llama.cpp

fbddb262 - ggml-cuda : use i and j instead of i0 and i in vec_dot_tq2_0_q8_1