build wheels with CUDA 13.0.x, test wheels against mix of CTK versions

[jameslamb]

Contributes to rapidsai/build-planning#257

• builds CUDA 13 wheels with the 13.0 CTK
• ensures wheels ship with a runtime dependency of nvidia-nvjitlink>={whatever-minor-version-they-were-built-against}

Contributes to rapidsai/build-planning#256

• updates wheel tests to cover a range of CTK versions (we previously, accidentally, were only testing the latest 12.x and 13.x)

Notes for Reviewers

Doesn't this need conda changes too?

I don't think so.

cuML doesn't directly use libnvjitlink in conda packages, and it dynamically links to libcuvs:

cuml/cpp/CMakeLists.txt

Line 64 in 95b7be3

option(CUML_USE_CUVS_STATIC "Build and statically link the CUVS library" OFF)

https://github.com/rapidsai/cuml/blob/main/conda/recipes/libcuml/recipe.yaml#L118-L126

Wheels need a runtime dependency on nvidia-nvjitlink because they statically link cuVS:

cuml/python/libcuml/CMakeLists.txt

Line 54 in 95b7be3

set(CUML_USE_CUVS_STATIC ON)

[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.

[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: Pro

Run ID: 004c79f2-e11c-4e7c-9089-20758920ae3f

📥 Commits

Reviewing files that changed from the base of the PR and between af81dc5 and a6fb39b.

📒 Files selected for processing (3)

• dependencies.yaml
• python/cuml/pyproject.toml
• python/libcuml/pyproject.toml

🚧 Files skipped from review as they are similar to previous changes (2)

• python/libcuml/pyproject.toml
• dependencies.yaml

---

📝 Walkthrough

Summary by CodeRabbit

• Chores

  • CI/build now sources install constraints from an environment variable and consolidates constraint usage.
  • Pip installs prefer binary wheels and simplify test extras handling.

• Dependencies

  • Separated nvjitlink into its own dependency and pinned CUDA toolkit to 13.* where applicable.
  • Added explicit support matrix for additional CUDA versions (12.2–13.1) with improved architecture-specific handling.

Walkthrough

Replace local constraint files with an environment-provided PIP_CONSTRAINT, add --prefer-binary to several pip installs, split nvjitlink out of cuda-toolkit into a separate nvidia-nvjitlink dependency, and expand/explicitly pin CUDA-version matrices in dependencies.yaml.

Changes

Cohort / File(s)	Summary
CI test scripts ci/test_wheel.sh, ci/test_wheel_dask.sh, ci/test_wheel_integrations.sh	Switch constraint sourcing from ./constraints.txt to the PIP_CONSTRAINT environment variable; add --prefer-binary to pip install commands; simplify extras usage (e.g., use [test]) and consolidate constraint flags.
Dependency matrix dependencies.yaml	Introduce explicit CUDA minor-version entries (12.2, 12.5, 12.8, 12.9, 13.0, 13.1) and use_cuda_wheels selector branches; pin cuda-toolkit per-version selectors and add explicit nvidia-nvjitlink requirements where appropriate; document selector behavior and nvJitLink ↔ CTK minor-version relationship.
Project manifests python/cuml/pyproject.toml, python/libcuml/pyproject.toml	Remove nvjitlink from cuda-toolkit[...] extras and change cuda-toolkit constraint to ==13.*; add a separate top-level dependency nvidia-nvjitlink>=13.0,<14 (aligning nvjitlink versioning with CTK 13.x).

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

Possibly related PRs

• Add nvjitlink to cuda-toolkit pip extras for cusparse compatibility #7794: Related changes to how nvjitlink is handled vs. cuda-toolkit (adds/removes nvjitlink from extras).
• Update to 26.06 #7879: Related dependency/version-pin adjustments in project manifests and dependency matrices.

Suggested labels

CUDA/C++

Suggested reviewers

• gforsyth
• bdice

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly describes the main changes: building wheels with CUDA 13.0.x and testing them against multiple CTK versions, which aligns with the changeset.
Description check	✅ Passed	The description provides clear context for the changes: CUDA 13 wheel building, nvjitlink runtime dependencies, and expanded CTK version testing, all matching the actual changeset.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

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

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

📝 Coding Plan

• Generate coding plan for human review comments

---

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

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@ci/test_wheel_integrations.sh`:
- Line 33: The BERTopic install call using the command "rapids-pip-retry install
--prefer-binary bertopic" omits the pip constraint that other wheel tests use;
either append the constraint flag (--constraint "${PIP_CONSTRAINT}") to that
install invocation so it uses the same generated constraints, or add a short
inline comment above the "rapids-pip-retry install --prefer-binary bertopic"
line explaining why constraints are intentionally excluded for this integration
test (and reference that PIP_CONSTRAINT is set by rapids-init-pip) so the
deviation is explicit and auditable.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: ddae4d85-e5cf-4f8a-a961-c0b0f6a7c7d2

📥 Commits

Reviewing files that changed from the base of the PR and between 5a00a4d and af81dc5.

📒 Files selected for processing (6)

• ci/test_wheel.sh
• ci/test_wheel_dask.sh
• ci/test_wheel_integrations.sh
• dependencies.yaml
• python/cuml/pyproject.toml
• python/libcuml/pyproject.toml

[divyegala]

The solver does not find an issue with this vs the tight pinning in the cudatoolkit metapackage?

[jameslamb]

Right, if the environment doesn't have requirement cuda-toolkit[nvjitlink] anywhere, then cuda-toolkit's == pin on nvidia-nvjitlink doesn't matter.

If someone were to pin cuda-toolkit[nvjitlink]=={specific-version} that's also find, as long as {specific-version}>=13.0 (taking advantage of the fact the nvJitLink major.minor happens to follow the CTK's major.minor).

[jameslamb]

/merge