[tests] use the correct n_gpu in TrainerIntegrationTest::test_train_and_eval_dataloaders for XPU

[faaany]

What does this PR do?

The below test fails on intel GPU with 8 cards:

___________________________________ TrainerIntegrationTest.test_train_and_eval_dataloaders ___________________________________

self = <tests.trainer.test_trainer.TrainerIntegrationTest testMethod=test_train_and_eval_dataloaders>

    def test_train_and_eval_dataloaders(self):
        n_gpu = max(1, backend_device_count(torch_device))
        trainer = get_regression_trainer(learning_rate=0.1, per_device_train_batch_size=16)
>       self.assertEqual(trainer.get_train_dataloader().total_batch_size, 16 * n_gpu)
E       AssertionError: 16 != 128

tests/trainer/test_trainer.py:1035: AssertionError

This is because the n_gpu is 8, but the default _n_gpus in TrainingArguments is 1 as can be seen here, leading to a total_batch_size of 16. This test works on CUDA, because CUDA supports DataParallel and would set _n_gpus to a value larger than 1 as can be seen here. But on all other devices except CUDA and CPU, this test will fail.

To fix this issue, we need to add a conditional check to the test to make this test work on other devices except GPU. Pls have a review: @muellerzr and @pacman100

[faaany]

Hi @ArthurZucker , thanks for reviewing my other PRs. Would you like to take care of this one as well? Like other tests, this one is also written with only GPU in mind and fails on other devices like XPU. Thanks a lot!

[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]

Fine by me, cc @muellerzr feel free to merge if this is alright with you as well

[faaany]

@muellerzr can we proceed with this merge? thanks a lot!

[muellerzr]

Good for me as well!