llama.cpp

f64d44a9 - CUDA: Fixed OpenLLaMA 3b mmq, reduced compile time (#2590)