Fix tokenizer_class: LlamaTokenizer -> LlamaTokenizerFast (root cause of garbled v2 output, per Dipankar's byte-decoder analysis)

tokenizer_config.json

@@ -188,7 +188,7 @@
   "model_max_length": 16384,
   "pad_token": "<｜end▁of▁sentence｜>",
   "sp_model_kwargs": {},
-  "tokenizer_class": "LlamaTokenizer",
+  "tokenizer_class": "LlamaTokenizerFast",
   "unk_token": null,
   "use_default_system_prompt": false
-}
+}