Async scheduling splits rollout and training onto different GPU groups to run in parallel, at the cost of policy lag, since the training data comes from a slightly earlier version of the model, which may harm convergence stability and final performance.