Function main

funasr/bin/train_ds.py:64–255

Main. Args: **kwargs: Additional keyword arguments.

(**kwargs)

def main(**kwargs):

    # set random seed
    """Main.
    
        Args:
            **kwargs: Additional keyword arguments.
        """
    set_all_random_seed(kwargs.get("seed", 0))
    torch.backends.cudnn.enabled = kwargs.get("cudnn_enabled", torch.backends.cudnn.enabled)
    torch.backends.cudnn.benchmark = kwargs.get("cudnn_benchmark", torch.backends.cudnn.benchmark)
    torch.backends.cudnn.deterministic = kwargs.get("cudnn_deterministic", True)
    # open tf32
    torch.backends.cuda.matmul.allow_tf32 = kwargs.get("enable_tf32", True)

    rank = int(os.environ.get("RANK", 0))
    local_rank = int(os.environ.get("LOCAL_RANK", 0))
    world_size = int(os.environ.get("WORLD_SIZE", 1))

    if local_rank == 0:
        tables.print()

    use_ddp = world_size > 1
    use_fsdp = kwargs.get("use_fsdp", False)
    use_deepspeed = kwargs.get("use_deepspeed", False)
    if use_deepspeed:
        logging.info(f"use_deepspeed: {use_deepspeed}")
        deepspeed.init_distributed(dist_backend=kwargs.get("backend", "nccl"))
    elif use_ddp or use_fsdp:
        logging.info(f"use_ddp: {use_ddp}, use_fsdp: {use_fsdp}")
        dist.init_process_group(
            backend=kwargs.get("backend", "nccl"),
            init_method="env://",
        )
        torch.cuda.set_device(local_rank)

    # rank = dist.get_rank()

    logging.info("Build model, frontend, tokenizer")
    device = kwargs.get("device", "cuda")
    kwargs["device"] = "cpu"
    model = AutoModel(**kwargs)

    # save config.yaml
    if rank == 0:
        prepare_model_dir(**kwargs)

    # parse kwargs
    kwargs = model.kwargs
    kwargs["device"] = device
    tokenizer = kwargs["tokenizer"]
    frontend = kwargs["frontend"]
    model = model.model
    del kwargs["model"]

    # freeze_param
    freeze_param = kwargs.get("freeze_param", None)
    if freeze_param is not None: