RAdamSchedulefree give great results but errors after 1 epoch

Hello, RAdam has insanely smooth progress throughout the first epoch of training, but after the epoch finishes, it stops working with an error:

Optimizer was not in train mode when step is called.

This can be prevented by instead of using epochs, you can use repeats but i still think this is a bug.

Start by reproducing the RAdamSchedulefree training flow for one epoch and capture the point where "Optimizer was not in train mode when step is called" occurs. Compare it with the repeats-based flow described in the issue; done means epoch-based training completes without the optimizer mode error.

Written by the indexing model from the issue text.