Fix DAC conversion script

[ebezzam]

What does this PR do

1. Fix DAC conversion:

• Most notably, performing weight norm removal on GPU instead of on CPU (otherwise get differences for layers with weight norm when applying models on GPU)
• Missing feature extractor parameters
• Correctly casting sampling rate

2. More consistent add/remove weight norm functions
3. Update explanation of high tolerances during testing. We now know it comes from weight norm removal on CPU (instead of GPU) and different implementations of Snake1d (their version uses JIT). Nevertheless, we stick with current models on the Hub, as differences are minimal.

Reproducer to show weight norm difference when doing weight removal on a different device: https://gist.github.com/ebezzam/c83f186dcfeaab8cac040c960eb474cd

[ebezzam]

Updated (definite) reason for high tolerances

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[vasqu]

We should not update the conversion if we don't change the hub. Having a legacy path is unideal and makes it confusing for the average user as hub differs from the script here.

There are two option imo:

• Change to new conversion (no extra flags) and update hub weights
• Only leave the description where the differences stem from

I'd prefer option 1 even if it was breaking tbh. Would wait on Eustache here tbh

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: dac

[ebezzam]

thanks @vasqu!

🚨 @eustlb (when you're back), @vasqu and I spoke offline that it would be better to:

• ask Descript to update model weights (with conversion done on GPU)
• switch to Snake 1D with JIT

Main reason being that several models are depending on DAC (XCodec, Dia, Higgs Boson, maybe more), and it would be better that they don't depend on a model with minor output differences. As model addition/integration will be trickier since we may not be able to isolate if differences are coming from DAC or from implementing the new model.