pytorch

a8a10903 - Perform appropriate CUDA stream synchronization in distributed autograd. (#53929)

Perform appropriate CUDA stream synchronization in distributed autograd. (#53929) Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/53929 The local autograd engine performs appropriate stream synchronization between autograd nodes in the graph to ensure a consumer's stream is synchronized with the producer's stream before executing the consumer. However in case of distributed autograd, the SendRpcBackward function receives gradients over the wire and TensorPipe uses its own pool of streams for this purpose. As a result, the tensors are received on TensorPipe's stream pool but SendRpcBackward runs on a different stream during the backward pass and there is no logic to synchronize these streams. To fix this, I've enhanced DistEngine to synchronize these streams appropriately when it receives grads over the wire. ghstack-source-id: 124055277 (Note: this ignores all push blocking failures!) Test Plan: 1) Added unit test which reproduced the issue. 2) waitforbuildbot. Reviewed By: walterddr, wanchaol Differential Revision: D27025307 fbshipit-source-id: 2944854e688e001cb3989d2741727b30d9278414