train_lora_auto.py

"""
TFMF LoRA 训练脚本 - 新版自动化入口

这个脚本不会覆盖现有的 train_lora.py，
用于从一个 JSONL 数据目录或单个 JSONL 文件批量加载训练样本，
自动生成 LoRA 适配器并保存到指定输出目录。

格式要求：每条训练数据必须至少包含字段：system, user, assistant
示例：
{ "system": "...", "user": "...", "assistant": "..." }
"""

import argparse
import glob
import json
import os
import random
import sys

import torch
from datasets import Dataset
from peft import LoraConfig, get_peft_model
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    DataCollatorForSeq2Seq,
    Trainer,
    TrainingArguments,
)

DEFAULT_MODEL_PATH = "/data/coding/TFMF"
DEFAULT_DATA_DIR = "./TFMF/dataset"
DEFAULT_OUTPUT_DIR = "./TFMF/lora_adapters/teacher_chinese_auto"
DEFAULT_LORA_TARGET_MODULES = [
    "q_proj",
    "k_proj",
    "v_proj",
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj",
]


def parse_args():
    parser = argparse.ArgumentParser(description="自动化 LoRA 训练脚本（新版本）")
    parser.add_argument(
        "--model-path",
        type=str,
        default=DEFAULT_MODEL_PATH,
        help="基座模型目录，例如 /data/coding/TFMF",
    )
    parser.add_argument(
        "--data-dir",
        type=str,
        default=DEFAULT_DATA_DIR,
        help="训练数据目录，脚本会加载目录下所有 .jsonl 文件",
    )
    parser.add_argument(
        "--data-path",
        type=str,
        default=None,
        help="单个训练文件路径，优先于 --data-dir",
    )
    parser.add_argument(
        "--output-dir",
        type=str,
        default=DEFAULT_OUTPUT_DIR,
        help="LoRA 输出目录",
    )
    parser.add_argument("--epochs", type=int, default=3, help="训练轮数")
    parser.add_argument("--batch-size", type=int, default=8, help="每卡 batch 大小")
    parser.add_argument(
        "--gradient-accumulation-steps",
        type=int,
        default=2,
        help="梯度累积步数",
    )
    parser.add_argument("--learning-rate", type=float, default=2e-4, help="学习率")
    parser.add_argument(
        "--max-seq-length",
        type=int,
        default=2048,
        help="最大序列长度",
    )
    parser.add_argument("--save-steps", type=int, default=50, help="保存间隔步数")
    parser.add_argument("--logging-steps", type=int, default=10, help="日志输出步数")
    parser.add_argument("--warmup-steps", type=int, default=10, help="warmup 步数")
    parser.add_argument(
        "--validation-split",
        type=float,
        default=0.1,
        help="验证集比例，0.0 表示不切分",
    )
    parser.add_argument("--seed", type=int, default=42, help="随机种子")
    parser.add_argument(
        "--lora-r",
        type=int,
        default=32,
        help="LoRA r 值",
    )
    parser.add_argument(
        "--lora-alpha",
        type=int,
        default=64,
        help="LoRA alpha 值",
    )
    parser.add_argument(
        "--lora-dropout",
        type=float,
        default=0.05,
        help="LoRA dropout",
    )
    return parser.parse_args()


def set_seed(seed: int):
    random.seed(seed)
    torch.manual_seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)


def find_jsonl_files(data_dir: str):
    if not os.path.isdir(data_dir):
        raise FileNotFoundError(f"数据目录不存在: {data_dir}")
    file_paths = sorted(glob.glob(os.path.join(data_dir, "*.jsonl")))
    if not file_paths:
        raise FileNotFoundError(f"未在数据目录中找到 .jsonl 文件: {data_dir}")
    return file_paths


def load_samples_from_file(path: str):
    if not os.path.isfile(path):
        raise FileNotFoundError(f"数据文件不存在: {path}")
    samples = []
    with open(path, "r", encoding="utf-8") as f:
        for line in f:
            if not line.strip():
                continue
            item = json.loads(line)
            if not all(k in item for k in ("system", "user", "assistant")):
                raise ValueError(
                    f"样本缺少必要字段 system/user/assistant: {path}\n行内容: {line[:200]}"
                )
            samples.append(item)
    return samples


def format_sample(sample: dict, tokenizer):
    messages = [
        {"role": "system", "content": sample["system"]},
        {"role": "user", "content": sample["user"]},
        {"role": "assistant", "content": sample["assistant"]},
    ]
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=False,
    )
    return {"text": text}


def tokenize_samples(examples, tokenizer, max_length):
    return tokenizer(
        examples["text"],
        truncation=True,
        max_length=max_length,
        padding=False,
    )


def build_dataset(samples, tokenizer, max_length):
    dataset = Dataset.from_list(samples).map(
        lambda x: format_sample(x, tokenizer),
        batched=False,
    )
    dataset = dataset.map(
        lambda x: tokenize_samples(x, tokenizer, max_length),
        batched=True,
        remove_columns=["text"],
    )
    dataset = dataset.map(lambda x: {"labels": x["input_ids"]})
    return dataset


