WIP: wheels: test on mix of CTK versions

[jameslamb]

Description

Similar to #21671, but without the changes to building wheels with CTK 13.0. Using this to test whether cuDF wheels exhibit similar issues to rapidsai/cugraph#5443 in a runtime environment with an older nvidia-nvjitlink.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[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.

[jameslamb]

/ok to test

[jameslamb]

aha!

libcudf wheels currently have a requirement like nvidia-nvjitlink>=13.0

cudf/dependencies.yaml

Line 614 in daa44ee

- nvidia-nvjitlink>=13.0

That floor is wrong... it should be >=13.1 because we build against CTK 13.1 (as I've been testing in rapidsai/build-planning#256 and rapidsai/build-planning#257).

This nvidia-nvjitlink==13.0.* pin in constraints forced out the runtime issues users would see if they installed cudf-cu13 and nvidia-nvjitlink==13.0.* together (as might happen in an environment with torch wheels):

FAILED test_transform.py::test_transform_udf - RuntimeError: Jitify fatal error: Dynamically loaded libnvJitLink.so.13 is too old; it is version 13.0, need at least version 13.1
FAILED io/test_parquet.py::test_read_parquet_filters_jit[a.txt-pa_filter0-plc_filter0-True] - RuntimeError: Jitify fatal error: Dynamically loaded libnvJitLink.so.13 is too old; it is version 13.0, need at least version 13.1
FAILED io/test_parquet.py::test_read_parquet_filters_jit[a.txt-pa_filter1-plc_filter1-True] - RuntimeError: Jitify fatal error: Dynamically loaded libnvJitLink.so.13 is too old; it is version 13.0, need at least version 13.1
FAILED io/test_parquet.py::test_read_parquet_filters_jit[binary_source_or_sink1-pa_filter0-plc_filter0-True] - RuntimeError: Jitify fatal error: Uninitialized
FAILED io/test_parquet.py::test_read_parquet_filters_jit[binary_source_or_sink1-pa_filter1-plc_filter1-True] - RuntimeError: Jitify fatal error: Uninitialized
FAILED io/test_parquet.py::test_read_parquet_filters_jit[BytesIO-pa_filter0-plc_filter0-True] - RuntimeError: Jitify fatal error: Uninitialized
FAILED io/test_parquet.py::test_read_parquet_filters_jit[BytesIO-pa_filter1-plc_filter1-True] - RuntimeError: Jitify fatal error: Uninitialized

(build link)

[jameslamb]

And notice this only happens on the CUDA 13.0 jobs (not 13.1):

[jameslamb]

This was just for testing, to get the error messages in #21808 (comment) confirming that our nvidia-nvjitlink wheel pins here were too permissive.

Closing this, see #21671