llama.cpp

cb27fe9c - opencl: use flat mv q5_k when weight exceeds image1d_buffer_t limit (#26880)