A high-throughput and memory-efficient inference and serving engine for LLMs