Comment by amanda99

You would use a KV cache to cache a significant chunk of the inference work.