llama.cpp

c3ea58ac - CUDA: remove DMMV, consolidate F16 mult mat vec (#10318)