Use new parametrization based weight norm if available

[ezyang]

What does this PR do?

In pytorch/pytorch#103001 I introduce a new
parametrization based version of weight_norm. One big benefit of
the new API is that the resulting model is deepcopy'able; today, you can't
deepcopy Wav2Vec2 models.

Since the new API isn't even in PyTorch main yet, I'd like to feature
gate it here, so that it gets used whenever PyTorch is recent enough to
support it. It would be a big help for me if you could take this change
earlier rather than later; otherwise I will have to patch transformers
in our own CI to get our benchmark harness working on Wav2Vec2.

Signed-off-by: Edward Z. Yang [email protected]

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

cc @sanchit-gandhi

[sgugger]

LGTM as it's properly gated. @sanchit-gandhi can you have a second look?
You'll just need to run make style and make fix-copies on your branch to make the CI happy.

[HuggingFaceDocBuilderDev]

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

[ezyang]

PyTorch side PR has landed!

[sanchit-gandhi]

LGTM too - thanks for the PR @ezyang! Here's the PR in the PyTorch repo for reference: pytorch/pytorch#103001

[ydshieh]

@ezyang

With nightly pytorch, we get

AttributeError: encoder.pos_conv_embed.conv.weight_v not found in PyTorch model

when trying to load a pytorch model into a TF model.

The TF model is looking for encoder.pos_conv_embed.conv.weight_v but we now have encoder.pos_conv_embed.conv.parametrizations.weight.original0. (This is from our (TF)Wav2Vec2Model model class).

Question: In your PR pytorch/pytorch#103001, is this part

def _weight_norm_compat_hook()

that deals with the backward compatibility? If so, we will copy it :-)

[ezyang]

Yep. The change here is not FC so the ingester needs updating.