llama.cpp

62532c05 - cuda : do warp-based block reduce