Arthur

@SY-Xuan let's try to be clear when we say: - results are correct / wrong: what is wrong for you? You did not share generation, nor did you provide a...

No worries! 🤗 I'll close this as completed

LGTM, mostly concerned with quanto specific function addition, that should leave in quanto.py not in the modeling (as much as possible of course).

Hey ! You are using custom code, we probably won't have the bandwidth to debug it for you! Best recommendation is to put a debug breakpoint and see what is...

Kwargs are nice but we need explicit configs / explicit arguments, I am not 100% convinced kwargs is the way to go. We could have de `PreTrainedConfig` for that matter,...

#30615 superseeds this! Feel free to close

cc @younesbelkada I think we have seen something similar recently?

cc @pacman100 and @SunMarc

I think a good way to se if it is worth the shot is to benchmark your code and check if you have speedups in different contexts!

Sounds like a good addition then! I'll let @fxmarty review and will be doing the final pass!