avoid calling gc.collect and cuda.empty_cache

[ydshieh]

What does this PR do?

Let's avoid calling gc.collect and cuda.empty_cache while the tests are running on CPU:

• those operations are slow
• (actually, in most cases, they are only relevant for integration tests which use large models)

Running on GPT2 tests,

60 seconds on main, 20 seconds on this PR

[Rocketknight1]

Yes, this seems like a good speed fix! cc @LysandreJik @ArthurZucker for core maintainer review

[LysandreJik]

Smart! Should a helper method be made that only runs on CPU?

Both the gc.collect and the torch device checks could be moved into the backend_empty_cache method (or an other method that wraps both)

[ydshieh]

Yes, a helper method is nice. Will update

[ydshieh]

updated.

So far it doesn't call gc.collect() at all (default value False).
I would like to see if this would cause issue.
In general, we don't need to call it after each test method (so in tearDown) as it is slow.

[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.