LoRA微调7B大模型：从数据准备到推理效果对比的完整流程记录

# 核心依赖版本
torch==2.4.0+cu121
transformers==4.46.3
peft==0.13.2
bitsandbytes==0.44.1
trl==0.12.1
accelerate==1.1.1
datasets==3.1.0

# LoRA配置
lora_config = LoraConfig(
    r=16,                    # 秩，16在7B模型上足够
    lora_alpha=32,           # 缩放因子，通常设为2*r
    target_modules=[         # 注意力层的QKV和输出投影
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj"
    ],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

# 量化配置
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",           # NF4量化，比FP4更稳
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True       # 双重量化，再省0.4GB
)

{
  "instruction": "二甲双胍在肾功能不全患者中如何调整剂量？",
  "input": "患者eGFR为35ml/min/1.73m²",
  "output": "当eGFR在30-45ml/min/1.73m²时，二甲双胍应减量至500mg/日，并密切监测肾功能..."
}

def format_example(example):
    if example.get("input"):
        prompt = f"### 指令：{example['instruction']}\n### 输入：{example['input']}\n### 回答："
    else:
        prompt = f"### 指令：{example['instruction']}\n### 回答："
    return {"text": prompt + example["output"] + tokenizer.eos_token}

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

model_name = "Qwen/Qwen2.5-7B-Instruct"

# 1. 加载tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

# 2. 4-bit量化加载模型
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=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)
model.config.use_cache = False

# 3. 注入LoRA
lora_config = LoraConfig(
    r=16, lora_alpha=32, lora_dropout=0.05, bias="none",
    target_modules=["q_proj","k_proj","v_proj","o_proj",
                    "gate_proj","up_proj","down_proj"],
    task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# 输出: trainable params: 40,370,176 || all params: 7,655,986,688 || trainable%: 0.5273

# 4. 数据
dataset = load_dataset("json", data_files={"train": "train.json", "validation": "val.json"})
def format_example(ex):
    if ex.get("input"):
        p = f"### 指令：{ex['instruction']}\n### 输入：{ex['input']}\n### 回答："
    else:
        p = f"### 指令：{ex['instruction']}\n### 回答："
    return {"text": p + ex["output"] + tokenizer.eos_token}
dataset = dataset.map(format_example)

# 5. 训练参数
training_args = TrainingArguments(
    output_dir="./qwen2.5-7b-medical-lora",
    per_device_train_batch_size=4,
    gradient_accumulation_steps=2,
    num_train_epochs=3,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    logging_steps=10,
    save_strategy="epoch",
    evaluation_strategy="epoch",
    bf16=True,
    optim="paged_adamw_8bit",
    gradient_checkpointing=True,
    max_grad_norm=0.3,
    report_to="none"
)

# 6. 训练
trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["validation"],
    tokenizer=tokenizer,
    dataset_text_field="text",
    max_seq_length=1024,
    packing=False
)
trainer.train()
trainer.save_model("./qwen2.5-7b-medical-lora/final")

from peft import PeftModel
import torch

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "./qwen2.5-7b-medical-lora/final")
model = model.merge_and_unload()  # 合并LoRA权重，推理更快
model.eval()

def generate(instruction, input_text=""):
    if input_text:
        prompt = f"### 指令：{instruction}\n### 输入：{input_text}\n### 回答："
    else:
        prompt = f"### 指令：{instruction}\n### 回答："
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs, max_new_tokens=512,
            temperature=0.3, top_p=0.9, do_sample=True,
            repetition_penalty=1.05
        )
    return tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)