttentionMetadata
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return (1, 1, 1, 1, 1)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        return
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        return
+
+
+@dataclass
+class PlaceholderAttentionMetadata(AttentionMetadata):
+    """Attention metadata for prefill and decode batched together."""
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]]
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+
+    # Maximum query length in the batch.
+    max_query_len: Optional[int]
+
+    # Max number of query tokens among request in the batch.
+    max_decode_query_len: Optional[int]
+
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Maximum sequence length among decode batch. 0 if there are prefill
+    # requests only.
+    max_decode_seq_len: int
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    query_start_loc: Optional[torch.Tensor]
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor]
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor]
+
+    # (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]
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+    use_cuda_graph: bool
+
+    _cached_prefill_metadata: Optional["PlaceholderAttentionMetadata"] = None
+    _cached_decode_metadata: Optional["PlaceholderAttentionMetadata"] = None
+
+    @property
+    def prefill_metadata(self) -> Optional["PlaceholderAttentionMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            return self._cached_prefill_metadata
+
+        assert self.seq_lens is not None
+        assert self.seq_lens_tensor is not None
+        assert self.query_start_loc is not None
+        assert self.context_lens_tensor is not None
+        assert self.seq_start_loc is not None
+
+        # Placeholders
+        slot_mapping = torch.empty(0)
+        block_tables = torch.empty(0)
+
+        self._cached_prefill_metadata = PlaceholderAttentionMetadata(
+            num_prefills=self.num_prefills,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=0,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=self.
+            multi_modal_placeholder_index_maps,
+            enable_kv_scales_calculation=self.enable_kv_scales_calculation,
+            seq_lens=self.seq_lens[:self.num_prefills],
+            seq_lens_tensor=self.seq_lens_tensor[:self.num_prefills],
+            max_decode_query_len=0,
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=self.max_prefill_seq_len,
+            max_decode_seq_len=0,
+            query_start_loc=self.query_start_loc[:self.num_prefills + 1],
+            seq_start_loc=self.seq_start_loc[:self.num_prefills + 1],
+            context_lens_tensor=self.context_lens_tensor[:self.num_prefills],
+            block_tables=block_tables,
+            use_cuda_graph=False,
+        )
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> Optional["PlaceholderAttentionMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            return self._cached_decode_metadata
+        assert self.seq_lens_tensor is not None
+
+        # Placeholders
+        slot_mapping = torch.empty(0)
+        block_tables = torch.empty(0)
+
+        self._cached_decode_metadata = PlaceholderAttentionMetadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=self.num_decode_tokens,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=None,
+            enable_kv_scales_calculation=True,
+            seq_lens=None,
+            seq_lens_tensor=self.seq_lens_tensor[self.num_prefills:],
+            max_decode_query_len=self.max_decode_query_len,
+            max_query_len=None,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.max_decode_seq_len,
+            query_start_loc=None,
+            seq_start_loc=None,
+            context_lens_tensor=None,
+            block_tables=block_tables,
+            use_cuda_graph=self.use_cuda_graph,
+        )
+        return self._cached_decode_metadata
+
+    def advance_step(self,
+                     model_input: "ModelInputForGPUWithSamplingMetadata",
+                     sampled_token_ids: Optional[torch.Tensor],
+                     block_size: int,
+                     num_seqs: int,
+                     num_queries: int,
+                     turn_prefills_into_decodes: bool = False):
+        """
+        Update metadata in-place to advance one decode step.
+        """
+        # When using cudagraph, the num_seqs is padded to the next captured
+        # batch sized, but num_queries tracks the actual number of requests in
+        # the batch. For --enforce-eager mode, num_seqs == num_queries
+        if num_seqs != num_queries:
+            assert num_seqs > num_queries
+            assert self.use_cuda_graph
+
+        assert not turn_prefills_into_decodes, \
+            ("Multi-Step + Chunked-Prefill is not supported for attention-free"
+             "models. turn_prefills_into_decodes is a "
+             "Multi-Step + Chunked-Prefill specific parameter.")
+
+        assert self.seq_lens is not None
+        assert self.max_decode_seq_len == max(self.seq_lens)
+
+        assert self.num_prefills == 0
+        assert self.num_prefill_tokens == 0
+        assert self.num_decode_tokens == num_seqs
+
+        assert self.seq_lens is not None
+        assert len(self.seq_lens) == num_seqs
+        assert self.seq_lens_tensor is not None
+        assert self.seq_lens_tensor.shape == (num_seqs, )
+        assert self.max_query_len == 1
+        assert self.max_prefill_seq_len == 0
+
+        assert self.query_start_loc is not None
+        assert self.query_start_loc.shape == (num_queries + 1, )
+        assert self.seq_start_loc is not None
+        assert self.seq_start_loc.shape == (num_seqs + 1, )
+
+        assert self.context_lens_tensor is not None
+        assert self.context_lens_tensor.shape == (num_queries, )
+
+        assert self.block_tables is not None
+
+        # Update query lengths. Note that we update only queries and not seqs,
+        # since tensors may be padded due to captured cuda graph batch size
+        for i in range(num_queries):
+            self.seq_lens[i] += 1
+        self.max_decode_seq_len = max(self.seq_lens)
+
+        # Update sequences, masking off entries greater than num_queries
+        device = self.seq_lens_tensor.device
+        mask = torch.arange(self.seq_lens_tensor.size(0),
+                            device=device) < num_queries
+        self.seq_lens_tensor += mask.to(self.seq_lens_tensor.dtype)
+        if sampled_token_ids is not None:
+            model_input.input_tokens.masked_scatter_(
+                mask, sampled_token_ids[:num_queries])
+
+
+class PlaceholderAttentionMetadataBuilder(
+        AttentionMetadataBuilder[PlaceholderAttentionMetadata]):
+
+    def __init__(self, input_builder: "ModelInputForGPUBuilder"):
+
+        self.input_builder = input_builder
+        self.runner = input_builder.runner
+
+    def prepare(self):
+        self.prefill_seq_lens: List[int] = []
+        self.context_lens: List[int] = []
+        self.curr_seq_lens: List[int] = []
+        self.multimodal_placeholder_maps: Dict[
+            str,
+            MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
+        self.num_prefills = 0
+        self.num_prefill_tokens = 0
+        self.num_decode_tokens = 0
+
+    def _add_seq_group(
+            self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
+            chunked_prefill_enabled: bool):
+        """Add a sequence group to the metadata. Specifically update/append
+        1. context length.
+        """
+        is_prompt = inter_data.is_prompt
+
+        for (seq_id, token_len, seq_len, curr_seq_len, query_len, context_len,
+             curr_sliding_window_block) in zip(
+                 inter_data.seq_ids, [len(t) for t in inter_data.input_tokens],
+                 inter_data.orig_seq_lens, inter_data.seq_lens,
+                 inter_data.query_lens, inter_data.context_lens,
+                 inter_data.curr_sliding_window_blocks):
+            self.context_lens.append(context_len)
+
+            if is_prompt:
+                mm_maps = inter_data.multi_modal_placeholder_maps
+                if mm_maps:
+                    for modality, placeholders in mm_maps.items():
+                        self.multimodal_placeholder_maps[modality].extend(
+                            placeholders)
+
+                self.num_prefills += 1
+                self.num_prefill_tokens += token_len
+                self.prefill_seq_lens.append(seq_len)
+            else:
+                assert query_len == 1, (
+                    "seq_len: {}, context_len: {}, query_len: {}".format(
+                        seq_len, context_len, query_len))
+                self.num_decode_tokens += query_len
+                self.curr_seq_lens.append(curr_seq_len)
+
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int):
+        """Build attention metadata with on-device tensors.
+
+        Args:
+            seq_lens: The maybe padded sequence lengths of the input sequences.
+            query_lens: The query lengths of the input sequences.
+            cuda_graph_pad_size: The padding size for cuda graph.
+                                 -1 if cuda graph is not used.
+            batch_size: The maybe padded batch size.
+        """
+        for inter_data in self.input_builder.inter_data_list:
+            self._add_seq_group(inter_data,
+                                self.input_builder.chunked_prefill_enabled)
+
+        device = self.runner.device
+        use_captured_graph = cuda_graph_pad_size != -1
+
+        logits_soft_cap = getattr(self.runner.model_config.hf_config,
+                                  "attn_logit_softcapping", None)
+        if logits_soft_cap is not None:
+            raise ValueError(
+                "Please use Flashinfer backend for models with logits_soft_cap"
+                " (i.e., Gemma-2). Otherwise, the output might be wrong."
+                " Set Flashinfer backend by "
+                "export VLLM_ATTENTION_BACKEND=FLASHINFER.")
+
+        max_query_len = max(query_lens)
+        decode_query_lens = query_lens[self.num_prefills:]
+        if len(decode_query_lens) > 0:
+            max_decode_query_len = max(decode_query_lens)
+        else:
+            max_decode_query_len = 1
+        max_prefill_seq_len = max(self.prefill_seq_lens, default=0)
+        max_decode_seq_len = max(self.curr_seq_lens, default=0)
+        num_decode_tokens = self.num_decode_tokens
+
+        if use_captured_graph:
+            num_decode_tokens = batch_size
+
+        assert max_query_len > 0, ("query_lens: {}".format(query_lens))
+
+        context_lens_tensor = torch.tensor(self.context_lens,
+                                           dtype=torch.int,
+                                           device=device)
+        seq_lens_tensor = torch.tensor(seq_lens,
+                                       dtype=torch.int,
+                                       device=device)
+        query_lens_tensor = torch.tensor(query_lens,
+                                         dtype=torch.long,
+                                         device=device)
+        query_start_loc = torch.zeros(query_lens_tensor.shape[0] + 1,
+                                      dtype=torch.int32,
+                                      device=device)
+        seq_start_loc = torch.zeros(seq_lens_tensor.shape[0] + 1,
+                                    dtype=torch.int32,
+                                    device=device)
+        placeholder_index_maps = {
+            modality: placeholder_map.index_map()
+            for modality, placeholder_map in
+            self.multimodal_placeholder_maps.items()
+        }
+        torch.cumsum(seq_lens_tensor,
+                     dim=0,
+                     dtype=seq_start_loc.dtype,
+                     out=seq_start_loc[1:])
+        torch.cumsum(query_lens_tensor,
+                     dim=0,
+                     dtype=query_start_loc.dtype,
+                     out=query_start_loc[1:])
+
+        # Placeholders
+        slot_mapping = torch.empty(0)
+        block_tables = torch.empty(0)
+
+        return PlaceholderAttentionMetadata(
+            num_prefills=self.num_prefills,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=placeholder_index_maps,
+            enable_kv_scales_calculation=True,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=num_decode_tokens,
+            seq_lens=seq_lens,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=max_query_len,
+            max_decode_query_len=max_decode_query_len,
+            max_prefill_seq_len=max_prefill_seq_len,
+            max_decode_seq_len=max_decode_seq_len,
+            query_start_loc=query_start_loc,
+            seq_start_loc=seq_start_loc,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=use_captured_graph,
+        )
+
+
+class PlaceholderAttentionImpl(AttentionImpl):
+
+    def __init__(self, *args, **kwargs) -> None:
+        return
+
+    def forward(self, *args, **kwargs) -> torch.Tensor:
+        raise NotImplementedError

.venv/lib/python3.11/site-packages/vllm/attention/backends/rocm_flash_attn.py

@@ -0,0 +1,891 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Attention layer ROCm GPUs."""
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type
+
+import torch
+
+import vllm.envs as envs
+from vllm import _custom_ops as ops
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionLayer,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import (CommonAttentionState,
+                                           CommonMetadataBuilder)
+from vllm.attention.ops.paged_attn import (PagedAttention,
+                                           PagedAttentionMetadata)
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import ModelInputForGPUWithSamplingMetadata
+
+logger = init_logger(__name__)
+
+_PARTITION_SIZE_ROCM = 512
+_GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
+_ON_NAVI = "gfx1" in _GPU_ARCH
+_ON_MI250_MI300 = any(arch in _GPU_ARCH
+                      for arch in ["gfx90a", "gfx940", "gfx941", "gfx942"])
+
+
+class ROCmFlashAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "ROCM_FLASH"
+
+    @staticmethod
+    def get_impl_cls() -> Type["ROCmFlashAttentionImpl"]:
+        return ROCmFlashAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return ROCmFlashAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["ROCmFlashAttentionMetadataBuilder"]:
+        return ROCmFlashAttentionMetadataBuilder
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class ROCmFlashAttentionMetadata(AttentionMetadata, PagedAttentionMetadata):
+    """Metadata for FlashAttentionBackend.
+
+    NOTE: Any python object stored here is not updated when it is
+    cuda-graph replayed. If you have values that need to be changed
+    dynamically, it should be stored in tensor. The tensor has to be
+    updated from `CUDAGraphRunner.forward` API.
+    """
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]]
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Maximum sequence length among decode batch. 0 if there are prefill
+    # requests only.
+    max_decode_seq_len: int
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+    use_cuda_graph: bool
+
+    # NOTE(sang): Definition of context_len, query_len, and seq_len.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ----------------------|
+    #                                   |-- query_len ---|
+
+    # Maximum query length in the batch. None for decoding.
+    max_query_len: Optional[int] = None
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    query_start_loc: Optional[torch.Tensor] = None
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor] = None
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor] = None
+
+    # Max number of query tokens among request in the batch.
+    max_decode_query_len: Optional[int] = None
+
+    _cached_prefill_metadata: Optional["ROCmFlashAttentionMetadata"] = None
+    _cached_decode_metadata: Optional["ROCmFlashAttentionMetadata"] = None
+
+    # Begin encoder attn & enc/dec cross-attn fields...
+
+    # Encoder sequence lengths representation
+    encoder_seq_lens: Optional[List[int]] = None
+    encoder_seq_lens_tensor: Optional[torch.Tensor] = None
+
+    # Maximum sequence length among encoder sequences
+    max_encoder_seq_len: Optional[int] = None
+
+    # Number of tokens input to encoder
+    num_encoder_tokens: Optional[int] = None
+
+    # Cross-attention memory-mapping data structures: slot mapping
+    # and block tables
+    cross_slot_mapping: Optional[torch.Tensor] = None
+    cross_block_tables: Optional[torch.Tensor] = None
+
+    @property
+    def prefill_metadata(self) -> Optional["ROCmFlashAttentionMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            return self._cached_prefill_metadata
+
+        assert self.seq_lens is not None
+        assert self.seq_lens_tensor is not None
+        assert self.block_tables is not None
+
+        self._cached_prefill_metadata = ROCmFlashAttentionMetadata(
+            num_prefills=self.num_prefills,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=0,
+            slot_mapping=self.slot_mapping[:self.num_prefill_tokens],
+            multi_modal_placeholder_index_maps=self.
+            multi_modal_placeholder_index_maps,
+            enable_kv_scales_calculation=self.enable_kv_scales_calculation,
+            seq_lens=self.seq_lens[:self.num_prefills],
+            seq_lens_tensor=self.seq_lens_tensor[:self.num_prefills],
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=self.max_prefill_seq_len,
+            max_decode_seq_len=0,
+            query_start_loc=None if self.query_start_loc is None else
+            self.query_start_loc[:self.num_prefills + 1],
+            seq_start_loc=None if self.seq_start_loc is None else
+            self.seq_start_loc[:self.num_prefills + 1],
+            context_lens_tensor=None if self.context_lens_tensor is None else
+            self.context_lens_tensor[:self.num_prefills],
+            block_tables=self.block_tables[:self.num_prefills],
+            use_cuda_graph=False,
+            # Begin encoder & cross attn fields below...
+            encoder_seq_lens=self.encoder_seq_lens,
+            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
+            max_encoder_seq_len=self.max_encoder_seq_len,
+            cross_slot_mapping=self.cross_slot_mapping,
+            cross_block_tables=self.cross_block_tables)
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> Optional["ROCmFlashAttentionMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            return self._cached_decode_metadata
+        assert self.block_tables is not None
+        assert self.seq_lens_tensor is not None
+
+        self._cached_decode_metadata = ROCmFlashAttentionMetadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=self.num_decode_tokens,
+            slot_mapping=self.slot_mapping[self.num_prefill_tokens:],
+            multi_modal_placeholder_index_maps=None,
+            enable_kv_scales_calculation=True,
+            seq_lens=None,
+            seq_lens_tensor=self.seq_lens_tensor[self.num_prefills:],
+            max_query_len=None,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.max_decode_seq_len,
+            query_start_loc=None,
+            seq_start_loc=None,
+            context_lens_tensor=None,
+            block_tables=self.block_tables[self.num_prefills:],
+            use_cuda_graph=self.use_cuda_graph,
+            # Begin encoder & cross attn fields below...
+            encoder_seq_lens=self.encoder_seq_lens,
+            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
+            max_encoder_seq_len=self.max_encoder_seq_len,
+            cross_slot_mapping=self.cross_slot_mapping,
+            cross_block_tables=self.cross_block_tables)
+        # Batch may be composed of prefill|decodes, adjust query start indices
+        # to refer to the start of decodes when the two are split apart.
+        # E.g. in tokens:[3 prefills|6 decodes], query_start_loc=[3,9] => [0,6].
