Two different tricks for fast LLM inference