This paper proposes an advanced architecture that mitigates difficulties of recurrent matrix multiplications by decomposing A-multiplications into several teams and optimizing positional encoding by Grouped Finite Impulse Reaction (FIR) filtering, and incorporates the same mechanism to reinforce The soundness and general performance of your product