How does this work in the context of a pipeline? One of the steps is a multi gpu training that requires accelerate.