llama.cpp

42b53d19 - CUDA: revise q8_1 data layout for mul_mat_q (#7824)