Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs F i n d i n g s - E M N L P