[tests] add require_torch_sdpa for test that needs sdpa support

[faaany]

What does this PR do?

I got this test failed in my XPU environment. After debugging, I found that the EXPECTED_LOGITS in this test are computed with the assumption that _attn_implementation is set to sdpa. The code for setting the _attn_implementation are this line and this line. I experimented to downgrade my torch version to make _attn_implementation value to eager and it fails on CUDA. So we neeed to add the maker require_torch_sdpa to this test.

@ArthurZucker and @amyeroberts

[amyeroberts]

Thanks for digging into this!

Looks good to me! Would be great to add the bnb requirement for the other test too

[amyeroberts]

It looks like the test about should have require_bitsandbytes too 👀

@ydshieh If we have this requirement, will this test be run as part of the slow overnight tests?

[faaany]

@amyeroberts I pushed a commit to add the require_bitsandbytes flag for test_batched_4bit, test_batched_small_model_logits is fine without bitsandbytes.

[ydshieh]

@amyeroberts Well, I guess I made a huge mistake in #29208 where I removed

# Add bitsandbytes for mixed int8 testing
RUN python3 -m pip install --no-cache-dir bitsandbytes

because we aimed to move the quantization tests outside the main model CI job.

@younesbelkada @SunMarc I think we should add the above back to

docker/transformers-all-latest-gpu/Dockerfile

as there are some model tests using bnb

class InstructBlipModelIntegrationTest(unittest.TestCase):
    @require_bitsandbytes
    @require_accelerate
    def test_inference_vicuna_7b(self):

WDYT?

[younesbelkada]

Yes sounds great ! #30427

[SunMarc]

I see that there are around 20 models with bnb tests. So yeah, let's add bitsandbytes package in the main CI.

[amyeroberts]

Thanks for handling this and updating!