Attention backend registry
_Backend
Bases: Enum
Source code in vllm/attention/backends/registry.py
| class _Backend(enum.Enum):
FLASH_ATTN = enum.auto()
TRITON_ATTN = enum.auto()
XFORMERS = enum.auto()
ROCM_FLASH = enum.auto()
ROCM_AITER_MLA = enum.auto()
ROCM_AITER_FA = enum.auto() # used for ViT attn backend
TORCH_SDPA = enum.auto()
FLASHINFER = enum.auto()
FLASHINFER_MLA = enum.auto()
TRITON_MLA = enum.auto()
CUTLASS_MLA = enum.auto()
FLASHMLA = enum.auto()
FLASH_ATTN_MLA = enum.auto()
PALLAS = enum.auto()
IPEX = enum.auto()
NO_ATTENTION = enum.auto()
FLEX_ATTENTION = enum.auto()
TREE_ATTN = enum.auto()
ROCM_ATTN = enum.auto()
|
CUTLASS_MLA class-attribute
instance-attribute
FLASHINFER class-attribute
instance-attribute
FLASHINFER_MLA class-attribute
instance-attribute
FLASHMLA class-attribute
instance-attribute
FLASH_ATTN class-attribute
instance-attribute
FLASH_ATTN_MLA class-attribute
instance-attribute
FLEX_ATTENTION class-attribute
instance-attribute
IPEX class-attribute
instance-attribute
NO_ATTENTION class-attribute
instance-attribute
PALLAS class-attribute
instance-attribute
ROCM_AITER_FA class-attribute
instance-attribute
ROCM_AITER_MLA class-attribute
instance-attribute
ROCM_ATTN class-attribute
instance-attribute
ROCM_FLASH class-attribute
instance-attribute
TORCH_SDPA class-attribute
instance-attribute
TREE_ATTN class-attribute
instance-attribute
TRITON_ATTN class-attribute
instance-attribute
TRITON_MLA class-attribute
instance-attribute