WarpDrive runs the entire MADRL workflow end-to-end on a single GPU, thereby using a single store of data for simulation roll-outs, inference, and training.