modify one line to support full sft and inference of llamafactory

modeling_minicpmo.py

@@ -751,7 +751,7 @@ class MiniCPMO(MiniCPMOPreTrainedModel):
         input_ids=None,
         pixel_values=None,
         tgt_sizes=None,
-        audio_features=None,
+        audio_features=[],
         audio_feature_lens=None,
         image_bound=None,
         audio_bounds=None,
@@ -3256,4 +3256,4 @@ def prepare_inputs_for_generation(
             "attention_mask": attention_mask,
         }
     )
-    return model_inputs
+    return model_inputs