Autoregressive decoding recomputes what it already knows. We build a KV cache from scratch on GPT-2, prove it’s exact, break it three ways on purpose, and measure the price you pay for it.
--- title: "Maitreya - Notes" listing: contents: posts feed: true sort: "date desc" type: default categories: true sort-ui: false filter-ui: false page-layout: full title-block-banner: true ---
