SD-Forge - macOS: Incompatibilities When Refiner Is Enabled

[x1ziad]

When enabling the Refiner model under the Forge backend on macOS, generation fails or leads to unexpected behavior such as:
Blank outputs or black images
Sudden crashes or hangs
TypeError or Tensor mismatches in logs
CUDA-related or torch.compile/inductor backend issues (even though CUDA isn’t on macOS)

[AnabelleKM]

Interesting that you mention this. (Though on Windows)

Yesterday I was experimenting with img2img and ControlNet. Which seemed to work until I tried using a refiner. Then Controlnet didn't seem to work as expected anymore. But I didn't experience crashes, hangs or black images.

[StrongerXi]

@x1ziad can you share how you are running this? I'm curious where torch.compile gets triggered. If it's CUDA related, maybe you had one of your tensors with device='cuda'? Can you try running without torch.compile?

Moreover, torch.compile support for Mac MPS is tracked here: pytorch/pytorch#150121.