llama.cpp

c824d2e3 - ggml : avoid conv 2d kernel round up