I’m trying to speed up both training phase and inference time. While QAT gives promising results, but when I add AMP I get RuntimeError: expected scalar type Float but found Half. It seems fake quant only support float. Is this intended or is there work around?