[2409.17066] VPTQ: Extreme Low-bit Vector Post-Training Quantization for Large Language Models