llama.cpp

ggml-cuda: use passed ops instead of hardcoded ops

#16712

Merged