llama.cpp

c80b8a2b - llama : remove mirrors, perform Device -> Host when partial offload