[V1] port xformers backend to v1 (#21342)

Signed-off-by: Giancarlo Delfin <gdelfin@meta.com>
This commit is contained in:
Giancarlo Delfin
2025-08-05 10:04:46 -07:00
committed by GitHub
parent ae87ddd040
commit 469b3ffaaa
6 changed files with 438 additions and 1 deletions
+4
View File
@@ -271,6 +271,7 @@ class CudaPlatformBase(Platform):
TRITON_ATTN_VLLM_V1 = "vllm.v1.attention.backends.triton_attn.TritonAttentionBackend" # noqa: E501
FLASH_ATTN_V1 = "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend" # noqa: E501
TREE_ATTN_V1 = "vllm.v1.attention.backends.tree_attn.TreeAttentionBackend" # noqa: E501
XFORMERS_V1 = "vllm.v1.attention.backends.xformers.XFormersAttentionBackend" # noqa: E501
if selected_backend == _Backend.FLASHINFER:
logger.info_once("Using FlashInfer backend on V1 engine.")
@@ -291,6 +292,9 @@ class CudaPlatformBase(Platform):
elif selected_backend == _Backend.TREE_ATTN:
logger.info_once("Using Tree Attention backend on V1 engine.")
return TREE_ATTN_V1
elif selected_backend == _Backend.XFORMERS_VLLM_V1:
logger.info_once("Using XFormers backend on V1 engine.")
return XFORMERS_V1
from vllm.attention.selector import is_attn_backend_supported
+1
View File
@@ -63,6 +63,7 @@ class _Backend(enum.Enum):
NO_ATTENTION = enum.auto()
FLEX_ATTENTION = enum.auto()
TREE_ATTN = enum.auto()
XFORMERS_VLLM_V1 = enum.auto()
class PlatformEnum(enum.Enum):