Conda-forge package

[brynpickering]

I'd like to create cuPDLPx conda-forge feedstock so it's easier for users to create a working environment. I think it should be straightforward:

• cuda, cuda-nvcc, gcc, and cmake as build dependencies
• run the cmake build

Even better would be to use the PyPI distro so that it comes with a Python API (it isn't necessary for a conda-forge package to have a Python API, but it's a nice addition). I think this would then be:

• cuda, cuda-nvcc, gcc, and cmake as build dependencies
• pip install cupdlpx

The first option is working for me locally, the second isn't (does not appear to contain CMakeLists.txt. style error that I'm yet to successfully debug).

Is this something you've looked into at all and would have pointers on?

[ZedongPeng]

Hi @brynpickering, thanks so much for looking into this. There seems to be an issue with the current PyPI installation path — I ran into the same does not appear to contain CMakeLists.txt error on my side.

For the time being, I’d suggest building and installing cuPDLPx from source instead, via
pip install git+https://github.com/MIT-Lu-Lab/cuPDLPx.git@v0.2.5

[ZedongPeng]

Hi @brynpickering. We just published cuPDLPx v0.2.7 on PyPI, and this issue should now be fixed.

[LucasBoTang]

Hi @brynpickering, thanks again for looking into this.

The issue should now be fully resolved. Please try installing the latest release:

pip install cupdlpx==0.2.7

[brynpickering]

Started work on it here

[brynpickering]

After quite some trial and error, I have found a recipe that builds successfully. @LucasBoTang would you be able to review the recipe and the most recent CUDA build logs in that PR to ensure I haven't missed something critical?

[brynpickering]

And perhaps I add you also as a recipe maintainer?

[LucasBoTang]

@brynpickering Thanks a lot for working through this!

I’ll take a careful look at the recipe and get back to you if I spot anything concerning.

And yes, I’d be happy to be added as a recipe maintainer.