Arthur
Arthur
Of course!
Hey! Sure I was off for a bit but will have a look
@gante from my personal test, changing the inv_freq to float32 can increase performances on MMLU of about 20points. There are a few things to test: - `inv_freq` is a buffer,...
super nice 🤗
Also linked to #29285
I think this was fixed! For llama at least and gemma 🤗
Wip, not urgent! Can already be done but it's not save / loaded
I don't have bandwidth yet so nice if you want to do ti!
Hey! Thanks, you are probably right. Would you like to open a PR to change this and make it more friendly?