Yeah, on "round trip tests" it worked fine on encode/decode since it would use MPS on both, but stumbled across the issue when doing some other test and the LLM used CPU to decode and caught it in the default/test case.