Joao Gante

@WissamAntoun thank you for pointing it out, I completely missed it in the original thread! 🙏 Will make the change EDIT: this change also makes it ~10% faster 👍

Hi @tmoroder 👋 Thank you for adding all that information to the issue

Interesting. Since `transformers==4.20.1`, there are only two DeBERTa PRs: 1. https://github.com/huggingface/transformers/pull/17940 (should have no impact at all here) 2. https://github.com/huggingface/transformers/pull/18256 (what should have been a TPU-friendly `take_along_axis`) As @Rocketknight1 said,...

@tmoroder I can confirm that I can run your example with `jit_compile=True` (i.e. XLA compilation) on `model.compile()`, using a GPU, if the two changes you made in your [third TPU...

(linking issues -- the Tile issue is also present in the following unsolved issue: https://github.com/huggingface/transformers/issues/14058)

The cause is trivial (the `multiple` argument of `tf.tile` can't have dynamic shapes), but the fix may be not :D Will look into it

@tmoroder the dynamic shape in question is the batch size. I may be able to get an alternative to `tf.tile`, but I highly doubt that it will make a difference...

Wooo nice! 🎊 I'm closing this issue since the problem seems to be solved for now. Feel free to reopen if you run into new related issues. Also, if you...

@Narsil @KMFODA I'm in favor of moving it to a `StoppingCriteria`, so that all conditions that can terminate generation fall under the same class. However, it should be noted that...

Hey @ariG23498 👋 Regarding failing tests: - The tests in `run_tests_torch_and_tf` seem to need a slightly higher tolerance. Let's leave those as their are for now (failing), as we might...