llama-fine-tuning-QLoRA.py

import os

# 切换到当前文件所在的目录
os.chdir(os.path.dirname(__file__))

# 导入必要的库
import torch
from datasets import load_dataset, Dataset
from transformers import (
    AutoModelForCausalLM,  # 用于加载预训练的语言模型
    AutoTokenizer,  # 用于加载与模型相匹配的分词器
    BitsAndBytesConfig,  # 用于配置4-bit量化
    HfArgumentParser,  # 用于解析命令行参数
    TrainingArguments,  # 用于设置训练参数
    pipeline,  # 用于创建模型的pipeline
    logging,  # 用于记录日志
)
from peft import LoraConfig, PeftModel  # 用于配置和加载QLoRA模型
from trl import SFTTrainer  # 用于执行监督式微调的Trainer

# 设置预训练模型的名称
model_name = "meta-llama/Llama-3.1-8B-Instruct"

# 设置微调后模型的名称
new_model = "Llama-3.1-8b-Instruct-fine-tuned"

# LoRA的注意力维度
lora_r = 64

# Alpha参数用于LoRA缩放
lora_alpha = 16

# LoRA层的dropout概率
lora_dropout = 0.1

# 激活4-bit精度的基础模型加载
use_4bit = True

# 4-bit基础模型的计算数据类型
bnb_4bit_compute_dtype = "float16"

# 4-bit量化类型（fp4或nf4）
bnb_4bit_quant_type = "nf4"

# 激活4-bit基础模型的嵌套量化（双重量化）
use_nested_quant = False

# 输出目录，用于存储模型预测和检查点
output_dir = "./results"

# 训练周期数
num_train_epochs = 1

# 是否启用fp16/bf16训练（在A100上将bf16设置为True）
fp16 = False
bf16 = True

# GPU上每个训练批次的样本数
per_device_train_batch_size = 4

# GPU上每个评估批次的样本数
per_device_eval_batch_size = 4

# 累积梯度的更新步骤数
gradient_accumulation_steps = 1

# 是否启用梯度检查点
gradient_checkpointing = True

# 最大梯度归一化（梯度裁剪）
max_grad_norm = 0.3

# 初始学习率（AdamW优化器）
learning_rate = 2e-4

# 权重衰减，应用于全部layer（不包括bias/LayerNorm的权重）
weight_decay = 0.001

# 优化器
optim = "paged_adamw_32bit"

# 学习率计划
lr_scheduler_type = "cosine"

# 训练步数（覆盖num_train_epochs）
max_steps = -1

# 线性预热的步数比率（从0到学习率）
warmup_ratio = 0.03

# 按长度分组序列
group_by_length = True

# 每X更新步骤保存检查点
save_steps = 0

# 每X更新步骤记录日志
logging_steps = 25

# SFT参数配置
# 最大序列长度
max_seq_length = None

# 打包多个短示例到同一输入序列以提高效率
packing = False

# 将整个模型加载到 GPU 0
device_map = {"": 0}

# 加载数据集
dataset = load_dataset(path="json", data_dir="./num_list", data_files="num_list_500_per_sample_100_length.json")
fine_tune_dataset = []
print("Loading dataset...")
for instance in dataset["train"]:
    prompt = instance["system_prompt"] + "\n\n" + instance["description"] + "\nQuestion: " + instance["data"]["question"] + "\nData: " + instance["data"]["struct_data"]
    answer = instance["data"]["answer"]
    completion = f"The answer is {answer}."
    fine_tune_dataset.append({"prompt": prompt, "completion": completion})

fine_tune_dataset = Dataset.from_list(fine_tune_dataset)

compute_dtype = getattr(torch, bnb_4bit_compute_dtype)

bnb_config = BitsAndBytesConfig(
    load_in_4bit=use_4bit,
    bnb_4bit_quant_type=bnb_4bit_quant_type,
    bnb_4bit_compute_dtype=compute_dtype,
    bnb_4bit_use_double_quant=use_nested_quant,
)

if compute_dtype == torch.float16 and use_4bit:
    major, _ = torch.cuda.get_device_capability()
    if major >= 8:
        print("GPU支持bfloat16")

# 加载模型
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map=device_map
)
model.config.use_cache = False
model.config.pretraining_tp = 1

# 加载分词器
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"  # 修复fp16训练中的溢出问题

peft_config = LoraConfig(
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    r=lora_r,
    bias="none",
    task_type="CAUSAL_LM",
)

training_arguments = TrainingArguments(
    output_dir=output_dir,
    num_train_epochs=num_train_epochs,
    per_device_train_batch_size=per_device_train_batch_size,
    gradient_accumulation_steps=gradient_accumulation_steps,
    optim=optim,
    save_steps=save_steps,
    logging_steps=logging_steps,
    learning_rate=learning_rate,
    weight_decay=weight_decay,
    fp16=fp16,
    bf16=bf16,
    max_grad_norm=max_grad_norm,
    max_steps=max_steps,
    warmup_ratio=warmup_ratio,
    group_by_length=group_by_length,
    lr_scheduler_type=lr_scheduler_type,
    report_to="tensorboard",
)

# 设置监督式微调参数
trainer = SFTTrainer(
    model=model,
    train_dataset=fine_tune_dataset,
    peft_config=peft_config,
    dataset_text_field="text",
    max_seq_length=max_seq_length,
    tokenizer=tokenizer,
    args=training_arguments,
    packing=packing,
)

# 训练模型
trainer.train()

trainer.model.save_pretrained(new_model)