llama.cpp

01b6f68a - backend : group nodes in a single compute when user don't need them