[bugfix] fix grpo-padding-free get_logps(#6275)

Author: hjh0119

swift/trainers/rlhf_trainer/grpo_trainer.py

@@ -1838,6 +1838,9 @@ def _get_per_token_logps_and_entropies_single(self,
         use_local_entropy = not hasattr(super(), '_get_per_token_logps_and_entropies') and compute_entropy
 
         can_use_super = (not self.is_multimodal and 'logits_to_keep' in parameters and not use_local_entropy)
+        if 'attention_mask' not in inputs:
+            # when set padding_free true, the attention_mask is not in inputs
+            can_use_super = False
 
         if can_use_super:
             # save memory