llama.cpp

04aaae1d - add avx2 for dot_q8_0_q8_0, 2x faster than scalar (#1211)