LoRa weights are not merged back in to the base model ?

[n0w0f]

Hi crystal-text-llm team,

Thanks for this interesting work !

I was going through this code base and noticed that in the prepare_model_and_tokenizer function in llama-sample.py script, LoRa weights are not merged back in to the base model. Is there any reason for not doing so ?

model = PeftModel.from_pretrained(model, args.model_path, device_map="auto")
    
 return model, tokenizer