7B模型LoRA/QLoRA微调全记录：显存从24G降到6G，效果不减反增

Python 3.10.12
torch 2.1.2+cu118
transformers 4.36.2
peft 0.7.1
bitsandbytes 0.41.3
accelerate 0.25.0
datasets 2.16.1

import torch
from transformers import (
    AutoModelForCausalLM, AutoTokenizer, 
    TrainingArguments, BitsAndBytesConfig
)
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from datasets import load_dataset
from trl import SFTTrainer

# 量化配置 - QLoRA核心
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16
)

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)
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=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

# 数据加载
dataset = load_dataset("json", data_files="device_logs.jsonl", split="train")

def format_fn(example):
    text = f"### 设备型号: {example['device']}\\n### 错误码: {example['error_code']}\\n### 维修建议: {example['advice']}"
    return {"text": text}

dataset = dataset.map(format_fn)

# 训练参数 - 关键配置
training_args = TrainingArguments(
    output_dir="./qlora-device-logs",
    per_device_train_batch_size=4,
    per_device_eval_batch_size=4,
    gradient_accumulation_steps=4,
    num_train_epochs=3,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    bf16=True,
    logging_steps=50,
    eval_strategy="steps",
    eval_steps=200,
    save_steps=500,
    max_grad_norm=0.3,
    remove_unused_columns=False,
    report_to="tensorboard",
)

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    args=training_args,
    train_dataset=dataset,
    dataset_text_field="text",
    max_seq_length=1024,
)

trainer.train()

from peft import PeftModel
import torch

# 加载原始模型和LoRA权重
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "./qlora-device-logs/checkpoint-2000")

def generate(device, error_code):
    prompt = f"### 设备型号: {device}\\n### 错误码: {error_code}\\n### 维修建议:"
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    outputs = model.generate(
        **inputs,
        max_new_tokens=128,
        temperature=0.7,
        do_sample=True,
        top_p=0.95
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# 对比测试
print("=== 基础模型输出 ===")
print(base_generate("RackServer-X3", "E-4012"))
print("\\n=== QLoRA微调后输出 ===")
print(generate("RackServer-X3", "E-4012"))