+        if self._cached_decode_metadata.query_start_loc is not None:
+            qs = self._cached_decode_metadata.query_start_loc
+            self._cached_decode_metadata.query_start_loc = qs - qs[0]
+        return self._cached_decode_metadata
+
+    def advance_step(self,
+                     model_input: "ModelInputForGPUWithSamplingMetadata",
+                     sampled_token_ids: Optional[torch.Tensor],
+                     block_size: int,
+                     num_seqs: int,
+                     num_queries: int,
+                     turn_prefills_into_decodes: bool = False):
+        """
+        Update metadata in-place to advance one decode step.
+        """
+
+        assert not turn_prefills_into_decodes, \
+            ("Chunked prefill is not supported with rocm_flash_attn yet."
+             "turn_prefills_into_decodes is a Multi-Step + Chunked-Prefill "
+             "specific parameter.")
+
+        # When using cudagraph, the num_seqs is padded to the next captured
+        # batch sized, but num_queries tracks the actual number of requests in
+        # the batch. For --enforce-eager mode, num_seqs == num_queries
+        if num_seqs != num_queries:
+            assert num_seqs > num_queries
+            assert self.use_cuda_graph
+
+        assert self.num_prefills == 0
+        assert self.num_prefill_tokens == 0
+        assert self.num_decode_tokens == num_seqs
+        assert self.slot_mapping.shape == (num_seqs, )
+
+        assert self.seq_lens is not None
+        assert len(self.seq_lens) == num_seqs
+        assert self.seq_lens_tensor is not None
+        assert self.seq_lens_tensor.shape == (num_seqs, )
+        assert self.max_query_len == 1
+        assert self.max_prefill_seq_len == 0
+        assert self.max_decode_seq_len == max(self.seq_lens)
+
+        assert self.query_start_loc is not None
+        assert self.query_start_loc.shape == (num_queries + 1, )
+        assert self.seq_start_loc is not None
+        assert self.seq_start_loc.shape == (num_seqs + 1, )
+
+        assert self.context_lens_tensor is not None
+        assert self.context_lens_tensor.shape == (num_queries, )
+
+        assert self.block_tables is not None
+        assert self.block_tables.shape[0] == num_seqs
+
+        # Update query lengths. Note that we update only queries and not seqs,
+        # since tensors may be padded due to captured cuda graph batch size
+        for i in range(num_queries):
+            self.seq_lens[i] += 1
+        self.max_decode_seq_len = max(self.seq_lens)
+
+        ops.advance_step_flashattn(num_seqs=num_seqs,
+                                   num_queries=num_queries,
+                                   block_size=block_size,
+                                   input_tokens=model_input.input_tokens,
+                                   sampled_token_ids=sampled_token_ids,
+                                   input_positions=model_input.input_positions,
+                                   seq_lens=self.seq_lens_tensor,
+                                   slot_mapping=self.slot_mapping,
+                                   block_tables=self.block_tables)
+
+
+class ROCmFlashAttentionMetadataBuilder(
+        CommonMetadataBuilder[ROCmFlashAttentionMetadata]):
+
+    _metadata_cls = ROCmFlashAttentionMetadata
+
+
+def _make_alibi_bias(alibi_slopes: torch.Tensor,
+                     dtype: torch.dtype,
+                     seq_lens: Optional[List[int]],
+                     make_attn_mask: bool = True) -> List[torch.Tensor]:
+    attn_biases = []
+    if seq_lens:
+        for seq_len in seq_lens:
+            bias = torch.arange(seq_len, dtype=dtype)
+            # NOTE(zhuohan): HF uses
+            #     `bias = bias[None, :].repeat(seq_len, 1)`
+            # here. We find that both biases give the same results, but
+            # the bias below more accurately follows the original ALiBi
+            # paper.
+            bias = bias[None, :] - bias[:, None]
+
+            num_heads = alibi_slopes.shape[0]
+            bias = bias[None, :].repeat(
+                (num_heads, 1, 1)).to(alibi_slopes.device)
+            bias.mul_(alibi_slopes[:, None, None])
+            if make_attn_mask:
+                inf_mask = torch.empty(
+                    (1, seq_len, seq_len),
+                    dtype=bias.dtype).fill_(-torch.inf).triu_(diagonal=1).to(
+                        alibi_slopes.device)
+                attn_biases.append((bias + inf_mask).to(dtype))
+            else:
+                attn_biases.append(bias.to(dtype))
+
+    return attn_biases
+
+
+def _get_seq_len_block_table_args(
+    attn_metadata: ROCmFlashAttentionMetadata,
+    attn_type: str,
+) -> tuple:
+    '''
+    The particular choice of sequence-length
+    attributes which should be extracted from attn_metadata is dependent
+    on the type of attention operation.
+
+    Decoder attn -> select entirely decoder self-attention-related fields
+    Encoder/decoder cross-attn -> select encoder sequence lengths
+    Encoder attn -> select encoder sequence lengths fields
+    
+    Arguments:
+
+    * attn_metadata: Attention metadata structure associated with attention op
+    * attn_type: encoder attention, decoder self-attention,
+                encoder/decoder cross-attention
+
+    Returns:
+
+    * Appropriate sequence-lengths tensors for query and key
+    * Appropriate max sequence-length scalar
+    '''
+
+    partial_prefix_sum = 0
+    if attn_type == AttentionType.ENCODER:
+        assert attn_metadata.encoder_seq_lens is not None
+        assert attn_metadata.encoder_seq_lens_tensor is not None
+        query_seq_start_loc = torch.tensor(
+            [0] + [
+                partial_prefix_sum := partial_prefix_sum + i
+                for i in attn_metadata.encoder_seq_lens
+            ],
+            device=attn_metadata.encoder_seq_lens_tensor.device,
+            dtype=attn_metadata.encoder_seq_lens_tensor.dtype)
+        causal_mask = False
+
+        # No block tables associated with encoder attention
+        return (query_seq_start_loc, attn_metadata.max_encoder_seq_len,
+                query_seq_start_loc, attn_metadata.max_encoder_seq_len,
+                attn_metadata.encoder_seq_lens, causal_mask)
+    elif attn_type == AttentionType.DECODER:
+        # Decoder self-attention
+        # Choose max_seq_len based on whether we are in prompt_run
+        assert attn_metadata.seq_lens is not None
+        assert attn_metadata.seq_lens_tensor is not None
+        query_seq_start_loc = torch.tensor(
+            [0] + [
+                partial_prefix_sum := partial_prefix_sum + i
+                for i in attn_metadata.seq_lens
+            ],
+            device=attn_metadata.seq_lens_tensor.device,
+            dtype=attn_metadata.seq_lens_tensor.dtype)
+        max_seq_len = attn_metadata.max_prefill_seq_len
+        causal_mask = True
+
+        return (query_seq_start_loc, max_seq_len, query_seq_start_loc,
+                max_seq_len, attn_metadata.seq_lens, causal_mask)
+    elif attn_type == AttentionType.ENCODER_DECODER:
+        assert attn_metadata.seq_lens is not None
+        assert attn_metadata.encoder_seq_lens_tensor is not None
+        query_start_loc = torch.tensor(
+            [0] + [
+                partial_prefix_sum := partial_prefix_sum + i
+                for i in attn_metadata.seq_lens
+            ],
+            device=attn_metadata.encoder_seq_lens_tensor.device,
+            dtype=attn_metadata.encoder_seq_lens_tensor.dtype)
+
+        partial_prefix_sum = 0
+        assert attn_metadata.encoder_seq_lens is not None
+        assert attn_metadata.seq_lens_tensor is not None
+        key_seq_start_loc = torch.tensor(
+            [0] + [
+                partial_prefix_sum := partial_prefix_sum + i
+                for i in attn_metadata.encoder_seq_lens
+            ],
+            device=attn_metadata.seq_lens_tensor.device,
+            dtype=attn_metadata.seq_lens_tensor.dtype)
+        causal_mask = False
+
+        # Enc/dec cross-attention KVs match encoder sequence length;
+        # cross-attention utilizes special "cross" block tables
+        return (query_start_loc, attn_metadata.max_prefill_seq_len,
+                key_seq_start_loc, attn_metadata.max_encoder_seq_len,
+                attn_metadata.seq_lens, causal_mask)
+    else:
+        raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+
+class ROCmFlashAttentionImpl(AttentionImpl):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prompt_tokens -------------->|
+    |<--prompt_0-->|<--prompt_1-->|...|<--prompt_N-1-->|
+
+    Otherwise, the layout is as follows:
+    |<------------------ num_generation_tokens (M) ----------------->|
+    |<--generation_0-->|..........|<--generation_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+
+    If chunked prefill is enabled, prefill tokens and decode tokens can be
+    batched together in a flattened 1D query.
+
+    |<----- num_prefill_tokens ---->|<------- num_decode_tokens ----------->|	
+    |<-prompt_0->|...|<-prompt_N-1->|<-generation_0->|...|<-generation_M-1->|
+
+    Currently, cuda graph is disabled for chunked prefill, meaning there's no
+    padding between prefill and decode tokens.
+    """
+
+    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: str = AttentionType.DECODER,
+    ) -> None:
+        if blocksparse_params is not None:
+            raise ValueError(
+                "ROCmFlashAttention does not support blocksparse attention.")
+
+        if logits_soft_cap is None:
+            # In flash-attn, setting logits_soft_cap as 0 means no soft cap.
+            self.logits_soft_cap = 0.0
+        else:
+            self.logits_soft_cap = logits_soft_cap
+        self.attn_type = attn_type
+        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
+        self.sliding_window = ((sliding_window, sliding_window)
+                               if sliding_window is not None else (-1, -1))
+        self.kv_cache_dtype = kv_cache_dtype
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        supported_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in supported_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {supported_head_sizes}.")
+
+        self.use_naive_attn = False
+        # NOTE: Allow for switching between Triton and CK. Defaulting to triton.
+        self.use_triton_flash_attn = envs.VLLM_USE_TRITON_FLASH_ATTN
+        if self.use_triton_flash_attn:
+            if logits_soft_cap is not None:
+                raise ValueError(
+                    "ROCm Triton FlashAttention does not support attention"
+                    "logits soft capping."
+                    " please try using the ROCm CK "
+                    "FA backend instead by setting the env var "
+                    "`VLLM_USE_TRITON_FLASH_ATTN=0`")
+
+            from vllm.attention.ops.triton_flash_attention import (  # noqa: F401
+                triton_attention)
+            self.attn_func = triton_attention
+            logger.debug("Using Triton FA in ROCmBackend")
+            if self.sliding_window != (-1, -1):
+                logger.warning("ROCm Triton FA does not currently support "
+                               "sliding window attention. If using half "
+                               "precision, please try using the ROCm CK "
+                               "FA backend instead by setting the env var "
+                               "`VLLM_USE_TRITON_FLASH_ATTN=0`")
+        else:
+            # if not using triton, navi3x/navi21/navi10 do not use flash-attn
+            # either
+            if not current_platform.has_device_capability(90):
+                self.use_naive_attn = True
+            else:
+                try:
+                    from flash_attn import flash_attn_varlen_func  # noqa: F401
+                    self.attn_func = flash_attn_varlen_func
+                    logger.debug("Using CK FA in ROCmBackend")
+                except ModuleNotFoundError:
+                    self.use_naive_attn = True
+
+            if self.use_naive_attn:
+                if logits_soft_cap is not None:
+                    raise ValueError(
+                        "ROCm Naive FlashAttention does not support"
+                        "attention logits soft capping.")
+
+                self.attn_func = _sdpa_attention
+                logger.debug("Using naive (SDPA) attention in ROCmBackend")
+
+    def repeat_kv(self, x: torch.Tensor, n_rep: int) -> torch.Tensor:
+        """torch.repeat_interleave(x, dim=1, repeats=n_rep)"""
+        tokens, n_kv_heads, head_dim = x.shape
+        return (x[:, :,
+                  None, :].expand(tokens, n_kv_heads, n_rep,
+                                  head_dim).reshape(tokens, n_kv_heads * n_rep,
+                                                    head_dim))
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: ROCmFlashAttentionMetadata,
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Forward pass with FlashAttention and PagedAttention.
+
+        For decoder-only models: query, key and value must be non-None.
+
+        For encoder/decoder models:
+        * ROCmFlashAttentionImpl.forward() may be invoked for both self- and 
+            cross-attention layers.
+        * For self-attention: query, key and value must be non-None.
+        * For cross-attention:
+            * Query must be non-None
+            * During prefill, key and value must be non-None; key and value
+              get cached for use during decode.
+            * During decode, key and value may be None, since:
+              (1) key and value tensors were cached during prefill, and
+              (2) cross-attention key and value tensors do not grow during
+                  decode
+        
+        A note on how the attn_type (attention type enum) argument impacts
+        attention forward() behavior:
+    
+            * DECODER: normal decoder-only behavior;
+                use decoder self-attention block table
+            * ENCODER: no KV caching; pass encoder sequence
+                attributes (encoder_seq_lens/encoder_seq_lens_tensor/
+                max_encoder_seq_len) to kernel, in lieu of decoder
+                sequence attributes (seq_lens/seq_lens_tensor/max_seq_len)
+            * ENCODER_DECODER: cross-attention behavior;
+                use cross-attention block table for caching KVs derived
+                from encoder hidden states; since KV sequence lengths
+                will match encoder sequence lengths, pass encoder sequence
+                attributes to kernel (encoder_seq_lens/encoder_seq_lens_tensor/
+                max_encoder_seq_len)
+
+        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]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+            attn_type: Select attention type, between encoder attention,
+                       decoder self-attention, or encoder/decoder cross-
+                       attention. Defaults to decoder self-attention,
+                       which is the vLLM default generally
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        query = query.view(-1, self.num_heads, self.head_size)
+        if key is not None:
+            assert value is not None
+            key = key.view(-1, self.num_kv_heads, self.head_size)
+            value = value.view(-1, self.num_kv_heads, self.head_size)
+        else:
+            assert value is None
+
+        if self.attn_type != AttentionType.ENCODER and kv_cache.numel() > 0:
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            if key is not None and value is not None:
+                # Reshape the input keys and values and store them in the
+                # cache. If kv_cache is not provided, the new key and value
+                # tensors are not cached. This happens during the initial
+                # memory profiling run.
+                PagedAttention.write_to_paged_cache(
+                    key,
+                    value,
+                    key_cache,
+                    value_cache,
+                    attn_metadata.slot_mapping
+                    if self.attn_type != AttentionType.ENCODER_DECODER else
+                    attn_metadata.cross_slot_mapping,
+                    self.kv_cache_dtype,
+                    layer._k_scale,
+                    layer._v_scale,
+                )
+
+        if self.attn_type != AttentionType.ENCODER:
+            num_prefill_tokens = attn_metadata.num_prefill_tokens
+        else:
+            assert attn_metadata.num_encoder_tokens is not None
+            num_prefill_tokens = attn_metadata.num_encoder_tokens
+
+        output = torch.empty_like(query)
+        # Query for decode. KV is not needed because it is already cached.
+        decode_query = query[num_prefill_tokens:]
+        # QKV for prefill.
+        query = query[:num_prefill_tokens]
+
+        if key is not None and value is not None \
+            and self.attn_type != AttentionType.ENCODER_DECODER:
+            key = key[:num_prefill_tokens]
+            value = value[:num_prefill_tokens]
+
+        if prefill_meta := attn_metadata.prefill_metadata:
+            # Prompt run.
+            # normal attention and DECODER
+            if self.attn_type == AttentionType.DECODER and (
+                    kv_cache.numel() == 0 or prefill_meta.block_tables is None
+                    or prefill_meta.block_tables.numel() == 0):
+                (query_seq_start_loc, query_max_seq_len, key_seq_start_loc,
+                 key_max_seq_len, seq_lens,
+                 causal_mask) = (prefill_meta.seq_start_loc,
+                                 prefill_meta.max_prefill_seq_len,
+                                 prefill_meta.seq_start_loc,
+                                 prefill_meta.max_prefill_seq_len,
+                                 attn_metadata.seq_lens, True)
+            # prefix-enabled attention and ENCODER/ENCODER_DECODER
+            else:
+                (query_seq_start_loc, query_max_seq_len, key_seq_start_loc,
+                 key_max_seq_len, seq_lens,
+                 causal_mask) = _get_seq_len_block_table_args(
+                     prefill_meta, self.attn_type)
+            # Prompt run.
+            if kv_cache.numel() == 0 or prefill_meta.block_tables.numel() == 0:
+                # triton attention
+                # When block_tables are not filled, it means q and k are the
+                # prompt, and they have the same length.
+                attn_masks = None
+                if self.use_triton_flash_attn:
+                    if self.alibi_slopes is not None:
+                        attn_masks = _make_alibi_bias(
+                            self.alibi_slopes,
+                            query.dtype,
+                            seq_lens,
+                            make_attn_mask=False)  # type: ignore
+                    out, _ = self.attn_func(
+                        query,
+                        key,
+                        value,
+                        None,
+                        query_seq_start_loc,
+                        key_seq_start_loc,
+                        query_max_seq_len,
+                        key_max_seq_len,
+                        causal_mask,
+                        self.scale,
+                        attn_masks[0][None]
+                        if attn_masks is not None else None,
+                    )
+                elif self.use_naive_attn:
+                    if self.num_kv_heads != self.num_heads:
+                        # Interleave for MQA workaround.
