tensorflowWithoutCuda

[hacker1024]

The tensorflowWithoutCuda derivation changes slightly when config.cudaSupport is enabled, as some of its dependencies use CUDA.

Can this variant be added as a CI job?

[SomeoneSerge]

Sure, can you please submit the patch? Can you also open a nixpkgs issue about the attribute not being stable under config.cudaSupport? It shouldn't change, ideally

[hacker1024]

I believe it should change, as TensorFlow has dependencies that use CUDA. If it were to override those, it might lead to conflicts with other packages in a larger CUDA-enabled Python environment.

[SomeoneSerge]

Yes, it's likely that tensorflowWithoutCuda would end up being non-composable with many other python packages. It might still work for python "applications". It'll also become less of a concern once adisbladis's work (requiredPythonModules) gets merged. We might want to reconsider whether tensorflowWithoutCuda is making a promise it can't actually keep.

All of those are nixpkgs concerns