llama.cpp

580991bb - Squeeze out about 5% more performance in Q4_1 inference