Figure: the cover page for the Technical Reference "KV Cache Offloading for LLM Inference", version 4.0, August 2026, by Sam Pooni of San Jose, CA. It presents a distributed endpoint architecture with intelligent caching, attention-aware eviction and CXL.mem acceleration, built on three standards — CXL 3.0, UCIe and UEC — and three techniques: per-head eviction, EMA attention scoring and RoPE-aware prefetch. The animated rings, nodes and flow lines behind the title are decorative. All values are an analytical model, not measured.