RSQ: Learning from Important Tokens Leads to Better Quantized LLMs