haochen wang

Hi, thanks for your excellent work! When I increase the batch_size, the forward time linearly increases. I know I should reduce the accumulate_step to keep the same training setting, but...