or should i create a new PR. What I want is an easy capability of training on a single node with multiple gpus