llama.cpp

58062860 - ggml : use WARP_SIZE/2 for argmax reduction offset (#18092)