vllm.v1.attention.backends.flex_attention

class FlexAttentionBackend(AttentionBackend):
    accept_output_buffer: bool = True

    @classmethod
    def get_supported_dtypes(cls) -> list[torch.dtype]:
        return [torch.float16, torch.bfloat16, torch.float32]

    @classmethod
    def validate_head_size(cls, head_size: int) -> None:
        return  # FlexAttention supports any head size

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

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

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

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

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

    @staticmethod
    def use_cascade_attention(*args, **kwargs) -> bool:
        return False

class FlexAttentionImpl(AttentionImpl):
    sliding_window: Optional[tuple[int, int]]
    alibi_slopes: Optional[torch.Tensor]
    logits_soft_cap: Optional[float]

    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,
        logits_soft_cap: Optional[float] = None,
        attn_type: AttentionType = AttentionType.DECODER,
        kv_sharing_target_layer_name: Optional[str] = None,
        **kwargs,
    ) -> None:
        self.num_heads = num_heads
        self.head_size = head_size
        self.scale = float(scale)
        self.num_kv_heads = num_kv_heads
        self.attn_type = attn_type

        if attn_type not in (AttentionType.ENCODER_ONLY,
                             AttentionType.DECODER):
            raise NotImplementedError(
                f"FlexAttention does not support {attn_type} attention")

        if alibi_slopes is not None:
            raise NotImplementedError(
                "FlexAttention does not support alibi slopes yet.")
        else:
            self.alibi_slopes = None
        if sliding_window is not None:
            raise NotImplementedError(
                "FlexAttention does not support sliding window yet.")
        else:
            self.sliding_window = (-1, -1)
        self.kv_cache_dtype = kv_cache_dtype
        self.logits_soft_cap = logits_soft_cap
        if self.logits_soft_cap is not None:
            raise NotImplementedError(
                "FlexAttention does not support logits soft cap yet.")

        assert self.num_heads % self.num_kv_heads == 0
        self.num_queries_per_kv = self.num_heads // self.num_kv_heads

        if kv_sharing_target_layer_name is not None:
            raise NotImplementedError(
                "FlexAttention does not support kv sharing yet.")

        FlexAttentionBackend.validate_head_size(head_size)
        if is_quantized_kv_cache(self.kv_cache_dtype):
            raise NotImplementedError(
                "FlexAttention does not support quantized kv-cache. Yet")

    @staticmethod
    def view_as_4d(tensor: torch.Tensor) -> torch.Tensor:
        """View a 3d tensor as 4D."""
        if tensor.ndim == 4:
            return tensor
        assert tensor.ndim == 3
        return tensor[None, :, :, :]

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

        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]
        """
        assert output is not None, "Output tensor must be provided."
        if output_scale is not None or output_block_scale is not None:
            raise NotImplementedError(
                "fused output quantization is not yet supported"
                " for FlexAttentionImpl")

        enable_gqa = self.num_kv_heads != self.num_heads

        if attn_metadata is None:
            # Profiling run.
            return output
            # query = self.view_as_4d(query).permute(0, 2, 1, 3)
            # return torch.empty_like(query)

        num_actual_tokens = attn_metadata.num_actual_tokens

        if not attn_metadata.causal:
            assert self.attn_type == AttentionType.ENCODER_ONLY

            query, key_tensor, value_tensor = map(
                lambda x: self.view_as_4d(x).permute(0, 2, 1, 3),
                (query, key, value),
            )

        else:
            assert self.attn_type == AttentionType.DECODER
            key_cache, value_cache = kv_cache.unbind(0)

            torch.ops._C_cache_ops.reshape_and_cache_flash(
                key,
                value,
                key_cache,
                value_cache,
                attn_metadata.slot_mapping,
                self.kv_cache_dtype,
                layer._k_scale,
                layer._v_scale,
            )

            # View out the block_size dim
            key_cache = key_cache.view(-1, self.num_kv_heads, self.head_size)
            value_cache = value_cache.view(-1, self.num_kv_heads,
                                           self.head_size)
            query, key_tensor, value_tensor = map(
                lambda x: self.view_as_4d(x).permute(0, 2, 1, 3),
                (query, key_cache, value_cache),
            )

        query = query[:, :, :num_actual_tokens, :]
        # Doesn't work for now -> constraint violation
        # torch._dynamo.try_mark_dynamic(query, 2)

        assert attn_metadata.block_mask is not None
        block_m, block_n = attn_metadata.block_mask.BLOCK_SIZE

        kernel_options = get_kernel_options(query, block_m, block_n,
                                            attn_metadata.direct_build)
        out = flex_attention_compiled(
            query,
            key_tensor,
            value_tensor,
            attn_metadata.transformed_score_mod,
            attn_metadata.block_mask,
            self.scale,
            enable_gqa=enable_gqa,
            kernel_options=kernel_options,
        )

