Fix past_key_values

modeling_chatglm.py

@@ -932,9 +932,10 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
             standardize_cache_format: bool = False,
     ) -> Dict[str, Any]:
         # update past_key_values
-        model_kwargs["past_key_values"] = self._extract_past_from_model_output(
+        cache_name, cache = self._extract_past_from_model_output(
             outputs, standardize_cache_format=standardize_cache_format
         )
+        model_kwargs[cache_name] = cache
 
         # update attention mask
         if "attention_mask" in model_kwargs: