Why is "padding_mask_p1"

[jackfsuia]

Hi, trl community. In the PPO, trl has the following code, which i think might be problematic, because the reward is added on one position after the last EOS token, not on the EOS token like it use to be. You can see the actual_end below doesn't seem to be right?

            # be very careful with `padding_mask_p1`; see https://excalidraw.com/#json=LWnzG4w2k5DjF_EOL_xPt,e2w3a-hFJ_gX5vOfeyXGTw
                response_idxs = torch.arange(responses.shape[1], device=responses.device).repeat(responses.shape[0], 1)
                padding_mask = response_idxs > sequence_lengths.unsqueeze(1)
                logprobs = torch.masked_fill(logprobs, padding_mask, INVALID_LOGPROB)
                ref_logprobs = torch.masked_fill(ref_logprobs, padding_mask, INVALID_LOGPROB)
                sequence_lengths_p1 = sequence_lengths + 1
                padding_mask_p1 = response_idxs > (sequence_lengths_p1.unsqueeze(1))
                values = torch.masked_fill(values, padding_mask_p1, 0)

                # 4. compute rewards
                kl = logprobs - ref_logprobs
                non_score_reward = -args.kl_coef * kl
                rewards = non_score_reward.clone()
                actual_start = torch.arange(rewards.size(0), device=rewards.device)
                actual_end = torch.where(sequence_lengths_p1 < rewards.size(1), sequence_lengths_p1, sequence_lengths)
                rewards[[actual_start, actual_end]] += scores

What does padding_mask_p1 mean ? The link attached above does not seem to be a detailed explanation, with just one graph. Thank you for your help.