vllm.v1.attention.backends.registry ¶

class AttentionBackendEnum(Enum, metaclass=_AttentionBackendEnumMeta):
    """Enumeration of all supported attention backends.

    The enum value is the default class path, but this can be overridden
    at runtime using register_backend().

    To get the actual backend class (respecting overrides), use:
        backend.get_class()
    """

    FLASH_ATTN = "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend"
    FLASH_ATTN_DIFFKV = (
        "vllm.v1.attention.backends.flash_attn_diffkv.FlashAttentionDiffKVBackend"
    )
    TRITON_ATTN = "vllm.v1.attention.backends.triton_attn.TritonAttentionBackend"
    ROCM_ATTN = "vllm.v1.attention.backends.rocm_attn.RocmAttentionBackend"
    ROCM_AITER_MLA = "vllm.v1.attention.backends.mla.rocm_aiter_mla.AiterMLABackend"
    ROCM_AITER_TRITON_MLA = (
        "vllm.v1.attention.backends.mla.aiter_triton_mla.AiterTritonMLABackend"
    )
    ROCM_AITER_FA = (
        "vllm.v1.attention.backends.rocm_aiter_fa.AiterFlashAttentionBackend"
    )
    ROCM_AITER_MLA_SPARSE = (
        "vllm.v1.attention.backends.mla.rocm_aiter_mla_sparse.ROCMAiterMLASparseBackend"
    )
    XPU_MLA_SPARSE = "vllm.v1.attention.backends.mla.xpu_mla_sparse.XPUMLASparseBackend"
    TORCH_SDPA = ""  # this tag is only used for ViT
    FLASHINFER = "vllm.v1.attention.backends.flashinfer.FlashInferBackend"
    FLASHINFER_MLA = (
        "vllm.v1.attention.backends.mla.flashinfer_mla.FlashInferMLABackend"
    )
    TOKENSPEED_MLA = (
        "vllm.v1.attention.backends.mla.tokenspeed_mla.TokenspeedMLABackend"
    )
    FLASHINFER_MLA_SPARSE = (
        "vllm.v1.attention.backends.mla.flashinfer_mla_sparse."
        "FlashInferMLASparseBackend"
    )
    TRITON_MLA = "vllm.v1.attention.backends.mla.triton_mla.TritonMLABackend"
    CUTLASS_MLA = "vllm.v1.attention.backends.mla.cutlass_mla.CutlassMLABackend"
    FLASHMLA = "vllm.v1.attention.backends.mla.flashmla.FlashMLABackend"
    FLASHMLA_SPARSE = (
        "vllm.v1.attention.backends.mla.flashmla_sparse.FlashMLASparseBackend"
    )
    # DeepSeek V4 sparse MLA backends (model-driven; selected via the V4 layer).
    FLASHMLA_SPARSE_DSV4 = (
        "vllm.models.deepseek_v4.sparse_mla.DeepseekV4FlashMLABackend"
    )
    FLASHINFER_MLA_SPARSE_DSV4 = (
        "vllm.models.deepseek_v4.nvidia.flashinfer_sparse."
        "DeepseekV4FlashInferMLASparseBackend"
    )
    ROCM_FLASHMLA_SPARSE_DSV4 = (
        "vllm.models.deepseek_v4.amd.rocm.DeepseekV4ROCMAiterMLASparseBackend"
    )
    FLASH_ATTN_MLA = "vllm.v1.attention.backends.mla.flashattn_mla.FlashAttnMLABackend"
    NO_ATTENTION = "vllm.v1.attention.backends.no_attention.NoAttentionBackend"
    FLEX_ATTENTION = "vllm.v1.attention.backends.flex_attention.FlexAttentionBackend"
    ROCM_AITER_UNIFIED_ATTN = (
        "vllm.v1.attention.backends.rocm_aiter_unified_attn."
        "RocmAiterUnifiedAttentionBackend"
    )
    CPU_ATTN = "vllm.v1.attention.backends.cpu_attn.CPUAttentionBackend"
    TURBOQUANT = "vllm.v1.attention.backends.turboquant_attn.TurboQuantAttentionBackend"
    # Placeholder for third-party/custom backends - must be registered before use
    # set to None to avoid alias with other backend, whose value is an empty string
    CUSTOM = None

    def get_path(self, include_classname: bool = True) -> str:
        """Get the class path for this backend (respects overrides).

