KV Cache - Page 98

A KV cache, or key-value cache, speeds up autoregressive transformer inference by storing attention information calculated for tokens that have already been processed. Without the cache, the model would recompute those keys and values each time it generates another token. Reusing them reduces repeated work and improves response latency, especially for long prompts and outputs. The tradeoff is memory: cache size grows with sequence length, batch size, layer count, and model dimensions. Production systems manage this cost with paging, quantization, eviction, prompt caching, and request scheduling. KV caching changes inference efficiency, not the model’s learned knowledge, and cached data must be protected when prompts contain sensitive information.