Degradation in audio and image classification samples with new transformers package

[ljaljushkin]

Audio and image classification samples fail with error AttributeError: <> object has no attribute 'no_sync'

For example, Command line for finetuning of wave2vec with QAT fails with

AttributeError: 'Wav2Vec2ForSequenceClassification' object has no attribute 'no_sync'

It's reproduced with latest version of transformers: 4.30.0
There's no such issue with 4.22.0 version

[ljaljushkin]

@helena-intel @AlexKoff88 @echarlaix @vuiseng9 @yujiepan-work

[AlexKoff88]

@ljaljushkin, it seems like this is not an Optimum-related problem but rather a Transformer problem. Some relevant issues: https://discuss.huggingface.co/t/distributeddataparallel-object-has-no-attribute-no-sync/5469
I think it should be worked around in the sample.

[helena-intel]

@ljaljushkin thanks for reporting this. The examples are definitely expected to work out of the box with the version of transformers we install as dependency. Could you try running with torchrun like you changed for some other runs in #315 ?

[ljaljushkin]

@ljaljushkin thanks for reporting this. The examples are definitely expected to work out of the box with the version of transformers we install as dependency. Could you try running with torchrun like you changed for some other runs in #315 ?

Yes, I had verified that. The issue is reproduced for mentioned samples only.

[ljaljushkin]

@helena-intel this is not reproduced on the main branch now. But audio sample still doesn't work #423