+                        key = self.repeat_kv(key, self.num_queries_per_kv)
+                        value = self.repeat_kv(value, self.num_queries_per_kv)
+                    if self.alibi_slopes is not None:
+                        attn_masks = _make_alibi_bias(
+                            self.alibi_slopes,
+                            query.dtype,
+                            attn_metadata.seq_lens,
+                            make_attn_mask=True)  # type: ignore
+                    query = query.movedim(0, query.dim() - 2)
+                    key = key.movedim(0, key.dim() - 2)
+                    value = value.movedim(0, value.dim() - 2)
+                    # sdpa math backend attention
+                    out = self.attn_func(
+                        query,
+                        key,
+                        value,
+                        query_seq_start_loc,
+                        num_prefill_tokens,
+                        self.num_heads,
+                        self.head_size,
+                        self.scale,
+                        causal_mask,
+                        attn_masks,
+                    )
+                else:
+                    out = self.attn_func(
+                        q=query,
+                        k=key,
+                        v=value,
+                        cu_seqlens_q=query_seq_start_loc,
+                        cu_seqlens_k=key_seq_start_loc,
+                        max_seqlen_q=prefill_meta.max_prefill_seq_len,
+                        max_seqlen_k=key_max_seq_len,
+                        softmax_scale=self.scale,
+                        causal=True,
+                        window_size=self.sliding_window,
+                        alibi_slopes=self.alibi_slopes,
+                        softcap=self.logits_soft_cap,
+                    )
+
+                # common code for prefill
+                assert output[:num_prefill_tokens].shape == out.shape
+                if output.shape[0] > num_prefill_tokens:
+                    output[:num_prefill_tokens] = out
+                else:
+                    output = out
+            else:
+                # prefix-enabled attention
+                output[:num_prefill_tokens] = PagedAttention.forward_prefix(
+                    query,
+                    key,
+                    value,
+                    self.kv_cache_dtype,
+                    key_cache,
+                    value_cache,
+                    prefill_meta.block_tables,
+                    prefill_meta.query_start_loc,
+                    prefill_meta.seq_lens_tensor,
+                    prefill_meta.context_lens_tensor,
+                    prefill_meta.max_query_len,
+                    self.alibi_slopes,
+                    self.sliding_window[0],
+                    layer._k_scale,
+                    layer._v_scale,
+                )
+
+        if decode_meta := attn_metadata.decode_metadata:
+            # Decoding run.
+            # Whether to use rocm custom paged attention or not
+            num_seqs, num_heads, head_size = decode_query.shape
+            block_size = value_cache.shape[3]
+            gqa_ratio = num_heads // self.num_kv_heads
+            use_custom = _use_rocm_custom_paged_attention(
+                decode_query.dtype, head_size, block_size, gqa_ratio,
+                decode_meta.max_decode_seq_len)
+            if use_custom:
+                max_seq_len = (decode_meta.max_decode_seq_len if self.attn_type
+                               != AttentionType.ENCODER_DECODER else
+                               decode_meta.max_encoder_seq_len)
+                assert max_seq_len is not None
+                max_num_partitions = (
+                    (max_seq_len + _PARTITION_SIZE_ROCM - 1) //
+                    _PARTITION_SIZE_ROCM)
+                assert _PARTITION_SIZE_ROCM % 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)
+                if num_prefill_tokens > 0:
+                    out = output[num_prefill_tokens:]
+                else:
+                    out = output
+                ops.paged_attention_rocm(
+                    out,
+                    exp_sums,
+                    max_logits,
+                    tmp_output,
+                    decode_query,
+                    key_cache,
+                    value_cache,
+                    self.num_kv_heads,
+                    self.scale,
+                    decode_meta.block_tables
+                    if self.attn_type != AttentionType.ENCODER_DECODER else
+                    decode_meta.cross_block_tables,
+                    decode_meta.seq_lens_tensor
+                    if self.attn_type != AttentionType.ENCODER_DECODER else
+                    decode_meta.encoder_seq_lens_tensor,
+                    block_size,
+                    max_seq_len,
+                    self.alibi_slopes,
+                    self.kv_cache_dtype,
+                    layer._k_scale,
+                    layer._v_scale,
+                )
+            else:
+                output[num_prefill_tokens:] = PagedAttention.forward_decode(
+                    decode_query,
+                    key_cache,
+                    value_cache,
+                    decode_meta.block_tables
+                    if self.attn_type != AttentionType.ENCODER_DECODER else
+                    decode_meta.cross_block_tables,
+                    decode_meta.seq_lens_tensor
+                    if self.attn_type != AttentionType.ENCODER_DECODER else
+                    decode_meta.encoder_seq_lens_tensor,
+                    decode_meta.max_decode_seq_len
+                    if self.attn_type != AttentionType.ENCODER_DECODER else
+                    decode_meta.max_encoder_seq_len,
+                    self.kv_cache_dtype,
+                    self.num_kv_heads,
+                    self.scale,
+                    self.alibi_slopes,
+                    layer._k_scale,
+                    layer._v_scale,
+                )
+
+        # Reshape the output tensor.
+        return output.view(-1, self.num_heads * self.head_size)
+
+
+def _sdpa_attention(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    seq_lens: List[int],
+    num_tokens: int,
+    num_heads: int,
+    head_size: int,
+    scale: float,
+    attn_masks: Optional[List[torch.Tensor]] = None,
+) -> torch.Tensor:
+    start = 0
+    output = torch.empty((num_tokens, num_heads, head_size),
+                         dtype=query.dtype,
+                         device=query.device)
+
+    for i, seq_len in enumerate(seq_lens):
+        end = start + seq_len
+        with torch.backends.cuda.sdp_kernel(enable_math=True,
+                                            enable_flash=False,
+                                            enable_mem_efficient=False):
+            sub_out = torch.nn.functional.scaled_dot_product_attention(
+                query[:, start:end, :],
+                key[:, start:end, :],
+                value[:, start:end, :],
+                dropout_p=0.0,
+                is_causal=attn_masks is None,
+                attn_mask=attn_masks[i] if attn_masks else None,
+                scale=scale).movedim(query.dim() - 2, 0)
+            output[start:end, :, :] = sub_out
+            start = end
+
+    return output
+
+
+def _use_rocm_custom_paged_attention(qtype: torch.dtype, head_size: int,
+                                     block_size: int, gqa_ratio: int,
+                                     max_seq_len: int) -> bool:
+    # rocm custom page attention not support on navi (gfx1*)
+    return (_ON_MI250_MI300 and not _ON_NAVI
+            and (qtype == torch.half or qtype == torch.bfloat16)
+            and (head_size == 64 or head_size == 128)
+            and (block_size == 16 or block_size == 32)
+            and (gqa_ratio >= 1 and gqa_ratio <= 16) and max_seq_len <= 32768)

.venv/lib/python3.11/site-packages/vllm/attention/backends/torch_sdpa.py

@@ -0,0 +1,681 @@
+# SPDX-License-Identifier: Apache-2.0
+""" Attention layer with torch scaled_dot_product_attention
+    and PagedAttention."""
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+from torch.nn.functional import scaled_dot_product_attention
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionLayer,
+                                              AttentionMetadata,
+                                              AttentionMetadataBuilder,
+                                              AttentionType)
+from vllm.attention.backends.utils import CommonAttentionState
+from vllm.attention.ops.ipex_attn import PagedAttention
+from vllm.attention.ops.paged_attn import PagedAttentionMetadata
+from vllm.logger import init_logger
+from vllm.utils import make_tensor_with_pad
+from vllm.worker.cpu_model_runner import ModelInputForCPUBuilder
+
+logger = init_logger(__name__)
+
+
+class TorchSDPABackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "TORCH_SDPA"
+
+    @staticmethod
+    def get_impl_cls() -> Type["TorchSDPABackendImpl"]:
+        return TorchSDPABackendImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return TorchSDPAMetadata
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_builder_cls() -> Type["TorchSDPAMetadataBuilder"]:
+        return TorchSDPAMetadataBuilder
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class TorchSDPAMetadata(AttentionMetadata, PagedAttentionMetadata):
+    """Metadata for TorchSDPABackend.
+    """
+    # Currently, input sequences can only contain all prompts
+    # or all decoding. True if all sequences are prompts.
+    chunked_prefill: bool
+    seq_lens: Optional[List[int]] = None  # For non-chunked prefill
+
+    # For chunked prefill only
+    max_query_len: Optional[int] = None
+    max_kv_len: Optional[int] = None
+    query_start_loc: Optional[torch.Tensor] = None
+    kv_start_loc: Optional[torch.Tensor] = None
+    prefill_block_tables: Optional[torch.Tensor] = None
+
+    # Begin encoder attn & enc/dec cross-attn fields...
+    # Encoder sequence lengths representation
+    encoder_seq_lens: Optional[List[int]] = None
+    encoder_seq_lens_tensor: Optional[torch.Tensor] = None
+
+    # Maximum sequence length among encoder sequences
+    max_encoder_seq_len: Optional[int] = None
+
+    # Number of tokens input to encoder
+    num_encoder_tokens: Optional[int] = None
+
+    # Cross-attention memory-mapping data structures: slot mapping
+    # and block tables
+    cross_slot_mapping: Optional[torch.Tensor] = None
+    cross_block_tables: Optional[torch.Tensor] = None
+
+    def __post_init__(self):
+        # Set during the execution of the first attention op.
+        # It is a list because it is needed to set per prompt
+        # when alibi slopes is used. It is because of the limitation
+        # from xformer API.
+        # will not appear in the __repr__ and __init__
+        self.attn_bias: Optional[List[torch.Tensor]] = None
+        self.encoder_attn_bias: Optional[List[torch.Tensor]] = None
+        self.cross_attn_bias: Optional[List[torch.Tensor]] = None
+
+    @property
+    def is_all_encoder_attn_metadata_set(self):
+        '''
+        All attention metadata required for encoder attention is set.
+        '''
+        return ((self.encoder_seq_lens is not None)
+                and (self.encoder_seq_lens_tensor is not None)
+                and (self.max_encoder_seq_len is not None))
+
+    @property
+    def is_all_cross_attn_metadata_set(self):
+        '''
+        All attention metadata required for enc/dec cross-attention is set.
+
+        Superset of encoder attention required metadata.
+        '''
+        return (self.is_all_encoder_attn_metadata_set
+                and (self.cross_slot_mapping is not None)
+                and (self.cross_block_tables is not None))
+
+    @property
+    def prefill_metadata(self) -> Optional["TorchSDPAMetadata"]:
+        if self.num_prefill_tokens == 0:
+            return None
+        return self
+
+    @property
+    def decode_metadata(self) -> Optional["TorchSDPAMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+        return self
+
+    def get_seq_lens(
+        self,
+        attn_type: str,
+    ):
+        '''
+        Extract appropriate sequence lengths from attention metadata
+        according to attention type.
+
+        Arguments:
+
+        * attn_metadata: Attention metadata structure associated with attention
+        * attn_type: encoder attention, decoder self-attention,
+                    encoder/decoder cross-attention
+
+        Returns:
+        * Appropriate sequence lengths tensor for query
+        * Appropriate sequence lengths tensor for key & value
+        '''
+
+        if (attn_type == AttentionType.DECODER
+                or attn_type == AttentionType.ENCODER_ONLY):
+            seq_lens_q = self.seq_lens
+            seq_lens_kv = self.seq_lens
+        elif attn_type == AttentionType.ENCODER:
+            seq_lens_q = self.encoder_seq_lens
+            seq_lens_kv = self.encoder_seq_lens
+        elif attn_type == AttentionType.ENCODER_DECODER:
+            seq_lens_q = self.seq_lens
+            seq_lens_kv = self.encoder_seq_lens
+        else:
+            raise AttributeError(f"Invalid attention type {str(attn_type)}")
+        return seq_lens_q, seq_lens_kv
+
+    def get_attn_bias(
+        self,
+        attn_type: str,
+    ) -> Optional[List[torch.Tensor]]:
+        '''
+        Extract appropriate attention bias from attention metadata
+        according to attention type.
+
+        Arguments:
+
+        * attn_metadata: Attention metadata structure associated with attention
+        * attn_type: encoder attention, decoder self-attention,
+                    encoder/decoder cross-attention
+
+        Returns:
+        * Appropriate attention bias value given the attention type
+        '''
+
+        if (attn_type == AttentionType.DECODER
+                or attn_type == AttentionType.ENCODER_ONLY):
+            return self.attn_bias
+        elif attn_type == AttentionType.ENCODER:
+            return self.encoder_attn_bias
+        elif attn_type == AttentionType.ENCODER_DECODER:
+            return self.cross_attn_bias
+        else:
+            raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+    def set_attn_bias(
+        self,
+        attn_bias: List[torch.Tensor],
+        attn_type: str,
+    ) -> None:
+        '''
+        Update appropriate attention bias field of attention metadata,
+        according to attention type.
+
+        Arguments:
+
+        * attn_metadata: Attention metadata structure associated with attention
+        * attn_bias: The desired attention bias value
+        * attn_type: encoder attention, decoder self-attention,
+                    encoder/decoder cross-attention
+        '''
+
+        if (attn_type == AttentionType.DECODER
+                or attn_type == AttentionType.ENCODER_ONLY):
+            self.attn_bias = attn_bias
+        elif attn_type == AttentionType.ENCODER:
+            self.encoder_attn_bias = attn_bias
+        elif attn_type == AttentionType.ENCODER_DECODER:
+            self.cross_attn_bias = attn_bias
+        else:
+            raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+    def get_seq_len_block_table_args(
+        self,
+        attn_type: str,
+    ) -> tuple:
+        '''
+        The particular choice of sequence-length- and block-table-related
+        attributes which should be extracted from attn_metadata is dependent
+        on the type of attention operation.
+
+        Decoder attn -> select entirely decoder self-attention-related fields
+        Encoder/decoder cross-attn -> select encoder sequence lengths &
+                                    cross-attn block-tables fields
+        Encoder attn -> select encoder sequence lengths fields & no block tables
+
+        Arguments:
+
+        * attn_metadata: Attention metadata structure associated with attention
+        * is_prompt: True if prefill, False otherwise
+        * attn_type: encoder attention, decoder self-attention,
+                    encoder/decoder cross-attention
+
+        Returns:
+
+        * Appropriate sequence-lengths tensor
+        * Appropriate max sequence-length scalar
+        * Appropriate block tables (or None)
+        '''
+
+        if (attn_type == AttentionType.DECODER
+                or attn_type == AttentionType.ENCODER_ONLY):
+            # Decoder self-attention
+            # Choose max_seq_len based on whether we are in prompt_run
+            return (self.seq_lens_tensor, self.max_decode_seq_len,
+                    self.block_tables)
+        elif attn_type == AttentionType.ENCODER_DECODER:
+            # Enc/dec cross-attention KVs match encoder sequence length;
+            # cross-attention utilizes special "cross" block tables
+            return (self.encoder_seq_lens_tensor, self.max_encoder_seq_len,
+                    self.cross_block_tables)
+        elif attn_type == AttentionType.ENCODER:
+            # No block tables associated with encoder attention
+            return (self.encoder_seq_lens_tensor, self.max_encoder_seq_len,
+                    None)
+        else:
+            raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+
+class TorchSDPAMetadataBuilder(AttentionMetadataBuilder[TorchSDPAMetadata]):
+
+    def __init__(self, input_builder: ModelInputForCPUBuilder) -> None:
+        self.chunked_prefill = input_builder.chunked_prefill
+        self.input_builder = input_builder
+
+    def prepare(self):
+        self.input_data = self.input_builder.input_data
+
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int) -> TorchSDPAMetadata:
+        input_data = self.input_data
+        prefill_seq_lens = seq_lens[0:input_data.num_prefills]
+        prefill_query_lens = query_lens[0:input_data.num_prefills]
+        slot_mapping = torch.tensor(input_data.slot_mapping,
+                                    dtype=torch.long,
+                                    device="cpu")
+
+        # For chunked-prefill
+        if self.chunked_prefill and input_data.num_prefill_tokens != 0:
+            prefill_block_tables = make_tensor_with_pad(
+                self.input_data.prefill_block_tables,
+                pad=0,
+                dtype=torch.int32,
+                device="cpu",
+            )
+            query_lens_tensor = torch.tensor(prefill_query_lens,
+                                             dtype=torch.int32,
+                                             device="cpu")
+            kv_lens_tensor = torch.tensor(prefill_seq_lens,
+                                          dtype=torch.int32,
+                                          device="cpu")
+            query_start_loc = torch.zeros(input_data.num_prefills + 1,
+                                          dtype=torch.int32,
+                                          device="cpu")
+            kv_start_loc = torch.zeros(input_data.num_prefills + 1,
+                                       dtype=torch.int32,
+                                       device="cpu")
+            torch.cumsum(query_lens_tensor,
+                         dim=0,
+                         dtype=torch.int32,
+                         out=query_start_loc[1:])
+            torch.cumsum(kv_lens_tensor,
+                         dim=0,
+                         dtype=torch.int32,
+                         out=kv_start_loc[1:])
+            max_query_len = max(prefill_query_lens)
+            max_kv_len = max(prefill_seq_lens)
+        else:
+            prefill_block_tables = None
+            query_start_loc = None
+            kv_start_loc = None
+            max_query_len = None
+            max_kv_len = None
+
+        # For paged attention
+        if input_data.num_decode_tokens != 0:
+            seq_lens_tensor = torch.tensor(
+                input_data.seq_lens[input_data.num_prefills:],
+                dtype=torch.int32,
+                device="cpu",
+            )
+            block_tables = make_tensor_with_pad(
+                self.input_data.decode_block_tables,
+                pad=0,
+                dtype=torch.int32,
+                device="cpu",
+            )
+        else:
+            block_tables = torch.tensor([])
+            seq_lens_tensor = torch.tensor(
+                input_data.seq_lens[:input_data.num_prefills],
+                dtype=torch.int32,
+                device="cpu",
+            )
+
+        # For multi-modal models
+        placeholder_index_maps = None
+        if len(input_data.multi_modal_inputs_list) != 0:
+            placeholder_index_maps = {
+                modality: placeholder_map.index_map()
+                for modality, placeholder_map in
+                input_data.multi_modal_placeholder_maps.items()
+            }
+
+        attn_metadata = TorchSDPAMetadata(
+            chunked_prefill=self.chunked_prefill,
+            seq_lens=prefill_seq_lens,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=max_query_len,
+            max_kv_len=max_kv_len,
+            query_start_loc=query_start_loc,
+            kv_start_loc=kv_start_loc,
+            max_decode_seq_len=input_data.max_decode_seq_len,
+            num_prefills=input_data.num_prefills,
+            num_prefill_tokens=input_data.num_prefill_tokens,
+            num_decode_tokens=input_data.num_decode_tokens,
+            block_tables=block_tables,
+            prefill_block_tables=prefill_block_tables,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=placeholder_index_maps,
+            enable_kv_scales_calculation=False,
+        )
+
+        return attn_metadata
+
+
+class TorchSDPABackendImpl(AttentionImpl[TorchSDPAMetadata]):
+
+    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: str = AttentionType.DECODER,
+    ) -> None:
+        if blocksparse_params is not None:
+            raise ValueError(
+                "Torch SPDA does not support block-sparse attention.")
