llama.cpp

9c42b171 - CUDA: fix logic for V100 + GGML_CUDA_FORCE_MMQ (#12098)