Deprecate probability parameter on SVC and LinearSVC

[switch527]

Closes #7982

Mirrors sklearn PR #32050 on cuml.SVC and cuml.LinearSVC. Sentinel + _effective_X property pattern, same as PR #7958. FutureWarning fires from fit when the user passes an explicit value.

Also fixes a latent bug in the accel proxy where _gpu_fit was reading self.probability truthily. Default SVC() on small data would have routed through the probability code path since the sentinel string is truthy.

[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.

[csadorf]

@switch527 Thank you for the contribution! Would you mind cleaning up the PR description a bit?

[switch527]

Sorry for the original PR body. It had template-y ### Changes / ### Verification sections that I should have caught before opening. I pushed a rewrite a few minutes ago, and the live body should be much cleaner now if you reload.

One label note, the PR is currently tagged breaking + improvement, but the deprecation should be non-breaking since the sentinel default preserves prior behavior (existing code just gets a warning). Could you swap when you have a moment?

[csadorf]

Thanks for updating, but the description is still overly verbose, just focus on the high-level changes we don't need a detailed prose of all changes. We label PRs that introduce a deprecation also as breaking.

[switch527]

Got it, cut it down. New body is up.

The breaking tag is noted. Thanks for the clarification.

[csadorf]

/ok to test 9a758ee

[switch527]

@jcrist, Heads-up: validate-release-target is failing because the PR's Release
field still says 26.06. Could you bump it to 26.08?

[switch527]

@jcrist , code currently says 26.10. Did you mean the next release (now 26.10 after the 26.08 bump) or literally 26.08?

[csadorf]

@jcrist , code currently says 26.10. Did you mean the next release (now 26.10 after the 26.08 bump) or literally 26.08?

I've updated it.

[switch527]

Thanks. So removal stays at 26.10 (one cycle after the 26.08 deprecation), code already reads that way, leaving the docstrings as-is.

[csadorf]

Actually, I think we should get this into the 26.06 release. Please rebase this branch on release/26.06 and update the relevant deprecation notices.

[switch527]

[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: 4a3cc590-d8e6-4768-8643-baf6d6eeaf76

📥 Commits

Reviewing files that changed from the base of the PR and between 553f48e and 9c43ad3.

📒 Files selected for processing (1)

• python/cuml/cuml_accel_tests/upstream/pytest.ini

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

• python/cuml/cuml_accel_tests/upstream/pytest.ini

---

📝 Walkthrough

Summary by CodeRabbit

• Deprecated Features

  • The probability parameter in SVC and LinearSVC is deprecated; setting it will emit a FutureWarning. Use CalibratedClassifierCV(..., ensemble=False) instead.

• Behavior

  • GPU acceleration no longer applies when probability=True; handling of the deprecation sentinel is version-aware to preserve compatibility.

• Documentation

  • Limitations updated to recommend the calibrated-classifier alternative and note non-acceleration for probability=True.

• Tests

  • Test suite updated to filter and validate the new deprecation warnings and adjusted behavior.

Walkthrough

Deprecates probability for SVC and LinearSVC with a "deprecated" sentinel, introduces _effective_probability and SKLEARN_19 handling, removes accel per-class sample-count restriction, exports effective value to sklearn internals, updates docs, and adjusts tests and warning filters.

Changes

SVC/LinearSVC probability parameter deprecation

Layer / File(s)	Summary
LinearSVC probability deprecation python/cuml/cuml/svm/linear_svc.py	Constructor default changed to probability="deprecated", adds _effective_probability mapping sentinel→False, fit() emits FutureWarning when probability explicitly set, and all internal logic (solver arg, predict branching, @available_if) now uses _effective_probability.
SVC probability deprecation with sklearn version handling python/cuml/cuml/svm/svc.py	Adds SKLEARN_19 detection, constructor default probability="deprecated", _effective_probability property, fit() warns only when explicit, pins inner clones to sentinel, and updates CPU↔GPU param translation (from_sklearn rejects probability=True, as_sklearn chooses value based on SKLEARN_19).
SVMBase export and accel override updates python/cuml/cuml/svm/svm_base.pyx, python/cuml/cuml/accel/_overrides/sklearn/svm.py	Export fitted _effective_probability instead of raw probability; accel override removes per-class sample-count restriction for probability=True and updates sentinel-handling comments.
Documentation: limitations and migration guidance docs/source/cuml-accel/limitations.rst	Replaces prior sample-count fallback text with deprecation notice and recommends sklearn.calibration.CalibratedClassifierCV(SVC(), ensemble=False) as the supported alternative to avoid CPU fallback.
Sklearn import/export integration test updates python/cuml/tests/test_sklearn_import_export.py	test_svc now excludes probability from round-trip checks; when probability=True it validates predict_proba via converted sklearn model and expects UnsupportedOnGPU when importing sklearn SVC with probability=True.
Deprecation warning test suite python/cuml/tests/test_svm.py	Adds module-level warning filter and new tests asserting FutureWarning emission semantics: explicit use warns, default sentinel does not, and warning happens once per outer fit.
Test warning filter updates multiple test modules and CI config (python/cuml/cuml_accel_tests/..., python/cuml/tests/*, pytest.ini)	Add pytest.mark.filterwarnings and module-level pytestmark entries to suppress the probability deprecation FutureWarning during test runs; TODO comments mark future removal.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

Possibly related PRs

• rapidsai/cuml#8126: Both PRs update the same test_svc_probability warning filters in python/cuml/cuml_accel_tests/integration/test_svc.py to suppress scikit-learn/cuML SVC(probability=...) deprecation FutureWarnings.
• rapidsai/cuml#7981: Both PRs modify the same SVM probability deprecation handling by recognizing scikit-learn’s "deprecated" sentinel and wiring it through cuML via _effective_probability, with corresponding test updates.

