Recently in MXNet there's an issue about random model output when given fixed model and input, if computation happens on CPU with OpenBLAS: https://github.com/apache/incubator-mxnet/issues/11853
We found it is reproducible with OpenBLAS 0.3.1, but not reproducible with 0.2.20. It is easier to reproduce on AWS p3.16xlarge instance with Intel(R) Xeon(R) CPU E5-2686 v4.
I don't have enough knowledge to track down to the breaking point. Instead I searched a bit and now guess it may be related with https://github.com/xianyi/OpenBLAS/issues/1666 and https://github.com/JuliaLang/julia/issues/27960.
If you read julia discussion in your own link it is not reproducible with 0.3.0 and devel, i.e can you confirm that only 0.3.1 is broken?
0.3.2 should be released this weekend. You could try the current "develop" branch (or just apply the changes from PR #1667 to 0.3.1 - you can get them in patch format by appending ".diff" to the PR number in the github interface) to see if the fix from #1666 actually solves your problem.
@martin-frbg Thanks for the information. I tested it and seems 0.3.2.dev is good.
@martin-frbg looking forward to the release
Closing as 0.3.2 is released now.
Most helpful comment
0.3.2 should be released this weekend. You could try the current "develop" branch (or just apply the changes from PR #1667 to 0.3.1 - you can get them in patch format by appending ".diff" to the PR number in the github interface) to see if the fix from #1666 actually solves your problem.