llama.cpp

061f0eff - ggml-cuda: use passed ops instead of hardcoded ops (#16712)