Emphasize all DDP forward() outputs must participate in computing loss (#20586)
Summary:
CC borguz chenyangyu1988
Pull Request resolved: https://github.com/pytorch/pytorch/pull/20586
Reviewed By: ezyang
Differential Revision: D15373674
Pulled By: mrshenli
fbshipit-source-id: b986918b3592616a9bcc88fba1b8fd53016f68d7