llama.cpp

be55695e - ggml-backend : fix async copy from CPU (#8897)

ggml-backend : fix async copy from CPU (#8897) * ggml-backend : fix async copy from CPU * cuda : more reliable async copy, fix stream used when the devices are the same