Sylvain Gugger

Looking into it, thanks for the repro!

The problem should be fixed on master. We'll make a patch release on Monday with the fix.

I don't think there is any point @Forpee

I can reproduce. Will try to have a look later today or early next week. Thanks for the report!

Should be fixed by the PR linked abov.

cc @pacman100

@pacman100 friendly ping

Thanks a lot for the analysis and your fix suggestion. The idea of having a `skip_keys` argument in `dispatch_model` could definitely live in Accelerate and Transformers could then set it...

@emvw7yf I started to draft something in the PR linked above. I am not seeing your speed-ups on two GPUs, but I have an nvlink so it might be why...

Ok, support should be coming natively in Transformers with the PR above then.