Improving Efficiency For Gpus With Decoupled Delegate