The KV Cache Is Just Memoization

To write each next word, an LLM looks back over the whole sentence and recomputes every key and value from scratch, every single step. But a token's key and value never change once it exists, while its query is used exactly once and then thrown away. So you cache the keys and values, discard the queries, and turn O(N^2) recompute into O(N) append-and-reuse. The KV cache is just memoization for attention. #shorts *Follow Me* ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ 🐦 X: @datamlistic https://x.com/datamlistic 📸 Instagram: @datamlistic https://www.instagram.com/datamlistic 📱 TikTok: @datamlistic https://www.tiktok.com/@datamlistic 👔 Linkedin: https://www.linkedin.com/company/datamlistic *Channel Support* ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ The best way to support the channel is to share the content. ;) If you'd like to also support the channel financially, donating the price of a coffee is always warmly welcomed! (completely optional and voluntary) ► Patreon: https://www.patreon.com/datamlistic ► Bitcoin (BTC): 3C6Pkzyb5CjAUYrJxmpCaaNPVRgRVxxyTq ► Ethereum (ETH): 0x9Ac4eB94386C3e02b96599C05B7a8C71773c9281 ► Cardano (ADA): addr1v95rfxlslfzkvd8sr3exkh7st4qmgj4ywf5zcaxgqgdyunsj5juw5 ► Tether (USDT): 0xeC261d9b2EE4B6997a6a424067af165BAA4afE1a #shorts #machinelearning #llm