Change cache = true in config.json to significantly boost inference performance

config.json

@@ -19,6 +19,6 @@
   "tie_word_embeddings": false,
   "torch_dtype": "bfloat16",
   "transformers_version": "4.30.0.dev0",
-  "use_cache": false,
+  "use_cache": true,
   "vocab_size": 32000
 }