Gemma-2B 部署与微调

一、部署 Gemma-2B

1. 配置环境

conda create -n py38 python=3.8
conda activate py38

# 升级pip
python -m pip install --upgrade pip
# 更换 pypi 源加速库的安装
pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple

pip install fastapi==0.110.2
pip install uvicorn==0.29.0
pip install requests==2.31.0
pip install modelscope==1.11.0
pip install transformers==4.40.0
pip install accelerate==0.29.3

pip install peft
pip install SentencePiece

2. 模型下载

使用 modelscope 中的 snapshot_download 函数下载模型，第一个参数为模型名称，参数 cache_dir 为模型的下载路径。

import torch
from modelscope import snapshot_download, AutoModel, AutoTokenizer
import os
model_dir = snapshot_download('Lucachen/gemma2b', cache_dir='/mnt/disk/yqh/autodl-tmp', revision='master')

3. 代码准备

在 /mnt/disk/yqh/autodl-tmp 路径下新建 api.py 文件

from fastapi import FastAPI, Request
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import uvicorn
import json
import datetime
import torch

# 设置设备参数
DEVICE = "cuda"  # 使用CUDA
DEVICE_ID = "0"  # CUDA设备ID，如果未设置则为空
CUDA_DEVICE = f"{DEVICE}:{DEVICE_ID}" if DEVICE_ID else DEVICE  # 组合CUDA设备信息

# 清理GPU内存函数
def torch_gc():
    if torch.cuda.is_available():  # 检查是否可用CUDA
        with torch.cuda.device(CUDA_DEVICE):  # 指定CUDA设备
            torch.cuda.empty_cache()  # 清空CUDA缓存
            torch.cuda.ipc_collect()  # 收集CUDA内存碎片


# 创建FastAPI应用
app = FastAPI()

# 处理POST请求的端点
@app.post("/")
async def create_item(request: Request):
    global model, tokenizer, pipeline # 声明全局变量以便在函数内部使用模型和分词器
    json_post_raw = await request.json()  # 获取POST请求的JSON数据
    json_post = json.dumps(json_post_raw)  # 将JSON数据转换为字符串
    json_post_list = json.loads(json_post)  # 将字符串转换为Python对象
    prompt = json_post_list.get('prompt')  # 获取请求中的提示
    history = json_post_list.get('history', [])  # 获取请求中的历史记录

    messages = [
            # {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": prompt}
    ]

    # 调用模型进行对话生成
    prompt = pipeline.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    outputs = pipeline(
    prompt,
    max_new_tokens=1024,
    add_special_tokens=True,
    do_sample=True,
    temperature=0.7,
    top_k=50,
    top_p=0.95
    )
    response = outputs[0]["generated_text"][len(prompt):]

    now = datetime.datetime.now()  # 获取当前时间
    time = now.strftime("%Y-%m-%d %H:%M:%S")  # 格式化时间为字符串
    # 构建响应JSON
    answer = {
        "response": response,
        "status": 200,
        "time": time
    }
    # 构建日志信息
    log = "[" + time + "] " + '", prompt:"' + prompt + '", response:"' + repr(response) + '"'
    print(log)  # 打印日志
    torch_gc()  # 执行GPU内存清理
    return answer  # 返回响应

# 主函数入口
if __name__ == '__main__':
    # 加载预训练的分词器和模型
    model_name_or_path = '/mnt/disk/yqh/autodl-tmp/Lucachen/gemma2b'
    tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
    model = AutoModelForCausalLM.from_pretrained(model_name_or_path, device_map="auto", torch_dtype=torch.bfloat16).cuda()
    pipeline = pipeline("text-generation",model=model_name_or_path,model_kwargs={"torch_dtype": torch.bfloat16},device="cuda")

    # 启动FastAPI应用
    # 用6006端口可以将autodl的端口映射到本地，从而在本地使用api
    uvicorn.run(app, host='127.0.0.1', port=6006, workers=1)  # 在指定端口和主机上启动应用

    

 新建 test_api.py

import requests
import json

def get_completion(prompt):
    headers = {'Content-Type': 'application/json'}
    data = {"prompt": prompt}
    response = requests.post(url='http://127.0.0.1:6006', headers=headers, data=json.dumps(data))
    return response.json()['response']

if __name__ == '__main__':
    print(get_completion('你好'))

