zixuanweeei commented on a change in pull request #17075: [MKLDNN] mkldnn RNN operator enhancement URL: https://github.com/apache/incubator-mxnet/pull/17075#discussion_r358067041
########## File path: docs/static_site/src/pages/api/faq/env_var.md ########## @@ -349,6 +349,10 @@ If ctypes is used, it must be `mxnet._ctypes.ndarray.NDArrayBase`. - Values: 0(false) or 1(true) ```(default=1)``` - If this variable is set, MXNet will simplify the computation graph, eliminating duplicated operations on the same inputs. +* MXNET_USE_MKLDNN_RNN + - Values: 0(false) or 1(true) ```(default=1)``` + - This variable controls whether to use the MKL-DNN backend in fused RNN operator for CPU context. There are two fusion implementations of RNN operator in MXNet. The MKL-DNN implementation has a better performance than the naive one, but the latter is more stable in the backward operation currently. Review comment: I think it is not stable in the backward pass. I have trained the bucketing model (https://github.com/apache/incubator-mxnet/tree/master/example/rnn/bucketing) with the backend of MKL-DNN RNN Backward. It resulted in a convergent optimizing curve. But it has not been tested in other applications for training a model. So I provided an env variable for users to switch to the naive implementation. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
