LoRA / QLoRA — 低资源微调

LoRA vs QLoRA

方案	显存需求（7B）	训练速度	效果
全量微调	~80GB	慢	最好
LoRA (fp16)	~14GB	中	好
QLoRA (4bit)	~6GB	较慢	接近 LoRA

QLoRA 完整流程

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

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

# 2. 加载模型
model_name = "Qwen/Qwen2-7B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)
model = prepare_model_for_kbit_training(model)

# 3. LoRA 配置
peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

# 4. 准备数据集
def format_chat(example):
    return {
        "text": f"<|im_start|>system\n你是金融助手<|im_end|>\n"
                f"<|im_start|>user\n{example['instruction']}<|im_end|>\n"
                f"<|im_start|>assistant\n{example['output']}<|im_end|>"
    }

raw_data = [
    {"instruction": "解释不良贷款率", "output": "不良贷款率是..."},
    {"instruction": "分析信贷风险", "output": "信贷风险主要包括..."},
]
dataset = Dataset.from_list(raw_data).map(format_chat)

# 5. 训练
training_args = TrainingArguments(
    output_dir="./qlora-output",
    num_train_epochs=3,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    learning_rate=2e-4,
    bf16=True,
    logging_steps=10,
    save_strategy="epoch",
    warmup_ratio=0.05,
    lr_scheduler_type="cosine",
    report_to="none"
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=dataset,
    peft_config=peft_config,
    dataset_text_field="text",
    max_seq_length=2048,
    tokenizer=tokenizer
)

trainer.train()
trainer.save_model("./qlora-weights")

合并权重并导出

from peft import PeftModel

# 加载基础模型（fp16）
base_model = AutoModelForCausalLM.from_pretrained(
    model_name, torch_dtype=torch.float16, device_map="auto"
)

# 加载并合并 LoRA
model = PeftModel.from_pretrained(base_model, "./qlora-weights")
merged_model = model.merge_and_unload()

# 保存合并后的模型
merged_model.save_pretrained("./merged-model")
tokenizer.save_pretrained("./merged-model")

# 可直接用 vLLM 加载
# python -m vllm.entrypoints.openai.api_server --model ./merged-model