merge_script.py

import torch
from peft import PeftModel  # Ensure you have 'peft' library or modify according to your setup
import os
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
import argparse
from utils import get_logger  # Ensure this is implemented in your environment
import json

logger = get_logger("merge", "info")

def smart_tokenizer_and_embedding_resize(tokenizer, model, custom_tokens_path=None):
    """Resize tokenizer and embedding to accommodate new tokens."""
    special_tokens_dict = {
        "pad_token": "[PAD]",
        "eos_token": "</s>",
        "bos_token": "<s>",
        "unk_token": "<unk>"
    }
    
    # Load custom tokens if specified
    custom_tokens = []
    if custom_tokens_path is not None:
        with open(custom_tokens_path, 'r') as file:
            custom_tokens = [line.strip() for line in file.readlines()]

    num_added_toks = tokenizer.add_special_tokens(special_tokens_dict)
    if custom_tokens:
        num_added_toks += tokenizer.add_tokens(custom_tokens, special_tokens=True)
    
    model.resize_token_embeddings(len(tokenizer))
    logger.info(f"Resized tokenizer and model embeddings. Added {num_added_toks} tokens.")

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("-bm", "--base_model", type=str, default="meta-llama/Llama-2-7b-chat-hf", help="Base model name or path")
    parser.add_argument("-lm", "--lora_model", type=str, required=True, help="Path to the Lora model directory")
    parser.add_argument("-o", "--output", type=str, required=True, help="Output directory for the merged model")
    parser.add_argument("--custom_tokens", type=str, default=None, help="Path to a file containing custom tokens")
    args = parser.parse_args()

    if not os.path.exists(args.lora_model):
        raise FileNotFoundError(f"LoRA model directory {args.lora_model} not found.")
    
    os.makedirs(args.output, exist_ok=True)
    
    # Load the base model and tokenizer
    model = AutoModelForCausalLM.from_pretrained(args.base_model)
    tokenizer = AutoTokenizer.from_pretrained(args.base_model)

    # Adjust tokenizer and model for any additional tokens
    smart_tokenizer_and_embedding_resize(tokenizer, model, args.custom_tokens)

    # Load and merge the LoRA model
    logger.info("Loading and merging the LoRA model...")
    lora_model = PeftModel.from_pretrained(model, args.lora_model, merge_with_base=True)
    
    # Save the merged model and tokenizer
    lora_model.save_pretrained(args.output)
    tokenizer.save_pretrained(args.output)
    logger.info(f"Merged model saved to {args.output}")

if __name__ == "__main__":
    main()