my bad, didn't realize it was not public yet. We'll work on publishing it asap. We basically have a stateful Tensor subclass and a Float8Linear layer that work together to get to working Float8 training.

→ View original post on X — @soumithchintala