RIS-Kernel: Running 64k context LLMs on CPU via sparse attentiongithub.com/santosardr2 pointssantosardr22 days ago