In the field of autonomous driving, both single-task and multi-task frameworks are commonly used. The multi-task framework incorporates a feature that can be applied to multiple tasks using multiple heads, thus minimizing processing costs compared to the single-task framework. We propose a unified multi-task framework for scene-based and instance-based tasks by utilizing the BEV features and instance query, respectively. Scene-based tasks, like point cloud forecasting, employ temporal BEV features to predict future point clouds. During instance-based tasks such as object tracking and motion forecasting, each head can leverage instance queries generated in object detection. Our results achieved the top score in the Argoverse leaderboard and comparable results with the state-of-the-art in the NuScene dataset.