Tokenizers are supposed to be framework (PyTorch/TensorFlow/FLAX) agnostic so we probably don't want to do in that direction.

Gotcha. Is this going to be the case for all tokenizers in the future? Because currently they default to PyTorch except for Bart's.
Also, I think the docstring for Bart tokenizer's return_tensors needs to be updated then since it says: optional, defaults to "pt"

The fact that the BART-like tokenizers have return_tensors="pt" is a mistake. The tokenizers should be framework-agnostic.

We will have to update this, which will be a breaking change, so we'll try to put it in the v4.0.0. Do you want to open a PR to fix the issue?

Sure. I'll close this then and make a new PR for that.

Hi @Mehrad0711! We're rushing to transformers@v4.0.0, and we don't want that in this release. I've taken the liberty of fixing it in #8599. Sorry about that, I hope you had not started your development.

If you have, you can push your fixes and open a PR and I'll incorporate those changes in my PR and mark you as co-author.

No problem @LysandreJik . Thanks for fixing it!