dependencies: keras-nlp<0.14 pin

[gante]

What does this PR do?

On the latest keras-nlp release -- "KerasNLP no longer supports Keras 2.". This means we don't support this version, as we only support Keras 2

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[amyeroberts]

Fine with pinning to keep things working for now, but would be good to have a plan for future keras updates. @Rocketknight1 could we remove keras-nlp?

also cc @ydshieh for reference

[gante]

@amyeroberts @ydshieh I followed our docs for version updates, but the resulting PR doesn't look right.

What's the expected procedure to update breaking versions? 🤔

I've confirmed that:

• having keras-nlp==0.14.0 causes these test errors
• calling pip install "keras-nlp<0.14" -U fixes it

However, since the docker images are not updated in this PR, CI will continue to be red

[amyeroberts]

@gante Tbh, I'm not sure. @ArthurZucker is the best person to ping here, as he wrote the docs

[ArthurZucker]

Let's go on slack @gante !

[gante]

@amyeroberts the dev-ci commit ran against the :dev docker images, which stem from the latest commit here. The failed test is unrelated and reported on slack, the TF issues are fixed as it can be seen :)

If you approve these changes, after the hub failures are resolved:

1. I will add a commit [push-ci-image] to the other PR, which will push the updated docker images. After the CI on this commit runs, all models will run their CI with that docker image, fixing the issues on CI.
2. merge this PR, which fixes the issue on local installs and ensures future docker image updates include this change.
3. Update the CI docker docs with the learnings in this process, to make them more clear to everyone :)

[amyeroberts]

Thanks!

[Rocketknight1]

Hey @amyeroberts, we currently don't have a transition plan to move Transformers to Keras 3 because of how disruptive it would be to backwards compatibility. A lot of our TF/Keras usage is users who use it for production workflows with older software versions, so we think it'll be quite a while before we're ready to deprecate all of that and migrate to Keras 3.

For the forseeable future, we're going to keep Transformers on Keras 2, and support Keras 3 by having libraries like KerasNLP and KerasCV pick up the slack and save/load safetensors checkpoints that Transformers can read. This has already started thanks to @ariG23498's work, and more architectures will be supported very soon.

Therefore, it's totally okay to pin Keras and Keras-associated libraries on the last version that supported Keras 2, and that pin will probably be maintained indefinitely.

[julien-c]

(btw cc @Wauplin for viz)

[amyeroberts]

Thanks for the detailed explanation @Rocketknight1 !

[Wauplin]

Thanks for the ping, good to keep in mind. And thanks @Rocketknight1 for the full context!

[fxmarty]

LGTM

[gante]

Docker image pushed in #31687 -- merging 🤞