Question about token alignment

Hi! According to the paper, it seems that we should use per_step_index with highest logit to do minimal editing distance. However, in the code, you directly used input_ids, which might be counterintuitive. Could you explain that?