Merge branch 'release/3.0-beta2' of https://github.com/PaddlePaddle/P…

…addleNLP into origin_release_3.0-beta2
DesmonDay · Nov 8, 2024 · 37f4bc5 · 37f4bc5
2 parents 32701d4 + d02c406
commit 37f4bc5
Show file tree

Hide file tree

Showing 3 changed files with 11 additions and 9 deletions.
diff --git 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
@@ -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
@@ -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 (