MARS: Unleashing the Power of Variance Reduction for Training Large Models | Read Paper on Bytez