Is there any point in setting `fp16_full_eval=True` if training in `fp16`? I train a Huggingface model with fp16=True, e.g.:
