You can’t cheaply recompute without re-running the whole model – so KV cache starts piling up Feature Large language model ...