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 initial GPU load when novram is used. (#8294)

comfyanonymous89a84e32d2a7
diff --git a/comfy/model_management.py b/comfy/model_management.py
index 1e6e8553f..f5b37e68e 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -695,7 +695,7 @@ def unet_inital_load_device(parameters, dtype):
return torch_dev
cpu_dev = torch.device("cpu")
- if DISABLE_SMART_MEMORY:
+ if DISABLE_SMART_MEMORY or vram_state == VRAMState.NO_VRAM:
return cpu_dev
model_size = dtype_size(dtype) * parameters