Support for Gemma 2?

[shan23chen]

Hi,

I am wondering whether this support training for second gen Gemma based llava.

When we trained with the new Gemma with this repo, we ran into an error of:

TypeError: '<' not supported between instances of 'NoneType' and 'int'

And debugging lead us to a weir huggingface past key cache bug. Wondering whether you guys have any idea how to solve it.

Thanks!