+        if logits_soft_cap is not None:
+            logger.warning_once("Torch SPDA does not support logits soft cap. "
+                                "Outputs may be slightly off.")
+        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
+        self.sliding_window = sliding_window
+        self.kv_cache_dtype = kv_cache_dtype
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+        self.need_mask = (self.alibi_slopes is not None
+                          or self.sliding_window is not None)
+
+        supported_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in supported_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {supported_head_sizes}.")
+        if kv_cache_dtype != "auto":
+            raise NotImplementedError(
+                "Torch SDPA backend does not support FP8 KV cache. "
+                "Please use xFormers backend instead.")
+        self.attn_type = attn_type
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: TorchSDPAMetadata,  # type: ignore
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Forward pass with torch SDPA and PagedAttention.
+
+        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]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        attn_type = self.attn_type
+        if (attn_type == AttentionType.ENCODER
+                and (not attn_metadata.is_all_encoder_attn_metadata_set)):
+            raise AttributeError("Encoder attention requires setting "
+                                 "encoder metadata attributes.")
+        elif (attn_type == AttentionType.ENCODER_DECODER
+              and (not attn_metadata.is_all_cross_attn_metadata_set)):
+            raise AttributeError("Encoder/decoder cross-attention "
+                                 "requires setting cross-attention "
+                                 "metadata attributes.")
+
+        # Reshape the query, key, and value tensors.
+        query = query.view(-1, self.num_heads, self.head_size)
+        if key is not None:
+            assert value is not None
+            key = key.view(-1, self.num_kv_heads, self.head_size)
+            value = value.view(-1, self.num_kv_heads, self.head_size)
+        else:
+            assert value is None
+
+        if (attn_type != AttentionType.ENCODER and kv_cache.numel() > 0):
+            # KV-cache during decoder-self- or
+            # encoder-decoder-cross-attention, but not
+            # during encoder attention.
+            #
+            # Even if there are no new key/value pairs to cache,
+            # we still need to break out key_cache and value_cache
+            # i.e. for later use by paged attention
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            if (key is not None) and (value is not None):
+                if attn_type == AttentionType.ENCODER_DECODER:
+                    # Update cross-attention KV cache (prefill-only)
+                    # During cross-attention decode, key & value will be None,
+                    # preventing this IF-statement branch from running
+                    updated_slot_mapping = attn_metadata.cross_slot_mapping
+                else:
+                    # Update self-attention KV cache (prefill/decode)
+                    updated_slot_mapping = attn_metadata.slot_mapping
+
+                PagedAttention.write_to_paged_cache(
+                    key, value, key_cache, value_cache, updated_slot_mapping,
+                    self.kv_cache_dtype, layer._k_scale, layer._v_scale)
+
+        if attn_type != AttentionType.ENCODER:
+            # Decoder self-attention supports chunked prefill.
+            # Encoder/decoder cross-attention requires no chunked
+            # prefill (100% prefill or 100% decode tokens, no mix)
+            num_prefill_tokens = attn_metadata.num_prefill_tokens
+            num_decode_tokens = attn_metadata.num_decode_tokens
+        else:
+            # Encoder attention - chunked prefill is not applicable;
+            # derive token-count from query shape & and treat them
+            # as 100% prefill tokens
+            assert attn_metadata.num_encoder_tokens is not None
+            num_prefill_tokens = attn_metadata.num_encoder_tokens
+            num_decode_tokens = 0
+
+        if attn_type == AttentionType.DECODER:
+            # Only enforce this shape-constraint for decoder
+            # self-attention
+            assert key.shape[0] == num_prefill_tokens + num_decode_tokens
+            assert value.shape[0] == num_prefill_tokens + num_decode_tokens
+
+        output = torch.empty_like(query)
+        if prefill_meta := attn_metadata.prefill_metadata:
+            assert attn_metadata.seq_lens is not None
+            if not prefill_meta.prefill_metadata.chunked_prefill:  # type: ignore
+                self._run_sdpa_forward(output,
+                                       query,
+                                       key,
+                                       value,
+                                       prefill_meta,
+                                       attn_type=attn_type)
+            else:
+                # prefix-enabled attention
+                assert not self.need_mask
+                import intel_extension_for_pytorch.llm.modules as ipex_modules
+                output = torch.empty_like(query)
+                ipex_modules.PagedAttention.flash_attn_varlen_func(
+                    output[:prefill_meta.num_prefill_tokens, :, :],
+                    query[:prefill_meta.num_prefill_tokens, :, :],
+                    key_cache,
+                    value_cache,
+                    prefill_meta.query_start_loc,
+                    prefill_meta.kv_start_loc,
+                    prefill_meta.max_query_len,
+                    prefill_meta.max_kv_len,
+                    self.scale,
+                    True,
+                    prefill_meta.prefill_block_tables,
+                    self.alibi_slopes,
+                )
+
+        if decode_meta := attn_metadata.decode_metadata:
+            assert attn_type != AttentionType.ENCODER_ONLY, (
+                "Encoder-only models should not have decode metadata.")
+            # Decoding run.
+            (
+                seq_lens_arg,
+                max_seq_len_arg,
+                block_tables_arg,
+            ) = decode_meta.get_seq_len_block_table_args(attn_type)
+
+            PagedAttention.forward_decode(
+                output[attn_metadata.num_prefill_tokens:, :, :],
+                query[attn_metadata.num_prefill_tokens:, :, :],
+                key_cache,
+                value_cache,
+                block_tables_arg,
+                seq_lens_arg,
+                max_seq_len_arg,
+                self.kv_cache_dtype,
+                self.num_kv_heads,
+                self.scale,
+                self.alibi_slopes,
+                layer._k_scale,
+                layer._v_scale,
+            )
+
+        # Reshape the output tensor.
+        return output.view(-1, self.num_heads * self.head_size)
+
+    def _run_sdpa_forward(
+        self,
+        output: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: TorchSDPAMetadata,
+        attn_type: str = AttentionType.DECODER,
+    ) -> None:
+        if self.num_kv_heads != self.num_heads:
+            key = key.repeat_interleave(self.num_queries_per_kv, dim=1)
+            value = value.repeat_interleave(self.num_queries_per_kv, dim=1)
+
+        attn_masks = attn_metadata.get_attn_bias(attn_type)
+        if attn_masks is None:
+            if self.alibi_slopes is not None:
+                attn_masks = _make_alibi_bias(
+                    self.alibi_slopes, query.dtype,
+                    attn_metadata.seq_lens)  # type: ignore
+            elif self.sliding_window is not None:
+                assert attn_metadata.seq_lens is not None
+                attn_masks = _make_sliding_window_bias(
+                    attn_metadata.seq_lens, self.sliding_window,
+                    query.dtype)  # type: ignore
+            else:
+                seq_lens, _ = attn_metadata.get_seq_lens(attn_type)
+                attn_masks = [None] * len(seq_lens)
+            attn_metadata.set_attn_bias(attn_masks, attn_type)
+
+        query = query.movedim(0, query.dim() - 2)
+        key = key.movedim(0, key.dim() - 2)
+        value = value.movedim(0, value.dim() - 2)
+
+        causal_attn = (attn_type == AttentionType.DECODER)
+
+        seq_lens_q, seq_lens_kv = attn_metadata.get_seq_lens(attn_type)
+        start_q, start_kv = 0, 0
+        for seq_len_q, seq_len_kv, mask in zip(seq_lens_q, seq_lens_kv,
+                                               attn_masks):
+            end_q = start_q + seq_len_q
+            end_kv = start_kv + seq_len_kv
+            sub_out = scaled_dot_product_attention(
+                query[None, :, start_q:end_q, :],
+                key[None, :, start_kv:end_kv, :],
+                value[None, :, start_kv:end_kv, :],
+                attn_mask=mask,
+                dropout_p=0.0,
+                is_causal=causal_attn and mask is None,
+                scale=self.scale).squeeze(0).movedim(query.dim() - 2, 0)
+            output[start_q:end_q, :, :] = sub_out
+            start_q, start_kv = end_q, end_kv
+
+
+def _make_alibi_bias(
+    alibi_slopes: torch.Tensor,
+    dtype: torch.dtype,
+    seq_lens: List[int],
+) -> List[torch.Tensor]:
+    attn_biases: List[torch.Tensor] = []
+    for seq_len in seq_lens:
+        bias = torch.arange(seq_len, dtype=dtype)
+        # NOTE(zhuohan): HF uses
+        #     `bias = bias[None, :].repeat(seq_len, 1)`
+        # here. We find that both biases give the same results, but
+        # the bias below more accurately follows the original ALiBi
+        # paper.
+        bias = bias[None, :] - bias[:, None]
+
+        num_heads = alibi_slopes.shape[0]
+        bias = bias[None, :].repeat((num_heads, 1, 1))
+        bias.mul_(alibi_slopes[:, None, None]).unsqueeze_(0)
+        inf_mask = torch.empty(
+            (1, seq_len, seq_len),
+            dtype=bias.dtype).fill_(-torch.inf).triu_(diagonal=1)
+        attn_biases.append((bias + inf_mask).to(dtype))
+
+    return attn_biases
+
+
+def _make_sliding_window_bias(
+    seq_lens: List[int],
+    window_size: Optional[int],
+    dtype: torch.dtype,
+) -> List[torch.Tensor]:
+    attn_biases: List[torch.Tensor] = []
+    for seq_len in seq_lens:
+        tensor = torch.full(
+            (1, seq_len, seq_len),
+            dtype=dtype,
+            fill_value=1,
+        )
+        shift = 0
+        mask = torch.tril(tensor, diagonal=shift).to(dtype)  # type: ignore
+        if window_size is not None:
+            mask = torch.triu(mask, diagonal=shift - window_size + 1)
+        mask = torch.log(mask)
+        attn_biases.append(mask.to(dtype))
+
+    return attn_biases

.venv/lib/python3.11/site-packages/vllm/attention/backends/triton_mla.py

@@ -0,0 +1,746 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from collections import defaultdict
+from contextlib import contextmanager
+from dataclasses import dataclass
+from itertools import accumulate
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type
+
+from vllm.multimodal import MultiModalPlaceholderMap
+
+try:
+    from flashinfer import BatchDecodeMlaWithPagedKVCacheWrapper
+    FLASHINFER_WORKSPACE_BUFFER_SIZE = 256 * 1024 * 1024
+except ImportError:
+    BatchDecodeMlaWithPagedKVCacheWrapper = None
+    FLASHINFER_WORKSPACE_BUFFER_SIZE = 0
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.attention.backends.abstract import (AttentionBackend,
+                                              AttentionMetadata,
+                                              AttentionMetadataBuilder,
+                                              AttentionState, AttentionType)
+from vllm.attention.backends.mla.utils import MLACommonImpl, MLACommonMetadata
+from vllm.attention.backends.utils import (PAD_SLOT_ID, compute_slot_mapping,
+                                           compute_slot_mapping_start_idx,
+                                           is_block_tables_empty)
+from vllm.attention.ops.triton_decode_attention import decode_attention_fwd
+from vllm.utils import async_tensor_h2d, make_tensor_with_pad
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import (ModelInputForGPUBuilder,
+                                          ModelInputForGPUWithSamplingMetadata)
+
+
+class TritonMLABackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "TRITON_MLA"
+
+    @staticmethod
+    def get_impl_cls() -> Type["TritonMLAImpl"]:
+        return TritonMLAImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return TritonMLAMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["TritonMLAMetadataBuilder"]:
+        return TritonMLAMetadataBuilder
+
+    @staticmethod
+    def get_state_cls() -> Type["TritonMLAState"]:
+        return TritonMLAState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,  # assumed to be 1 for MLA
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return (num_blocks, block_size, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        ops.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        ops.copy_blocks_mla(kv_caches, src_to_dists)
+
+    @staticmethod
+    def get_supported_head_sizes() -> List[int]:
+        return [576]
+
+
+class TritonMLAState(AttentionState):
+
+    def __init__(self, runner):
+        self.runner = runner
+        self._is_graph_capturing = False
+
+    @contextmanager
+    def graph_capture(self, max_batch_size: int):
+        self._is_graph_capturing = True
+
+        self._graph_slot_mapping = torch.full((max_batch_size, ),
+                                              PAD_SLOT_ID,
+                                              dtype=torch.long,
+                                              device=self.runner.device)
+        self._graph_seq_lens = torch.ones(max_batch_size,
+                                          dtype=torch.int32,
+                                          device=self.runner.device)
+        self._graph_block_tables = torch.from_numpy(
+            self.runner.graph_block_tables).to(device=self.runner.device)
+
+        self._positions = torch.zeros((max_batch_size, ),
+                                      dtype=torch.long,
+                                      device=self.runner.device)
+
+        yield
+
+        self._is_graph_capturing = False
+        del self._graph_slot_mapping
+        del self._graph_seq_lens
+        del self._graph_block_tables
+        del self._positions
+
+    def graph_clone(self, batch_size: int):
+        assert self._is_graph_capturing
+        return self.__class__(self.runner)
+
+    def graph_capture_get_metadata_for_batch(
+            self, batch_size: int, is_encoder_decoder_model: bool = False):
+        assert self._is_graph_capturing
+
+        attn_metadata = self.runner.attn_backend.make_metadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=batch_size,
+            slot_mapping=self._graph_slot_mapping[:batch_size],
+            multi_modal_placeholder_index_maps=None,
+            enable_kv_scales_calculation=True,
+            seq_lens=None,
+            seq_lens_tensor=self._graph_seq_lens[:batch_size],
+            max_query_len=1,
+            max_decode_query_len=1,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.runner.max_seq_len_to_capture,
+            query_start_loc=None,
+            seq_start_loc=None,
+            context_lens_tensor=None,
+            block_tables=self._graph_block_tables[:batch_size],
+            use_cuda_graph=True,
+            input_positions=self._positions[:batch_size],
+            head_dim=self.runner.model_config.get_head_size())
+
+        if is_encoder_decoder_model:
+            raise NotImplementedError(
+                "TritonMLAState does not support encoder/decoder yet")
+
+        return attn_metadata
+
+    def get_graph_input_buffers(self,
+                                attn_metadata,
+                                is_encoder_decoder_model: bool = False):
+        input_buffers = {
+            "slot_mapping": attn_metadata.slot_mapping,
+            "seq_lens_tensor": attn_metadata.decode_metadata.seq_lens_tensor,
+            "block_tables": attn_metadata.decode_metadata.block_tables,
+            "input_positions": attn_metadata.decode_metadata.input_positions,
+        }
+        if is_encoder_decoder_model:
+            raise NotImplementedError(
+                "TritonMLAState does not support encoder/decoder yet")
+
+        return input_buffers
+
+    def prepare_graph_input_buffers(self,
+                                    input_buffers,
+                                    attn_metadata,
+                                    is_encoder_decoder_model: bool = False):
+        input_positions = attn_metadata.input_positions
+        num_positions = input_positions.shape[0]
+        input_buffers["seq_lens_tensor"].copy_(
+            attn_metadata.decode_metadata.seq_lens_tensor, non_blocking=True)
+        input_buffers["block_tables"].copy_(
+            attn_metadata.decode_metadata.block_tables, non_blocking=True)
+        # CUDA graph buffer is padded so only perform a partial copy based on
+        # num_positions
+        input_buffers["input_positions"][:num_positions].copy_(
+            input_positions, non_blocking=True)
+        if is_encoder_decoder_model:
+            raise NotImplementedError(
+                "TritonMLAState does not support encoder/decoder yet")
+
+    def begin_forward(self, model_input):
+        return
+
+
+@dataclass
+class TritonMLAMetadata(MLACommonMetadata):
+    """Metadata for TritonMLAMetadata.
