赞
踩
目录
记录ChatGLM3-6B部署及官方Lora微调示例详细步骤及如何使用微调后的模型进行推理
使用git clone 命令下载源码
git clone https://github.com/THUDM/ChatGLM3.git
如图所示
模型权重文件从魔塔进行下载,不需要翻墙。权重文件比较大,所以花费时间也比较长,请耐心等待。
- git lfs install
- git clone https://www.modelscope.cn/ZhipuAI/chatglm3-6b.git
使用pwd命令获取模型路径,这个路径后面需要用到:
- pwd
- /mnt/workspace/chatglm3-6b
使用命令方式启动,启动之前需要修改模型地址配置。在路径 ChatGLM3/basic_demo 下找到文件 cli_demo.py 文件,修改MODEL_PATH,修改后的路径就是第二步【下载模型】后使用 pwd 命令查询出来的路径。
启动之前安装依赖
- cd 到 ChatGLM3 路径下
-
- pip install -r requirements.txt
使用一下命令启动并验证,第一次启动会略慢
- cd 到 basic_demo 路径下
-
- python cli_demo.py
经过多次尝试,微调的GPU显存应不小于24G,不然容易报OOM等错误。微调参数意义参考: ChatGLM3/finetune_demo at main · THUDM/ChatGLM3 (github.com)
首先先安装微调的依赖
- cd 到目录 ChatGLM3/finetune_demo
-
- pip install -r requirements.txt
上传数据
转换数据,调整为标准的对话格式
- import json
- from typing import Union
- from pathlib import Path
-
-
- def _resolve_path(path: Union[str, Path]) -> Path:
- return Path(path).expanduser().resolve()
-
-
- def _mkdir(dir_name: Union[str, Path]):
- dir_name = _resolve_path(dir_name)
- if not dir_name.is_dir():
- dir_name.mkdir(parents=True, exist_ok=False)
-
-
- def convert_adgen(data_dir: Union[str, Path], save_dir: Union[str, Path]):
- def _convert(in_file: Path, out_file: Path):
- _mkdir(out_file.parent)
- with open(in_file, encoding='utf-8') as fin:
- with open(out_file, 'wt', encoding='utf-8') as fout:
- for line in fin:
- dct = json.loads(line)
- sample = {'conversations': [{'role': 'user', 'content': dct['content']},
- {'role': 'assistant', 'content': dct['summary']}]}
- fout.write(json.dumps(sample, ensure_ascii=False) + '\n')
-
- data_dir = _resolve_path(data_dir)
- save_dir = _resolve_path(save_dir)
-
- train_file = data_dir / 'train.json'
- if train_file.is_file():
- out_file = save_dir / train_file.relative_to(data_dir)
- _convert(train_file, out_file)
-
- dev_file = data_dir / 'dev.json'
- if dev_file.is_file():
- out_file = save_dir / dev_file.relative_to(data_dir)
- _convert(dev_file, out_file)
-
-
- convert_adgen('data/AdvertiseGen', 'data/AdvertiseGen_fix')
得到转换后的训练和验证数据:
使用以下命令开始训练, data/AdvertiseGen_fix - 微调数据路径; /mnt/workspace/chatglm3-6b - 模型权重路径
- cd 到 finetune_demo 目录下
-
- CUDA_VISIBLE_DEVICES=0 NCCL_P2P_DISABLE="1" NCCL_IB_DISABLE="1" python finetune_hf.py data/AdvertiseGen_fix /mnt/workspace/chatglm3-6b configs/lora.yaml
训练中,根据数据量和参数设置的不同而花费的时间不同,我大概花了1个小时
验证微调后的效果
CUDA_VISIBLE_DEVICES=0 NCCL_P2P_DISABLE="1" NCCL_IB_DISABLE="1" python inference_hf.py output/checkpoint-3000/ --prompt "类型#裙*版型#显瘦*材质#网纱*风格#性感*裙型#百褶*裙下摆#压褶*裙长#连衣裙*裙衣门襟#拉链*裙衣门襟#套头*裙款式#拼接*裙款式#拉链*裙款式#木耳边*裙款式#抽褶*裙款式#不规则"
如果想要在 basic_demo 路径下的各demo中结合使用微调后的模型,需要修改 basic_demo/ 下的*_demo.py代码,即使用 finetune_demo/inference_hf 中的 方法 load_model_and_tokenizer 替换各demo里面获取 model 和 tokenizer的方法
- def load_model_and_tokenizer(
- model_dir: Union[str, Path], trust_remote_code: bool = True
- ) -> tuple[ModelType, TokenizerType]:
- model_dir = _resolve_path(model_dir)
- if (model_dir / 'adapter_config.json').exists():
- model = AutoPeftModelForCausalLM.from_pretrained(
- model_dir, trust_remote_code=trust_remote_code, device_map='auto'
- )
- tokenizer_dir = model.peft_config['default'].base_model_name_or_path
- else:
- model = AutoModelForCausalLM.from_pretrained(
