Integration Testing for Quantized ONNX export

[dbogunowicz]

Asana ticket: https://app.asana.com/0/1201735099598270/1201949554543506/f

There are some discrepancies that are probably harmless and expected. For instance, when comparing two pruned models, the sequences of last operations (before logits) differ slightly (this is probably due to how the newer PyTorch version handles the model-->onnx export pipeline).

The prediction head for the original onnx model ("before")

The prediction head for the retrieved onnx model ("after").

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[bfineran]

looks like a good first pass, going to suggest that we refactor the tests into a class that can test parts of the pipeline and pass on (ie apply recipe -> load weights -> export ...)
(see https://docs.pytest.org/en/7.1.x/getting-started.html#group-multiple-tests-in-a-class)

[dbogunowicz]

We do not have yet system in place to catch this error in CI/CD, but this PR will not be complete unless we also merge:
neuralmagic/sparsezoo#141

[bfineran]

@dbogunowicz any reason not to use load_numpy_list? that should handle tar'd directories right?

[dbogunowicz]

@bfineran
Oh, only now I have realized that it load_numpy_list calls load_numpy_from_tar. You are absolutely right.

[KSGulin]

@dbogunowicz @bfineran Should transformers testing be added to GHA? It seems those tests are getting skipped, as they are a new category not yet included