Fix KNN neighbor buffer layout bugs

[nethum529]

Summary

• Copy callable weights into aligned C-contiguous CuPy storage before the KNN kernel reads them.
• Convert sparse self-edge indices to contiguous int64 before the raw kernel while keeping explicit sparse query results as int32.
• Add regressions for strided weights, unaligned weights, and duplicate sparse rows.

Testing

• 14 passed for the KNN weights test selection.
• Compute Sanitizer reports ERROR SUMMARY: 0 errors for the unaligned callable path.
• Changed-file pre-commit checks pass.

Fixes #8362

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[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: 7a356767-15b0-4c4d-8174-5d525c0183ce

📥 Commits

Reviewing files that changed from the base of the PR and between 6607a3e and 565dcf3.

📒 Files selected for processing (1)

• python/cuml/cuml/neighbors/nearest_neighbors.pyx

🚧 Files skipped from review as they are similar to previous changes (1)

• python/cuml/cuml/neighbors/nearest_neighbors.pyx

---

📝 Walkthrough

Summary by CodeRabbit

• Bug Fixes

  • Improved KNN regressor reliability with custom callable weights by ensuring returned GPU weights are consistently float32 and C-order contiguous.
  • Fixed sparse nearest-neighbor self-query behavior by ensuring indices are correctly formatted, excluding self-edges, and returning consistent distance and index data.

• Tests

  • Added coverage for callable weight functions returning non-contiguous GPU arrays, verifying predictions match scikit-learn.
  • Added a sparse nearest-neighbor regression test validating self-edge exclusion and result equivalence to scikit-learn.

Walkthrough

The changes normalize callable KNN weights to C-contiguous float32 arrays and normalize sparse self-edge indices to contiguous int64 arrays. New regression tests cover non-contiguous callable weights and sparse self-query behavior.

Changes

Neighbors layout fixes

Layer / File(s)	Summary
Align callable weights python/cuml/cuml/neighbors/weights.py, python/cuml/tests/test_kneighbors_regressor.py	Callable weight results are converted to C-contiguous float32 storage, with coverage for non-contiguous CuPy outputs.
Normalize sparse self-edge indices python/cuml/cuml/neighbors/nearest_neighbors.pyx, python/cuml/tests/test_nearest_neighbors.py	Self-edge removal converts indices to contiguous int64 storage, with sparse self-query coverage against scikit-learn results.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Suggested reviewers: jcrist

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 60.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly summarizes the primary KNN buffer layout fixes.
Description check	✅ Passed	The description accurately covers the KNN buffer fixes, regression tests, testing results, and linked issue.
Linked Issues check	✅ Passed	The changes fix callable-weight contiguity and sparse self-edge index-width issues and add relevant regression tests for issue [#8362].
Out of Scope Changes check	✅ Passed	The code and regression tests stay within the linked issue scope; copyright header updates are incidental metadata changes.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[viclafargue]

LGTM for the most part, but I have one comment.

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@python/cuml/cuml/neighbors/weights.py`:
- Around line 59-61: Update the callable-weights conversion in weights.py to use
an unconditional allocating CuPy conversion, such as cp.array with copy=True,
while preserving float32 dtype and C order so the resulting pointer is aligned
for kneighbors_regressor.pyx. Add coverage using an unaligned C-contiguous
device-array result from the callable and verify the conversion produces a safe
aligned allocation.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: e206c94b-e742-476f-ad69-1d3114bec701

📥 Commits

Reviewing files that changed from the base of the PR and between ee12f2c and 0d98baa.

📒 Files selected for processing (2)

• python/cuml/cuml/neighbors/weights.py
• python/cuml/tests/test_kneighbors_regressor.py

💤 Files with no reviewable changes (1)

• python/cuml/tests/test_kneighbors_regressor.py

[nethum529]

Pushed the simplification and merged latest upstream main. The whole alignment-handling block is replaced by the one-liner as suggested, and the regression tests pass unchanged since they only assert behavior.

[viclafargue]

Thanks! Please update the PR description to remove the claim that unaligned weights are handled. Otherwise, this looks good to me.

It looks like cuPy can successfully convert and reorder unaligned memory. However, if the type and orders are correct the unaligned pointer is preserved throughout.

[nethum529]

Will do & Happy to help :)

[viclafargue]

/ok to test d1c8ef1

[viclafargue]

/ok to test 6607a3e

[viclafargue]

/ok to test 565dcf3