Or at least use a modern llama.cpp with KV activation rotation.