How to install cudatoolkit package on the WinPython distribution

[cristipurdel]

Just saw it on stackoverflow
https://stackoverflow.com/questions/55027544/how-to-install-cudatoolkit-package-on-the-winpython-distribution
It would be really helpful if I could install cudatoolkit fron numba via pip, in order to have a portable python distribution.
Any plans on adding this support?

[stuartarchibald]

Thanks for the request. The cudatoolkit packages are just a redistribution of the distributable parts of the NVIDIA cudatoolkit binaries and are built from the conda recipe in this repo. I presume what you are looking for is to be able to do:

pip install cudatoolkit

?
Which means some group has to take responsibility for providing and maintaining this. I'll raise it at the next core developer meeting for Numba org.

[seibert]

Note that distributing packages with no Python in them (like the cudatoolkit) is not really in the spirit of what wheels are for, although if such a package were hosted elsewhere than PyPI, anything is possible.

On a different note, we should ensure that Numba will automatically attempt to use the CUDA toolkit installed at the system level if it doesn't find one inside its environment. We currently require the use of awkward (and anachronistically named) environment variables in this case, which no longer makes sense. If Numba were more streamlined this way, then we could tell non-conda users to install CUDA from the NVIDIA developer site, as many other CUDA applications do.

[cristipurdel]

pip install cudatoolkit

This is exactly what I want to do.
It will help also with testing various cudatoolkit versions, even portable when there is no cudatoolkit installed due to admin rights.

[EMJzero]

I just bumped in the same frustrating thing...so isn't there any way so to get around anaconda in this case? I already have some hotswappable versions of the toolkit on my machine, i'm not quite in the mood of installing others only for numba :(

[seibert]

We will fix the next release of Numba to detect cudatoolkit in the system if none is found in the conda environment.

[cristipurdel]

I would prefer to have the cudatoolkit as wheel in order to be portable and run without admin rights.
If the system (no admin rights) does not have cuda installed, I cannot neither install conda, nor cuda.
You could maintain for example only the lowest cudatoolkit supported by current numba version.

[EMJzero]

We will fix the next release of Numba to detect cudatoolkit in the system if none is found in the conda environment.

Perfect, I'll look forward to that, thanks!

[Zhiwei-Zhai]

I had a similar problem. Now I find a way to solve it.

For windows system, you can download and install Cuda Toolkit from CUDA Toolkit Archive.
and install numba by pip install numba
According to the searching order of Numba for a CUDA toolkit installation, you could add an enviroment of CUDA_HOME=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.2.

Then it works for me.

[vinaydandwani]

@chushan89 It worked. Thanks a lot