+
+    NOTE: Any python object stored here is not updated when it is
+    cuda-graph replayed. If you have values that need to be changed
+    dynamically, it should be stored in tensor. The tensor has to be
+    updated from `CUDAGraphRunner.forward` API.
+    """
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]]
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+
+    # NOTE(sang): Definition of context_len, query_len, and seq_len.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ---------------------|
+    #                                   |-- query_len ---|
+
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Maximum sequence length among decode batch. 0 if there are prefill
+    # requests only.
+    max_decode_seq_len: int
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor]
+
+    # (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]
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+
+    use_cuda_graph: bool
+
+    # Maximum query length in the batch.
+    max_query_len: Optional[int] = None
+
+    # Max number of query tokens among request in the batch.
+    max_decode_query_len: Optional[int] = None
+
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    query_start_loc: Optional[torch.Tensor] = None
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor] = None
+
+    _cached_prefill_metadata: Optional["TritonMLAMetadata"] = None
+    _cached_decode_metadata: Optional["TritonMLAMetadata"] = None
+
+    num_prefill_tokens: int
+
+    num_kv_splits: int = 4  # TODO(lucas) add heuristic
+    attn_logits: Optional[torch.Tensor] = None
+    req_idx: Optional[torch.Tensor] = None
+
+    # The dimension of the attention heads
+    head_dim: Optional[int] = None
+
+    def __post_init__(self):
+        supported_head_sizes = TritonMLABackend.get_supported_head_sizes()
+        if self.head_dim is not None and self.head_dim \
+                not in supported_head_sizes:
+            raise ValueError(
+                f"Only {supported_head_sizes} are supported for head_dim,",
+                f"received {self.head_dim}.")
+
+    @property
+    def prefill_metadata(self) -> Optional["TritonMLAMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            return self._cached_prefill_metadata
+
+        assert self.seq_lens is not None
+        assert self.seq_lens_tensor is not None
+
+        # Compute some attn_metadata fields which default to None
+        query_start_loc = (None if self.query_start_loc is None else
+                           self.query_start_loc[:self.num_prefills + 1])
+        slot_mapping = (None if self.slot_mapping is None else
+                        self.slot_mapping[:self.num_prefill_tokens])
+        seq_lens = (None if self.seq_lens is None else
+                    self.seq_lens[:self.num_prefills])
+        seq_lens_tensor = (None if self.seq_lens_tensor is None else
+                           self.seq_lens_tensor[:self.num_prefills])
+        seq_start_loc = (None if self.seq_start_loc is None else
+                         self.seq_start_loc[:self.num_prefills + 1])
+        context_lens_tensor = (None if self.context_lens_tensor is None else
+                               self.context_lens_tensor[:self.num_prefills])
+        block_tables = (None if self.block_tables is None else
+                        self.block_tables[:self.num_prefills])
+        input_positions = (None if self.input_positions is None else
+                           self.input_positions[:self.num_prefill_tokens])
+
+        self._cached_prefill_metadata = TritonMLAMetadata(
+            num_prefills=self.num_prefills,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=0,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=self.
+            multi_modal_placeholder_index_maps,
+            enable_kv_scales_calculation=self.enable_kv_scales_calculation,
+            input_positions=input_positions,
+            seq_lens=seq_lens,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=self.max_prefill_seq_len,
+            max_decode_query_len=0,
+            max_decode_seq_len=0,
+            query_start_loc=query_start_loc,
+            seq_start_loc=seq_start_loc,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=False,
+            head_dim=self.head_dim)
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> Optional["TritonMLAMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            return self._cached_decode_metadata
+        assert self.seq_lens_tensor is not None
+
+        # Compute some attn_metadata fields which default to None
+        slot_mapping = (None if self.slot_mapping is None else
+                        self.slot_mapping[self.num_prefill_tokens:])
+        seq_lens_tensor = (None if self.seq_lens_tensor is None else
+                           self.seq_lens_tensor[self.num_prefills:])
+        block_tables = (None if self.block_tables is None else
+                        self.block_tables[self.num_prefills:])
+        input_positions = (None if self.input_positions is None else
+                           self.input_positions[self.num_prefill_tokens:])
+
+        self._cached_decode_metadata = TritonMLAMetadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=self.num_decode_tokens,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=None,
+            enable_kv_scales_calculation=True,
+            seq_lens=None,
+            seq_lens_tensor=seq_lens_tensor,
+            max_decode_query_len=self.max_decode_query_len,
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.max_decode_seq_len,
+            # Batch may be composed of prefill|decodes, adjust query start
+            # indices to refer to the start of decodes. E.g.
+            # in tokens:[3 prefills|6 decodes], query_start_loc=[3,9] => [0,6].
+            query_start_loc=(self.query_start_loc[self.num_prefills:] -
+                             self.query_start_loc[self.num_prefills])
+            if self.query_start_loc is not None else None,
+            seq_start_loc=self.seq_start_loc[self.num_prefills:]
+            if self.seq_start_loc is not None else None,
+            context_lens_tensor=None,
+            block_tables=block_tables,
+            use_cuda_graph=self.use_cuda_graph,
+            input_positions=input_positions,
+            head_dim=self.head_dim)
+        return self._cached_decode_metadata
+
+    def advance_step(self,
+                     model_input: "ModelInputForGPUWithSamplingMetadata",
+                     sampled_token_ids: Optional[torch.Tensor],
+                     block_size: int,
+                     num_seqs: int,
+                     num_queries: int,
+                     turn_prefills_into_decodes: bool = False):
+        """
+        Update metadata in-place to advance one decode step.
+        """
+        # When using cudagraph, the num_seqs is padded to the next captured
+        # batch sized, but num_queries tracks the actual number of requests in
+        # the batch. For --enforce-eager mode, num_seqs == num_queries
+        if num_seqs != num_queries:
+            assert num_seqs > num_queries
+            assert self.use_cuda_graph
+
+        if turn_prefills_into_decodes:
+            # When Mutli-Step is enabled with Chunked-Prefill, prefills and
+            # decodes are scheduled together. In the first step, all the
+            # prefills turn into decodes. This update reflects that
+            # conversion.
+            assert self.num_decode_tokens + self.num_prefills == num_seqs
+            self.num_decode_tokens += self.num_prefills
+            self.num_prefills = 0
+            self.num_prefill_tokens = 0
+            self.max_prefill_seq_len = 0
+            self.max_query_len = 1
+
+            self.slot_mapping = self.slot_mapping[:num_seqs]
+        else:
+            assert self.seq_lens is not None
+            assert self.max_decode_seq_len == max(self.seq_lens)
+
+        assert self.num_prefills == 0
+        assert self.num_prefill_tokens == 0
+        assert self.num_decode_tokens == num_seqs
+        assert self.slot_mapping.shape == (num_seqs, )
+
+        assert self.seq_lens is not None
+        assert len(self.seq_lens) == num_seqs
+        assert self.seq_lens_tensor is not None
+        assert self.seq_lens_tensor.shape == (num_seqs, )
+        assert self.max_query_len == 1
+        assert self.max_prefill_seq_len == 0
+
+        assert self.query_start_loc is not None
+        assert self.query_start_loc.shape == (num_queries + 1, )
+        assert self.seq_start_loc is not None
+        assert self.seq_start_loc.shape == (num_seqs + 1, )
+
+        assert self.context_lens_tensor is not None
+        assert self.context_lens_tensor.shape == (num_queries, )
+
+        assert self.block_tables is not None
+        assert self.block_tables.shape[0] == num_seqs
+
+        # Update query lengths. Note that we update only queries and not seqs,
+        # since tensors may be padded due to captured cuda graph batch size
+        for i in range(num_queries):
+            self.seq_lens[i] += 1
+        self.max_decode_seq_len = max(self.seq_lens)
+
+        ops.advance_step_flashattn(num_seqs=num_seqs,
+                                   num_queries=num_queries,
+                                   block_size=block_size,
+                                   input_tokens=model_input.input_tokens,
+                                   sampled_token_ids=sampled_token_ids,
+                                   input_positions=model_input.input_positions,
+                                   seq_lens=self.seq_lens_tensor,
+                                   slot_mapping=self.slot_mapping,
+                                   block_tables=self.block_tables)
+
+
+class TritonMLAMetadataBuilder(AttentionMetadataBuilder[TritonMLAMetadata]):
+
+    def __init__(self, input_builder: "ModelInputForGPUBuilder"):
+        self.input_builder = input_builder
+        self.runner = input_builder.runner
+        self.sliding_window = input_builder.sliding_window
+        self.block_size = input_builder.block_size
+
+    def prepare(self):
+        self.slot_mapping: List[int] = []
+        self.prefill_seq_lens: List[int] = []
+        self.context_lens: List[int] = []
+        self.block_tables: List[List[int]] = []
+        self.curr_seq_lens: List[int] = []
+        self.input_positions: List[int] = []
+        self.multimodal_placeholder_maps: Dict[
+            str,
+            MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
+        self.num_prefills = 0
+        self.num_prefill_tokens = 0
+        self.num_decode_tokens = 0
+        self.has_prefix_cache_hit = False
+
+    def _add_seq_group(
+            self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
+            chunked_prefill_enabled: bool, prefix_cache_hit: bool):
+        """Add a sequence group to the metadata. Specifically update/append
+        1. context length.
+        2. block table.
+        3. slot mapping.
+        """
+        is_prompt = inter_data.is_prompt
+        block_tables = inter_data.block_tables
+
+        for (seq_id, token_len, seq_len, curr_seq_len, query_len, context_len,
+             curr_sliding_window_block, input_positions) in zip(
+                 inter_data.seq_ids, [len(t) for t in inter_data.input_tokens],
+                 inter_data.orig_seq_lens, inter_data.seq_lens,
+                 inter_data.query_lens, inter_data.context_lens,
+                 inter_data.curr_sliding_window_blocks,
+                 inter_data.input_positions):
+            self.input_positions.extend(input_positions)
+            self.context_lens.append(context_len)
+            if is_prompt:
+                mm_maps = inter_data.multi_modal_placeholder_maps
+                if mm_maps:
+                    for modality, placeholders in mm_maps.items():
+                        self.multimodal_placeholder_maps[modality].extend(
+                            placeholders)
+
+                self.num_prefills += 1
+                self.num_prefill_tokens += token_len
+                self.prefill_seq_lens.append(seq_len)
+            else:
+                self.num_decode_tokens += query_len
+                self.curr_seq_lens.append(curr_seq_len)
+
+            # Compute block table.
+            # TODO(sang): Combine chunked prefill and prefix caching by
+            # only allowing multiple of block_size chunk size.
+            # NOTE: This only works for oooooooxxx style attention.
+            block_table = []
+            if prefix_cache_hit:
+                # NOTE(woosuk): For flash-attn, the block table should
+                # include the entries for the incoming prefill tokens.
+                block_table = block_tables[seq_id]
+            elif ((chunked_prefill_enabled or not is_prompt)
+                  and block_tables is not None):
+                if curr_sliding_window_block == 0:
+                    block_table = block_tables[seq_id]
+                else:
+                    block_table = block_tables[seq_id][
+                        -curr_sliding_window_block:]
+            self.block_tables.append(block_table)
+
+            # Compute slot mapping.
+            is_profile_run = is_block_tables_empty(block_tables)
+            start_idx = compute_slot_mapping_start_idx(is_prompt, query_len,
+                                                       context_len,
+                                                       self.sliding_window)
+            compute_slot_mapping(is_profile_run, self.slot_mapping, seq_id,
+                                 seq_len, context_len, start_idx,
+                                 self.block_size, inter_data.block_tables)
+
+    def _get_graph_runner_block_tables(
+            self, num_seqs: int,
+            block_tables: List[List[int]]) -> torch.Tensor:
+        # The shape of graph_block_tables is
+        # [max batch size, max context len // block size].
+        max_batch_size, max_blocks = self.runner.graph_block_tables.shape
+        assert max_batch_size >= num_seqs
+
+        graph_block_tables = self.runner.graph_block_tables[:num_seqs]
+        for i, block_table in enumerate(block_tables):
+            if block_table:
+                num_blocks = len(block_table)
+                if num_blocks <= max_blocks:
+                    graph_block_tables[i, :num_blocks] = block_table
+                else:
+                    # It may be possible to have more blocks allocated due
+                    # to lookahead slots of multi-step, however, they are
+                    # not used anyway, so can be safely ignored.
+                    graph_block_tables[
+                        i, :max_blocks] = block_table[:max_blocks]
+
+        return torch.from_numpy(graph_block_tables).to(
+            device=self.runner.device, non_blocking=True)
+
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int):
+        """Build attention metadata with on-device tensors.
+
+        Args:
+            seq_lens: The maybe padded sequence lengths of the input sequences.
+            query_lens: The query lengths of the input sequences.
+            cuda_graph_pad_size: The padding size for cuda graph.
+                                 -1 if cuda graph is not used.
+            batch_size: The maybe padded batch size.
+        """
+        prefix_cache_hit = any([
+            inter_data.prefix_cache_hit
+            for inter_data in self.input_builder.inter_data_list
+        ])
+        for inter_data in self.input_builder.inter_data_list:
+            self._add_seq_group(inter_data,
+                                self.input_builder.chunked_prefill_enabled,
+                                prefix_cache_hit)
+
+        device = self.runner.device
+        use_captured_graph = cuda_graph_pad_size != -1
+
+        max_query_len = max(query_lens)
+        decode_query_lens = query_lens[self.num_prefills:]
+        if len(decode_query_lens) > 0:
+            max_decode_query_len = max(decode_query_lens)
+        else:
+            max_decode_query_len = 1
+        max_prefill_seq_len = max(self.prefill_seq_lens, default=0)
+        max_decode_seq_len = max(self.curr_seq_lens, default=0)
+        num_decode_tokens = self.num_decode_tokens
+        query_start_loc = list(accumulate(query_lens, initial=0))
+        seq_start_loc = list(accumulate(seq_lens, initial=0))
+
+        num_seqs = len(seq_lens)
+        if use_captured_graph:
+            self.slot_mapping.extend([PAD_SLOT_ID] * cuda_graph_pad_size)
+            self.block_tables.extend([] * cuda_graph_pad_size)
+            num_decode_tokens = batch_size - self.num_prefill_tokens
+            block_tables = self._get_graph_runner_block_tables(
+                num_seqs, self.block_tables)
+        else:
+            block_tables = make_tensor_with_pad(
+                self.block_tables,
+                pad=0,
+                dtype=torch.int,
+                device=device,
+            )
+        assert max_query_len > 0, ("query_lens: {}".format(query_lens))
+
+        assert device is not None
+        context_lens_tensor = async_tensor_h2d(self.context_lens, torch.int,
+                                               device, self.runner.pin_memory)
+        seq_lens_tensor = async_tensor_h2d(seq_lens, torch.int, device,
+                                           self.runner.pin_memory)
+        input_positions = async_tensor_h2d(self.input_positions, torch.long,
+                                           device, self.runner.pin_memory)
+        slot_mapping_tensor = async_tensor_h2d(self.slot_mapping, torch.long,
+                                               device, self.runner.pin_memory)
+        query_start_loc_tensor = async_tensor_h2d(query_start_loc, torch.int32,
+                                                  device,
+                                                  self.runner.pin_memory)
+        seq_start_loc_tensor = async_tensor_h2d(seq_start_loc, torch.int32,
+                                                device, self.runner.pin_memory)
+        placeholder_index_maps = {
+            modality: placeholder_map.index_map()
+            for modality, placeholder_map in
+            self.multimodal_placeholder_maps.items()
+        }
+
+        return TritonMLAMetadata(
+            num_prefills=self.num_prefills,
+            slot_mapping=slot_mapping_tensor,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=num_decode_tokens,
+            seq_lens=seq_lens,
+            multi_modal_placeholder_index_maps=placeholder_index_maps,
+            enable_kv_scales_calculation=True,
+            input_positions=input_positions,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=max_query_len,
+            max_decode_query_len=max_decode_query_len,
+            max_prefill_seq_len=max_prefill_seq_len,
+            max_decode_seq_len=max_decode_seq_len,
+            query_start_loc=query_start_loc_tensor,
+            seq_start_loc=seq_start_loc_tensor,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=use_captured_graph,
+            num_kv_splits=4,  # TODO(lucas) add heuristic
+            head_dim=self.runner.model_config.get_head_size(),
+        )
+
+
+class TritonMLAImpl(MLACommonImpl[TritonMLAMetadata]):
+
+    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]],
+            logits_soft_cap: Optional[float],
+            attn_type: str,
+            # MLA Specific Arguments
+            **kwargs) -> None:
+        super().__init__(num_heads, head_size, scale, num_kv_heads,
+                         alibi_slopes, sliding_window, kv_cache_dtype,
+                         blocksparse_params, logits_soft_cap, attn_type,
+                         **kwargs)
+
+        unsupported_features = [
+            alibi_slopes, sliding_window, blocksparse_params, logits_soft_cap
+        ]
+        if any(unsupported_features):
+            raise NotImplementedError(
+                "TritonMLAImpl does not support one of the following: "
+                "alibi_slopes, sliding_window, blocksparse_params, "
+                "logits_soft_cap")
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "TritonMLAImpl")
+
+    def _forward_prefill(
+        self,
+        q: torch.Tensor,
+        kv_c_normed: torch.Tensor,
+        k_pe: torch.Tensor,
+        attn_metadata: TritonMLAMetadata,
+    ) -> torch.Tensor:
+        assert isinstance(attn_metadata, TritonMLAMetadata)
+        return self._forward_prefill_flash(q, kv_c_normed, k_pe,
+                                           attn_metadata.seq_start_loc,
+                                           attn_metadata.max_prefill_seq_len)
+
+    def _forward_decode(
+        self,
+        q_nope: torch.Tensor,
+        q_pe: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: TritonMLAMetadata,
+    ) -> torch.Tensor:
+        assert kv_c_and_k_pe_cache.numel() > 0
+        if self.kv_cache_dtype.startswith("fp8"):
+            raise NotImplementedError("FP8 Triton MLA not yet supported")
+
+        decode_meta = attn_metadata.decode_metadata
+        assert decode_meta is not None
+        B = q_nope.shape[0]
+
+        q = torch.cat([q_nope, q_pe], dim=-1)
+        o = torch.zeros(B,
+                        self.num_heads,
+                        self.kv_lora_rank,
+                        dtype=q.dtype,
+                        device=q.device)
+
+        # TODO(lucas) Allocate ahead of time
+        attn_logits = torch.empty(
+            (
+                B,
+                self.num_heads,
+                attn_metadata.num_kv_splits,
+                # NOTE(lucas) idk why the +1 is here but sglang has it so we
+                # just mirror that
+                self.kv_lora_rank + 1,
+            ),
+            dtype=torch.float32,
+            device=q.device,
+        )
+
+        # Add a head dim of 1
+        kv_c_and_k_pe_cache = kv_c_and_k_pe_cache.unsqueeze(2)
+        kv_c_cache = kv_c_and_k_pe_cache[..., :self.kv_lora_rank]
+        PAGE_SIZE = kv_c_and_k_pe_cache.size(1)
+
+        # Run MQA
+        decode_attention_fwd(q, kv_c_and_k_pe_cache, kv_c_cache, o,
+                             decode_meta.block_tables,
+                             decode_meta.seq_lens_tensor, attn_logits,
+                             attn_metadata.num_kv_splits, self.scale,
+                             PAGE_SIZE)
+
+        return self._v_up_proj_and_o_proj(o)

.venv/lib/python3.11/site-packages/vllm/attention/backends/utils.py

@@ -0,0 +1,582 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Attention backend utils"""
+from collections import defaultdict
+from contextlib import contextmanager
+from itertools import accumulate
+from typing import TYPE_CHECKING, Any, Dict, List, Tuple, Type, TypeVar, Union
+
+import numpy as np
+import torch
+
+from vllm.attention import (AttentionMetadata, AttentionMetadataBuilder,
+                            AttentionState)
+from vllm.attention.backends.abstract import AttentionType
+from vllm.multimodal import MultiModalPlaceholderMap
+from vllm.utils import async_tensor_h2d, make_tensor_with_pad
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner_base import ModelRunnerBase
+
+# Error string(s) for encoder/decoder
+# unsupported attention scenarios
+STR_NOT_IMPL_ENC_DEC_ROCM_HIP = ("ROCm/HIP is not currently supported "
+                                 "with encoder/decoder models.")
