llama.cpp

ceef6b52 - ggml: avoid creating CUDA context during device init (#20595)