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.

Disable offload stream when torch compile. (#10961)

comfyanonymousf55c98a89f76
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 38c506df5..d8ce80010 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -1042,6 +1042,9 @@ def get_offload_stream(device):
if NUM_STREAMS == 0:
return None
+ if torch.compiler.is_compiling():
+ return None
+
if device in STREAMS:
ss = STREAMS[device]
#Sync the oldest stream in the queue with the current