Function load_scheduler

internlm/utils/model_checkpoint.py:244–267

(ckpt_path: str, lr_scheduler, optimizer, learning_rate, train_state: TrainState)

def load_scheduler(ckpt_path: str, lr_scheduler, optimizer, learning_rate, train_state: TrainState):
    scheduler_states = llm_load(os.path.join(ckpt_path, "schedulder.pt"))
    if learning_rate != scheduler_states["base_lrs"][0] and gpc.is_rank_for_log():
        logger.warning(
            f"Using new learning rate {learning_rate} to replace old learn rate {scheduler_states['base_lrs'][0]}."
        )

    base_lrs = copy.deepcopy(scheduler_states["base_lrs"])
    scheduler_states["base_lrs"] = [learning_rate] * len(scheduler_states["base_lrs"])
    if "after_scheduler_dict" in scheduler_states:
        scheduler_states["after_scheduler_dict"]["base_lrs"] = [learning_rate] * len(
            scheduler_states["after_scheduler_dict"]["base_lrs"]
        )

    lr_scheduler.load_state_dict(scheduler_states)
    lr_scheduler.last_epoch = train_state.step_count + 1

    ratios = [learning_rate / lr for lr in base_lrs]
    for idx, param_group in enumerate(optimizer.param_groups):
        param_group["lr"] = param_group["lr"] * ratios[idx]
    torch.cuda.empty_cache()

    if gpc.is_rank_for_log():
        logger.info(f"reload load_scheduler:{lr_scheduler}")


class CheckpointManager: