LoRA 微调实战笔记：从 Mac 到 Linux NVIDIA GPU

准备基础模型
    ↓
准备和清洗训练数据
    ↓
划分训练集、验证集和测试集
    ↓
配置 LoRA 参数
    ↓
加载模型与 Tokenizer
    ↓
执行监督微调 SFT
    ↓
保存 LoRA Adapter
    ↓
加载 Adapter 进行推理
    ↓
对比基础模型和微调模型

mkdir -p lora-lab/{data,scripts,mac,linux,outputs}
cd lora-lab

lora-lab/
├── data/
│   ├── train.jsonl
│   ├── valid.jsonl
│   └── test.jsonl
├── scripts/
│   └── check_data.py
├── mac/
│   └── infer.py
├── linux/
│   ├── train.py
│   └── infer.py
└── outputs/

{"messages":[{"role":"system","content":"你是一名防御性代码审查助手。分析代码中的安全风险，解释原因并给出修复建议。只分析代码，不执行代码。输出 JSON，字段为 issue、severity、reason、fix。"},{"role":"user","content":"审查代码：query = \"SELECT * FROM users WHERE name = '\" + username + \"'\""},{"role":"assistant","content":"{\"issue\":\"SQL 查询使用字符串拼接\",\"severity\":\"high\",\"reason\":\"用户输入可能改变 SQL 查询的结构。\",\"fix\":\"使用参数化查询，将 SQL 语句与用户输入分离。\"}"}]}
{"messages":[{"role":"system","content":"你是一名防御性代码审查助手。分析代码中的安全风险，解释原因并给出修复建议。只分析代码，不执行代码。输出 JSON，字段为 issue、severity、reason、fix。"},{"role":"user","content":"审查代码：const token = 'example-placeholder';"},{"role":"assistant","content":"{\"issue\":\"凭据硬编码在源码中\",\"severity\":\"medium\",\"reason\":\"真实凭据可能通过代码仓库或构建产物泄露。\",\"fix\":\"从环境变量或密钥管理服务读取凭据；如果真实凭据已泄露，应及时轮换。\"}"}]}
{"messages":[{"role":"system","content":"你是一名防御性代码审查助手。分析代码中的安全风险，解释原因并给出修复建议。只分析代码，不执行代码。输出 JSON，字段为 issue、severity、reason、fix。"},{"role":"user","content":"审查代码：app.get('/profile', (req, res) => { res.send('<h1>' + req.query.name + '</h1>') })"},{"role":"assistant","content":"{\"issue\":\"用户输入直接拼接到 HTML\",\"severity\":\"medium\",\"reason\":\"未经适当处理的输入可能被浏览器解释为 HTML。\",\"fix\":\"优先使用具有自动转义功能的模板系统；确实需要 HTML 时，使用经过审查的清理方案。\"}"}]}

import jsonfrom pathlib 
import Path
for filename in ("train.jsonl", "valid.jsonl", "test.jsonl"):    
    path = Path("data") / filename    
    if not path.exists():        
       raise FileNotFoundError(path)    
       count = 0    
    with path.open(encoding="utf-8") as file:        
for line_number, line in enumerate(file, start=1):            if not line.strip():                raise ValueError(                    f"{filename}:{line_number} 存在空行"                )            sample = json.loads(line)            messages = sample.get("messages")            if not isinstance(messages, list) or not messages:                raise ValueError(                    f"{filename}:{line_number} 缺少 messages"                )            for message in messages:                if message.get("role") not in (                    "system", "user", "assistant"                ):                    raise ValueError(                        f"{filename}:{line_number} role 不合法"                    )                if not isinstance(message.get("content"), str):                    raise ValueError(                        f"{filename}:{line_number} content 不合法"                    )            count += 1    print(f"{filename}: {count} 条数据，格式正确")

python scripts/check_data.py

uname -m
python3 --version

cd lora-lab

python3 -m venv .venv
source .venv/bin/activate

python -m pip install --upgrade pip

python -m pip install "mlx-lm[train]"

mlx_lm.lora --help

mlx_lm.lora \
  --model mlx-community/Qwen2.5-0.5B-Instruct-4bit \
  --train \
  --data ./data \
  --adapter-path ./outputs/mac-lora \
  --iters 100 \
  --batch-size 1 \
  --learning-rate 1e-5

