libtorch build failure about vcpkg

[jimwang118]

Please provide the steps to reproduce the issue.

[dannypike]

Unfortunately, I cannot. It has taken me about six hours to get everything on my machine back to running with CUDA 11.8. I thought that I explained this, above. To reinstall CUDA 12 in order to break it again would take too long; I don't want to do that. :)

Would it be possible to create some kind of verification-tool that checks what versions of libraries libtorch thinks it can "see" and whether they are compatible with each other? It is a very complex build and, because of the versioning issues that come with using a Python thing, it's easy to get in a mess,

That would have been a very helpful diagnostic tool.

Thanks.

[jimwang118]

In addition to installing cuda, you also need to install cudnn, and their two versions need to be matched. You can check whether the two software versions on your computer meet the requirements. Could you please provide the log of libtorch installation failure?

[dannypike]

Thanks for the quick response. I'll try that out when I next have a spare weekend.

I don't want to break the setup that I have right now, because it'll stop me from working. At the moment, libtorch is happy to build with CUDA 11.8.

[jimwang118]

Your issue is solved, I will close this issue. If you have other questions, please submit a new issue.