pytorch

c758d9ec - [optim][adamax] group tensors in foreach to maximize perf