What should be shifted for decoder input for Bart

I think this is copy pasted from the causal language modeling doc (for which your labels are your inputs shifted to the right).