大模型 lora 训练后的权重合并

本文将介绍如下内容：

• 一、Lora 合并的环境依赖
• 二、Lora 合并代码实现

一、Lora 合并的环境依赖

autoawq                           0.2.5
peft                              0.14.0
transformers                      4.46.3

二、Lora 合并代码实现

import torch
from peft import PeftModel
from transformers import AutoTokenizer, AutoModelForCausalLM, LlamaTokenizer
from transformers.generation.utils import GenerationConfig


def apply_lora(model_name_or_path, output_path, lora_path):
    print(f"Loading the base model from {model_name_or_path}")
    base_tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=False, trust_remote_code=True)
    # base = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.bfloat16, trust_remote_code=True)
    base = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.float32, trust_remote_code=True)

    print(f"Loading the LoRA adapter from {lora_path}")

    lora_model = PeftModel.from_pretrained(
        base,
        lora_path,
        torch_dtype=torch.float32,
    )

    print("Applying the LoRA")
    model = lora_model.merge_and_unload()

    print(f"Saving the target model to {output_path}")
    model.save_pretrained(output_path)
    base_tokenizer.save_pretrained(output_path)


if __name__ == "__main__":
    model_name_or_path = "/nasdata/zhanjie/models/Qwen2.5-32B-Instruct"
    output_path = "./output"
    lora_path = "/nasdata/zhanjie/models/qwen2_5-32B-lora-1-0-7"
    apply_lora(model_name_or_path, output_path, lora_path)
    print("Done")