KV caching means that when you have 10k prompt, all follow up questions return immediately - this is standard with all inference engines. Now if you are not ha

KV caching means that when you have 10k prompt, all follow up questions return immediately - this is standard with all inference engines. Now if you are not happy with the last answer, you maybe want to simply regenerate it or change your last question - this is branching of the conversation. Llama.cpp is capable of re-using the KV cache up to that point while MLX does not (I am using MLX server from MLX community project). I haven't tried with LMStudio. Maybe worth a try, thanks for the heads-up.