Yeah. It might be possible in llama.cpp soon, but the vulkan implementation may or may not be fast on the IGP.