Pre-compiled tests failing

C++
1

Sorry I did not see this posted before and no mention in the installation instructions.

In both the release and debug distributions of the libtorch libraries there are test directories with pre-compiled executable programs.

  1. Is it expected that a fair number of the test programs are not passing?
  2. is it expected that ones that fail differ between the release and debug versions of the libtorch
    library?
  3. some issue debug assertions (e.g. “list iterator not decrementable” in GraphTest.exe) pointing to programming errors rather than test failures

My system has Windows 10, CUDA 10.0, cuDNN v7.6.2, libtorch 1.2 and a Nvidia 1080 graphics card.

I’m asking because I have general issues getting anything downstream to work, and I’m trying to troubleshoot each component separately. So, if anyone has the tests working, particularly on a similar system, please let me know.

A feature request is adding a “tests passing|failing” icon generated by CI to the git repository so we know what to expect

2

Yes, they are currently missing in the Windows tests. Would you please create a ticket at https://github.com/pytorch/pytorch/issues. I tried running the tests locally and most of the tests passed. The only one failed is c10_metaprogramming_test.exe. I didn’t try the debug versions.
About the icon, there is already one in https://github.com/pytorch/pytorch or https://ci.pytorch.org/jenkins/job/pytorch-builds/job/pytorch-win-ws2016-cuda9-cudnn7-py3-trigger.

3

Thanks, this was very helpful. The solution to getting these running is the same solution you provided that fixed my other issues. the libtorch-release/lib/*.dlls need to be copied into the same directory as the exe files (or the environment variable PATH needs to have a path to the dlls). I didn’t check all of the tests, but I tested some that were failing before.