llama.cpp

73955f7d - CUDA: no FP16 arithmetic for vector FA kernel (#17558)