local_batch_size: int = 8
    """Local batch size (i.e., per-device batch size)"""

    global_batch_size: int = -1
    """
    Global batch size (defaults to `training.local_batch_size * data-parallel degree`)
    """

        # calculate gradient accumulation steps
        self.gradient_accumulation_steps = global_batch_size // (
            job_config.training.local_batch_size * dp_degree
        )
        assert self.gradient_accumulation_steps > 0
        self.loss_fn = rescale_accumulated_loss(
            self.loss_fn, self.gradient_accumulation_steps
        )