fix padding in dpo trainer

[pacman100]

What does this PR do?

1. The padding token in DPO trainer is wrongly set to 0 when tokenizer uses a different token id for padding and tokenizer is passed. This happens because the conditional statement is checking if padding value is None but if the default is 0, it is never None.

[pacman100]

cc @younesbelkada

[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.

[younesbelkada]

Thanks, in principle this looks good - I would love though a second look from @kashif - let me know what do you think and once that's good for you, feel free to merge!

[kashif]

great catch thanks!