A deep dive into the memory bottlenecks of LLM inference, exploring how the KV Cache works and how PagedAttention solves severe memory fragmentation.