I am trying to build CUDA code using Circle CI on both Windows and Linux.
I don’t want to run the CUDA code, just build it. Therefore I need the CUDA toolkit and compilers installed, but I don’t need a GPU. I have previously got the build to work locally on a Windows VM with no GPU.
Based on the docs at Execution environments overview - CircleCI I have tried the following config
jobs: build-windows: machine: image: 'windows-server-2019-cuda:current' shell: 'powershell.exe -ExecutionPolicy Bypass' resource_class: 'windows.medium' ...
But this fails with the error ‘Job was rejected because resource class medium, image windows-server-2019-cuda:current is not a valid resource class’
Is it possible to allow the windows-server-2019-cuda image to run on non GPU machines please? Or alternatively is it possible to get the CUDA runtime added to the standard Windows images? Currently what I’m having to do instead is install CUDA manually using ‘choco install cuda -y’ but this adds 10 minutes plus to every build. To reduce this time I’m looking at caching the CUDA runtime install dir, but I’m not sure it will work as after restoring from the cache I don’t think that the CUDA integration with Visual Studio will be setup correctly which I think might disrupt the MSBuild run. Although to be fair, I haven’t tried it yet.
Any help or advice here would be much appreciated.