diff --git a/internlm/core/engine.py b/internlm/core/engine.py
index b92b7f8..eb33e35 100644
--- a/internlm/core/engine.py
+++ b/internlm/core/engine.py
@@ -186,7 +186,7 @@ class Engine:
         if to_gpu:
             batch_data = move_to_device(batch_data)
 
-        # For packed-dataset, batch_data is (micro_num, micro_num*micro_bsz),
+        # For packed-dataset, batch_data is (micro_num, micro_bsz*seq_len),
         #   therefore 'batch_size' is equal to 'micro_num'
         # For nopacked-dataset, batch_data is (micro_num*micro_bsz, seq_len),
         #   therefore 'batch_size' is equal to 'micro_num*micro_bsz'
diff --git a/internlm/core/scheduler/pipeline_scheduler.py b/internlm/core/scheduler/pipeline_scheduler.py
index 5a6da61..5b864ff 100644
--- a/internlm/core/scheduler/pipeline_scheduler.py
+++ b/internlm/core/scheduler/pipeline_scheduler.py
@@ -193,6 +193,9 @@ class PipelineScheduler(BaseScheduler):
         # Pipeline schedule just puts data in memory,
         batch_data, actual_batch_size = engine.load_batch(data_iter, to_gpu=False)
 
+        # Even if 'use_flash_attn' is False, the data seen when the 'load_batch' is called is still packed,
+        # because internlm's current train dataset is packed, even using dummy data.
+        # The unpack operation is performed in load_micro_batch().
         if check_data_is_packed(batch_data):
             micro_num = actual_batch_size
         else: