Sylvain Gugger

Sorry, I didn't this until now. The blog post you're referring too is a bit old now, and it was when we were just grasping with super-convergence. Now we can...

This is a flaky test, don't worry. Thanks again for your contribution!

That might because your new job did not run ;-) You need to add at the end [here](https://github.com/huggingface/transformers/blob/d7e2d7b40b1070cddfe878e13705725f49a2cf1f/.circleci/config.yml#L1000) for the one at each commit and [there](https://github.com/huggingface/transformers/blob/d7e2d7b40b1070cddfe878e13705725f49a2cf1f/.circleci/config.yml#L1024) for the nigthly one...

I think it's preferable to use the more formal option, but in most cases, I'd prefer to reformulate the sentences to use the first person plural (wir) unless the sentence...

Let me know when you're done and thanks a lot for diving into German translation! (Sorry should have begun with that!)

Merged the PR in Accelerate! Don't forget to add some documentation and also setup some tests for this so it doesn't get broken by future PRs :-)

The PR is #18534, but nothing will beat using offline mode with the model cached, since you are then doing 0 calls to the API.

If the revision is specified as a commit sha, then yes, the cache should be used. This is not implemented by #18534 however, but could be some follow up work....

Yes there is probably some way to cache that the file does not exist for a given commit sha. Pinged a few people internally to see if they like that...

Ideally it should mean that all models work/all tests pass apart from functionality explicitly having versions tests (like CUDA bfloat16 or torch FX where we test against a specific PyTorch...