Unified Model for Multimodal Understanding and Generation