CPU only install

[torch-points3d]

CPU only install

[vlaminckaxel]

Can we tag a new release with these improvements?

[nicolas-chaulet]

There is no way to force CPU only at the moment. If you have a GPU it will compile the CUDA kernels. Is this a problem?

[DLwbm123]

Yes, I tried to install with torch==1.1.0 and CUDA 9.0 and failed with the same error("TORCH_CHECK") in #37 . The installation is successful on my laptop without GPU. So I wonder if torch-points-kernels can be installed with CPU only.

[nicolas-chaulet]

Not at the moment no... You could clone the repo and deactivate CUDA install manually before installing from source.

[vlaminckaxel]

We also could just install the cuda ops when the gpu version of torch is installed.
https://github.com/nicolas-chaulet/torch-points-kernels/pull/60

[nicolas-chaulet]

Do you think that would work when installing the package in a docker image (for example during CI)?

[vlaminckaxel]

If we have a docker with Cuda installed it should be possible.
But to test not only the compilation of the ops we also should have a GPU instance.

I will upgrade the cicd pipeline.

[nicolas-chaulet]

@CCInc This was solved right?

[CCInc]

With the new version, I think the CPU version of torch will only install the CPU ops automatically. Existing docker images will need to be updated to use FORCE_CUDA=1 to compile the GPU ops. In any case, you can force CPU only installation with FORCE_ONLY_CPU=1.

[nicolas-chaulet]

Yes doing that now!

[nicolas-chaulet]

Ok, 0.7.0 should be published now! Let me know if that works for you and we can close this issue.

[Stakhan]

Hi there,

I was also trying to install torch-points-kernels (to further install torch-points3d) with CPU only version and I can confirm it works on my side (Ubuntu 20.04) with version 0.7.0.

[nicolas-chaulet]

Thanks!