[MM Encoder] Default to use TORCH_SDPA backend for ViT on Volta/Turing GPU (#36472)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2026-03-09 03:43:44 -07:00
committed by GitHub
parent aaf5fa9abf
commit b0906d8b02
2 changed files with 30 additions and 7 deletions
+15 -7
View File
@@ -413,12 +413,20 @@ class CudaPlatformBase(Platform):
@classmethod
def get_supported_vit_attn_backends(cls) -> list["AttentionBackendEnum"]:
return [
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.TRITON_ATTN,
AttentionBackendEnum.TORCH_SDPA,
AttentionBackendEnum.FLASHINFER,
]
if cls.has_device_capability(80):
return [
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.TRITON_ATTN,
AttentionBackendEnum.TORCH_SDPA,
AttentionBackendEnum.FLASHINFER,
]
else:
return [
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.TORCH_SDPA,
AttentionBackendEnum.TRITON_ATTN,
AttentionBackendEnum.FLASHINFER,
]
@classmethod
def get_vit_attn_backend(
@@ -438,7 +446,7 @@ class CudaPlatformBase(Platform):
cc = cls.get_device_capability()
for vit_attn_backend in cls.get_supported_vit_attn_backends():
if vit_attn_backend == AttentionBackendEnum.TORCH_SDPA:
continue
return vit_attn_backend
try:
backend_class = vit_attn_backend.get_class()
is_backend_supported = backend_class.supports_head_size(