llama.cpp

952f3953 - ggml : allow CUDA graphs when using pipeline parallelism (#13814)