[tests] enable automatic speech recognition pipeline tests on XPU

[faaany]

What does this PR do?

"cuda" and "cuda:0" are hard-coded in test_whisper_longform and test_seamless_v2 of test_pipelines_automatic_speech_recognition.py. So when running the unit tests on XPU, I get 2 failing tests:

================================================= short test summary info ==================================================
SKIPPED [1] tests/pipelines/test_pipelines_automatic_speech_recognition.py:269: test requires TensorFlow
FAILED tests/pipelines/test_pipelines_automatic_speech_recognition.py::AutomaticSpeechRecognitionPipelineTests::test_seamless_v2 - AssertionError: Torch not compiled with CUDA enabled
FAILED tests/pipelines/test_pipelines_automatic_speech_recognition.py::AutomaticSpeechRecognitionPipelineTests::test_whisper_longform - AssertionError: Torch not compiled with CUDA enabled
============================= 2 failed, 40 passed, 1 skipped, 31 warnings in 203.95s (0:03:23) =============================

This PR proposes to use the device-agnostic variable torch_device instead of the hard-coded "cuda" device. After the fix,
all unit tests in test_pipelines_automatic_speech_recognition.py run through on XPU as shown below:

================================================= short test summary info ==================================================
SKIPPED [1] tests/pipelines/test_pipelines_automatic_speech_recognition.py:269: test requires TensorFlow
================================== 42 passed, 1 skipped, 32 warnings in 219.40s (0:03:39) ==================================

Who can review?

@Narsil @ArthurZucker

[ArthurZucker]

LGTM but we might need to add require_torch_gpu as it was tested on GPU

[faaany]

LGTM but we might need to add require_torch_gpu as it was tested on GPU

If I change require_torch to require_torch_gpu, this test will be skipped on Intel GPU. But I don't see any reason why this test should only work on NV GPU. Maybe we can change require_torch to require_torch_accelerator to let CPU save this test, what do you think?

[ArthurZucker]

Alright, I just want to make sure our CI runs it with cuda because the values were obtained with it and not 100% sure you always get the same. Though it is not logits some device can get it differently, we'll leave that and I'll merge thanks for the PR

[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.