amanda99·в прошлом году·discussYou would use a KV cache to cache a significant chunk of the inference work.