pytorch

c86c7fb7 - [optim][asgd] group tensors in foreach to maximize perf