+
+PAD_SLOT_ID = -1
+
+# Switch to numpy implementation of compute_slot_mapping
+# if we have at least this many elements. Could be tuned further.
+_COMPUTE_SLOT_MAPPING_NUMPY_NUMEL = 256
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import ModelInputForGPUBuilder
+
+
+def is_block_tables_empty(block_tables: Union[None, Dict]):
+    """
+    Check if block_tables is None or a dictionary with all None values.
+    """
+    if block_tables is None:
+        return True
+    return (isinstance(block_tables, dict)
+            and all(value is None for value in block_tables.values()))
+
+
+def compute_slot_mapping_start_idx(is_prompt: bool, query_len: int,
+                                   context_len: int, sliding_window: int):
+    """
+    Compute the start index of slot mapping.
+    """
+    start_idx = 0
+    if is_prompt and sliding_window is not None:
+        start_idx = max(0, query_len - sliding_window)
+    return start_idx
+
+
+def _compute_slot_mapping_python(slot_mapping: List[int],
+                                 block_table: List[int], range_start: int,
+                                 range_end: int, block_size: int):
+    for i in range(range_start, range_end):
+        block_number = block_table[i // block_size]
+        block_offset = i % block_size
+        slot = block_number * block_size + block_offset
+        slot_mapping.append(slot)
+
+
+def _compute_slot_mapping_numpy(slot_mapping: List[int],
+                                block_table: List[int], range_start: int,
+                                range_end: int, block_size: int):
+    block_table_array = np.array(block_table)
+    idx = np.arange(range_start, range_end)
+    block_offset = idx % block_size
+    idx //= block_size
+    seq_slot_mapping_array = block_table_array[idx]
+    seq_slot_mapping_array *= block_size
+    seq_slot_mapping_array += block_offset
+    slot_mapping.extend(seq_slot_mapping_array)
+
+
+def compute_slot_mapping(is_profile_run: bool, slot_mapping: List[int],
+                         seq_id: int, seq_len: int, context_len: int,
+                         start_idx: int, block_size: int,
+                         block_tables: Dict[int, List[int]]):
+    """
+    Compute slot mapping.
+    """
+    if is_profile_run:
+        # During memory profiling, the block tables are not
+        # initialized yet. In this case, we just use a dummy
+        # slot mapping.
+        # In embeddings, the block tables are {seq_id: None}.
+        slot_mapping.extend([PAD_SLOT_ID] * seq_len)
+        return
+
+    # Mask the [0, start_idx) tokens of the prompt with
+    # PAD_SLOT_ID, where start_idx is max(0, seq_len -
+    # sliding_window). For example, if the prompt len is 10,
+    # sliding window is 8, and block size is 4, the first two
+    # tokens are masked and the slot mapping will be
+    # [-1, -1, 2, 3, 4, 5, 6, 7, 0, 1].
+    padding_mask_len = max(0, start_idx - context_len)
+    slot_mapping.extend([PAD_SLOT_ID] * padding_mask_len)
+
+    range_start = max(start_idx, context_len)
+    range_end = seq_len
+    numel = range_end - range_start
+    block_table = block_tables[seq_id]
+
+    # numpy implementation will be faster than python if we have
+    # many elements, otherwise it will be slower.
+    if numel < _COMPUTE_SLOT_MAPPING_NUMPY_NUMEL:
+        _compute_slot_mapping_python(slot_mapping, block_table, range_start,
+                                     range_end, block_size)
+    else:
+        _compute_slot_mapping_numpy(slot_mapping, block_table, range_start,
+                                    range_end, block_size)
+
+
+TAttentionMetadata = TypeVar("TAttentionMetadata", bound='AttentionMetadata')
+
+
+class CommonMetadataBuilder(AttentionMetadataBuilder[TAttentionMetadata]):
+
+    _metadata_cls: Type[TAttentionMetadata]
+
+    def __init__(self, input_builder: "ModelInputForGPUBuilder"):
+        self.input_builder = input_builder
+        self.runner = input_builder.runner
+
+        self.sliding_window = input_builder.sliding_window
+        self.block_size = input_builder.block_size
+
+    def prepare(self):
+        self.slot_mapping: List[int] = []
+        self.prefill_seq_lens: List[int] = []
+        self.context_lens: List[int] = []
+        self.block_tables: List[List[int]] = []
+        self.curr_seq_lens: List[int] = []
+        self.multimodal_placeholder_maps: Dict[
+            str,
+            MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
+        self.num_prefills = 0
+        self.num_prefill_tokens = 0
+        self.num_decode_tokens = 0
+
+    def _add_seq_group(
+            self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
+            chunked_prefill_enabled: bool):
+        is_prompt = inter_data.is_prompt
+        block_tables = inter_data.block_tables
+
+        for (seq_id, token_len, seq_len, curr_seq_len, query_len, context_len,
+             curr_sliding_window_block) in zip(
+                 inter_data.seq_ids, [len(t) for t in inter_data.input_tokens],
+                 inter_data.orig_seq_lens, inter_data.seq_lens,
+                 inter_data.query_lens, inter_data.context_lens,
+                 inter_data.curr_sliding_window_blocks):
+            self.context_lens.append(context_len)
+            if is_prompt:
+                mm_maps = inter_data.multi_modal_placeholder_maps
+                if mm_maps:
+                    for modality, placeholders in mm_maps.items():
+                        self.multimodal_placeholder_maps[modality].extend(
+                            placeholders)
+
+                self.num_prefills += 1
+                self.num_prefill_tokens += token_len
+                self.prefill_seq_lens.append(seq_len)
+            else:
+                assert query_len == 1, (
+                    "seq_len: {}, context_len: {}, query_len: {}".format(
+                        seq_len, context_len, query_len))
+                self.num_decode_tokens += query_len
+                self.curr_seq_lens.append(curr_seq_len)
+
+            # Compute block table.
+            # TODO(sang): Combine chunked prefill and prefix caching by
+            # only allowing multiple of block_size chunk size.
+            # NOTE: This only works for oooooooxxx style attention.
+            block_table = []
+            if inter_data.prefix_cache_hit:
+                block_table = block_tables[seq_id]
+            elif ((chunked_prefill_enabled or not is_prompt)
+                  and block_tables is not None):
+                if curr_sliding_window_block == 0:
+                    block_table = block_tables[seq_id]
+                else:
+                    block_table = block_tables[seq_id][
+                        -curr_sliding_window_block:]
+            self.block_tables.append(block_table)
+
+            # Compute slot mapping.
+            is_profile_run = is_block_tables_empty(block_tables)
+            start_idx = compute_slot_mapping_start_idx(is_prompt, query_len,
+                                                       context_len,
+                                                       self.sliding_window)
+            compute_slot_mapping(is_profile_run, self.slot_mapping, seq_id,
+                                 seq_len, context_len, start_idx,
+                                 self.block_size, inter_data.block_tables)
+
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int):
+        """Build attention metadata with on-device tensors.
+
+        Args:
+            seq_lens: The maybe padded sequence lengths of the input sequences.
+            query_lens: The query lengths of the input sequences.
+            cuda_graph_pad_size: The padding size for cuda graph.
+                                 -1 if cuda graph is not used.
+            batch_size: The maybe padded batch size.
+        """
+        for inter_data in self.input_builder.inter_data_list:
+            self._add_seq_group(inter_data,
+                                self.input_builder.chunked_prefill_enabled)
+
+        device = self.runner.device
+        use_captured_graph = cuda_graph_pad_size != -1
+
+        max_query_len = max(query_lens)
+        max_prefill_seq_len = max(self.prefill_seq_lens, default=0)
+        max_decode_seq_len = max(self.curr_seq_lens, default=0)
+        num_decode_tokens = self.num_decode_tokens
+        query_start_loc = list(accumulate(query_lens, initial=0))
+        seq_start_loc = list(accumulate(seq_lens, initial=0))
+
+        if use_captured_graph:
+            self.slot_mapping.extend([PAD_SLOT_ID] * cuda_graph_pad_size)
+            self.block_tables.extend([] * cuda_graph_pad_size)
+            num_decode_tokens = batch_size
+
+            # The shape of graph_block_tables is
+            # [max batch size, max context len // block size].
+            input_block_tables = self.runner.graph_block_tables[:batch_size]
+            for i, block_table in enumerate(self.block_tables):
+                if block_table:
+                    input_block_tables[i, :len(block_table)] = block_table
+            block_tables = torch.from_numpy(input_block_tables).to(
+                device, non_blocking=True)
+        else:
+            block_tables = make_tensor_with_pad(
+                self.block_tables,
+                pad=0,
+                dtype=torch.int,
+                device=device,
+            )
+        assert max_query_len > 0, "query_lens: {}".format(query_lens)
+
+        assert device is not None
+        context_lens_tensor = async_tensor_h2d(self.context_lens, torch.int,
+                                               device, self.runner.pin_memory)
+        seq_lens_tensor = async_tensor_h2d(seq_lens, torch.int, device,
+                                           self.runner.pin_memory)
+        slot_mapping_tensor = async_tensor_h2d(self.slot_mapping, torch.long,
+                                               device, self.runner.pin_memory)
+        query_start_loc_tensor = async_tensor_h2d(query_start_loc, torch.int32,
+                                                  device,
+                                                  self.runner.pin_memory)
+        seq_start_loc_tensor = async_tensor_h2d(seq_start_loc, torch.int32,
+                                                device, self.runner.pin_memory)
+        placeholder_index_maps = {
+            modality: placeholder_map.index_map()
+            for modality, placeholder_map in
+            self.multimodal_placeholder_maps.items()
+        }
+
+        return self._metadata_cls(  # type: ignore
+            num_prefills=self.num_prefills,
+            slot_mapping=slot_mapping_tensor,
+            multi_modal_placeholder_index_maps=placeholder_index_maps,
+            enable_kv_scales_calculation=True,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=num_decode_tokens,
+            seq_lens=seq_lens,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=max_query_len,
+            max_prefill_seq_len=max_prefill_seq_len,
+            max_decode_seq_len=max_decode_seq_len,
+            query_start_loc=query_start_loc_tensor,
+            seq_start_loc=seq_start_loc_tensor,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=use_captured_graph,
+        )
+
+
+class CommonAttentionState(AttentionState):
+
+    def __init__(self, runner: "ModelRunnerBase"):
+        self.runner = runner
+        self._is_graph_capturing = False
+
+    @contextmanager
+    def graph_capture(self, max_batch_size: int):
+
+        self._is_graph_capturing = True
+
+        self._graph_slot_mapping = torch.full((max_batch_size, ),
+                                              PAD_SLOT_ID,
+                                              dtype=torch.long,
+                                              device=self.runner.device)
+        self._graph_seq_lens = torch.ones(max_batch_size,
+                                          dtype=torch.int32,
+                                          device=self.runner.device)
+        self._graph_block_tables = torch.from_numpy(
+            self.runner.graph_block_tables).to(device=self.runner.device)
+
+        yield
+
+        self._is_graph_capturing = False
+        del self._graph_slot_mapping
+        del self._graph_seq_lens
+        del self._graph_block_tables
+
+    def graph_clone(self, batch_size: int) -> "CommonAttentionState":
+        assert self._is_graph_capturing
+        return self.__class__(self.runner)
+
+    def graph_capture_get_metadata_for_batch(
+            self, batch_size: int, is_encoder_decoder_model: bool = False):
+        assert self._is_graph_capturing
+        attn_metadata = self.runner.attn_backend.make_metadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=batch_size,
+            slot_mapping=self._graph_slot_mapping[:batch_size],
+            multi_modal_placeholder_index_maps=None,
+            enable_kv_scales_calculation=True,
+            seq_lens=None,
+            seq_lens_tensor=self._graph_seq_lens[:batch_size],
+            max_query_len=1,
+            max_decode_query_len=1,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.runner.max_seq_len_to_capture,
+            query_start_loc=None,
+            seq_start_loc=None,
+            context_lens_tensor=None,
+            block_tables=self._graph_block_tables[:batch_size],
+            use_cuda_graph=True,
+        )
+        if is_encoder_decoder_model:
+            # The encoder decoder model works only with XFormers and
+            # Flash Attention backend. Assert the same.
+            assert self.runner.attn_backend.get_name() in\
+                ["XFORMERS", "FLASH_ATTN"], \
+                f"Expected attn_backend name to be either 'XFORMERS' or " \
+                f"'FLASH_ATTN', but "\
+                f"got '{self.runner.attn_backend.get_name()}'"
+            self._update_captured_metadata_for_enc_dec_model(
+                batch_size=batch_size, attn_metadata=attn_metadata)
+
+        return attn_metadata
+
+    def get_graph_input_buffers(
+            self,
+            attn_metadata,
+            is_encoder_decoder_model: bool = False) -> Dict[str, Any]:
+        input_buffers = {
+            "slot_mapping": attn_metadata.slot_mapping,
+            "seq_lens_tensor": attn_metadata.decode_metadata.seq_lens_tensor,
+            "block_tables": attn_metadata.decode_metadata.block_tables,
+        }
+        if is_encoder_decoder_model:
+            # The encoder decoder model works only with XFormers and
+            # Flash Attention backend. Assert the same.
