ConvTranspose3d cuDNN bug non-contiguous input continues to persist

Continue to have the problem

which appears to be a hot topic in GitHub even today:

Can anything be done, for example, contact NVIDIA?

Hi Nicolo,

I couldn’t reproduce the issue with the last code snippet posted here. Is this code raising this exception on your system?
If not, could you provide an executable code snippet so that we can reproduce this issue?

We are currently working on a similar fix to this issue, which is also related to large inputs/outputs.


I think the problem is due to large matrices. I don’t think it’s a problem of torch but rather of CuDNN, the problem is that working high resolution with volumes here is impossible.