fix grpo multi turn tp (#3837)

Co-authored-by: hjh <[email protected]>

Author: hjh0119

swift/trainers/rlhf_trainer/grpo_trainer.py

@@ -762,7 +762,7 @@ def _fast_infer(self, inputs: InputsType) -> Tuple[InputsType, OutputsType]:
             outputs = []
         outputs = gather_object(outputs)
         if self.args.tensor_parallel_size > 1:
-            outputs = [item for output in outputs for item in output]
+            outputs = [[item] for output in outputs for item in output]
         outputs = self.reorder_outputs(outputs, distributed_idx)
         if self.args.sleep_level > 0 and self.infer_rank >= 0:
             self.engine.engine.sleep(level=self.args.sleep_level)