llama.cpp

38aca9e1 - llama : factor out tensor offloading outside the build call (wip)