# 续写同一文件：training_project/src/readingtorch/engine.py
    if resume:
        saved = load_checkpoint(resume)
        if saved.get("kind") != "training" or saved["config"] != config:
            raise ValueError("恢复时请使用原来的任务、种子、批量大小与学习率")
        if saved["source_hash"] != bundle.source_hash:
            raise ValueError("数据文件已经改变，请另开一次训练")
        model.load_state_dict(saved["model"])
        optimizer.load_state_dict(saved["optimizer"])
        history, start = saved["history"], saved["epoch"]
        best_score, best_epoch, best_state = (
            saved["best_score"],
            saved["best_epoch"],
            saved["best_model"],
        )
        generator.set_state(saved["loader_rng"])
        # 接回随机层的下一次抽样位置，模型初始化已在恢复之前完成。
        torch.set_rng_state(saved["torch_rng"])
        if device == "cuda" and saved["cuda_rng"]:
            torch.cuda.set_rng_state_all(saved["cuda_rng"])
    if epochs <= start:
        raise ValueError("目标总轮数应大于检查点已经完成的轮数")
    train_loader = make_loader(bundle, "train", batch_size, generator)
    val_loader = make_loader(bundle, "val", batch_size)
    loss_fn = nn.MSELoss() if task == "regression" else nn.CrossEntropyLoss()
    scale = bundle.preprocess["y_scale"].item() if task == "regression" else 1.0