diff --git a/src/rlhf/reward.py b/src/rlhf/reward.py
index cd03bef515333cf505158b39e8ae793ca70c3552..fd12d9dd563ff645f2a6a28169b848b57703236d 100644
--- a/src/rlhf/reward.py
+++ b/src/rlhf/reward.py
@@ -181,6 +181,10 @@ class RewardModel(pl.LightningModule):
         return reward
     
     def forward(self, x_p, x_a, m_p, m_a):
+        # 因为前向传播的时候，需要过两次模型。所以反馈的时候需要冻结其中一次的参数
+        # 不然梯度会被计算两次，在包含 deepspeed 框架下会报错
+        # 报错信息：Gradient computed twice for this partition.
+        
         with torch.enable_grad():
             prefer_reward = self.single_forward(x_p, prompt_mask=m_p)
         with torch.no_grad():