Add an option to reduce compile() console spam

[Rocketknight1]

This is a very simple PR to add an option to reduce the console spam from our compile() method.

Briefly, Keras model expect you to pass a loss function to compile(). However, transformers models usually compute loss internally. The solution we used was that if the user didn't pass a loss argument to compile(), we would read the model's loss output and use that as the loss. This was non-standard Keras behaviour, though, so we added a warning to make sure users knew what was going on.

Now that we've been doing it for a while, though, that warning is probably just more console spam. This PR adds the option to specify loss="auto", which has exactly the same behaviour but eliminates the warning. The warning also includes a line telling users that they can do that.

[sgugger]

Or maybe we just make the warning an info?

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[Rocketknight1]

I think we tried that, but then most users would miss it! I think we definitely want to make users aware of what's going on (because this is the one big place we diverge from the Keras training standard), but give them an option to disable the warning once they know.

[sgugger]

I think proper documentation is a better solution than throwing a warning all the time. Users are pretty mad at us with those already.

[Rocketknight1]

This could be a good opportunity to document stuff, actually! Do you prefer a sidebar tutorial, or some text in the docstring that gets added to all of our TF models?

[sgugger]

I think in the official examples where we actually use torch.compile is probably the best place.

[Rocketknight1]

This is TF compilation, not torch.compile()! All of our TF examples use it - you can't call model.fit() in Keras before calling model.compile()

[sgugger]

Sorry I meant model.compile. And yes, all the examples sounds about right.

[Rocketknight1]

Done! I added a comment when we call compile() in the example scripts. I also fixed up one unnecessary use of run_eagerly in run_mlm.py - oops!

I moved the warning message to logging.info, so it should be invisible to most users now, and they can still pass loss="auto" to disable it entirely if desired.

[sgugger]

Those and the task examples in the doc too if you don't mind :-)

[sgugger]

Thanks!