Merge efd8073450 into 3f1a7cc25b

trainer/train_agent.py

@@ -494,4 +494,6 @@ if __name__ == "__main__":
         else:
             rl_train_epoch(epoch, loader, len(loader), rollout_engine, ref_model, reward_model, 0, wandb, use_sglang = (args.rollout_engine == "sglang"))
 
-    if dist.is_initialized(): dist.destroy_process_group()
+    if dist.is_initialized():
+        dist.barrier()
+        dist.destroy_process_group()

trainer/train_distillation.py

@@ -250,4 +250,6 @@ if __name__ == "__main__":
             train_epoch(epoch, loader, len(loader), teacher_model, lm_config_student, 0, wandb, args.alpha, args.temperature)
     
     # ========== 9. 清理分布进程 ==========
-    if dist.is_initialized(): dist.destroy_process_group()
+    if dist.is_initialized():
+        dist.barrier()
+        dist.destroy_process_group()

trainer/train_dpo.py

@@ -230,4 +230,6 @@ if __name__ == "__main__":
             train_epoch(epoch, loader, len(loader), ref_model, lm_config, 0, wandb, args.beta)
     
     # ========== 9. 清理分布进程 ==========
-    if dist.is_initialized(): dist.destroy_process_group()
+    if dist.is_initialized():
+        dist.barrier()
+        dist.destroy_process_group()

trainer/train_full_sft.py

@@ -175,4 +175,6 @@ if __name__ == "__main__":
             train_epoch(epoch, loader, len(loader), 0, wandb)
     
     # ========== 9. 清理分布进程 ==========
-    if dist.is_initialized(): dist.destroy_process_group()
+    if dist.is_initialized():
+        dist.barrier()
+        dist.destroy_process_group()

trainer/train_grpo.py

@@ -336,4 +336,6 @@ if __name__ == "__main__":
             grpo_train_epoch(epoch, loader, len(loader), rollout_engine, ref_model, reward_model, 0, wandb, use_sglang = (args.rollout_engine == "sglang"))
     
     # ========== 9. 清理分布进程 ==========
-    if dist.is_initialized(): dist.destroy_process_group()
+    if dist.is_initialized():
+        dist.barrier()
+        dist.destroy_process_group()

trainer/train_lora.py

@@ -188,4 +188,6 @@ if __name__ == "__main__":
             train_epoch(epoch, loader, len(loader), lora_params, 0, wandb)
     
     # ========== 10. 清理分布进程 ==========
-    if dist.is_initialized(): dist.destroy_process_group()
+    if dist.is_initialized():
+        dist.barrier()
+        dist.destroy_process_group()

trainer/train_ppo.py

@@ -439,4 +439,6 @@ if __name__ == "__main__":
             ppo_train_epoch(epoch, loader, len(loader), rollout_engine, ref_model, actor_scheduler, critic_scheduler, reward_model, 0, wandb, use_sglang = (args.rollout_engine == "sglang"))
     
     # ========== 9. 清理分布进程 ==========
-    if dist.is_initialized(): dist.destroy_process_group()
+    if dist.is_initialized():
+        dist.barrier()
+        dist.destroy_process_group()

trainer/train_pretrain.py

@@ -173,4 +173,6 @@ if __name__ == "__main__":
             train_epoch(epoch, loader, len(loader), 0, wandb)
     
     # ========== 9. 清理分布进程 ==========
-    if dist.is_initialized(): dist.destroy_process_group()
+    if dist.is_initialized():
+        dist.barrier()
+        dist.destroy_process_group()