llama.cpp

a818f302 - CUDA: use MMQ instead of cuBLAS by default (#8075)