解决deepspeed框架的bug：不保存调度器状态，模型训练重启时学习率从头开始

本文链接：https://blog.csdn.net/tongjingqi_/article/details/132700846

deepspeed存在一个bug，即在训练时不保存调度器状态，因此如果训练中断后再重新开始训练，调度器还是会从头开始而不是接着上一个checkpoint的调度器状态来训练。这个bug在deepspeed的github中也有其他人提出：https://github.com/microsoft/DeepSpeed/issues/3875
因此我们需要写一个保存调度器状态的代码，才可以解决这个问题。
具体方法是加一个callback类，专门负责保存调度器的状态以及在训练重新开始时加载调度器的状态：
先在训练文件中给trainer加一个callback

from smoe.callbacks.save_model import SchedulerStateCallback
trainer.add_callback(SchedulerStateCallback)

class SchedulerStateCallback(TrainerCallback):
    def on_save(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs):
        if os.environ.get("RANK", "0") == "0":
            #scheduler = kwargs['lr_scheduler']
            scheduler = kwargs.get("lr_scheduler")
            if scheduler is None:
                return 
            scheduler_state = scheduler.state_dict()
            #save_path = os.path.join(args.output_dir, SCHEDULER_NAME)
             # 使用 PREFIX_CHECKPOINT_DIR 和 global_step 创建检查点目录名
            checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}"
            # 完整的检查点目录路径
            checkpoint_path = os.path.join(args.output_dir, checkpoint_folder)
            # 如果目录不存在，则创建它
            if not os.path.exists(checkpoint_path):
                os.makedirs(checkpoint_path)
            # 完整的保存路径
            save_path = os.path.join(checkpoint_path, SCHEDULER_NAME)
            # 保存scheduler状态
            torch.save(scheduler_state, save_path)

    def on_train_begin(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs):
        # 如果resume_from_checkpoint设置了有效路径
        if args.resume_from_checkpoint is not None:
            load_path = os.path.join(args.resume_from_checkpoint, SCHEDULER_NAME)
            # 如果该路径下有保存的调度器状态，则加载它
            if os.path.exists(load_path):
                #scheduler = kwargs['lr_scheduler']
                scheduler = kwargs.get("lr_scheduler")
                if scheduler is None:
                    return 
                scheduler_state = torch.load(load_path)
                scheduler.load_state_dict(scheduler_state)