llama.cpp

1e9c5443 - llama : refactor tensor offloading as callback