Towards MoE Deployment: Mitigating Inefficiencies in Mixture-of-Expert (MoE) Inference