llama.cpp

ebd062bc - cuda : use 512 threads for soft_max instead of 32