fix: use triton_attn as default vision attention on B300 (SM103) (#25570)
This commit is contained in:
@@ -951,7 +951,7 @@ class VisionAttention(nn.Module):
|
|||||||
major, minor = get_device_capability()
|
major, minor = get_device_capability()
|
||||||
if major == 9:
|
if major == 9:
|
||||||
backend = "fa3"
|
backend = "fa3"
|
||||||
elif major == 10:
|
elif major == 10 and minor != 3:
|
||||||
backend = "fa4"
|
backend = "fa4"
|
||||||
else:
|
else:
|
||||||
backend = "triton_attn"
|
backend = "triton_attn"
|
||||||
|
|||||||
Reference in New Issue
Block a user