llama.cpp

016c1d91 - CUDA: use highest compiled arch in get_mmvq_mmid_max_batch (#31679)