llama.cpp

69071d3b - Squeeze out about 5% more performance in Q4_1 inference