There are a lot of shared prefixes from user prompts. You can save by first looking into the cache to find the longest prefix for a prompt. Their MLA makes the KV cache particularly efficient.