RomanNum3ral
/
ComfyUI
mirror of https://github.com/Comfy-Org/ComfyUI.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Fallback to cudnn attention on linux if flash attention doesn't work. (#15146)

This commit is contained in:
comfyanonymous 2026-07-29 18:10:28 -07:00 committed by GitHub
parent 4f874c5e3a
commit f73e8cde88
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
comfy/ops.py
View File

@ -41,7 +41,7 @@ def scaled_dot_product_attention(q, k, v, *args, **kwargs):
try:
if torch.cuda.is_available() and comfy.model_management.WINDOWS:
if torch.cuda.is_available():
from torch.nn.attention import SDPBackend, sdpa_kernel
import inspect
if "set_priority" in inspect.signature(sdpa_kernel).parameters:
@ -51,7 +51,10 @@ try:
SDPBackend.MATH,
]
SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
if comfy.model_management.WINDOWS:
SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
else:
SDPA_BACKEND_PRIORITY.insert(1, SDPBackend.CUDNN_ATTENTION)
def scaled_dot_product_attention(q, k, v, *args, **kwargs):
if q.nelement() < 1024 * 128: # arbitrary number, for small inputs cudnn attention seems slower