Publish a conda package to conda-forge

[leofang]

[bdice]

We already have a cccl conda package. What relationship will exist between cccl and cuda-cooperative? Does one depend on another at build time/run time? Do we ship cuda-cooperative features as a part of cccl?

https://github.com/conda-forge/cccl-feedstock/blob/main/recipe/meta.yaml

[leofang]

cuda.cooperative is a Python module that offers callable CUB warp/block routines for Numba CUDA kernels (https://github.com/NVIDIA/cccl/pull/1973). In this case, CCCL headers are a run-time dependency of cuda-cooperative.

Currently for the wheel build, we opt to keep a private copy of CCCL headers: https://github.com/NVIDIA/cccl/blob/2600135d12e74e5a022eb3038a90d71a9e5fe2ac/python/cuda/setup.py#L52-L58 but for conda we probably can just depend on the cccl package (and not cuda-cccl, since the intention was to follow the trunk as closely as possible).

[cpcloud]

@leofang It looks like there's already a cuda-cccl package (https://github.com/conda-forge/cuda-cccl-feedstock).

What needs to be done for this issue to be resolved?

[cpcloud]

This is done.

[leofang]

xref https://github.com/conda-forge/staged-recipes/pull/30963

[leofang]

As of v0.3.3 / https://github.com/conda-forge/cccl-python-feedstock/pull/19, the support matrix of conda/pip are on par: Linux (x64/arm64) & Windows (x64).