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.

Add backend option to TorchCompileModel.

comfyanonymous83b01f960ac7
diff --git a/comfy_extras/nodes_torch_compile.py b/comfy_extras/nodes_torch_compile.py
index 1d914fa93..1fe6f42c7 100644
--- a/comfy_extras/nodes_torch_compile.py
+++ b/comfy_extras/nodes_torch_compile.py
@@ -4,6 +4,7 @@ class TorchCompileModel:
@classmethod
def INPUT_TYPES(s):
return {"required": { "model": ("MODEL",),
+ "backend": (["inductor", "cudagraphs"],),
}}
RETURN_TYPES = ("MODEL",)
FUNCTION = "patch"
@@ -11,9 +12,9 @@ class TorchCompileModel:
CATEGORY = "_for_testing"
EXPERIMENTAL = True
- def patch(self, model):
+ def patch(self, model, backend):
m = model.clone()
- m.add_object_patch("diffusion_model", torch.compile(model=m.get_model_object("diffusion_model")))
+ m.add_object_patch("diffusion_model", torch.compile(model=m.get_model_object("diffusion_model"), backend=backend))
return (m, )
NODE_CLASS_MAPPINGS = {