llama.cpp

CUDA: use MMQ instead of cuBLAS by default

#8075

Merged