A bug in the implementation of the top-p sampling

https://github.com/lucidrains/PaLM-rlhf-pytorch/blob/6b02ee329106baff78e293afa7d1d2e6dd4e5ca2/palm_rlhf_pytorch/utils.py#L60

Using the sorted indices to index the sorted indices does not make sense. I think it may be return logits.scatter(1, sorted_indices, sorted_logits)