DeepSeek-R1 speeds up llama.cpp code by x2