Accelerating MoE Model Inference with Expert Sharding