[tests] skip bnb-related tests instead of failing on xpu

[faaany]

What does this PR do?

When running pytest tests/test_accelerator.py on XPU, there are 4 tests failed:

FAILED tests/test_accelerator.py::AcceleratorTester::test_accelerator_bnb - RuntimeError: No GPU found. A GPU is needed for quantization.
FAILED tests/test_accelerator.py::AcceleratorTester::test_accelerator_bnb_cpu_error - RuntimeError: No GPU found. A GPU is needed for quantization.
FAILED tests/test_accelerator.py::AcceleratorTester::test_accelerator_bnb_multi_device - ValueError: xpu is not supported in test_accelerator_bnb_multi_device.
FAILED tests/test_accelerator.py::AcceleratorTester::test_accelerator_bnb_multi_device_no_distributed - RuntimeError: No GPU found. A GPU is needed for quantization.

The reason for all 4 tests is the same: all tests use load_in_8bit=True when loading the model with AutoModelForCausalLM.from_pretrained, which will lead to the torch.cuda.is_available() check in hf_quantizer.validate_environment, e.g.

___________________________________________ AcceleratorTester.test_accelerator_bnb ____________________________________________

self = <test_accelerator.AcceleratorTester testMethod=test_accelerator_bnb>

    @require_non_torch_xla
    @slow
    @require_bnb
    def test_accelerator_bnb(self):
        """Tests that the accelerator can be used with the BNB library."""
        from transformers import AutoModelForCausalLM
    
>       model = AutoModelForCausalLM.from_pretrained(
            "EleutherAI/gpt-neo-125m",
            load_in_8bit=True,
            device_map={"": 0},
        )

tests/test_accelerator.py:439: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
../../miniforge3/envs/acc-ut-ww23/lib/python3.9/site-packages/transformers/models/auto/auto_factory.py:563: in from_pretrained
    return model_class.from_pretrained(
../../miniforge3/envs/acc-ut-ww23/lib/python3.9/site-packages/transformers/modeling_utils.py:3202: in from_pretrained
    hf_quantizer.validate_environment(
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

self = <transformers.quantizers.quantizer_bnb_8bit.Bnb8BitHfQuantizer object at 0x7f2cef87ceb0>, args = ()
kwargs = {'device_map': {'': 0}, 'from_flax': False, 'from_tf': False, 'torch_dtype': None}

    def validate_environment(self, *args, **kwargs):
        if not torch.cuda.is_available():
>           raise RuntimeError("No GPU found. A GPU is needed for quantization.")
E           RuntimeError: No GPU found. A GPU is needed for quantization.

../../miniforge3/envs/acc-ut-ww23/lib/python3.9/site-packages/transformers/quantizers/quantizer_bnb_8bit.py:62: RuntimeError

Since BNB doesn't support XPU and NPU yet, we should skip these tests rather than let them fail. I also checked the original PR #2343, which makes those tests device-agnostic. These 4 tests are overlooked because the author doesn't install bitsandbytes. But when users install the libraries for testing using pip install -e ".[testing]", the bitsandbytes library will be installed. So we should change the test maker to require_cuda, instead of making them device-agnostic, at least before the time when XPU and NPU are supported in BNB.

@muellerzr @SunMarc

[faaany]

add one more test: tests/test_big_modeling.py::BigModelingTester::test_dipatch_model_fp4_simple

[muellerzr]

Thanks! cc @SunMarc

[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]

LGTM ! Left a few comments

[muellerzr]

Thanks!