llama.cpp

307bfa25 - ggml: disable CUDA graphs for unsupported DUP and CONT node types (#12891)