[Trainer] Add time statistics for nccl-connection. (PaddlePaddle#6673)

paddlenlp/trainer/training_args.py

@@ -20,6 +20,7 @@
 import json
 import math
 import os
+import time
 import types
 import warnings
 from dataclasses import asdict, dataclass, field
@@ -883,7 +884,12 @@ def __post_init__(self):
                             "The enable_stage1_tensor_fusion or enable_stage1_overlap is not supported "
                             "by current version of Paddle. Please try latest develop Paddle."
                         )
+                paddle.device.cuda.synchronize()
+                start_time = time.time()
                 fleet.init(is_collective=True, strategy=strategy)
+                paddle.device.cuda.synchronize()
+                elapsed = time.time() - start_time
+                logger.info("NCCL-Connection costs {:.2f} ms.".format(elapsed))
 
                 logger.info(strategy)