Migrate pynvml.py to cuda.core.system

[mdboom]

Migrates usage of pynvml.py to the new Cython and cybind-based cuda.core.system.

I was unable to get my environment set up to test this locally, so there may be issues.

[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: 6ffbb46f-bdc1-4d1c-90a5-1f7d73099775

📥 Commits

Reviewing files that changed from the base of the PR and between 2f3b843 and 1f7cf38.

📒 Files selected for processing (1)

• python/cuml/tests/conftest.py

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

• python/cuml/tests/conftest.py

---

📝 Walkthrough

Summary by CodeRabbit

• Dependencies
  • Updated CUDA packages to require cuda-bindings>=12.9.6 and cuda-core>=1.0.0, excluding CUDA 13.0/13.1; removed nvidia-ml-py.
• Tests
  • Test environment now uses native CUDA system APIs for GPU detection and memory queries, improving reliability across devices.

Walkthrough

Updates dependency declarations to add cuda-bindings>=12.9.6,!=13.0.*,!=13.1.* and cuda-core>=1.0.0 and remove nvidia-ml-py>=12; updates test utilities to use cuda.core.system for device/memory queries and introduces _get_nvml_device_handle.

Changes

pynvml to cuda.core.system migration

Layer / File(s)	Summary
Dependency configuration updates conda/environments/all_cuda-129_arch-aarch64.yaml, conda/environments/all_cuda-129_arch-x86_64.yaml, conda/environments/all_cuda-132_arch-aarch64.yaml, conda/environments/all_cuda-132_arch-x86_64.yaml, dependencies.yaml, python/cuml/pyproject.toml	Adds cuda-bindings>=12.9.6 (excluding 13.0.* and 13.1.*) and cuda-core>=1.0.0 to environment and test dependency lists; removes nvidia-ml-py>=12.
Test NVML helper implementation python/cuml/tests/conftest.py	Imports cuda.core.system, adds _get_nvml_device_handle(device_id) to obtain Device objects from index or UUID (handles MIG parent devices), and updates _get_gpu_memory() to compute GiB and return None when unsupported.

---

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~12 minutes

---

Possibly related issues

• Migrate NvmlGpuInfo from pynvml to cuda.core.system after 1.0 release rapidsai/rapids-cli#147 — similar migration from pynvml/nvidia-ml-py to cuda.core.system and removal of the nvidia-ml-py dependency.

---

Suggested labels

improvement, non-breaking

---

Suggested reviewers

• jameslamb
• csadorf
• betatim

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main change: migrating from pynvml.py to cuda.core.system, which is reflected across all modified files.
Description check	✅ Passed	The description is relevant to the changeset, explaining the migration from pynvml.py to cuda.core.system and acknowledging testing limitations.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ 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.}

[csadorf]

@mdboom Is it safe to assume that this can wait to be release in 26.08 rather than 26.06?

[mdboom]

@mdboom Is it safe to assume that this can wait to be release in 26.08 rather than 26.06?

Yes. There is no concrete timeline to deprecate pynvml, so no urgency on this.

[betatim]

One question about how things will work for packages built for different cuda versions. Otherwise looks fine to me

[betatim]

Do the versions here correspond to the version of CUDA that the user has installed?

Will this also work for the -cu13 package that we upload to pypi?

[betatim]

Can you merge the latest changes from main into your branch and regenerate the dependency files? That should make the conflicts go away and make the "recently updated" check happy