Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

main: force cudnn.benchmark to false (#14390)

rattus6d18f4adacea
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 9dc0a4e13..55ddaab8e 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -534,8 +534,10 @@ try:
except:
pass
-if torch.cuda.is_available() and torch.backends.cudnn.is_available() and PerformanceFeature.AutoTune in args.fast:
- torch.backends.cudnn.benchmark = True
+
+def set_cudnn_benchmark():
+ if torch.cuda.is_available() and torch.backends.cudnn.is_available():
+ torch.backends.cudnn.benchmark = PerformanceFeature.AutoTune in args.fast
try:
if torch_version_numeric >= (2, 5):
diff --git a/main.py b/main.py
index 7fcc8e97d..0ad660376 100644
--- a/main.py
+++ b/main.py
@@ -490,6 +490,11 @@ def start_comfyui(asyncio_loop=None):
init_custom_nodes=(not args.disable_all_custom_nodes) or len(args.whitelist_custom_nodes) > 0,
init_api_nodes=not args.disable_api_nodes
))
+
+ # Re-apply Comfy's cuDNN benchmark policy after custom-node imports. Benchmark
+ # mode can request near-card-sized autotune workspaces, and some custom nodes set it at import time.
+ comfy.model_management.set_cudnn_benchmark()
+
hook_breaker_ac10a0.restore_functions()
cuda_malloc_warning()