+            assert self.runner.attn_backend.get_name() in\
+                ["XFORMERS", "FLASH_ATTN"], \
+                f"Expected attn_backend name to be either 'XFORMERS' or "\
+                f"'FLASH_ATTN', but "\
+                f"got '{self.runner.attn_backend.get_name()}'"
+            self._add_additonal_input_buffers_for_enc_dec_model(
+                attn_metadata=attn_metadata, input_buffers=input_buffers)
+        return input_buffers
+
+    def prepare_graph_input_buffers(
+            self,
+            input_buffers,
+            attn_metadata,
+            is_encoder_decoder_model: bool = False) -> None:
+        input_buffers["seq_lens_tensor"].copy_(
+            attn_metadata.decode_metadata.seq_lens_tensor, non_blocking=True)
+        input_buffers["block_tables"].copy_(
+            attn_metadata.decode_metadata.block_tables, non_blocking=True)
+        if is_encoder_decoder_model:
+            # The encoder decoder model works only with XFormers and
+            # Flash Attention backend. Assert the same.
+            assert self.runner.attn_backend.get_name() in\
+                ["XFORMERS", "FLASH_ATTN"], \
+                f"Expected attn_backend name to be either 'XFORMERS' or "\
+                f"'FLASH_ATTN', but "\
+                f"got '{self.runner.attn_backend.get_name()}'"
+            self._prepare_input_buffers_for_enc_dec_model(
+                attn_metadata, input_buffers)
+
+    def begin_forward(self, model_input) -> None:
+        return
+
+    def _update_captured_metadata_for_enc_dec_model(self, batch_size: int,
+                                                    attn_metadata):
+        """
+        Updates the attention metadata parameters for CUDA graph capture in an
+        encoder-decoder model.
+
+        This method modifies attention-related tensors and metadata required
+        for CUDA graph capture in encoder-decoder models. Specifically, it
+        updates the cross-attention and encoder sequence tensors in the 
+        AttentionMetadata object.
+        """
+        # During decode phase the cross_slot_mapping will be empty. Hence set
+        # an empty tensor for CUDA Graph capture.
+        attn_metadata.cross_slot_mapping = torch.tensor(
+            [], dtype=torch.int).cuda()
+        attn_metadata.cross_block_tables = torch.full(
+            (batch_size, self.runner.get_max_block_per_batch()),
+            1,
+            dtype=torch.int).cuda()
+        attn_metadata.encoder_seq_lens = torch.full((batch_size, ),
+                                                    1,
+                                                    dtype=torch.int).cuda()
+        attn_metadata.encoder_seq_lens_tensor = torch.full(
+            (batch_size, ), 1, dtype=torch.int).cuda()
+        attn_metadata.max_encoder_seq_len = self.runner.max_seq_len_to_capture
+        attn_metadata.num_encoder_tokens = 0
+
+    def _add_additonal_input_buffers_for_enc_dec_model(
+            self, attn_metadata, input_buffers: Dict[str, Any]):
+        """
+        Saves additional input buffers specific to the encoder-decoder model
+        from the attention metadata.
+
+        This method extracts and stores encoder-decoder related input buffers
+        from the `attn_metadata` into the `input_buffers` dictionary. The
+        buffers include encoder sequence lengths, cross-slot mappings, and
+        cross-block tables, which are essential for the encoder-decoder model
+        during CUDA graph replay.
+        """
+        input_buffers["encoder_seq_lens_tensor"] = (
+            attn_metadata.decode_metadata.encoder_seq_lens_tensor)
+        input_buffers["cross_slot_mapping"] = (
+            attn_metadata.decode_metadata.cross_slot_mapping)
+        input_buffers["cross_block_tables"] = (
+            attn_metadata.decode_metadata.cross_block_tables)
+
+    def _prepare_input_buffers_for_enc_dec_model(self, attn_metadata,
+                                                 input_buffers: Dict[str,
+                                                                     Any]):
+        """
+        Populates input buffers with data from the encoder-decoder model's
+        attention metadata.
+
+        This method fills the input buffers with encoder-decoder specific
+        tensors. It copies data from the `attn_metadata` and keyword arguments
+        (`kwargs`) into corresponding buffers in the `input_buffers` dictionary.
+        The copied data includes attention-related metadata as well as input 
+        IDs and positional information for the encoder.
+        """
+        input_buffers["encoder_seq_lens_tensor"].copy_(
+            attn_metadata.decode_metadata.encoder_seq_lens_tensor,
+            non_blocking=True)
+        input_buffers["cross_slot_mapping"].copy_(
+            attn_metadata.decode_metadata.cross_slot_mapping,
+            non_blocking=True)
+        input_buffers["cross_block_tables"].copy_(
+            attn_metadata.decode_metadata.cross_block_tables,
+            non_blocking=True)
+
+
+def is_all_encoder_attn_metadata_set(attn_metadata):
+    '''
+    All attention metadata required for encoder attention is set.
+    '''
+    return ((attn_metadata.encoder_seq_lens is not None)
+            and (attn_metadata.encoder_seq_lens_tensor is not None)
+            and (attn_metadata.max_encoder_seq_len is not None))
+
+
+def is_all_cross_attn_metadata_set(attn_metadata):
+    '''
+    All attention metadata required for enc/dec cross-attention is set.
+
+    Superset of encoder attention required metadata.
+    '''
+    return (attn_metadata.is_all_encoder_attn_metadata_set
+            and (attn_metadata.cross_slot_mapping is not None)
+            and (attn_metadata.cross_block_tables is not None))
+
+
+def get_seq_len_block_table_args(
+    attn_metadata,
+    is_prompt: bool,
+    attn_type: str,
+) -> tuple:
+    '''
+    The particular choice of sequence-length- and block-table-related
+    attributes which should be extracted from attn_metadata is dependent
+    on the type of attention operation.
+
+    Decoder attn -> select entirely decoder self-attention-related fields
+    Encoder/decoder cross-attn -> select encoder sequence lengths & 
+                                  cross-attn block-tables fields
+    Encoder attn -> select encoder sequence lengths fields & no block tables
+    
+    Arguments:
+
+    * attn_metadata: Attention metadata structure associated with attention op
+    * is_prompt: True if prefill, False otherwise
+    * attn_type: encoder attention, decoder self-attention,
+                 encoder/decoder cross-attention
+
+    Returns:
+
+    * Appropriate sequence-lengths tensor
+    * Appropriate max sequence-length scalar
+    * Appropriate block tables (or None)
+    '''
+
+    if attn_type == AttentionType.DECODER:
+        # Decoder self-attention
+        # Choose max_seq_len based on whether we are in prompt_run
+        if is_prompt:
+            max_seq_len = attn_metadata.max_prefill_seq_len
+        else:
+            max_seq_len = attn_metadata.max_decode_seq_len
+        return (attn_metadata.seq_lens_tensor, max_seq_len,
+                attn_metadata.block_tables)
+    elif attn_type == AttentionType.ENCODER_DECODER:
+        # Enc/dec cross-attention KVs match encoder sequence length;
+        # cross-attention utilizes special "cross" block tables
+        return (attn_metadata.encoder_seq_lens_tensor,
+                attn_metadata.max_encoder_seq_len,
+                attn_metadata.cross_block_tables)
+    elif attn_type == AttentionType.ENCODER:
+        # No block tables associated with encoder attention
+        return (attn_metadata.encoder_seq_lens_tensor,
+                attn_metadata.max_encoder_seq_len, None)
+    else:
+        raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+
+def get_num_prefill_decode_query_kv_tokens(
+    attn_metadata,
+    attn_type: str,
+) -> Tuple[int, int, int]:
+    """
+    Calculate the number of prefill and decode tokens for query, key/value
+    based on the attention metadata and the specified attention type.
+
+    Args:
+        attn_metadata (FlashAttentionMetadata): Attention Metadata object.
+        attn_type (AttentionType): The type of attention being used.
+    Returns:
+        Tuple[int, int, int]: A tuple containing three integers:
+            - The number of prefill query tokens.
+            - The number of prefill key/value tokens.
+            - The number of decode query tokens.
+
+    Raises:
+        AssertionError: If the number of encoder tokens in `attn_metadata` 
+        is `None` when required for the calculations.
+    """
+    num_prefill_query_tokens = 0
+    num_decode_query_tokens = 0
+    num_prefill_kv_tokens = 0
+    if attn_type == AttentionType.ENCODER:
+        # Encoder attention is only invoked during prefill phase.
+        # The same input servers a both query and key.
+        assert attn_metadata.num_encoder_tokens is not None
+        num_prefill_query_tokens = attn_metadata.num_encoder_tokens
+        num_prefill_kv_tokens = attn_metadata.num_encoder_tokens
+        num_decode_query_tokens = 0
+    elif attn_type == AttentionType.ENCODER_DECODER:
+        assert attn_metadata.num_encoder_tokens is not None
+        num_prefill_query_tokens = attn_metadata.num_prefill_tokens
+        # The key is the encoder/cross-attention.
+        num_prefill_kv_tokens = attn_metadata.num_encoder_tokens
+        num_decode_query_tokens = attn_metadata.num_decode_tokens
+    else:  # attn_type == AttentionType.DECODER or
+        # attn_type == AttentionType.ENCODER_ONLY
+        num_prefill_query_tokens = attn_metadata.num_prefill_tokens
+        num_prefill_kv_tokens = attn_metadata.num_prefill_tokens
+        num_decode_query_tokens = attn_metadata.num_decode_tokens
+
+    return (num_prefill_query_tokens, num_prefill_kv_tokens,
+            num_decode_query_tokens)

.venv/lib/python3.11/site-packages/vllm/attention/backends/xformers.py

@@ -0,0 +1,794 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Attention layer with xFormers and PagedAttention."""
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+from xformers import ops as xops
+from xformers.ops.fmha.attn_bias import (AttentionBias,
+                                         BlockDiagonalCausalMask,
+                                         BlockDiagonalMask,
+                                         LowerTriangularMaskWithTensorBias)
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionLayer,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import (
+    CommonAttentionState, CommonMetadataBuilder,
+    get_num_prefill_decode_query_kv_tokens, get_seq_len_block_table_args,
+    is_all_cross_attn_metadata_set, is_all_encoder_attn_metadata_set)
+from vllm.attention.ops.paged_attn import (PagedAttention,
+                                           PagedAttentionMetadata)
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class XFormersBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "XFORMERS"
+
+    @staticmethod
+    def get_impl_cls() -> Type["XFormersImpl"]:
+        return XFormersImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return XFormersMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["XFormersMetadataBuilder"]:
+        return XFormersMetadataBuilder
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class XFormersMetadata(AttentionMetadata, PagedAttentionMetadata):
+    """Metadata for XFormersbackend.
+
+    NOTE: Any python object stored here is not updated when it is
+    cuda-graph replayed. If you have values that need to be changed
+    dynamically, it should be stored in tensor. The tensor has to be
+    updated from `CUDAGraphRunner.forward` API.
+    """
+
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ----------------------|
+    #                                   |-- query_len ---|
+
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+
+    # FIXME: It is for flash attn.
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Maximum sequence length among decode batch. 0 if there are prefill
+    # requests only.
+    max_decode_seq_len: int
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+    use_cuda_graph: bool
+
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]] = None
+
+    # FIXME: It is for flash attn.
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor] = None
+
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor] = None
+
+    # Maximum query length in the batch. None for decoding.
+    max_query_len: Optional[int] = None
+
+    # Max number of query tokens among request in the batch.
+    max_decode_query_len: Optional[int] = None
+
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    query_start_loc: Optional[torch.Tensor] = None
+
+    # Self-attention prefill/decode metadata cache
+    _cached_prefill_metadata: Optional["XFormersMetadata"] = None
+    _cached_decode_metadata: Optional["XFormersMetadata"] = None
+
+    # Begin encoder attn & enc/dec cross-attn fields...
+
+    # Encoder sequence lengths representation
+    encoder_seq_lens: Optional[List[int]] = None
+    encoder_seq_lens_tensor: Optional[torch.Tensor] = None
+    # FIXME: It is for flash attn.
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    encoder_seq_start_loc: Optional[torch.Tensor] = None
+
+    # Maximum sequence length among encoder sequences
+    max_encoder_seq_len: Optional[int] = None
+
+    # Number of tokens input to encoder
+    num_encoder_tokens: Optional[int] = None
+
+    # Cross-attention memory-mapping data structures: slot mapping
+    # and block tables
+    cross_slot_mapping: Optional[torch.Tensor] = None
+    cross_block_tables: Optional[torch.Tensor] = None
+
+    def __post_init__(self):
+        # Set during the execution of the first attention op.
+        # It is a list because it is needed to set per prompt
+        # when alibi slopes is used. It is because of the limitation
+        # from xformer API.
+        # will not appear in the __repr__ and __init__
+        self.attn_bias: Optional[List[AttentionBias]] = None
+        self.encoder_attn_bias: Optional[List[AttentionBias]] = None
+        self.cross_attn_bias: Optional[List[AttentionBias]] = None
+
+    @property
+    def is_all_encoder_attn_metadata_set(self):
+        '''
+        All attention metadata required for encoder attention is set.
+        '''
+        return is_all_encoder_attn_metadata_set(self)
+
+    @property
+    def is_all_cross_attn_metadata_set(self):
+        '''
+        All attention metadata required for enc/dec cross-attention is set.
+
+        Superset of encoder attention required metadata.
+        '''
+        return is_all_cross_attn_metadata_set(self)
+
+    @property
+    def prefill_metadata(self) -> Optional["XFormersMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            # Recover cached prefill-phase attention
+            # metadata structure
+            return self._cached_prefill_metadata
+
+        assert ((self.seq_lens is not None)
+                or (self.encoder_seq_lens is not None))
+        assert ((self.seq_lens_tensor is not None)
+                or (self.encoder_seq_lens_tensor is not None))
+
+        # Compute some attn_metadata fields which default to None
+        query_start_loc = (None if self.query_start_loc is None else
+                           self.query_start_loc[:self.num_prefills + 1])
+        seq_start_loc = (None if self.seq_start_loc is None else
+                         self.seq_start_loc[:self.num_prefills + 1])
+        slot_mapping = (None if self.slot_mapping is None else
+                        self.slot_mapping[:self.num_prefill_tokens])
+        seq_lens = (None if self.seq_lens is None else
+                    self.seq_lens[:self.num_prefills])
+        seq_lens_tensor = (None if self.seq_lens_tensor is None else
+                           self.seq_lens_tensor[:self.num_prefills])
+        context_lens_tensor = (None if self.context_lens_tensor is None else
+                               self.context_lens_tensor[:self.num_prefills])
+        block_tables = (None if self.block_tables is None else
+                        self.block_tables[:self.num_prefills])
+
+        # Construct & cache prefill-phase attention metadata structure
+        self._cached_prefill_metadata = XFormersMetadata(
+            num_prefills=self.num_prefills,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=0,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=self.
+            multi_modal_placeholder_index_maps,
+            enable_kv_scales_calculation=self.enable_kv_scales_calculation,
+            seq_lens=seq_lens,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=self.max_prefill_seq_len,
+            max_decode_seq_len=0,
+            query_start_loc=query_start_loc,
+            seq_start_loc=seq_start_loc,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=False,
+            # Begin encoder & cross attn fields below...
+            encoder_seq_lens=self.encoder_seq_lens,
+            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
+            max_encoder_seq_len=self.max_encoder_seq_len,
+            cross_slot_mapping=self.cross_slot_mapping,
+            cross_block_tables=self.cross_block_tables)
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> Optional["XFormersMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            # Recover cached decode-phase attention
+            # metadata structure
+            return self._cached_decode_metadata
+        assert ((self.seq_lens_tensor is not None)
+                or (self.encoder_seq_lens_tensor is not None))
+
+        # Compute some attn_metadata fields which default to None
+        slot_mapping = (None if self.slot_mapping is None else
+                        self.slot_mapping[self.num_prefill_tokens:])
+        seq_lens_tensor = (None if self.seq_lens_tensor is None else
+                           self.seq_lens_tensor[self.num_prefills:])
+        block_tables = (None if self.block_tables is None else
+                        self.block_tables[self.num_prefills:])
+
+        # Construct & cache decode-phase attention metadata structure
+        self._cached_decode_metadata = XFormersMetadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=self.num_decode_tokens,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=None,
+            enable_kv_scales_calculation=True,
+            seq_lens_tensor=seq_lens_tensor,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.max_decode_seq_len,
+            block_tables=block_tables,
+            use_cuda_graph=self.use_cuda_graph,
+            # Begin encoder & cross attn fields below...
+            encoder_seq_lens=self.encoder_seq_lens,
+            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
+            max_encoder_seq_len=self.max_encoder_seq_len,
+            cross_slot_mapping=self.cross_slot_mapping,
+            cross_block_tables=self.cross_block_tables)
+
+        # Batch may be composed of prefill|decodes, adjust query start indices
+        # to refer to the start of decodes when the two are split apart.
+        # E.g. in tokens:[3 prefills|6 decodes], query_start_loc=[3,9] => [0,6].
+        if self._cached_decode_metadata.query_start_loc is not None:
+            qs = self._cached_decode_metadata.query_start_loc
+            self._cached_decode_metadata.query_start_loc = qs - qs[0]
+        return self._cached_decode_metadata
+
+
+def _get_attn_bias(
+    attn_metadata: XFormersMetadata,
+    attn_type: str,
+) -> Optional[AttentionBias]:
+    '''
+    Extract appropriate attention bias from attention metadata
+    according to attention type.
