Stable Install Path Testing - Phase 1

[gforsyth]

In #226 @csadorf proposed a new workflow in rapidsai/integration for stable install path testing for RAPIDS.

For an initial phase, we want to create a workflow that tests that an automated
installation of RAPIDS works as per the documented commands that we advertise to
users.

I propose that as a first step, we implement the following:

Test that we can install and import all of the RAPIDS packages

Using the install commands from the install selector tool at https://docs.rapids.ai/install/ with the following variants:

• For pip with the pypi.nvidia.com index included:
  • Across all supported Python versions (3.10 - 3.13 as of writing)
  • Across supported CUDA versions (12 and 13 as of writing)
  • On x86₆₄ and arm64
• For pip with only the default PyPI index included:
  • Across all supported Python versions (3.10 - 3.13 as of writing)
  • Across supported CUDA versions (12 and 13 as of writing)
  • On x86₆₄ and arm64
  • Installing cudf, dask-cudf, cuml, pylibraft, raft-dask (these are the packages that have wheels on PyPI)
• For conda
  • Across all supported Python versions (3.10 - 3.13 as of writing)
  • Across supported CUDA versions (12 and 13 as of writing)
  • On x86₆₄ and arm64

Test Environments

• Use nvidia/cuda for the pip tests above:

  • For CUDA12 and CUDA13 ubuntu images:
    • Install curl and then uv
    • Create per-Python venv with e.g. uv venv --python=3.10 --seed
    • Activate venv and run install commands
    • Test that packages import without error (import each package in a separate Python session to flag up any symbol loading errors)

  This might be fast enough to just do in a single container, installing each Python version in succession.

• Use rapidsai/miniforge-cuda for the conda tests above (there shouldn’t be any configuration changes needed)

Workflows

• We can use custom-job shared workflow for node-selection and point it to the image required for pip vs. conda testing

[gforsyth]

Completed in rapidsai/integration#837