llama.cpp

ed54e325 - Workaround for subgroup arithmetic failing on MoltenVK with AMD GPUs (issue 15846) (#15886)