llama.cpp

54a241f5 - sched : fix possible use of wrong ids tensor when offloading moe prompt processing (#15488)