Rotary GPU: Exploring Local Execution for Large MoE Models Under Limited VRAM