Test: generate with torch.compile(model.forward) as a fast test

[gante]

What does this PR do?

Follow-up to #34464

This PR:

1. Converts test_generate_compile_model_forward to a fast test. This means we will check generate with torch.compile(model.forward) at each commit on ALL models that support StaticCache 💛
2. Fixes failing cases of test_generate_compile_model_forward whenever possible
3. Tags models with _supports_static_cache = False #Reason when the model doesn't support torch.compile(model.forward)

---

✅ py.test tests/models/ -k test_generate_compile is all green, takes ~2 mins to run on all models on my machine

[ydshieh]

Love this!

Q: Is it really fast ...?

Remark: I feel get_max_cache_length is a better name than get_max_cache_shape but OK I know not great to change name all the time.

[gante]

Q: Is it really fast ...?

@ydshieh yes :D

[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.

[ArthurZucker]

I don't mind, tho I don't think our priority should be this (full compile vs compile forward in generate!) + I don't see the test being run in the CI! 🤗

[ArthurZucker]

Could you just make sure it's run

[ydshieh2]

We need to remove @require_torch_gpu too for def test_generate_compile

[ydshieh]

Before merge, feel free to ping me for a check for (if there is any) flakyness :-) or anything you think I can double check again.

[ArthurZucker]

Thanks can ignore my comments and merge 🤗

[ArthurZucker]

is it possible for the HybridCache to inherit from Static cache?

[ArthurZucker]

We might just need an extra class that says CompileCompatible , someone wanted is_static attr˜!

[gante]

(sorry, the PR is not ready yet, a few cases are still failing 👀 I didn't mean to request a review)

[gante]

Now it's working on all models, including encoder-decoder + cache 🤗

It's not too heavy on our CI, it should add ~2 mins if all models are run. And it should prevent us from many headaches! As we can see in diff, we had compilation enabled for a bunch of models that don't support it.