tested after rebasing, fix importance sampling bug

applications/ColossalChat/examples/requirements.txt

@@ -1,4 +1,4 @@
 pandas>=1.4.1
 sentencepiece
-colossalai==0.4.7
+colossalai>=0.4.7
 prompt_toolkit

applications/ColossalChat/requirements.txt

@@ -1,4 +1,4 @@
-transformers==4.39.3
+transformers>=4.39.3
 tqdm
 datasets==2.14.7
 loralib

applications/ColossalChat/rl_example.py

@@ -152,7 +152,7 @@ if __name__ == "__main__":
         "-ei",
         "--eval-interval",
         type=int,
-        default=100,
+        default=-1,
         help="Interval for evaluation. Evaluate every ei training steps.",
     )
     parser.add_argument(

colossalai/shardformer/modeling/qwen3.py

@@ -273,6 +273,7 @@ class Qwen3PipelineForwards:
         hidden_states: Optional[torch.FloatTensor] = None,
         stage_index: Optional[List[int]] = None,
         shard_config: ShardConfig = None,
+        **kwargs
     ):
         r"""
         Args: