diff --git a/paddlenlp/trainer/trainer.py b/paddlenlp/trainer/trainer.py
index 700308013d0b..8d88e26ea37b 100644
--- a/paddlenlp/trainer/trainer.py
+++ b/paddlenlp/trainer/trainer.py
@@ -2764,7 +2764,9 @@ def _load_optimizer_and_scheduler(self, checkpoint):
                 else:
                     opt_state_dict = None
             else:
-                model = self.model_wrapped if self.args.enable_sharding_comm_overlap else self.model
+                model = self.model
+                if hasattr(self.args, "enable_sharding_comm_overlap") and self.args.enable_sharding_comm_overlap:
+                    model = self.model_wrapped
                 opt_state_dict = self.unified_checkpoint_handler.load_unified_optimizer(
                     model=model,
                     optimizer=self.optimizer,
diff --git a/paddlenlp/trl/dpo_criterion.py b/paddlenlp/trl/dpo_criterion.py
index 2af2a8ef2096..be454e2ce4d1 100644
--- a/paddlenlp/trl/dpo_criterion.py
+++ b/paddlenlp/trl/dpo_criterion.py
@@ -287,10 +287,10 @@ def forward(
         )
         loss = dpo_loss + sft_loss
         if self.use_infohub:
-            infohub.policy_chosen_logps.append(policy_chosen_logps)
-            infohub.policy_rejected_logps.append(policy_rejected_logps)
-            infohub.sft_loss.append(sft_loss)
-            infohub.dpo_loss.append(dpo_loss)
+            infohub.policy_chosen_logps.append(policy_chosen_logps.detach())
+            infohub.policy_rejected_logps.append(policy_rejected_logps.detach())
+            infohub.sft_loss.append(sft_loss.detach())
+            infohub.dpo_loss.append(dpo_loss.detach())
             return loss
         else:
             return policy_chosen_logps, policy_rejected_logps, sft_loss, dpo_loss, loss
diff --git a/paddlenlp/trl/kto_criterion.py b/paddlenlp/trl/kto_criterion.py
index 52745e996999..a6ca6c4c837a 100644
--- a/paddlenlp/trl/kto_criterion.py
+++ b/paddlenlp/trl/kto_criterion.py
@@ -247,10 +247,10 @@ def forward(
             reference_kl_logps,
         )
         if self.use_infohub:
-            infohub.policy_chosen_logps.append(policy_chosen_logps)
-            infohub.policy_rejected_logps.append(policy_rejected_logps)
-            infohub.policy_kl_logps.append(policy_kl_logps)
-            infohub.kl.append(kl)
+            infohub.policy_chosen_logps.append(policy_chosen_logps.detach())
+            infohub.policy_rejected_logps.append(policy_rejected_logps.detach())
+            infohub.policy_kl_logps.append(policy_kl_logps.detach())
+            infohub.kl.append(kl.detach())
             return loss
         else:
             return (