Accelerating Large Language Models through Partially Linear Feed-Forward Network