How can I control the value of the maximum adaptive batch size of Accelerator for my benchmark configuration?