[python-package] Remove dead Python 2 guard and sklearn.cross_validation fallback

[PavelGuzenfeld]

Summary

Remove two pieces of dead compatibility code from compat.py:

• assert sys.version_info[0] == 3 — XGBoost requires Python >= 3.10
  (pyproject.toml), so this check can never fail. Additionally, using
  assert for version checks is incorrect since assertions are stripped
  under python -O.

• sklearn.cross_validation import fallback — sklearn.cross_validation
  was removed in scikit-learn 0.20 (2018). Any scikit-learn version that
  supports Python >= 3.10 will have sklearn.model_selection, making
  the try/except fallback unreachable.

Files changed

• python-package/xgboost/compat.py — 7 lines removed, 1 added

[Copilot]

Pull request overview

Removes unreachable legacy compatibility code in compat.py now that XGBoost requires modern Python/scikit-learn versions.

Changes:

• Removed a dead Python 2 guard that could never trigger under the project’s Python >= 3.10 requirement.
• Removed an unreachable fallback import of sklearn.cross_validation.StratifiedKFold, importing only from sklearn.model_selection.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[RAMitchell]

Please address CI failures

[PavelGuzenfeld]

Fixed the pre-commit failure — ruff flagged two formatting issues:

• Missing blank line before def py_str (need 2 blank lines before top-level functions)
• Extra blank line between RegressorMixin and StratifiedKFold imports

The Windows test failure was unrelated (self-hosted runner lost connectivity).

[PavelGuzenfeld]

The only CI failure — Google Test (C++) CUDA 12 (mgpu) — is unrelated to this PR's changes. It failed during Docker container initialization due to an ECR authentication error:

Docker login for '492475357299.dkr.ecr.us-west-2.amazonaws.com' failed with exit code 1

This is a transient infrastructure issue (AWS ECR auth), not a code problem. A maintainer with admin access would need to re-trigger the failed job.

[PavelGuzenfeld]

Hi, friendly ping on this PR. It has been approved by @RAMitchell and all CI checks are passing except for a single infrastructure-related failure (AWS ECR Docker login on the CUDA mgpu job), which appears unrelated to the changes here. I attempted to re-run the failed job but don't have the permissions. Could a maintainer re-trigger that job or merge given the failure is infrastructure-related? Thank you!

[RAMitchell]

I prepared a fix for the macOS x86_64 wheel CI failure here: RAMitchell#2

It avoids pulling Homebrew LLVMs libunwind into the wheel repair step by preferring Condas llvm-openmp runtime for macOS wheel builds, preserving the current macOS 10.15 deployment target instead of bumping it to 14.0.

If useful, the fix commit is caa714533c0503ca7e90d73339557314908a5341.

[PavelGuzenfeld]

Thanks @RAMitchell! I've cherry-picked your fix commit (caa7145) into this branch to resolve the macOS x86_64 wheel CI failure. Let's see if CI passes now.

[PavelGuzenfeld]

Thanks @RAMitchell! I've cherry-picked your fix commit (caa7145) into this branch to resolve the macOS x86_64 wheel CI failure. Let's see if CI passes now.

Issue persists.

[RAMitchell]

We will try to fix it in #12108

[trivialfis]

Out of curiosity, is there something like the previous k price happening? Looking at your agent tracker with the difficulty classification, looks quite similar to the trend in that competition.

[RAMitchell]

Please remove bde4a3e (macos fix) and we will merge it.