fix bug: fsdp2 cannnot run with npu, because the hardcode with cuda

[frozenleaves]

What does this PR do?

fix the bug: when run fsdp2 on npu device, it will raise an error:

AssertionError: Torch not compiled with CUDA enabled.

[S1ro1]

Don't have to import here, accelerator is passed into the function, can take device from there.

[frozenleaves]

done. Because some packages depend on version 1.7.0, such as llama-factory, etc., we hope to fix this issue on this version.

[S1ro1]

Small nit, except of that lgtm, thank you for noticing!

[S1ro1]

cc @SunMarc for a patch, not sure how we want to approach this

[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.

[SunMarc]

@bot /style

[github-actions]

Style fix is beginning .... View the workflow run here.

[SunMarc]

Thanks ! I will probably do a patch in a few days

[SunMarc]

@bot /style

[github-actions]

Style fix is beginning .... View the workflow run here.

[SunMarc]

Can you fix the style ? the other failures are not related

[frozenleaves]

Can you fix the style ? the other failures are not related

ok, the fixed code have already pushed

[frozenleaves]

@SunMarc Hi, I have try to fix the CI problem, I noticed that in the CI of accelerate is dependent on transformers’ UT, and there was a test case as follows：

    @require_accelerate
    def test_executable_batch_size(self):
        batch_sizes = []

        @find_executable_batch_size(starting_batch_size=64, auto_find_batch_size=True)
        def mock_training_loop_function(batch_size):
            nonlocal batch_sizes
            batch_sizes.append(batch_size)
            if batch_size > 16:
                raise RuntimeError("CUDA out of memory.")

        mock_training_loop_function()
        self.assertEqual(batch_sizes, [64, 57, 51, 45, 40, 36, 32, 28, 25, 22, 19, 17, 15])

This test case still judged by reducing 10% each time in the main branch of transformers, but in the implementation logic of find_executable_batch_size in accelerate, it was reduced by 50% each time, which led to the UT failing.

Here, I modified the implementation related to accelerate to make it pass the UT test. Please consider whether such a fix is reasonable or fix the corresponding test code of transformers.

[SunMarc]

Oh the real issue is that you are trying to merge this pr into this specific branch huggingface:v1.7.0-release. Can you reopen the PR with the right target branch ?

[frozenleaves]

Oh the real issue is that you are trying to merge this pr into this specific branch huggingface:v1.7.0-release. Can you reopen the PR with the right target branch ?

sure, is my operation correct？

[SunMarc]

I mean you need to recreate a pull request with the target branch main. Right now, you are still trying to merge into another branch:
frozenleaves wants to merge 2 commits into huggingface:v1.7.0-release from frozenleaves:v1.7.0-release