app.py

from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
from peft import LoraConfig, get_peft_model, TaskType
from datasets import load_dataset
import torch
import os

def main():
    # 基础模型位置
    model_name = "dushuai112233/Qwen2-1.5B-Instruct"
    device = "cuda" if torch.cuda.is_available() else "cpu"

    # 加载分词器和模型
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)

    # Setup PEFT
    peft_config = LoraConfig(
        task_type=TaskType.CAUSAL_LM,
        target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
        inference_mode=False,
        r=8,
        lora_alpha=32,
        lora_dropout=0.1
    )
    model = get_peft_model(model, peft_config)

    # 加载数据集
    ds = load_dataset("dushuai112233/medical")
    train_dataset = ds["train"]
    val_dataset = ds["validation"]

    # 数据集预处理
    def tokenize_function(examples):
        encodings = tokenizer(examples['question'], padding='max_length', truncation=True, max_length=128)
        encodings['labels'] = encodings['input_ids'].copy()
        return encodings

    train_dataset = train_dataset.map(tokenize_function, batched=True)
    val_dataset = val_dataset.map(tokenize_function, batched=True)

    # 设置训练参数
    training_args = TrainingArguments(
        output_dir="./output",
        evaluation_strategy="epoch",
        per_device_train_batch_size=1,
        per_device_eval_batch_size=1,
        logging_dir="./logs",
        logging_steps=10,
        save_steps=100,                 # 每 100 步保存一次检查点
        save_total_limit=2,             # 限制最多保存 2 个检查点
        num_train_epochs=10,
        load_best_model_at_end=False,   # 是否在训练结束时加载最优模型
    )

    # 定义 Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset,
        eval_dataset=val_dataset,
        tokenizer=tokenizer,
    )

    # 检查是否有中断点
    checkpoint = None
    if os.path.exists("./output") and len(os.listdir("./output")) > 0:
        checkpoint = max([os.path.join("./output", ckpt) for ckpt in os.listdir("./output")], key=os.path.getmtime)
        print(f"Resuming training from checkpoint: {checkpoint}")

    # 开始训练
    trainer.train(resume_from_checkpoint=checkpoint)

    # 保存最终模型
    model.save_pretrained('./output')

if __name__ == '__main__':
    main()