[WhisperModel] fix bug in reshaping labels

[jonatasgrosman]

Currently, in the Whisper model's forward pass, the target labels are reshaped using the view method before being passed into the loss function:

transformers/src/transformers/models/whisper/modeling_whisper.py

Line 1214 in 1567bef

loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))

The view method requires the Torch Tensor to be contiguous, and certain operations are commonly performed on the labels that might cause them not to be contiguous.

So using the view can cause problems during model training. This issue has already been fixed on another model by the @sanchit-gandhi in a PR, and I'm just replicating the same solution (using reshape() instead of view()) here for the Whisper model.

[HuggingFaceDocBuilderDev]

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

[sgugger]

Also cc @ArthurZucker

[ArthurZucker]

LGTM thanks