vllm.model_executor.layers.attention.mm_encoder_attention ¶

@CustomOp.register("mm_encoder_attn")
class MMEncoderAttention(CustomOp):
    """Multi-headed attention without any cache, used for multimodal encoder."""

    # --8<-- [end:mm_encoder_attn]

    def __init__(
        self,
        num_heads: int,
        head_size: int,
        scale: float | None = None,
        num_kv_heads: int | None = None,
        prefix: str = "",
    ) -> None:
        """
        Args:
            num_heads: number of attention heads per partition.
            head_size: hidden_size per attention head.
            scale: scale factor.
            num_kv_heads: number of kv heads.
            prefix: This has no effect, it is only here to make it easier to
                    swap between Attention and MultiHeadAttention
        """
        super().__init__()

        self.num_heads = num_heads
        self.head_size = head_size
        self.scale = scale
        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
        self.layer_name = prefix

        assert self.num_heads % self.num_kv_heads == 0, (
            f"num_heads ({self.num_heads}) is not "
            f"divisible by num_kv_heads ({self.num_kv_heads})"
        )
        self.num_queries_per_kv = self.num_heads // self.num_kv_heads

        # During model initialization, the default dtype is set as the model
        # weight and activation dtype.
        dtype = torch.get_default_dtype()

        # Get device-specific vision attention backend.
        self.attn_backend = get_vit_attn_backend(
            head_size=head_size,
            dtype=dtype,
        )

        self.is_flash_attn_backend = self.attn_backend in {
            AttentionBackendEnum.FLASH_ATTN,
            AttentionBackendEnum.ROCM_AITER_FA,
        }

        self._fa_version = (
            get_flash_attn_version(head_size=head_size)
            if self.is_flash_attn_backend
            else None
        )

        logger.info_once(f"Using {self.attn_backend} for MMEncoderAttention.")

    @classmethod
    def enabled(cls) -> bool:
        return True

    def view_qkv_to_4d(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        bsz: int,
        q_len: int,
        kv_len: int,
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        Reshape query, key, value to 4D tensors:
        (batch_size, seq_len, num_heads, head_size)
        """
        query = query.view(bsz, q_len, self.num_heads, self.head_size)
        key = key.view(bsz, kv_len, self.num_kv_heads, self.head_size)
        value = value.view(bsz, kv_len, self.num_kv_heads, self.head_size)

        return query, key, value

    def _forward_sdpa(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        cu_seqlens: torch.Tensor | None = None,
    ) -> torch.Tensor:
        """Input shape:
        (batch_size x seq_len x hidden_size) or
        (batch_size x seq_len x num_heads x head_size)
        """
        bsz, q_len = query.size()[:2]
        kv_len = key.size(1)
        is_reshaped = query.dim() != 4

        query, key, value = self.view_qkv_to_4d(query, key, value, bsz, q_len, kv_len)

        output = vit_torch_sdpa_wrapper(
            q=query,
            k=key,
            v=value,
            scale=self.scale,
            cu_seqlens=cu_seqlens,
            enable_gqa=self.num_heads > self.num_kv_heads,
        )
        if is_reshaped:
            output = output.reshape(bsz, q_len, -1)
        return output

    def _forward_fa(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        cu_seqlens: torch.Tensor | None = None,
        max_seqlen: torch.Tensor | None = None,  # Only used for Flash Attention
    ) -> torch.Tensor:
        """Input shape:
        (batch_size x seq_len x hidden_size) or
        (batch_size x seq_len x num_heads x head_size)
        """
        assert (cu_seqlens is not None and max_seqlen is not None) or (
            cu_seqlens is None and max_seqlen is None
        ), "cu_seqlens and max_seqlen should be both set or both None."

        bsz, q_len = query.size()[:2]
        kv_len = key.size(1)
        is_reshaped = query.dim() != 4

        query, key, value = self.view_qkv_to_4d(query, key, value, bsz, q_len, kv_len)

        output = vit_flash_attn_wrapper(
            q=query,
            k=key,
            v=value,
            batch_size=bsz,
            is_rocm_aiter=(self.attn_backend == AttentionBackendEnum.ROCM_AITER_FA),
            fa_version=self._fa_version,
            scale=self.scale,
            cu_seqlens=cu_seqlens,
            max_seqlen=max_seqlen,
        )
        if is_reshaped:
            output = output.reshape(bsz, q_len, -1)
        return output

    def forward_native(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        cu_seqlens: torch.Tensor | None = None,
        max_seqlen: torch.Tensor | None = None,  # Only used for Flash Attention
    ) -> torch.Tensor:
        return self._forward_sdpa(query, key, value, cu_seqlens)

    def forward_cuda(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        cu_seqlens: torch.Tensor | None = None,
        max_seqlen: torch.Tensor | None = None,  # Only used for Flash Attention
    ) -> torch.Tensor:
        if self.is_flash_attn_backend:
            return self._forward_fa(query, key, value, cu_seqlens, max_seqlen)
        elif self.attn_backend == AttentionBackendEnum.TORCH_SDPA:
            return self._forward_sdpa(query, key, value, cu_seqlens)
        else:
            raise ValueError(
                f"Unsupported multi-modal encoder attention backend for CUDA: "
                f"{self.attn_backend}."
            )

    def forward_cpu(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        cu_seqlens: torch.Tensor | None = None,
        max_seqlen: torch.Tensor | None = None,  # Only used for Flash Attention
    ) -> torch.Tensor:
        return self._forward_sdpa(query, key, value, cu_seqlens)

    def forward_xpu(
        self,
        query: torch.Tensor,
        key: torch.Tensor,
        value: torch.Tensor,
        cu_seqlens: torch.Tensor | None = None,
        max_seqlen: torch.Tensor | None = None,  # Only used for Flash Attention
    ) -> torch.Tensor:
        assert self.is_flash_attn_backend, (
            "XPU only supports FLASH_ATTN for vision attention."
        )
        return self._forward_fa(query, key, value, cu_seqlens, max_seqlen)