llama.cpp

1b280c9f - CUDA: fix softmax compile for old CUDA versions (#4862)