PyCUDA

[jakirkham]

Would be nice to have PyCUDA added to defaults. Given it links against CUDA, this may be the only way to distribute it ATM. Though if there is a way to get it into conda-forge as well, would be happy to discuss over at issue ( conda-forge/staged-recipes#3178 ).

[jakirkham]

Copying you on this, @jjhelmus, in case you have any ideas on how to approach this.

[jjhelmus]

I don't believe there are any plans for adding PyCUDA to defaults at this time. Do you have a working recipe for the package? That would make the process of including it in defaults considerably easier.

[jakirkham]

I see. Thanks for the info.

We have a recipe in PR ( conda-forge/staged-recipes#1775 ). However I don't know whether or not it works as we don't have nvcc. So we can't build it at conda-forge currently. Maybe @inducer can advise? 😉

[inducer]

You hit the nail on the head. At first I was excited when I saw that cudatoolkit was packaged, but when I realized that that did not include nvcc, I gave up for the time being.

[jakirkham]

Reading between the lines a bit, I think these docker images are used to build things with CUDA support. They are basically like condaforge/linux-anvil used in conda-forge, but with a different base image (e.g. nvidia/cuda:7.5-cudnn5-devel-centos6).

We tried to do the same thing some time ago in conda-forge. For instance PR ( conda-forge/docker-images#21 ) is very similar (though outdated a bit), but there were concerns about whether packages could leave the Docker images and what was allowed to leave. Though I suppose we could just start that effort again given that Continuum has likely worked out the tricky points regarding redistribution with NVIDIA directly. Just unclear on what terms conda-forge could redistribute such things.