wheels: use system 'tileiras', search site-packages for CMake targets

[jameslamb]

Fixes #8520
Contributes to rapidsai/build-planning#324
Contributes to rapidsai/build-planning#325

• switches wheel builds to system-installed tileiras
• updates libcuml CMake to search for pip-installed libraries (to avoid e.g. rebuilding cuVS from source in pip devcontainers)
• other minor CMake cleanup

How I tested this

Relied on CI. Check it out... 9 minute pip devcontainer jobs now 😎 : https://github.com/NVIDIA/cuml/actions/runs/34501697447/job/102954296927?pr=8573

[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]

Similar things are used in other parts of RAPIDS to ensure CMake targets installed in site-packages/ can be found.

Things like this were added in NVIDIA/cuvs#2565, for example.

[jameslamb]

/ok to test

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 3b813e94-a3a7-4cf7-ae9b-b8b9f6dfc9e6

📥 Commits

Reviewing files that changed from the base of the PR and between 82aa0ce and b048a7c.

📒 Files selected for processing (1)

• dependencies.yaml

Included review availability: Your plan provides up to 12 included reviews per hour; 11 remain after this review.

---

📝 Summary

Summary by CodeRabbit

• Installation
  • Improved detection of RAPIDS libraries installed through Python wheels.
  • Enhanced support for locating libraries in standard 64-bit library directories.
  • Updated dependency configuration for improved compatibility with CUDA 13 environments.
  • Updated supported dependency versions, including CUDA 12 bindings and scikit-learn.
• Maintenance
  • Updated third-party dependency configuration to use NVIDIA-maintained repositories.
  • Simplified CUVS dependency configuration for more consistent builds.

Walkthrough

Changes

Build dependency configuration

Layer / File(s)	Summary
Python wheel prefix discovery cpp/CMakeLists.txt, python/cuml/CMakeLists.txt, python/libcuml/CMakeLists.txt	CMake discovers RAPIDS wheel prefixes through Python, appends them to CMAKE_PREFIX_PATH, and enables global lib64 library lookup.
NVIDIA dependency source configuration cpp/cmake/thirdparty/get_cuvs.cmake, cpp/cmake/thirdparty/get_raft.cmake	CUVS and RAFT now use NVIDIA repository forks. CUVS no longer accepts the COMPILE_LIBRARY argument.
cutile dependency selectors and pins dependencies.yaml	cutile_python now uses the CUDA 13 cuda-tile selector for requirements and pyproject outputs. CUDA 12 remains empty. CUDA bindings and scikit-learn pins are updated.

Priority: ➖ Normal

Estimated code review effort: 3 (Moderate) | ~20 minutes

Severity of issue fixed: Medium

Merge Risk: ⚪ Minimal · up to b048a

No actionable merge-blocking issue remains from the finalized review.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Out of Scope Changes check	⚠️ Warning	The main changes are in scope, but the scikit-learn version update, RAFT fork change, copyright-only edits, and some dependency metadata changes are not clearly required by issue #8520.	Provide linked issue requirements that justify these additional changes, or remove or split the unrelated dependency, fork, and copyright updates into a separate pull request.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the two primary changes: using system-installed tileiras and searching site-packages for CMake targets.
Description check	✅ Passed	The description directly covers the wheel, tileiras, site-packages, CMake, and linked issue changes.
Linked Issues check	✅ Passed	The CMake search-path changes and system tileiras configuration address issue #8520 by allowing pip devcontainer builds to use installed cuVS-related packages instead of rebuilding cuVS from source.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0…

• Fix all pre-merge checks with AI

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[jameslamb]

conda Python tests are all failing like this:

FAILED test_core.py::test_constraints_match_installed_versions - AssertionError: Please update `cuml.accel.core._CONSTRAINTS` to match the new runtime constraints.
  
  Currently the following constraints failed:
  
  - scikit-learn<=1.9.0,>=1.6.0
  
  leading to the following user-facing logs:
  
  [cuml.accel] Warning: `cuml.accel` version 26.10.00a117 was tested with scikit-learn<=1.9.0,>=1.6.0, but scikit-learn=1.9.1 was found. Things may work, but you may also encounter potentially subtle breakage.
assert False

(build link)

That'll be fixed by #861

oldest-deps wheel jobs are failing like this:

  File "/pyenv/versions/3.11.16/lib/python3.11/site-packages/cuml/internals/outputs.py", line 9, in <module>
    import cudf
  File "/pyenv/versions/3.11.16/lib/python3.11/site-packages/cudf/__init__.py", line 20, in <module>
    from cudf import api, core, datasets, testing
  File "/pyenv/versions/3.11.16/lib/python3.11/site-packages/cudf/api/__init__.py", line 4, in <module>
    from cudf.api import extensions, types
  File "/pyenv/versions/3.11.16/lib/python3.11/site-packages/cudf/api/extensions/__init__.py", line 6, in <module>
    from cudf.api.extensions.accessor import (
  File "/pyenv/versions/3.11.16/lib/python3.11/site-packages/cudf/api/extensions/accessor.py", line 6, in <module>
    from cudf.core.dataframe import DataFrame
  File "/pyenv/versions/3.11.16/lib/python3.11/site-packages/cudf/core/dataframe.py", line 38, in <module>
    import pylibcudf as plc
  File "/pyenv/versions/3.11.16/lib/python3.11/site-packages/pylibcudf/__init__.py", line 14, in <module>
    from . import (
ImportError: /pyenv/versions/3.11.16/lib/python3.11/site-packages/pylibcudf/rolling.abi3.so: undefined symbol: _ZN4cudf18make_range_windowsERKNS_10table_viewERKNS_11column_viewENS_5orderENS_10null_orderESt7variantIJNS_9unboundedENS_11current_rowENS_14bounded_closedENS_12bounded_openEEESD_N4cuda3__410stream_refENSE_2mr3__420__version_bump_ver4_12resource_refIJNSI_17device_accessibleEEEE

(build link)

Not sure if that's being looked into yet, I'll look into it.

[jameslamb]

Not sure if that's being looked into yet, I'll look into it.

This was NVIDIA/cudf#24116, see that issue for the resolution. No changes needed in this PR, just a re-run.

[jameslamb]

/merge