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.

mm: discard async errors from pinning failures (#10738)

rattus9ca7e143afe6
diff --git a/comfy/model_management.py b/comfy/model_management.py
index e5554e225..9fcb699bc 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -1126,6 +1126,16 @@ if not args.disable_pinned_memory:
PINNING_ALLOWED_TYPES = set(["Parameter", "QuantizedTensor"])
+def discard_cuda_async_error():
+ try:
+ a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
+ b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
+ _ = a + b
+ torch.cuda.synchronize()
+ except torch.AcceleratorError:
+ #Dump it! We already know about it from the synchronous return
+ pass
+
def pin_memory(tensor):
global TOTAL_PINNED_MEMORY
if MAX_PINNED_MEMORY <= 0:
@@ -1158,6 +1168,8 @@ def pin_memory(tensor):
PINNED_MEMORY[ptr] = size
TOTAL_PINNED_MEMORY += size
return True
+ else:
+ discard_cuda_async_error()
return False
@@ -1186,6 +1198,8 @@ def unpin_memory(tensor):
if len(PINNED_MEMORY) == 0:
TOTAL_PINNED_MEMORY = 0
return True
+ else:
+ discard_cuda_async_error()
return False