llama.cpp

01d89f9b - Reintroduces stricter check for CPU->CUDA backend async copy via