        # Flex doesn't have an out variant today, rely on epilogue fusion
        out = out.permute(0, 2, 1, 3).squeeze(0)
        output[:num_actual_tokens, :, :].copy_(out)
        return output

@dataclass
class FlexAttentionMetadata:
    causal: bool
    num_actual_tokens: int  # Number of tokens excluding padding.
    max_query_len: int
    query_start_loc: torch.Tensor
    max_seq_len: int
    seq_lens: torch.Tensor
    block_table: torch.Tensor
    slot_mapping: torch.Tensor

    use_cascade: bool
    common_prefix_len: int
    cu_prefix_query_lens: Optional[torch.Tensor]
    prefix_kv_lens: Optional[torch.Tensor]
    suffix_kv_lens: Optional[torch.Tensor]

    # Block info
    total_cache_tokens: int
    block_size: int
    max_possible_sequence_length: int
    num_reqs: int
    physical_to_logical: torch.Tensor
    decode_offset: torch.Tensor
    num_blocks_per_seq: torch.Tensor

    # For logging.
    num_input_tokens: int = 0  # Number of tokens including padding.

    # Flex Metadata
    num_blocks = 0
    block_mask: Optional[BlockMask] = None
    score_mod: Optional[_score_mod_signature] = None
    logical_mask_mod: _mask_mod_signature = causal_mask_mod
    doc_ids: Optional[torch.Tensor] = None
    direct_build: bool = True
    q_block_size: int = 16
    kv_block_size: int = 16
    transformed_score_mod: Optional[_score_mod_signature] = None

    def _convert_physical_to_logical(
        self,
        request_lookup: torch.Tensor,
        q_idx: torch.Tensor,
        physical_kv_idx: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Convert physical indices to logical indices for both query and kv.

        NB is_within_lower_bound: do sequences start on block_boundaries?

        Returns:
            tuple of (is_valid, logical_q_idx, logical_kv_idx)
        """
        # Map query indices to corresponding request indices
        q_req = request_lookup[q_idx]

        # Convert physical KV indices to logical indices
        physical_kv_block = physical_kv_idx // self.block_size
        physical_kv_offset = physical_kv_idx % self.block_size
        logical_block_idx = self.physical_to_logical[q_req, physical_kv_block]
        logical_kv_idx = (logical_block_idx * self.block_size +
                          physical_kv_offset)

        # Determine valid kv indices
        live_block = logical_block_idx >= 0
        within_upper_bound = logical_kv_idx < self.seq_lens[q_req]
        within_lower_bound = logical_kv_idx >= 0
        is_valid = live_block & within_upper_bound & within_lower_bound

        # Convert physical query indices to logical indices
        local_q_idx = q_idx - self.query_start_loc[q_req]
        logical_q_idx = local_q_idx + self.decode_offset[q_req]

        return is_valid, logical_q_idx, logical_kv_idx

    def get_causal_mask_mod(self) -> _mask_mod_signature:
        """Creates the mask_mod function for FlexAttention.

        This function creates the combined mask mod function that handles:
            1. The paged attention block mapping
            2. The mapping from packed query sequences to logical query entries

        It also by defaults adds the decoding offset to the query indices.
        With this info we create the "logical" indices that are passed to
        mask_mod functions. This allows mask mod functions to be agnostic to
        layout of the query and key/value tensors.
        """
        assert self.doc_ids is not None

        def final_mask_mod(
            b: torch.Tensor,
            h: torch.Tensor,
            q_idx: torch.Tensor,
            physical_kv_idx: torch.Tensor,
        ) -> torch.Tensor:
            (is_valid, logical_q_idx,
             logical_kv_idx) = self._convert_physical_to_logical(
                 self.doc_ids, q_idx, physical_kv_idx)
            # Apply mask modification only for valid indices
            return torch.where(
                is_valid,
                self.logical_mask_mod(b, h, logical_q_idx, logical_kv_idx),
                False,
            )

        return final_mask_mod

    def get_bidirectional_mask_mod(self) -> _mask_mod_signature:
        """Creates the encoder mask_mod function for FlexAttention.

        Since the encoder bidirectional attention doesn't run with
        KV cache, this function creates a mask based on the
        packed query sequences.
        """
        # Create a lookup mapping from query indices -> request number
        request_lookup = _offsets_to_doc_ids_tensor(self.query_start_loc)

        def final_mask_mod(
            b: torch.Tensor,
            h: torch.Tensor,
            q_idx: torch.Tensor,
            kv_idx: torch.Tensor,
        ) -> torch.Tensor:
            return request_lookup[q_idx] == request_lookup[kv_idx]

        return final_mask_mod

    def get_transformed_score_mod(self) -> Optional[_score_mod_signature]:
        """Creates the transformed score_mod function for FlexAttention.

