Fix the bug in generation config

If we follow the default generation configuration, it does not utilize the key-value cache during inference. This can cause the model to be too slow to generate text efficiently.

generation_config.json

@@ -1,5 +1,5 @@
 {
   "_from_model_config": true,
   "transformers_version": "4.34.1",
-  "use_cache": false
+  "use_cache": true
 }