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

…addleNLP into origin_release_3.0-beta2
PaddlePaddle · Nov 8, 2024 · 5a03c3b · 5a03c3b
2 parents 32701d4 + d02c406
commit 5a03c3b
Show file tree

Hide file tree

Showing 4 changed files with 13 additions and 11 deletions.
diff --git a/paddlenlp/trainer/trainer.py b/paddlenlp/trainer/trainer.py
@@ -1988,7 +1988,7 @@ def get_expected_keys(inputs, keys):
                 self.optimizer = mix_precision_utils.MixPrecisionOptimizer(self.optimizer)
             self.optimizer = fleet.distributed_optimizer(self.optimizer)
 
-            if self.args.enable_sharding_comm_overlap:
+            if hasattr(self.args, "enable_sharding_comm_overlap") and self.args.enable_sharding_comm_overlap:
                 model.register_sharding_comm_overlap_hook(self.optimizer)
 
         # No pipeline mode, sharding only
@@ -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/trainer/unified_checkpoint/sharding_split_param_utils.py b/paddlenlp/trainer/unified_checkpoint/sharding_split_param_utils.py
@@ -181,7 +181,7 @@ def load_unified_optimizer_split_param(args, model, optimizer, resume_from_check
     param_shape_info = {}
 
     comm_buffer_list = optimizer._inner_opt._comm_buffer_list
-    if args.enable_sharding_comm_overlap:
+    if hasattr(args, "enable_sharding_comm_overlap") and args.enable_sharding_comm_overlap:
         comm_buffer_list = list(chain(*model._chunk_2_comm_buffers.values()))
         model = unwrap_model(model)
 

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 (