Avoid initializing gradInput twice in the backward phase of replication (#48890)
Summary:
https://github.com/pytorch/pytorch/issues/48889
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48890
Reviewed By: zhangguanheng66
Differential Revision: D25375697
Pulled By: ezyang
fbshipit-source-id: fd6f6089be44e68c4557b923550c7cadb90d739a
