[VITS] Fix nightly tests

[sanchit-gandhi]

What does this PR do?

Fixes the tokenizer integration test and multi-GPU test that failed on the nightly run: https://github.com/huggingface/transformers/actions/runs/6068405445/job/16461410319

The tokenizer fix is trivial (needed to update the commit ID)!

The multi-GPU test was failing because the output sequence length for VITS is a function of the model inputs, rather than being a function of the input sequence lengths only.

Let's say we have 2 GPUs over which we want to run DP:

• GPU 1 outputs a sequence length of N, which is computed based on the input in the first element of the batch x
• GPU 2 outputs a sequence length of M, which is computed based on the input in the second element of the batch y

=> there is nothing to enforce that N = M, since the VITS output sequence length is a function of the inputs. Thus, we cannot concatenate the inputs after running the forward pass, since they have different dims.

# pseudo code for data parallelism
input_1, input_2 = torch.split(input, 2, dim=0)

output_1 = model(input_1) 
output_2 = model(input_2)

output = torch.concatenate([output_1, output_2], dim=0)  # breaks because input_1 and input_2 have different sequence lengths

The fix for the test is to pass the same inputs to both GPUs, and disable the stochastic duration predictor. This way, we get consistent outputs across our GPUs.

[HuggingFaceDocBuilderDev]

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

[ydshieh]

Thanks a lot. Just a few nits for the comments.

[ArthurZucker]

Thanks! The tokenizers integration tests should rather use k as padding as we discussed!