llama.cpp

1a0843c4 - cuda : utilize tensor cores with multiple GPU devices