Add a CI to test conda-forge/cupy-feedstock?

[leofang]

I am just throwing out some ideas for discussion. 🙂

As you may know, conda-forge's CIs do not have access to GPUs. So, we've been building packages in the dark and had hit bugs a few times. I am wondering if we could enable a CI in this repo that uses PFN's resources to do testing (as an extra safe).

I imagine we could do this:

• use one of conda-forge's docker image (such as condaforge/linux-anvil-cuda:10.2)
• install miniconda and all ingredients for conda-build
• build the recipe from conda-forge/cupy-feedstock's master
• test the built package locally by running CuPy's test suite

Likely we don't have to test the full build matrix; PY38 + CUDA 10.2 should provide a representative data point.

One thing I am wondering: If this could be set up, is it possible to invoke the CI here when conda-forge/cupy-feedstock gets new PRs?

Any thoughts?

cc: @jakirkham

[jakirkham]

FWIW the tests are actually included in the Conda package. So one can just run the tests like so.

conda build --test /datasets/jkirkham/miniconda/pkgs/cupy-7.3.0-py37h940342b_0.tar.bz2

[leofang]

I guess I was thinking testing the recipe against every PR.

[jakirkham]

Every PR against the conda-forge repo or every PR against CuPy?

[jakirkham]

There has been discussion periodically about uploading packages from PRs in conda-forge. Listed a few relevant threads below. There is also some work in other areas that may crossover with this need.

xref: conda-forge/conda-forge.github.io#124
xref: conda-forge/conda-forge.github.io#204
xref: conda-forge/conda-forge.github.io#647

[leofang]

Every CuPy PR if resource permits, if not then at least every CF PR. We don't have to upload from here; this just provides extra insurance and we can still use the current approach for uploading.

[leofang]

@jakirkham My impression is the situation has changed and now RAPIDS uses gpuCI somewhere to test the CuPy conda package? If so then this is no longer needed. (I don't even remember I opened this...)

[jakirkham]

Nope this is still an issue. Though conda-forge is exploring this issue

[leofang]

Thanks, John! Just saw this new issue: conda-forge/conda-forge.github.io#1272.