test: isolate library wheel smoke test

[bdice]

Description

Install the library wheel in a clean virtual environment and call its load_library() entry point before installing the high-level wheel or test extras. This exposes missing library-wheel runtime dependencies that the full test environment can otherwise mask.

xref: rapidsai/build-planning#307

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[bdice]

/ok to test 5825636

[bdice]

/ok to test 5d4d61d

[bdice]

/ok to test 21150af

[jameslamb]

CI got stuck here in the latest GitHub outage, just merged in latest main to trigger a new run. I'll check back in a bit.

[jameslamb]

This is failing only for CUDA 12 right now.

+ python -c 'import libcugraph; assert (libraries := libcugraph.load_library()) and all(libraries)'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
    import libcugraph; assert (libraries := libcugraph.load_library()) and all(libraries)
                              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError

(build link)

I could reproduce that locally.

docker run \
  --rm \
  --gpus all \
  -it rapidsai/citestwheel:26.10-cuda12.9.2-rockylinux8-py3.13 \
  bash

pip install \
  --extra-index-url https://pypi.anaconda.org/rapidsai-wheels-nightly/simple/ \
  'libcugraph-cu12>=26.10.0a0'

python -c "import libcugraph; assert (libraries := libcugraph.load_library()) and all(libraries)"

Picking through the internals of libcugraph.load_library(), the root error is from loading libcuvs.so.

>>> lib = "/pyenv/versions/3.13.15/lib/python3.13/site-packages/libcuvs/lib64/libcuvs.so"
... ctypes.CDLL(lib, ctypes.RTLD_LOCAL)
... 
Traceback (most recent call last):
  File "<python-input-2>", line 2, in <module>
    ctypes.CDLL(lib, ctypes.RTLD_LOCAL)
    ~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^
  File "/pyenv/versions/3.13.15/lib/python3.13/ctypes/__init__.py", line 361, in __init__
    self._handle = self._load_library(name, mode, handle, winmode)
                   ~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/pyenv/versions/3.13.15/lib/python3.13/ctypes/__init__.py", line 403, in _load_library
    return _dlopen(name, mode)
OSError: libnvrtc.so.12: cannot open shared object file: No such file or directory

libcuvs is missing cuda-pathfinder, which it needs to load nvRTC. I put up a fix here: NVIDIA/cuvs#2524

Once that's merged, we should see CI pass on a re-run here... in my local environment, just running pip install cuda-pathfinder was sufficient for libcugraph.load_library() to succeed.

[jameslamb]

Seeing jobs pass now that that cuVS fix was merged. I'll queue this to merge.

[jameslamb]

/merge