4. api 部署与测试

启动 api 服务，并使用 python 中的 requests 库进行调用测试。

cd /mnt/disk/yqh/autodl-tmp
python api.py

Python test_api.py

二、LoRA 微调

from datasets import Dataset
import pandas as pd
from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForSeq2Seq, TrainingArguments, Trainer, GenerationConfig
import torch
from peft import LoraConfig, TaskType, get_peft_model
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
from peft import PeftModel

# 将JSON文件转换为CSV文件
df = pd.read_json('./huanhuan.json')
ds = Dataset.from_pandas(df)
print(ds[:3])


# 处理数据集
tokenizer = AutoTokenizer.from_pretrained('/mnt/disk/yqh/autodl-tmp/Lucachen/gemma2b', use_fast=False, trust_remote_code=True)

def process_func(example):
    MAX_LENGTH = 384    # Llama分词器会将一个中文字切分为多个token，因此需要放开一些最大长度，保证数据的完整性
    input_ids, attention_mask, labels = [], [], []
    instruction = tokenizer(f"<start_of_turn>system\n现在你要扮演皇帝身边的女人--甄嬛<end_of_turn>\n<start_of_turn>user\n{example['instruction'] + example['input']}<end_of_turn>\n<start_of_turn>model\n", add_special_tokens=False)
    #response = tokenizer(f"{example['output']}", add_special_tokens=False)
    response = tokenizer(f"{example['output']}<end_of_turn>model", add_special_tokens=False)
    input_ids = instruction["input_ids"] + response["input_ids"]
    attention_mask = instruction["attention_mask"] + response["attention_mask"] + [1]  # 因为eos token咱们也是要关注的所以 补充为1
    labels = [-100] * len(instruction["input_ids"]) + response["input_ids"]
    if len(input_ids) > MAX_LENGTH:  # 做一个截断
        input_ids = input_ids[:MAX_LENGTH]
        attention_mask = attention_mask[:MAX_LENGTH]
        labels = labels[:MAX_LENGTH]
    return {
        "input_ids": input_ids,
        "attention_mask": attention_mask,
        "labels": labels
    }
    
tokenized_id = ds.map(process_func, remove_columns=ds.column_names)
print(tokenized_id)
print(tokenizer.decode(tokenized_id[0]['input_ids']))
tokenizer.decode(list(filter(lambda x: x != -100, tokenized_id[1]["labels"])))

# 创建模型
model = AutoModelForCausalLM.from_pretrained('/mnt/disk/yqh/autodl-tmp/Lucachen/gemma2b', device_map="auto",torch_dtype=torch.bfloat16)
print(model)
model.enable_input_require_grads() # 开启梯度检查点时，要执行该方法
print(model.dtype)

# LoRA
config = LoraConfig(
    task_type=TaskType.CAUSAL_LM, 
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    inference_mode=False, # 训练模式
    r=8, # Lora 秩
    lora_alpha=32, # Lora alaph，具体作用参见 Lora 原理
    lora_dropout=0.1# Dropout 比例
)
print(config)
model = get_peft_model(model, config)
print(config)
model.print_trainable_parameters()

# 配置训练参数
args = TrainingArguments(
    output_dir="./output/gemma2b",
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    logging_steps=10,
    num_train_epochs=3,
    save_steps=100,
    learning_rate=1e-4,
    save_on_each_node=True,
    gradient_checkpointing=True
)

trainer = Trainer(
    model=model,
    args=args,
    train_dataset=tokenized_id,
    data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),
)
trainer.train()

# 保存 LoRA 和 tokenizer 结果
peft_model_id="./gemma2b_lora"
trainer.model.save_pretrained(peft_model_id)
tokenizer.save_pretrained(peft_model_id)

# 加载 LoRA 权重推理
mode_path = '/mnt/disk/yqh/autodl-tmp/Lucachen/gemma2b'
lora_path = './gemma2b_lora'

# 加载tokenizer
tokenizer = AutoTokenizer.from_pretrained(mode_path)

# 加载模型
model = AutoModelForCausalLM.from_pretrained(mode_path, device_map="auto",torch_dtype=torch.bfloat16)

# 加载lora权重
model = PeftModel.from_pretrained(model, model_id=lora_path, config=config)

prompt = "你是谁？"
messages = [
    # {"role": "system", "content": "现在你要扮演皇帝身边的女人--甄嬛"},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

model_inputs = tokenizer([text], return_tensors="pt").to('cuda')

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

print(response)