llama.cpp

0b4be4c4 - CUDA: fix FTZ in FA for Gemma 3 (#13991)