making LLMs sparse at inference time