Suggested labels

cuml-accel, sklearn-api-compat, algo: svm

Suggested reviewers

• csadorf
• jcrist
• betatim

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 19.44% 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 and concisely describes the main change: deprecating the probability parameter on SVC and LinearSVC classes.
Description check	✅ Passed	The description is directly related to the changeset, explaining the deprecation mirrors sklearn PR #32050, uses a sentinel pattern from PR #7958, and notes a latent bug fix in the accel proxy.
Linked Issues check	✅ Passed	The PR successfully addresses issue #7982 by deprecating the probability parameter on SVC and LinearSVC, implementing the sentinel+_effective_probability pattern, emitting FutureWarnings, and fixing the accel proxy bug.
Out of Scope Changes check	✅ Passed	All changes are scoped to implementing the probability deprecation: updating documentation, adding sentinel defaults, implementing _effective_probability properties, updating tests, and fixing the related accel proxy bug.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[jcrist]

Thanks @switch527! This mostly looks good-to-go! I'm going to fix up the merge conflict, push up a few small tweaks, and try to shepherd this through CI to get it in for the upcoming release.

[jcrist]

Thanks for your work!

[coderabbitai]

Actionable comments posted: 2

🤖 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 `@docs/source/cuml-accel/limitations.rst`:
- Around line 510-514: Replace the incorrect reference to
sklearn.calibration.CalibratedClassifier with the correct public class
CalibratedClassifierCV in the documentation text; update the sentence that
currently mentions "wrapping SVC with sklearn.calibration.CalibratedClassifier"
to instead say "wrapping SVC with CalibratedClassifierCV" and ensure the example
remains CalibratedClassifierCV(SVC(), ensemble=False) so the class name is
consistent with the example and public API.

In `@python/cuml/cuml/svm/svc.py`:
- Around line 114-117: The docstring and user-facing deprecation warning for the
SVC `probability` parameter are inconsistent with the PR timeline; update all
occurrences that mention removal in "26.08" or the deprecated directive
"26.06"/"26.08" to the agreed removal version "26.10". Specifically edit the
deprecated directive and message strings in python/cuml/cuml/svm/svc.py (the SVC
class/docstring and any warnings triggered for the `probability` parameter) and
also change the matching strings around lines referenced (e.g., the other
instance at lines ~461-463) so both the `.. deprecated::` directive and the
warning text use "26.10".

🪄 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: c77bb940-48d8-40a1-bcc1-c1e20368a019

📥 Commits

Reviewing files that changed from the base of the PR and between 44f43bb and 1ae8806.

📒 Files selected for processing (14)

• docs/source/cuml-accel/limitations.rst
• python/cuml/cuml/accel/_overrides/sklearn/svm.py
• python/cuml/cuml/svm/linear_svc.py
• python/cuml/cuml/svm/svc.py
• python/cuml/cuml/svm/svm_base.pyx
• python/cuml/cuml_accel_tests/integration/test_svc.py
• python/cuml/cuml_accel_tests/upstream/pytest.ini
• python/cuml/tests/explainer/test_explainer_kernel_shap.py
• python/cuml/tests/explainer/test_explainer_permutation_shap.py
• python/cuml/tests/test_base.py
• python/cuml/tests/test_linear_svm.py
• python/cuml/tests/test_pickle.py
• python/cuml/tests/test_sklearn_import_export.py
• python/cuml/tests/test_svm.py

🚧 Files skipped from review as they are similar to previous changes (12)

• python/cuml/tests/test_pickle.py
• python/cuml/cuml/svm/svm_base.pyx
• python/cuml/tests/test_linear_svm.py
• python/cuml/tests/explainer/test_explainer_permutation_shap.py
• python/cuml/cuml_accel_tests/integration/test_svc.py
• python/cuml/tests/test_svm.py
• python/cuml/cuml/svm/linear_svc.py
• python/cuml/tests/test_base.py
• python/cuml/cuml/accel/_overrides/sklearn/svm.py
• python/cuml/cuml_accel_tests/upstream/pytest.ini
• python/cuml/tests/test_sklearn_import_export.py
• python/cuml/tests/explainer/test_explainer_kernel_shap.py

[jcrist]

/ok to test 553f48e

[jcrist]

/ok to test 9c43ad3

[jcrist]

/merge

[csadorf]

Thanks a lot to both of you for bringing this over the finish line!

[switch527]

Happy to help!