        This function wraps the user's score_mod to handle physical-to-logical
        index conversion, similar to how get_mask_mod works for mask functions.
        """
        if self.score_mod is None:
            return None

        # Create a lookup mapping from query indices -> request number
        request_lookup = _offsets_to_doc_ids_tensor(self.query_start_loc)
        user_score_mod = self.score_mod

        def transformed_score_mod(
            score: torch.Tensor,
            b: torch.Tensor,
            h: torch.Tensor,
            q_idx: torch.Tensor,
            physical_kv_idx: torch.Tensor,
        ) -> torch.Tensor:
            (is_valid, logical_q_idx,
             logical_kv_idx) = self._convert_physical_to_logical(
                 request_lookup, q_idx, physical_kv_idx)

            return torch.where(
                is_valid,
                user_score_mod(score,
                               b,
                               h,
                               logical_q_idx,
                               logical_kv_idx,
                               physical_q=q_idx), -float('inf'))

        return transformed_score_mod

    def _build_block_mask_direct(self) -> BlockMask:
        """Direct block mask construction for standard causal attention.

        This method constructs the block mask directly using
        BlockMask.from_kv_blocks which is much more efficient than the
        generic create_block_mask approach.

        The direct path works as follows:
        1. For each query token, fetch blocks from block_table using max_seq_len
           (this fetches more blocks than needed for shorter sequences)
        2. Group query tokens into chunks of q_block_size
        3. For each group, deduplicate the blocks using unique_static_unsorted
        4. Create BlockMask using the deduplicated block indices

        Over-estimation occurs when a group of q_block_size tokens contains
        multiple sequence IDs (doc_ids). In this case, we fetch ALL blocks for
        each sequence represented in the group, even though individual query
        tokens may only need a subset of those blocks based on causal masking
        and their position.

        """
        page_to_block_ratio = self.kv_block_size // self.block_size
        if page_to_block_ratio != 1:
            raise ValueError(
                f"FlexAttention currently requires the cache block size "
                f"({self.block_size}) to be equal to the kv_block_size "
                f"({self.kv_block_size}). Please check your model's "
                f"configuration.")

        used_pages = self.block_table[
            self.doc_ids, :cdiv(self.max_seq_len, self.block_size)]
        used_pages_padded = pad_to_multiple(used_pages,
                                            multiple=self.q_block_size,
                                            dim=0)
        used_pages_padded = used_pages_padded.reshape(
            used_pages_padded.shape[0] // self.q_block_size, -1)
        used_pages_padded = used_pages_padded // page_to_block_ratio
        kv_indices = unique_static_unsorted((used_pages_padded.long()),
                                            M=self.num_blocks).to(torch.int32)

        kv_num_blocks = (kv_indices >= 0).sum(dim=-1).to(torch.int32)
        block_mask_kwargs = {
            "seq_lengths": (self.num_actual_tokens, self.total_cache_tokens),
            "kv_num_blocks": kv_num_blocks[None, None],
            "kv_indices": kv_indices[None, None],
            "full_kv_num_blocks": None,
            "full_kv_indices": None,
            "BLOCK_SIZE": (self.q_block_size, self.kv_block_size),
            "mask_mod": self.mask_mod,
        }

        # compute_q_blocks parameter is available in PyTorch 2.9+
        if is_torch_equal_or_newer("2.9.0.dev0"):
            block_mask_kwargs["compute_q_blocks"] = False
        return BlockMask.from_kv_blocks(**block_mask_kwargs)

    def build_block_mask(self) -> BlockMask:
        if self.causal:
            mask_mod = self.get_causal_mask_mod()
            kv_len = self.total_cache_tokens
        else:
            mask_mod = self.get_bidirectional_mask_mod()
            kv_len = self.num_actual_tokens
        return create_block_mask_compiled(
            mask_mod,
            None,
            None,
            self.num_actual_tokens,
            kv_len,
            device=self.block_table.device,
            BLOCK_SIZE=(self.q_block_size, self.kv_block_size),
        )

    def __post_init__(self):
        assert self.use_cascade is False, "Not implemented yet."
        assert self.common_prefix_len == 0, "Not implemented yet."
        assert self.cu_prefix_query_lens is None, "Not implemented yet."
        assert self.prefix_kv_lens is None, "Not implemented yet."
        assert self.suffix_kv_lens is None, "Not implemented yet."
        # Create a lookup mapping from query indices -> request number
        self.doc_ids = _offsets_to_doc_ids_tensor(self.query_start_loc)
        self.num_blocks = self.total_cache_tokens // self.block_size

        if self.causal:
            self.mask_mod = self.get_causal_mask_mod()
        else:
            self.mask_mod = self.get_bidirectional_mask_mod()

        self.transformed_score_mod = self.get_transformed_score_mod()

        if self.direct_build and self.causal:
            self.block_mask = self._build_block_mask_direct()
        else:
            self.block_mask = self.build_block_mask()