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.

Empty cache on mps.

comfyanonymous94680732d32b
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 10a706793..60bcd786b 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -389,7 +389,10 @@ def should_use_fp16():
def soft_empty_cache():
global xpu_available
- if xpu_available:
+ global vram_state
+ if vram_state == VRAMState.MPS:
+ torch.mps.empty_cache()
+ elif xpu_available:
torch.xpu.empty_cache()
elif torch.cuda.is_available():
if torch.version.cuda: #This seems to make things worse on ROCm so I only do it for cuda