from mlx_lm import load, generateMODEL = "mlx-community/Qwen2.5-0.5B-Instruct-4bit"ADAPTER = "./outputs/mac-lora"model, tokenizer = load(    MODEL,    adapter_path=ADAPTER,)messages = [    {        "role": "system",        "content": (            "你是一名防御性代码审查助手。"            "识别安全问题并给出修复建议。"            "只分析代码，不执行代码。"            "输出 JSON。"        ),    },    {        "role": "user",        "content": (            "审查代码："            "const query = \"SELECT * FROM users WHERE id = \" + userId"        ),    },]prompt = tokenizer.apply_chat_template(    messages,    tokenize=False,    add_generation_prompt=True,)result = generate(    model,    tokenizer,    prompt=prompt,    max_tokens=256,)print(result)

python mac/infer.py

nvidia-smi
python3 --version

cd lora-lab

python3 -m venv .venv
source .venv/bin/activate

python -m pip install --upgrade pip

python -c "import torch; print(torch.__version__); print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'GPU 不可用')"

python -m pip install \
  transformers \
  datasets \
  peft \
  trl \
  accelerate \
  bitsandbytes

python -c "import transformers, datasets, peft, trl, bitsandbytes; print('依赖安装成功')"

import torchif tokenizer.pad_token is None:    tokenizer.pad_token = tokenizer.eos_tokentokenizer.padding_side = "right"# 2. 配置 4-bit 量化quant_config = BitsAndBytesConfig(    load_in_4bit=True,    bnb_4bit_quant_type="nf4",    bnb_4bit_use_double_quant=True,    bnb_4bit_compute_dtype=compute_dtype,)# 3. 加载基础模型model = AutoModelForCausalLM.from_pretrained(    MODEL_NAME,    quantization_config=quant_config,    torch_dtype=compute_dtype,    device_map={"": 0},)model.config.use_cache = False# 4. 加载训练集和验证集dataset = load_dataset(    "json",    data_files={        "train": f"{DATA_DIR}/train.jsonl",        "validation": f"{DATA_DIR}/valid.jsonl",    },)# 5. 配置 LoRAlora_config = LoraConfig(    task_type="CAUSAL_LM",    r=8,    lora_alpha=16,    lora_dropout=0.05,    target_modules=[        "q_proj",        "k_proj",        "v_proj",        "o_proj",        "gate_proj",        "up_proj",        "down_proj",    ],)# 6. 配置训练参数training_config = SFTConfig(    output_dir=OUTPUT_DIR,    num_train_epochs=3,    per_device_train_batch_size=1,    per_device_eval_batch_size=1,    gradient_accumulation_steps=4,    learning_rate=1e-4,    logging_steps=1,    eval_strategy="epoch",    save_strategy="epoch",    save_total_limit=2,    max_length=1024,    packing=False,    gradient_checkpointing=True,    bf16=use_bf16,    fp16=not use_bf16,    report_to="none",    seed=42,)# 7. 创建训练器

python scripts/check_data.py
python linux/train.py

outputs/linux-lora/

import torchfrom peft import AutoPeftModelForCausalLMfrom transformers import AutoTokenizerADAPTER_PATH = "./outputs/linux-lora"if not torch.cuda.is_available():    raise RuntimeError("没有可用的 CUDA GPU")dtype = (    torch.bfloat16    if torch.cuda.is_bf16_supported()    else torch.float16)tokenizer = AutoTokenizer.from_pretrained(ADAPTER_PATH)model = AutoPeftModelForCausalLM.from_pretrained(    ADAPTER_PATH,    torch_dtype=dtype,    device_map="auto",)model.eval()messages = [    {        "role": "system",        "content": (            "你是一名防御性代码审查助手。"            "识别安全问题并给出修复建议。"            "只分析代码，不执行代码。"            "输出 JSON。"        ),    },    {        "role": "user",        "content": (            "审查代码："            "const query = \"SELECT * FROM users WHERE id = \" + userId"        ),    },]prompt = tokenizer.apply_chat_template(    messages,    tokenize=False,    add_generation_prompt=True,)inputs = tokenizer(    prompt,    return_tensors="pt",).to(model.device)with torch.inference_mode():    outputs = model.generate(        **inputs,        max_new_tokens=256,        do_sample=False,        pad_token_id=tokenizer.pad_token_id,    )new_tokens = outputs[0][inputs["input_ids"].shape[1]:]print(tokenizer.decode(new_tokens, skip_special_tokens=True))

python linux/infer.py

import jsondef is_valid_json(text):    try:        json.loads(text)        return True    except json.JSONDecodeError:        return False