llama.cpp

207c22ec - ggml: Re-enable CUDA graphs in presence of CONT and DUP nodes (#12970)