[Tests] Fix failing 8bit test

[younesbelkada]

What does this PR do?

Fixes two failing tests in https://github.com/huggingface/transformers/actions/runs/5873964880/job/15928072870

• tests/quantization/bnb/test_mixed_int8.py::MixedInt8Test::test_get_keys_to_not_convert & tests/quantization/bnb/test_mixed_int8.py::MixedInt8GPT2Test::test_get_keys_to_not_convert

Context: #25105 added stronger checks to enable the correct quantization of models on the Hub. Therefore it added a test that checks if mpt-7b is correctly quantized. Since that model requires einops to be added as a dependency I propose to simply add einops in the docker image

cc @ydshieh

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[ydshieh]

Hi @younesbelkada ! Could you tell a bit more why MPT needs einops?

[ydshieh]

Also you probably need refresh your CircleCI token?

[younesbelkada]

for MPT as you can see from the remote model, einops is used in files that are imported in modeling_mpt.py such as attention.py here: https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L7 it seems to be a strong dependency (hence the error on the slow test attached)
It is also listed as a dependency in the requirements file together with their custom fork of triton: https://huggingface.co/mosaicml/mpt-7b/blob/main/requirements.txt#L2 but the triton dependency is optional

[ydshieh]

OK, I think I got confused by the fact we have a modeling file and we also use the same checkpoint name in our model tests. Now I get it, but a question: why we want to use the remote code rather than the code in transformers for this quantization test ..? Is it really necessary to use the remote code the quantization?

[younesbelkada]

Thanks!
Regarding your question, yes I think so, #25105 added the support for correct quantization for most remote code models, therefore it is crucial to test the remote code model + non remote model in the test

[ydshieh]

Thanks for explaining. Let's try it!

I am not sure what's the original purpose of testing the modeling code on the Hub - although we pin a revision, the code might change on the Hub, and our tests won't detect any failure if any for new code.

But as it is already there, let's keep it - just not to add any more remote model into the relevant tests 🙏

[ArthurZucker]

Quite aligned with @ydshieh here. Don't think we should be testing a code on the hub with trust_remote_code. Why not use the model now that it is on transformers?
EDIT: read the original PR, guess it's okay, no a big fan of testing remote codes 😅