        Returns:
            The fully qualified class path string

        Raises:
            ValueError: If Backend.CUSTOM is used without being registered
        """
        path = _ATTN_OVERRIDES.get(self, self.value)
        if not path:
            raise ValueError(
                f"Backend {self.name} must be registered before use. "
                f"Use register_backend(Backend.{self.name}, 'your.module.YourClass')"
            )
        if not include_classname:
            path = path.rsplit(".", 1)[0]
        return path

    def get_class(self) -> "type[AttentionBackend]":
        """Get the backend class (respects overrides).

        Returns:
            The backend class

        Raises:
            ImportError: If the backend class cannot be imported
            ValueError: If Backend.CUSTOM is used without being registered
        """
        return resolve_obj_by_qualname(self.get_path())

    def is_overridden(self) -> bool:
        """Check if this backend has been overridden.

        Returns:
            True if the backend has a registered override
        """
        return self in _ATTN_OVERRIDES

    def clear_override(self) -> None:
        """Clear any override for this backend, reverting to the default."""
        _ATTN_OVERRIDES.pop(self, None)

class MambaAttentionBackendEnum(Enum, metaclass=_AttentionBackendEnumMeta):
    """Enumeration of all supported mamba attention backends.

    The enum value is the default class path, but this can be overridden
    at runtime using register_backend().

    To get the actual backend class (respecting overrides), use:
        backend.get_class()
    """

    MAMBA1 = "vllm.v1.attention.backends.mamba1_attn.Mamba1AttentionBackend"
    MAMBA2 = "vllm.v1.attention.backends.mamba2_attn.Mamba2AttentionBackend"
    SHORT_CONV = "vllm.v1.attention.backends.short_conv_attn.ShortConvAttentionBackend"
    LINEAR = "vllm.v1.attention.backends.linear_attn.LinearAttentionBackend"
    GDN_ATTN = "vllm.v1.attention.backends.gdn_attn.GDNAttentionBackend"
    # Placeholder for third-party/custom backends - must be registered before use
    # set to None to avoid alias with other backend, whose value is an empty string
    CUSTOM = None

    def get_path(self, include_classname: bool = True) -> str:
        """Get the class path for this backend (respects overrides).

        Returns:
            The fully qualified class path string

        Raises:
            ValueError: If Backend.CUSTOM is used without being registered
        """
        path = _MAMBA_ATTN_OVERRIDES.get(self, self.value)
        if not path:
            raise ValueError(
                f"Backend {self.name} must be registered before use. "
                f"Use register_backend(Backend.{self.name}, 'your.module.YourClass')"
            )
        if not include_classname:
            path = path.rsplit(".", 1)[0]
        return path

    def get_class(self) -> "type[AttentionBackend]":
        """Get the backend class (respects overrides).

        Returns:
            The backend class

        Raises:
            ImportError: If the backend class cannot be imported
            ValueError: If Backend.CUSTOM is used without being registered
        """
        return resolve_obj_by_qualname(self.get_path())

    def is_overridden(self) -> bool:
        """Check if this backend has been overridden.

        Returns:
            True if the backend has a registered override
        """
        return self in _MAMBA_ATTN_OVERRIDES

    def clear_override(self) -> None:
        """Clear any override for this backend, reverting to the default."""
        _MAMBA_ATTN_OVERRIDES.pop(self, None)

class _AttentionBackendEnumMeta(EnumMeta):
    """Metaclass for AttentionBackendEnum to provide better error messages."""

    def __getitem__(cls, name: str):
        """Get backend by name with helpful error messages."""
        try:
            return super().__getitem__(name)
        except KeyError:
            members = cast("dict[str, Enum]", cls.__members__).keys()
            valid_backends = ", ".join(members)
            raise ValueError(
                f"Unknown attention backend: '{name}'. "
                f"Valid options are: {valid_backends}"
            ) from None