May 02, 2026 KV Cache Compression: How LLMs Handle Million-Token Contexts Without Running Out of Memory