How to use the fairseq.tasks.setup_task function in fairseq

# Initialize CUDA and distributed training
    if torch.cuda.is_available() and not args.cpu:
        torch.cuda.set_device(args.device_id)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    if init_distributed:
        args.distributed_rank = distributed_utils.distributed_init(args)

    if distributed_utils.is_master(args):
        checkpoint_utils.verify_checkpoint_directory(args.save_dir)

    # Print args
    print(args)

    # Setup task, e.g., translation, language modeling, etc.
    task = tasks.setup_task(args)

    # Load valid dataset (we load training data below, based on the latest checkpoint)
    for valid_sub_split in args.valid_subset.split(','):
        task.load_dataset(valid_sub_split, combine=False, epoch=0)

    # Build model and criterion
    model = task.build_model(args)
    criterion = task.build_criterion(args)
    print(model)
    print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))
    print('| num. model params: {} (num. trained: {})'.format(
        sum(p.numel() for p in model.parameters()),
        sum(p.numel() for p in model.parameters() if p.requires_grad),
    ))

    # Build trainer

# Initialize CUDA and distributed training
    if torch.cuda.is_available() and not args.cpu:
        torch.cuda.set_device(args.device_id)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    if init_distributed:
        args.distributed_rank = distributed_utils.distributed_init(args)

    if distributed_utils.is_master(args):
        checkpoint_utils.verify_checkpoint_directory(args.save_dir)

    # Print args
    print(args)

    # Setup task, e.g., translation, language modeling, etc.
    task = tasks.setup_task(args)

    # Load valid dataset (we load training data below, based on the latest checkpoint)
    for valid_sub_split in args.valid_subset.split(','):
        task.load_dataset(valid_sub_split, combine=False, epoch=0)

    # Build model and criterion
    model = task.build_model(args)
    criterion = task.build_criterion(args)
    print(model)
    print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))
    print('| num. model params: {} (num. trained: {})'.format(
        sum(p.numel() for p in model.parameters()),
        sum(p.numel() for p in model.parameters() if p.requires_grad),
    ))

    # Build trainer

models, args = checkpoint_utils.load_model_ensemble(
        parsed_args.path.split(':'),
        arg_overrides=eval(parsed_args.model_overrides),
        task=task,
    )

    for arg in vars(parsed_args).keys():
        if arg not in {
            'self_target', 'future_target', 'past_target', 'tokens_per_sample',
            'output_size_dictionary', 'add_bos_token',
        }:
            setattr(args, arg, getattr(parsed_args, arg))

    # reduce tokens per sample by the required context window size
    args.tokens_per_sample -= args.context_window
    task = tasks.setup_task(args)

    # Load dataset splits
    task.load_dataset(args.gen_subset)
    dataset = task.dataset(args.gen_subset)
    if args.context_window > 0:
        dataset = LMContextWindowDataset(
            dataset=dataset,
            tokens_per_sample=args.tokens_per_sample,
            context_window=args.context_window,
            pad_idx=task.source_dictionary.pad(),
        )
    print('| {} {} {} examples'.format(args.data, args.gen_subset, len(dataset)))

    # Optimize ensemble for generation and set the source and dest dicts on the model (required by scorer)
    for model in models:
        model.make_generation_fast_()

def main(args):
    check_args(args)
    import_user_module(args)

    if args.max_tokens is None and args.max_sentences is None:
        args.max_tokens = 30000
    logger.info(args)

    use_cuda = torch.cuda.is_available() and not args.cpu

    # Load dataset splits
    task = tasks.setup_task(args)
    task.load_dataset(args.gen_subset)
    logger.info(
        "| {} {} {} examples".format(
            args.data, args.gen_subset, len(task.dataset(args.gen_subset))
        )
    )

    # Set dictionary
    tgt_dict = task.target_dictionary

    logger.info("| decoding with criterion {}".format(args.criterion))

    # Load ensemble
    logger.info("| loading model(s) from {}".format(args.path))
    models, criterions, _model_args = load_models_and_criterions(
        args.path.split(":"),

# Initialize CUDA and distributed training
    if torch.cuda.is_available() and not args.cpu:
        torch.cuda.set_device(args.device_id)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    if init_distributed:
        args.distributed_rank = distributed_utils.distributed_init(args)

    if distributed_utils.is_master(args):
        checkpoint_utils.verify_checkpoint_directory(args.save_dir)

    # Print args
    print(args)

    # Setup task, e.g., translation, language modeling, etc.
    task = tasks.setup_task(args)

    # Load valid dataset (we load training data below, based on the latest checkpoint)
    for valid_sub_split in args.valid_subset.split(','):
        task.load_dataset(valid_sub_split, combine=False, epoch=0)

    # Build model and criterion
    model = task.build_model(args)
    criterion = task.build_criterion(args)
    print(model)
    print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))
    print('| num. model params: {} (num. trained: {})'.format(
        sum(p.numel() for p in model.parameters()),
        sum(p.numel() for p in model.parameters() if p.requires_grad),
    ))

