Sylvain Gugger

This is not a bug. You need to resume your training in the exact same distributed setup.

Since you didn't put your inputs on the GPU, the generation part after the model runs is done on the CPU (Accelerate makes the model return outputs on the same...

As the error tells you the inputs of the models are integers. Why are you converting them to float16?

I can I dived a bit in the issue as I didn't understand why you had the bug even for the right input dtype. Turns out it's a bug in...

Thanks for the very clear issue! Your proposed fix sounds right, do you want to open a PR with it?

Thanks a lot for flagging! Do you want to open a PR to fix this?

There is still one last styling error in the test. Make sure your version of black is 22.3.0 as maybe it comes from a different version?

No worries at all! But it now looks like you pushed a diff in way too many files. Are you sure you are using version 22.3?

Mishig the failure on the doc was due to a typo (comment is hidden now since the suggestion was accepted) nothing to do for you :-)

The failure is unrelated. As long as all models have been moved this PR can be merged.