beam_search.pyfile, the process function (in actual classes), have the follwoing signature
def process( self, input_ids: torch.LongTensor, next_scores: torch.FloatTensor, next_tokens: torch.LongTensor, next_indices: torch.LongTensor, pad_token_id: Optional[int] = None, eos_token_id: Optional[Union[int, List[int]]] = None, beam_indices: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor]:
even though it returns
UserDict. Any reason why not annotate with Dict, Mapping rather than Tuple?