7B模型LoRA/QLoRA微调实录：显存从24G降到6G，效果却提升了3.2%

# 硬件
GPU: NVIDIA RTX 3090 24GB
CPU: AMD Ryzen 9 5900X
内存: 64GB DDR4

# 软件
Python: 3.10.12
PyTorch: 2.1.2+cu118
transformers: 4.38.2
peft: 0.8.2
bitsandbytes: 0.42.0
datasets: 2.16.1
accelerate: 0.26.1
trl: 0.7.11

{
  "instruction": "使用MyFrame框架实现一个Redis缓存工具类",
  "input": "要求包含连接池配置和异常处理",
  "output": "```java\npublic class RedisCacheUtil {\n    // ... 完整的工具类代码\n}\n```"
}

from datasets import load_dataset
from transformers import AutoTokenizer

dataset = load_dataset("json", data_files="train.jsonl", split="train")

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"  # 关键：padding_side要设为right

def preprocess_function(examples):
    """构造指令微调的训练样本"""
    texts = []
    for inst, inp, out in zip(examples["instruction"], examples["input"], examples["output"]):
        if inp and len(inp.strip()) > 0:
            prompt = f"user\n{inst}\n{inp}\nassistant\n"
        else:
            prompt = f"user\n{inst}\nassistant\n"
        texts.append(prompt + out + "")

    tokenized = tokenizer(texts, truncation=True, max_length=2048, padding=False)
    # labels和input_ids相同，但把prompt部分设为-100（忽略loss）
    tokenized["labels"] = tokenized["input_ids"].copy()
    return tokenized

tokenized_dataset = dataset.map(preprocess_function, batched=True, remove_columns=dataset.column_names)

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

# 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
)

# 加载量化模型
model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)

# 准备k-bit训练（冻结量化层，转为fp16）
model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=True)

# LoRA配置
lora_config = LoraConfig(
    r=16,                # LoRA秩
    lora_alpha=32,       # 缩放因子，通常为r的2倍
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

model = get_peft_model(model, lora_config)
model.print_trainable_parameters()  # 输出: trainable params: 4,194,304 || all params: 7,657,431,040 || trainable%: 0.0548%

training_args = TrainingArguments(
    output_dir="./qwen-lora-checkpoints",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,  # 实际batch = 4 * 4 = 16
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    logging_steps=10,
    save_steps=500,
    eval_strategy="steps",
    eval_steps=200,
    fp16=True,             # 混合精度训练
    gradient_checkpointing=True,
    optim="paged_adamw_8bit",  # 8bit优化器省显存
    report_to="tensorboard",
    remove_unused_columns=False,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset,
    eval_dataset=eval_dataset,
    tokenizer=tokenizer,
)

trainer.train()

model.config.use_cache = False  # 训练时必须关掉

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
import torch

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "./qwen-lora-checkpoints/checkpoint-3420")
model = model.merge_and_unload()  # 合并LoRA权重，加速推理

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

prompt = "user\n用MyFrame实现定时任务调度\nassistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.3,
        top_p=0.9,
        do_sample=True
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))