llama.cpp

28a77807 - skip dmabuf_p2p when one device is nvidia, due to driver crashes