diff --git a/train_sft_ds.py b/train_sft_ds.py
index 4962edc..c1a4500 100644
--- a/train_sft_ds.py
+++ b/train_sft_ds.py
@@ -672,9 +672,15 @@ def main():
     trainer.add_callback(CsvLossLogger(csv_path=os.path.join(args.output_dir, "loss.csv")))
 
     # 无共享盘：各节点本地 output_dir 下是否已有 checkpoint-*
-    ckpt_exists = (os.path.isdir(args.output_dir)
-                   and any(n.startswith("checkpoint-") for n in os.listdir(args.output_dir)))
-    resume_flag = True if ckpt_exists else None
+    # ckpt_exists = (os.path.isdir(args.output_dir)
+    #                and any(n.startswith("checkpoint-") for n in os.listdir(args.output_dir)))
+    # resume_flag = True if ckpt_exists else None
+
+    ckpt_local = 1 if (os.path.isdir(args.output_dir) and any(n.startswith("checkpoint-") for n in os.listdir(args.output_dir))) else 0
+    ckpt_tensor = torch.tensor(ckpt_local, device=(f"cuda:{local_rank}" if torch.cuda.is_available() and local_rank >= 0 else "cpu"))
+    if dist.is_available() and dist.is_initialized():
+        dist.all_reduce(ckpt_tensor, op=dist.ReduceOp.MAX)
+    resume_flag = True if ckpt_tensor.item() > 0 else None   
 
     print_once(f"[host={socket.gethostname()}] Resume = {resume_flag is True}")
     print_once("***** Starting training *****")