7B模型LoRA/QLoRA微调全记录：从数据清洗到显存占用对比

Python: 3.10.12
CUDA: 12.1
PyTorch: 2.1.2+cu121
transformers: 4.40.1
peft: 0.10.0
bitsandbytes: 0.43.1
accelerate: 0.29.2
datasets: 2.19.0
trl: 0.9.6

{
  "instruction": "患者出现头晕、恶心、视物模糊，血压180/110mmHg，应如何处理？",
  "output": "立即给予硝苯地平10mg舌下含服，密切监测血压，同时排查高血压急症...",
  "category": "心血管内科"
}

from datasets import Dataset
import json

def format_chat(example):
    """将原始数据转为Qwen2的chat格式"""
    return {
        "text": f"system\n你是一位资深医疗专家，请根据用户描述给出专业建议。\n"
                f"user\n{example['instruction']}\n"
                f"assistant\n{example['output']}\n"
    }

# 加载并处理数据
with open('medical_qa.json', 'r', encoding='utf-8') as f:
    raw_data = json.load(f)

dataset = Dataset.from_list(raw_data)
dataset = dataset.map(format_chat, remove_columns=['instruction', 'output', 'category'])
dataset = dataset.train_test_split(test_size=0.1, seed=42)

print(f"训练集: {len(dataset['train'])} 条, 验证集: {len(dataset['test'])} 条")

import torch
from transformers import (
    AutoModelForCausalLM, 
    AutoTokenizer, 
    BitsAndBytesConfig,
    TrainingArguments,
    Trainer
)
from peft import (
    LoraConfig, 
    get_peft_model, 
    prepare_model_for_kbit_training
)

# QLoRA 4bit量化配置
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16
)

# 加载模型（关键：device_map="auto"）
model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2-7B-Instruct",
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-7B-Instruct")
tokenizer.pad_token = tokenizer.eos_token

# 冻结模型参数
model = prepare_model_for_kbit_training(model)

# LoRA配置
lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_dropout=0.1,
    bias="none",
    task_type="CAUSAL_LM"
)

model = get_peft_model(model, lora_config)
model.print_trainable_parameters()  # 应该显示约4.2M可训练参数

# 训练参数
training_args = TrainingArguments(
    output_dir="./qwen2-7b-medical-lora",
    num_train_epochs=3,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=4,  # 实际batch_size = 2*4 = 8
    learning_rate=2e-4,
    warmup_steps=100,
    logging_steps=10,
    eval_strategy="steps",
    eval_steps=200,
    save_strategy="steps",
    save_steps=500,
    fp16=True,
    max_grad_norm=0.3,
    report_to="tensorboard"
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["test"],
    data_collator=lambda data: tokenizer(
        [d["text"] for d in data],
        padding=True,
        truncation=True,
        max_length=1024,
        return_tensors="pt"
    )
)

trainer.train()

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2-7B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto"
)
model = prepare_model_for_kbit_training(model)  # 实际这里不需要，但保留也无妨