llama.cpp

a3d50bc0 - ggml-backend : only offload from host buffers (#11120)