llama.cpp

93b9baee - convert-hf : reduce stacked MoE conversion RAM usage by a third