support inference of llamafactory

modeling_minicpmo.py

@@ -595,7 +595,7 @@ class MiniCPMO(MiniCPMOPreTrainedModel):
         elif self.training:
             for i in range(bs):
                 # dummy audio_embeddings
-                input_embeddings += audio_embeddings[0].mean() * 0
+                input_embeddings = input_embeddings + audio_embeddings[0].mean() * 0
 
         return input_embeddings
 
@@ -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,
@@ -2655,6 +2655,7 @@ class ConditionalChatTTS(PreTrainedModel):
     """
 
     config_class = ConditionalChatTTSConfig
+    _no_split_modules = []
 
     def __init__(self, config: ConditionalChatTTSConfig):
         super().__init__(config)