- model_dir, trust_remote_code=trust_remote_code, device_map='auto'
- )
- tokenizer_dir = model_dir
- tokenizer = AutoTokenizer.from_pretrained(
- tokenizer_dir, trust_remote_code=trust_remote_code
- )
- return model, tokenizer
以 basic_demo/cli_demo.py 为例,暴力粘合后的代码如下:
- import os
- import platform
- from pathlib import Path
- from typing import Annotated, Union
- from peft import AutoPeftModelForCausalLM, PeftModelForCausalLM
- from transformers import (
- AutoModelForCausalLM,
- AutoTokenizer,
- PreTrainedModel,
- PreTrainedTokenizer,
- PreTrainedTokenizerFast,
- )
-
- ModelType = Union[PreTrainedModel, PeftModelForCausalLM]
- TokenizerType = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
- def _resolve_path(path: Union[str, Path]) -> Path:
- return Path(path).expanduser().resolve()
-
- MODEL_PATH = os.environ.get('MODEL_PATH', 'THUDM/chatglm3-6b')
- TOKENIZER_PATH = os.environ.get("TOKENIZER_PATH", MODEL_PATH)
-
- def load_model_and_tokenizer(model_dir: Union[str, Path]) -> tuple[ModelType, TokenizerType]:
- model_dir = _resolve_path(model_dir)
- if (model_dir / 'adapter_config.json').exists():
- model = AutoPeftModelForCausalLM.from_pretrained(
- model_dir, trust_remote_code=True, device_map='auto'
- )
- tokenizer_dir = model.peft_config['default'].base_model_name_or_path
- else:
- model = AutoModelForCausalLM.from_pretrained(
- model_dir, trust_remote_code=True, device_map='auto'
- )
- tokenizer_dir = model_dir
- tokenizer = AutoTokenizer.from_pretrained(
- tokenizer_dir, trust_remote_code=True
- )
- return model, tokenizer
-
- # tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_PATH, trust_remote_code=True)
- # model = AutoModel.from_pretrained(MODEL_PATH, trust_remote_code=True, device_map="auto").eval()
-
- finetune_path = '/mnt/workspace/ChatGLM3/finetune_demo/output/checkpoint-3000'
- model, tokenizer = load_model_and_tokenizer(finetune_path)
-
- # add .quantize(bits=4, device="cuda").cuda() before .eval() to use int4 model
- # must use cuda to load int4 model
-
- os_name = platform.system()
- clear_command = 'cls' if os_name == 'Windows' else 'clear'
- stop_stream = False
-
- welcome_prompt = "欢迎使用 ChatGLM3-6B 模型,输入内容即可进行对话,clear 清空对话历史,stop 终止程序"
-
-
-
- def build_prompt(history):
- prompt = welcome_prompt
- for query, response in history:
- prompt += f"\n\n用户:{query}"
- prompt += f"\n\nChatGLM3-6B:{response}"
- return prompt
-
-
- def main():
- past_key_values, history = None, []
- global stop_stream
- print(welcome_prompt)
- while True:
- query = input("\n用户:")
- if query.strip() == "stop":
- break
- if query.strip() == "clear":
- past_key_values, history = None, []
- os.system(clear_command)
- print(welcome_prompt)
- continue
- print("\nChatGLM:", end="")
- current_length = 0
- for response, history, past_key_values in model.stream_chat(tokenizer, query, history=history, top_p=1,
- temperature=0.01,
- past_key_values=past_key_values,
- return_past_key_values=True):
- if stop_stream:
- stop_stream = False
- break
- else:
- print(response[current_length:], end="", flush=True)
- current_length = len(response)
- print("")
-
-
- if __name__ == "__main__":
- main()
其中,注释掉的为原获取 model 和 tokenizer的方法
最后使用 python cli_demo.py执行测试
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。