[tests] enable test_pipeline_accelerate_top_p on XPU

[faaany]

What does this PR do?

Is there any particular reason why we use device_map="auto" in this test test_pipeline_accelerate_top_p? If not, I would suggest using the device-agnostic variable torch_device instead of device_map="auto" just like in all other tests e.g. test_small_model_fp16.

Another reason is that the device_map="auto" mechanism is still not mature on XPU, causing the model to be loaded on the CPU, rather than on XPU. With this fix, test_pipeline_accelerate_top_p will definitively work on all devices. Below is an evidence for XPU:

BEFORE

================================================== short test summary info ===================================================
SKIPPED [1] tests/pipelines/test_pipelines_text_generation.py:221: test requires TensorFlow
SKIPPED [1] tests/pipelines/test_pipelines_text_generation.py:378: test requires CUDA
SKIPPED [1] tests/pipelines/test_pipelines_text_generation.py:180: test requires TensorFlow
FAILED tests/pipelines/test_pipelines_text_generation.py::TextGenerationPipelineTests::test_small_model_fp16 - RuntimeError: could not execute a primitive
===================================== 1 failed, 5 passed, 3 skipped, 8 warnings in 7.18s =====================================

AFTER

================================================== short test summary info ===================================================
SKIPPED [1] tests/pipelines/test_pipelines_text_generation.py:221: test requires TensorFlow
SKIPPED [1] tests/pipelines/test_pipelines_text_generation.py:378: test requires CUDA
SKIPPED [1] tests/pipelines/test_pipelines_text_generation.py:180: test requires TensorFlow
========================================== 6 passed, 3 skipped, 8 warnings in 7.52s ==========================================

Pls have a review, thx! @Narsil @ArthurZucker

[ArthurZucker]

I think in this one we want to test accelerate support with sampling so it's probably on purpose!

[faaany]

I think in this one we want to test accelerate support with sampling so it's probably on purpose!

Could you elaborate more on the relationship between top_p sampling and device? I don't see the need to use "device_map='auto'" in this case, e.g. if I change "device_map='auto'" to "device=torch_device", will the test on top_p sampling fail?

[ArthurZucker]

The name of the test is test_pipeline_accelerate_top_p to make sure we test accelerate dispatching automatically.
I thought testing "auto" made sense but you're right here, we can test any device, as long as we make sure we test the accelerate support here!

[faaany]

The name of the test is test_pipeline_accelerate_top_p to make sure we test accelerate dispatching automatically. I thought testing "auto" made sense but you're right here, we can test any device, as long as we make sure we test the accelerate support here!

done, thanks for the review!

[ArthurZucker]

Thanks for iterating!