Skip to content

Handle zero-update numerical stagnation in the SVM solver - #8609

Open
viclafargue wants to merge 1 commit into
NVIDIA:mainfrom
viclafargue:fix-svm-zero-delta-stagnation
Open

Handle zero-update numerical stagnation in the SVM solver#8609
viclafargue wants to merge 1 commit into
NVIDIA:mainfrom
viclafargue:fix-svm-zero-delta-stagnation

Conversation

@viclafargue

@viclafargue viclafargue commented Sep 10, 2026

Copy link
Copy Markdown
Contributor

Closes #8579.

Reset the kernel-cache state when an SMO iteration produces no coefficient
updates, then raise an actionable numerical-stagnation error if convergence
has not been reached.

Adds C++ and Python regression tests.

@copy-pr-bot

copy-pr-bot Bot commented Sep 10, 2026

Copy link
Copy Markdown

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.

@github-actions github-actions Bot added Cython / Python Cython or Python issue CUDA/C++ labels Sep 10, 2026
@viclafargue viclafargue added bug Something isn't working non-breaking Non-breaking change labels Sep 11, 2026
@viclafargue
viclafargue marked this pull request as ready for review September 11, 2026 08:17
@viclafargue
viclafargue requested review from a team as code owners September 11, 2026 08:17
@viclafargue
viclafargue requested a review from csadorf September 11, 2026 08:17
@coderabbitai

coderabbitai Bot commented Sep 11, 2026

Copy link
Copy Markdown

Review Change StackReview Change Stack

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: 447ad060-59c4-46d8-a755-8188aeba38c3

📥 Commits

Reviewing files that changed from the base of the PR and between 2c037a5 and 5e4d9f5.

📒 Files selected for processing (4)
  • cpp/src/svm/kernelcache.cuh
  • cpp/src/svm/smosolver.cuh
  • cpp/tests/sg/svc_test.cu
  • python/cuml/tests/test_svm.py

Included review availability: Your plan provides up to 12 included reviews per hour; 10 remain after this review.

📝 Summary

Summary by CodeRabbit

  • Bug Fixes

    • Improved SVM solver handling when optimization makes no progress.
    • Added clearer errors for numerical precision or scaling issues, including guidance to use higher precision where appropriate.
    • Improved kernel-cache state handling when no coefficient updates occur.

  • Tests

    • Added coverage for completing and restarting kernel-cache working sets.
    • Added regression coverage for stalled float32 polynomial SVR training.

Walkthrough

The change adds KernelCache::FinishWorkingSet(), uses it when the SMO solver makes no alpha update, and raises precision-specific errors when optimization stagnates. C++ and Python tests cover working-set reuse and float32 SVR stagnation.

Changes

SVR working-set completion

Layer / File(s) Summary
Kernel cache working-set lifecycle
cpp/src/svm/kernelcache.cuh, cpp/tests/sg/svc_test.cu		 KernelCache::FinishWorkingSet() transitions an initialized working set to READY. Typed tests cover repeated initialization and completion.
Solver completion and stagnation handling
cpp/src/svm/smosolver.cuh, python/cuml/tests/test_svm.py		 The solver tracks alpha updates, completes inactive working sets, and raises precision-specific errors when it cannot continue without progress. The regression test checks the float32 error message.

Priority: ➖ Normal

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

Change: Bug fix · Severity of issue fixed: Medium

Merge Risk: ⚪ Minimal · up to 5e4d9

The SVR failure path now produces an actionable numerical-stagnation error rather than an internal kernel-cache exception.

🚥 Pre-merge checks | ✅ 3 | ❌ 2

❌ Failed checks (2 warnings)

Check name Status Explanation Resolution
Out of Scope Changes check ⚠️ Warning The smosolver.cuh copyright header change adds & AFFILIATES. All rights reserved.. This change has no demonstrated connection to the kernel-cache failure or its regression tests. The remaining imp… Remove the unrelated copyright-header change from cpp/src/svm/smosolver.cuh, unless a separate repository requirement justifies it.
Docstring Coverage ⚠️ Warning Docstring coverage is 16.67% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 6 functions across 2 files. (2 skipped: 2… Write docstrings for the functions missing them to satisfy the coverage threshold.
✅ Passed checks (3 passed)
Check name Status Explanation
Description check ✅ Passed The description accurately summarizes the kernel-cache reset, numerical-stagnation error, and regression tests in the changeset.
Title check ✅ Passed The title clearly and concisely describes the main change: handling zero-update numerical stagnation in the SVM solver.
Linked Issues check ✅ Passed The changes address #8579. FinishWorkingSet() closes an initialized cache when an SMO iteration has no coefficient updates. The solver then raises a clear SMO error with float-precision or kernel-…
Full details: Out of Scope Changes check

Explanation

The smosolver.cuh copyright header change adds & AFFILIATES. All rights reserved.. This change has no demonstrated connection to the kernel-cache failure or its regression tests. The remaining implementation and tests support #8579.

Full details: Docstring Coverage

Explanation

Docstring coverage is 16.67% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 6 functions across 2 files. (2 skipped: 2 unsupported.)

  • 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.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@csadorf csadorf Awaiting requested review from csadorf csadorf is a code owner automatically assigned from NVIDIA/cuml-python-codeowners

At least 1 approving review is required to merge this pull request.

Assignees

@viclafargue viclafargue

Labels

bug Something isn't working CUDA/C++ Cython / Python Cython or Python issue non-breaking Non-breaking change

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

[BUG] SVR with a degree=10 polynomial kernel fails with an internal working-set state error

2 participants

@viclafargue @GPUtester