    # Build trainer

if args.buffer_size < 1:
        args.buffer_size = 1
    if args.max_tokens is None and args.max_sentences is None:
        args.max_sentences = 1

    assert not args.sampling or args.nbest == args.beam, \
        '--sampling requires --nbest to be equal to --beam'
    assert not args.max_sentences or args.max_sentences <= args.buffer_size, \
        '--max-sentences/--batch-size cannot be larger than --buffer-size'

    print(args)

    use_cuda = torch.cuda.is_available() and not args.cpu

    # Setup task, e.g., translation
    task = tasks.setup_task(args)

    # Load ensemble
    print('| loading model(s) from {}'.format(args.path))
    model_paths = args.path.split(':')
    models, model_args = utils.load_ensemble_for_inference(model_paths, task, model_arg_overrides=eval(args.model_overrides))

    # Set dictionaries
    src_dict = task.source_dictionary
    tgt_dict = task.target_dictionary

    # Optimize ensemble for generation
    for model in models:
        model.make_generation_fast_(
            beamable_mm_beam_size=None if args.no_beamable_mm else args.beam,
            need_attn=args.print_alignment,
        )

# Initialize CUDA and distributed training
    if torch.cuda.is_available() and not args.cpu:
        torch.cuda.set_device(args.device_id)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    if init_distributed:
        args.distributed_rank = distributed_utils.distributed_init(args)

    if distributed_utils.is_master(args):
        checkpoint_utils.verify_checkpoint_directory(args.save_dir)

    # Print args
    print(args)

    # Setup task, e.g., translation, language modeling, etc.
    task = tasks.setup_task(args)

    # Load valid dataset (we load training data below, based on the latest checkpoint)
    for valid_sub_split in args.valid_subset.split(','):
        task.load_dataset(valid_sub_split, combine=False, epoch=0)

    # Build model and criterion
    model = task.build_model(args)
    criterion = task.build_criterion(args)
    print(model)
    print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))
    print('| num. model params: {} (num. trained: {})'.format(
        sum(p.numel() for p in model.parameters()),
        sum(p.numel() for p in model.parameters() if p.requires_grad),
    ))

    # Build trainer

# Initialize CUDA and distributed training
    if torch.cuda.is_available() and not args.cpu:
        torch.cuda.set_device(args.device_id)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    if init_distributed:
        args.distributed_rank = distributed_utils.distributed_init(args)

    if distributed_utils.is_master(args):
        checkpoint_utils.verify_checkpoint_directory(args.save_dir)

    # Print args
    print(args)

    # Setup task, e.g., translation, language modeling, etc.
    task = tasks.setup_task(args)

    # Load valid dataset (we load training data below, based on the latest checkpoint)
    for valid_sub_split in args.valid_subset.split(','):
        task.load_dataset(valid_sub_split, combine=False, epoch=0)

    # Build model and criterion
    model = task.build_model(args)
    criterion = task.build_criterion(args)
    print(model)
    print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))
    print('| num. model params: {} (num. trained: {})'.format(
        sum(p.numel() for p in model.parameters()),
        sum(p.numel() for p in model.parameters() if p.requires_grad),
    ))

    # Build trainer

model = char_source_hybrid.CharSourceHybridModel.build_model(
                checkpoint_data["args"], task
            )
        elif architecture == "dual_decoder_kd":
            model = dual_decoder_kd_model.DualDecoderKDModel.build_model(
                checkpoint_data["args"], task
            )
        elif architecture == "hybrid_dual_decoder_kd":
            model = hybrid_dual_decoder_kd_model.HybridDualDecoderKDModel.build_model(
                checkpoint_data["args"], task
            )
        elif "semi_supervised" in architecture:
            model_args = copy.deepcopy(checkpoint_data["args"])
            model_args.source_vocab_file = src_dict_filename
            model_args.target_vocab_file = dst_dict_filename
            task = tasks.setup_task(model_args)

            model = ARCH_MODEL_REGISTRY[model_args.arch].build_model(model_args, task)
        elif architecture == "latent_var_transformer":
            task = tasks.setup_task(checkpoint_data["args"])
            model = latent_var_models.LatentVarModel.build_model(
                checkpoint_data["args"], task
            )
        elif architecture == "fb_levenshtein_transformer":
            task = tasks.setup_task(checkpoint_data["args"])
            model = levenshtein_transformer.LevenshteinTransformerModel.build_model(
                checkpoint_data["args"], task
            )
        else:
            raise RuntimeError(f"Architecture not supported: {architecture}")
        model.load_state_dict(checkpoint_data["model"])
        if hasattr(model, "get_student_model"):

def main(args):
    if args.max_tokens is None:
        args.max_tokens = 6000
    print(args)

    if not torch.cuda.is_available():
        raise NotImplementedError('Training on CPU is not supported')
    torch.cuda.set_device(args.device_id)
    torch.manual_seed(args.seed)

    # Setup task, e.g., translation, language modeling, etc.
    task = tasks.setup_task(args)

    # Load dataset splits
    load_dataset_splits(args, task, ['train', 'valid'])

    # Build model and criterion
    model = task.build_model(args)
    criterion = task.build_criterion(args)
    print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))
    print('| num. model params: {}'.format(sum(p.numel() for p in model.parameters())))

    # Build trainer
    if args.fp16:
        trainer = FP16Trainer(args, task, model, criterion)
    else:
        if torch.cuda.get_device_capability(0)[0] >= 7:
            print('| NOTICE: your device may support faster training with --fp16')