def print_config(args):
    print("=" * 70)
    print("LoRA 训练脚本参数:")
    for name, value in vars(args).items():
        print(f"  {name}: {value}")
    print("=" * 70)


def main():
    args = parse_args()
    set_seed(args.seed)

    if args.data_path is None and args.data_dir is None:
        raise ValueError("请通过 --data-path 或 --data-dir 指定训练数据")

    print_config(args)

    if args.data_path is not None:
        print(f"加载单个 JSONL 文件: {args.data_path}")
        samples = load_samples_from_file(args.data_path)
    else:
        print(f"加载数据目录: {args.data_dir}")
        file_paths = find_jsonl_files(args.data_dir)
        samples = []
        for path in file_paths:
            print(f"  加载: {path}")
            samples.extend(load_samples_from_file(path))

    if not samples:
        raise ValueError("没有加载到任何训练样本，请检查数据文件")

    random.shuffle(samples)
    split_index = int(len(samples) * (1 - args.validation_split))
    train_samples = samples[:split_index]
    val_samples = samples[split_index:] if args.validation_split > 0 else []

    print(f"总样本数: {len(samples)}")
    print(f"训练集: {len(train_samples)}")
    print(f"验证集: {len(val_samples)}")

    print("加载 tokenizer 和基座模型...")
    tokenizer = AutoTokenizer.from_pretrained(args.model_path, trust_remote_code=True)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "right"

    model = AutoModelForCausalLM.from_pretrained(
        args.model_path,
        device_map="auto",
        dtype=torch.bfloat16,
        trust_remote_code=True,
        low_cpu_mem_usage=True,
    )
    model.config.use_cache = False

    print("配置 LoRA...")
    lora_config = LoraConfig(
        r=args.lora_r,
        lora_alpha=args.lora_alpha,
        target_modules=DEFAULT_LORA_TARGET_MODULES,
        lora_dropout=args.lora_dropout,
        bias="none",
        task_type="CAUSAL_LM",
    )
    model = get_peft_model(model, lora_config)

    trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
    total = sum(p.numel() for p in model.parameters())
    print(f"可训练参数: {trainable / 1e6:.2f}M")
    print(f"总参数: {total / 1e9:.2f}B")
    print(f"可训练参数占比: {trainable / total * 100:.2f}%")

    print("构建训练数据集...")
    train_dataset = build_dataset(train_samples, tokenizer, args.max_seq_length)
    eval_dataset = build_dataset(val_samples, tokenizer, args.max_seq_length) if val_samples else None

    os.makedirs(args.output_dir, exist_ok=True)
    final_output_dir = os.path.join(args.output_dir, "final")
    os.makedirs(final_output_dir, exist_ok=True)

    training_args = TrainingArguments(
        output_dir=args.output_dir,
        num_train_epochs=args.epochs,
        per_device_train_batch_size=args.batch_size,
        per_device_eval_batch_size=args.batch_size,
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        learning_rate=args.learning_rate,
        warmup_steps=args.warmup_steps,
        logging_steps=args.logging_steps,
        save_steps=args.save_steps,
        save_total_limit=3,
        bf16=True,
        fp16=False,
        gradient_checkpointing=True,
        max_grad_norm=0.3,
        weight_decay=0.01,
        report_to="none",
        eval_strategy="steps" if eval_dataset is not None else "no",
        save_strategy="steps",
        load_best_model_at_end=False,
        logging_dir=os.path.join(args.output_dir, "logs"),
    )

    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        data_collator=DataCollatorForSeq2Seq(tokenizer, padding=True),
    )

    print("开始训练...")
    trainer.train()

    print(f"保存最终 LoRA 适配器到: {final_output_dir}")
    model.save_pretrained(final_output_dir)
    tokenizer.save_pretrained(final_output_dir)

    metadata = {
        "base_model": args.model_path,
        "data_source": args.data_path or args.data_dir,
        "train_samples": len(train_samples),
        "validation_samples": len(val_samples),
        "epochs": args.epochs,
        "batch_size": args.batch_size,
        "gradient_accumulation_steps": args.gradient_accumulation_steps,
        "learning_rate": args.learning_rate,
        "max_seq_length": args.max_seq_length,
        "lora_r": args.lora_r,
        "lora_alpha": args.lora_alpha,
        "lora_dropout": args.lora_dropout,
    }
    with open(os.path.join(args.output_dir, "adapter_config.json"), "w", encoding="utf-8") as f:
        json.dump(metadata, f, indent=2, ensure_ascii=False)

    total_size = 0
    for root, _, files in os.walk(final_output_dir):
        for fname in files:
            total_size += os.path.getsize(os.path.join(root, fname))
    print(f"最终适配器大小: {total_size / 1024 / 1024:.1f} MB")
    print("训练完成，请使用 final 目录下的适配器进行加载。")


if __name__ == "__main__":
    main()