llama.cpp

e03fb8ee - Reintroduces stricter check for CPU->CUDA backend async copy via