### Motivation and description Just wondering if there is a way to do mixed precision training in Flux? ### Possible Implementation _No response_