Load model in 4bit with bitsandbytes

Use LlamaTokenizerFast in handler

Fresh handler with LlamaTokenizer

Delete old handler

Create handler.py