Save memory for training？

[StarsTesla]

Can this code add feature like another flux training code "ai-toolkit", to use quantizing the model for training, which saves alot gpu memory, and keep clean and small code of this repo.