We analyze the reasons behind Differential Transformer's success, based on which we propose an efficient adaptation method to enhance pretrained LLMs.