The models are modified in place before being used and unpatched after. I think this is better than monkeypatching since it might make it easier to use faster non pytorch unet inference in the future.
comfyanonymous