+
+    Arguments:
+
+    * attn_metadata: Attention metadata structure associated with attention
+    * attn_type: encoder attention, decoder self-attention,
+                 encoder/decoder cross-attention
+
+    Returns:
+    * Appropriate attention bias value given the attention type
+    '''
+
+    if (attn_type == AttentionType.DECODER
+            or attn_type == AttentionType.ENCODER_ONLY):
+        return attn_metadata.attn_bias
+    elif attn_type == AttentionType.ENCODER:
+        return attn_metadata.encoder_attn_bias
+    elif attn_type == AttentionType.ENCODER_DECODER:
+        return attn_metadata.cross_attn_bias
+    else:
+        raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+
+def _set_attn_bias(
+    attn_metadata: XFormersMetadata,
+    attn_bias: List[Optional[AttentionBias]],
+    attn_type: str,
+) -> None:
+    '''
+    Update appropriate attention bias field of attention metadata,
+    according to attention type.
+
+    Arguments:
+
+    * attn_metadata: Attention metadata structure associated with attention
+    * attn_bias: The desired attention bias value
+    * attn_type: encoder attention, decoder self-attention,
+                 encoder/decoder cross-attention
+    '''
+
+    if (attn_type == AttentionType.DECODER
+            or attn_type == AttentionType.ENCODER_ONLY):
+        attn_metadata.attn_bias = attn_bias
+    elif attn_type == AttentionType.ENCODER:
+        attn_metadata.encoder_attn_bias = attn_bias
+    elif attn_type == AttentionType.ENCODER_DECODER:
+        attn_metadata.cross_attn_bias = attn_bias
+    else:
+        raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+
+class XFormersMetadataBuilder(CommonMetadataBuilder[XFormersMetadata]):
+
+    _metadata_cls = XFormersMetadata
+
+
+class XFormersImpl(AttentionImpl[XFormersMetadata]):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prefill_tokens ----------------->|	
+    |<--prefill_0-->|<--prefill_1-->|...|<--prefill_N-1--->|
+
+    Otherwise, the layout is as follows:	
+    |<----------------- num_decode_tokens ------------------>|	
+    |<--decode_0-->|..........|<--decode_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+
+    If chunked prefill is enabled, prefill tokens and decode tokens can be
+    batched together in a flattened 1D query.
+
+    |<----- num_prefill_tokens ---->|<------- num_decode_tokens --------->|
+    |<-prefill_0->|...|<-prefill_N-1->|<--decode_0-->|...|<--decode_M-1-->|
+
+    Currently, cuda graph is disabled for chunked prefill, meaning there's no
+    padding between prefill and decode tokens.
+    """
+
+    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: str = AttentionType.DECODER,
+    ) -> None:
+        if blocksparse_params is not None:
+            raise ValueError(
+                "XFormers does not support block-sparse attention.")
+        if logits_soft_cap is not None:
+            logger.warning_once("XFormers does not support logits soft cap. "
+                                "Outputs may be slightly off.")
+        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
+        self.sliding_window = sliding_window
+        self.kv_cache_dtype = kv_cache_dtype
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        suppored_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in suppored_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {suppored_head_sizes}.")
+
+        self.attn_type = attn_type
+
+    def forward(
+        self,
+        layer: AttentionLayer,
+        query: torch.Tensor,
+        key: Optional[torch.Tensor],
+        value: Optional[torch.Tensor],
+        kv_cache: torch.Tensor,
+        attn_metadata: "XFormersMetadata",
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Forward pass with xFormers and PagedAttention.
+
+        For decoder-only models: query, key and value must be non-None.
+
+        For encoder/decoder models:
+        * XFormersImpl.forward() may be invoked for both self- and cross-
+          attention layers.
+        * For self-attention: query, key and value must be non-None.
+        * For cross-attention:
+            * Query must be non-None
+            * During prefill, key and value must be non-None; key and value
+              get cached for use during decode.
+            * During decode, key and value may be None, since:
+              (1) key and value tensors were cached during prefill, and
+              (2) cross-attention key and value tensors do not grow during
+                  decode
+        
+        A note on how the attn_type (attention type enum) argument impacts
+        attention forward() behavior:
+    
+            * DECODER: normal decoder-only behavior;
+                use decoder self-attention block table
+            * ENCODER: no KV caching; pass encoder sequence
+                attributes (encoder_seq_lens/encoder_seq_lens_tensor/
+                max_encoder_seq_len) to kernel, in lieu of decoder
+                sequence attributes (seq_lens/seq_lens_tensor/max_seq_len).
+                Used for encoder branch of encoder-decoder models.
+            * ENCODER_ONLY: no kv_caching, uses the normal attention 
+                attributes (seq_lens/seq_lens_tensor/max_seq_len).
+            * ENCODER_DECODER: cross-attention behavior;
+                use cross-attention block table for caching KVs derived
+                from encoder hidden states; since KV sequence lengths
+                will match encoder sequence lengths, pass encoder sequence
+                attributes to kernel (encoder_seq_lens/encoder_seq_lens_tensor/
+                max_encoder_seq_len)
+    
+        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]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+            attn_type: Select attention type, between encoder attention,
+                       decoder self-attention, or encoder/decoder cross-
+                       attention. Defaults to decoder self-attention,
+                       which is the vLLM default generally
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        attn_type = self.attn_type
+        # Check that appropriate attention metadata attributes are
+        # selected for the desired attention type
+        if (attn_type == AttentionType.ENCODER
+                and (not attn_metadata.is_all_encoder_attn_metadata_set)):
+            raise AttributeError("Encoder attention requires setting "
+                                 "encoder metadata attributes.")
+
+        elif (attn_type == AttentionType.ENCODER_DECODER
+              and (not attn_metadata.is_all_cross_attn_metadata_set)):
+            raise AttributeError("Encoder/decoder cross-attention "
+                                 "requires setting cross-attention "
+                                 "metadata attributes.")
+
+        query = query.view(-1, self.num_heads, self.head_size)
+        if key is not None:
+            assert value is not None
+            key = key.view(-1, self.num_kv_heads, self.head_size)
+            value = value.view(-1, self.num_kv_heads, self.head_size)
+        else:
+            assert value is None
+
+        # Self-attention vs. cross-attention will impact
+        # which KV cache memory-mapping & which
+        # seqlen datastructures we utilize
+
+        if (attn_type != AttentionType.ENCODER and kv_cache.numel() > 0):
+            # KV-cache during decoder-self- or
+            # encoder-decoder-cross-attention, but not
+            # during encoder attention.
+            #
+            # Even if there are no new key/value pairs to cache,
+            # we still need to break out key_cache and value_cache
+            # i.e. for later use by paged attention
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            if (key is not None) and (value is not None):
+
+                if attn_type == AttentionType.ENCODER_DECODER:
+                    # Update cross-attention KV cache (prefill-only)
+                    # During cross-attention decode, key & value will be None,
+                    # preventing this IF-statement branch from running
+                    updated_slot_mapping = attn_metadata.cross_slot_mapping
+                else:
+                    # Update self-attention KV cache (prefill/decode)
+                    updated_slot_mapping = attn_metadata.slot_mapping
+
+                # Reshape the input keys and values and store them in the cache.
+                # If kv_cache is not provided, the new key and value tensors are
+                # not cached. This happens during the initial memory
+                # profiling run.
+                PagedAttention.write_to_paged_cache(
+                    key, value, key_cache, value_cache, updated_slot_mapping,
+                    self.kv_cache_dtype, layer._k_scale, layer._v_scale)
+        (num_prefill_query_tokens, num_prefill_kv_tokens,
+        num_decode_query_tokens) = \
+            get_num_prefill_decode_query_kv_tokens(attn_metadata, attn_type)
+
+        output = torch.empty_like(query)
+        # Query for decode. KV is not needed because it is already cached.
+        decode_query = query[num_prefill_query_tokens:]
+        # QKV for prefill.
+        query = query[:num_prefill_query_tokens]
+        if key is not None and value is not None:
+            key = key[:num_prefill_kv_tokens]
+            value = value[:num_prefill_kv_tokens]
+
+        assert query.shape[0] == num_prefill_query_tokens
+        assert decode_query.shape[0] == num_decode_query_tokens
+
+        if prefill_meta := attn_metadata.prefill_metadata:
+            # Prompt run.
+            if kv_cache.numel() == 0 or prefill_meta.block_tables.numel() == 0:
+                # normal attention.
+                # block tables are empty if the prompt does not have a cached
+                # prefix.
+                out = self._run_memory_efficient_xformers_forward(
+                    query, key, value, prefill_meta, attn_type=attn_type)
+                assert out.shape == output[:num_prefill_query_tokens].shape
+                output[:num_prefill_query_tokens] = out
+            else:
+                assert attn_type != AttentionType.ENCODER_ONLY, (
+                    "Encoder-only models should not have prefix attention.")
+
+                assert prefill_meta.query_start_loc is not None
+                assert prefill_meta.max_query_len is not None
+
+                # prefix-enabled attention
+                # TODO(Hai) this triton kernel has regression issue (broke) to
+                # deal with different data types between KV and FP8 KV cache,
+                # to be addressed separately.
+                out = PagedAttention.forward_prefix(
+                    query,
+                    key,
+                    value,
+                    self.kv_cache_dtype,
+                    key_cache,
+                    value_cache,
+                    prefill_meta.block_tables,
+                    prefill_meta.query_start_loc,
+                    prefill_meta.seq_lens_tensor,
+                    prefill_meta.context_lens_tensor,
+                    prefill_meta.max_query_len,
+                    self.alibi_slopes,
+                    self.sliding_window,
+                    layer._k_scale,
+                    layer._v_scale,
+                )
+                assert output[:num_prefill_query_tokens].shape == out.shape
+                output[:num_prefill_query_tokens] = out
+
+        if decode_meta := attn_metadata.decode_metadata:
+            assert attn_type != AttentionType.ENCODER_ONLY, (
+                "Encoder-only models should not have decode metadata.")
+
+            (
+                seq_lens_arg,
+                max_seq_len_arg,
+                block_tables_arg,
+            ) = get_seq_len_block_table_args(decode_meta, False, attn_type)
+
+            output[num_prefill_query_tokens:] = PagedAttention.forward_decode(
+                decode_query,
+                key_cache,
+                value_cache,
+                block_tables_arg,
+                seq_lens_arg,
+                max_seq_len_arg,
+                self.kv_cache_dtype,
+                self.num_kv_heads,
+                self.scale,
+                self.alibi_slopes,
+                layer._k_scale,
+                layer._v_scale,
+            )
+
+        # Reshape the output tensor.
+        return output.view(-1, self.num_heads * self.head_size)
+
+    def _run_memory_efficient_xformers_forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: XFormersMetadata,
+        attn_type: str = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        """Attention for 1D query of multiple prompts. Multiple prompt
+        tokens are flattened in to `query` input.
+
+        See https://facebookresearch.github.io/xformers/components/ops.html
+        for API spec.
+
+        Args:
+            output: shape = [num_prefill_tokens, num_heads, head_size]
+            query: shape = [num_prefill_tokens, num_heads, head_size]
+            key: shape = [num_prefill_tokens, num_kv_heads, head_size]
+            value: shape = [num_prefill_tokens, num_kv_heads, head_size]
+            attn_metadata: Metadata for attention.
+            attn_type: Select attention type, between encoder attention,
+                       decoder self-attention, or encoder/decoder cross-
+                       attention. Defaults to decoder self-attention,
+                       which is the vLLM default generally
+        """
+
+        original_query = query
+        if self.num_kv_heads != self.num_heads:
+            # GQA/MQA requires the shape [B, M, G, H, K].
+            # Note that the output also has the same shape (which is different
+            # from a spec from the doc).
+            query = query.view(query.shape[0], self.num_kv_heads,
+                               self.num_queries_per_kv, query.shape[-1])
+            key = key[:, :,
+                      None, :].expand(key.shape[0], self.num_kv_heads,
+                                      self.num_queries_per_kv, key.shape[-1])
+            value = value[:, :,
+                          None, :].expand(value.shape[0], self.num_kv_heads,
+                                          self.num_queries_per_kv,
+                                          value.shape[-1])
+
+        # Set attention bias if not provided. This typically happens at
+        # the very attention layer of every iteration.
+        # FIXME(woosuk): This is a hack.
+        attn_bias = _get_attn_bias(attn_metadata, attn_type)
+        if attn_bias is None:
+            if self.alibi_slopes is None:
+
+                # Cross attention block of decoder branch of encoder-decoder
+                # model uses seq_lens for dec / encoder_seq_lens for enc
+                if (attn_type == AttentionType.ENCODER_DECODER):
+                    assert attn_metadata.seq_lens is not None
+                    assert attn_metadata.encoder_seq_lens is not None
+
+                    # Cross-attention mask is non-causal
+                    attn_bias = BlockDiagonalMask.from_seqlens(
+                        attn_metadata.seq_lens, attn_metadata.encoder_seq_lens)
+
+                # Encoder branch of encoder-decoder model uses
+                # attn_metadata.encoder_seq_lens
+                elif attn_type == AttentionType.ENCODER:
+
+                    assert attn_metadata.encoder_seq_lens is not None
+
+                    # Encoder self-attention mask is non-causal
+                    attn_bias = BlockDiagonalMask.from_seqlens(
+                        attn_metadata.encoder_seq_lens)
+
+                # Self-attention block of encoder-only model just
+                # uses the seq_lens directly.
+                elif attn_type == AttentionType.ENCODER_ONLY:
+                    assert attn_metadata.seq_lens is not None
+
+                    # Encoder self-attention mask is non-causal
+                    attn_bias = BlockDiagonalMask.from_seqlens(
+                        attn_metadata.seq_lens)
+
+                # Self-attention block of decoder branch just
+                # uses the seq_lens directly
+                elif attn_type == AttentionType.DECODER:
+                    assert attn_metadata.seq_lens is not None
+
+                    # Decoder self-attention mask is causal
+                    attn_bias = BlockDiagonalCausalMask.from_seqlens(
+                        attn_metadata.seq_lens)
+                else:
+                    raise ValueError("Unknown AttentionType: %s", attn_type)
+
+                if self.sliding_window is not None:
+                    attn_bias = attn_bias.make_local_attention(
+                        self.sliding_window)
+                attn_bias = [attn_bias]
+            else:
+                assert attn_type == AttentionType.DECODER
+                assert attn_metadata.seq_lens is not None
+                attn_bias = _make_alibi_bias(self.alibi_slopes,
+                                             self.num_kv_heads, query.dtype,
+                                             attn_metadata.seq_lens)
+
+            _set_attn_bias(attn_metadata, attn_bias, attn_type)
+
+        # No alibi slopes.
+        # TODO(woosuk): Too many view operations. Let's try to reduce
+        # them in the future for code readability.
+        if self.alibi_slopes is None:
+            # Add the batch dimension.
+            query = query.unsqueeze(0)
+            key = key.unsqueeze(0)
+            value = value.unsqueeze(0)
+            out = xops.memory_efficient_attention_forward(
+                query,
+                key,
+                value,
+                attn_bias=attn_bias[0],
+                p=0.0,
+                scale=self.scale)
+            return out.view_as(original_query)
+
+        # Attention with alibi slopes.
+        # FIXME(woosuk): Because xformers does not support dynamic sequence
+        # lengths with custom attention bias, we process each prompt one by
+        # one. This is inefficient, especially when we have many short prompts.
+        assert attn_metadata.seq_lens is not None
+        output = torch.empty_like(original_query)
+        start = 0
+        for i, seq_len in enumerate(attn_metadata.seq_lens):
+            end = start + seq_len
+            out = xops.memory_efficient_attention_forward(
+                query[None, start:end],
+                key[None, start:end],
+                value[None, start:end],
+                attn_bias=attn_bias[i],
+                p=0.0,
+                scale=self.scale)
+            # TODO(woosuk): Unnecessary copy. Optimize.
+            output[start:end].copy_(out.view_as(original_query[start:end]))
+            start += seq_len
+        return output
+
+
+def _make_alibi_bias(
+    alibi_slopes: torch.Tensor,
+    num_kv_heads: int,
+    dtype: torch.dtype,
+    seq_lens: List[int],
+) -> List[AttentionBias]:
+    attn_biases: List[AttentionBias] = []
+    for seq_len in seq_lens:
+        bias = torch.arange(seq_len, dtype=dtype)
+        # NOTE(zhuohan): HF uses
+        #     `bias = bias[None, :].repeat(seq_len, 1)`
+        # here. We find that both biases give the same results, but
+        # the bias below more accurately follows the original ALiBi
+        # paper.
+        # Calculate a matrix where each element represents ith element- jth
+        # element.
+        bias = bias[None, :] - bias[:, None]
+
+        padded_len = (seq_len + 7) // 8 * 8
+        num_heads = alibi_slopes.shape[0]
+        bias = torch.empty(
+            1,  # batch size
+            num_heads,
+            seq_len,
+            padded_len,
+            device=alibi_slopes.device,
+            dtype=dtype,
+        )[:, :, :, :seq_len].copy_(bias)
+        bias.mul_(alibi_slopes[:, None, None])
+        if num_heads != num_kv_heads:
+            bias = bias.unflatten(1, (num_kv_heads, num_heads // num_kv_heads))
+        attn_biases.append(LowerTriangularMaskWithTensorBias(bias))
+
+    return attn_biases