llama.cpp

68e02c0d - cuda : fix warp reduction initialization of shared mem