This paper proposes an advanced architecture that mitigates worries of recurrent matrix multiplications by decomposing A-multiplications into various teams and optimizing positional encoding by means of Grouped Finite Impulse Response (FIR) filtering, and incorporates a similar mechanism to improve