llama.cpp

a8eca04c - support cuda virtual devices