Seconding this. I’ve found that llama-cpp doesn’t completely evict all of the GPU memory with its idle sleep parameter configured.