llama.cpp

ggml: avoid creating CUDA context during device init

#20595

Merged