Avoid duplicate input kwargs in `_decode`

modeling_minicpmo.py

@@ -636,6 +636,8 @@ class MiniCPMO(MiniCPMOPreTrainedModel):
         return self.llm(input_ids=None, position_ids=position_ids, inputs_embeds=vllm_embedding, **kwargs)
 
     def _decode(self, inputs_embeds, tokenizer, attention_mask, **kwargs):
+        kwargs.pop("output_hidden_states", None)
+        kwargs.pop("return_dict_in_generate", None)
         terminators = [tokenizer.convert_tokens_to_ids(i) for i in self.terminators]
         outputs = self.llm.generate(
             inputs_embeds=inputs_embeds,
@@ -777,6 +779,7 @@ class MiniCPMO(MiniCPMOPreTrainedModel):
         tokenizer=None,
         vision_hidden_states=None,
         stream=False,
+        decode_text=True,
         **kwargs,
     ):
         assert input_ids is not None
@@ -815,6 +818,9 @@ class MiniCPMO(MiniCPMOPreTrainedModel):
 
                 result = self._decode_text(outputs.sequences, tokenizer)
 
+        if decode_text is False:
+            return outputs
+        
         return result, outputs
 
     def chat(