Layer-Wise High-Impact Parameter Ratio Optimization in Post-Training Quantization for Large Language Models