CI: Migrate custom jobs to RTX PRO 6000

[bdice]

Description

Migrates custom CI jobs from L4 to RTX PRO 6000 runners.

The packaged-Java matrix retains L4 for ARM64 because RTX PRO 6000 runners are only available for AMD64.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[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.

[vyasr]

Thanks!

[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: 33eb02c9-cc91-429d-a411-b16381fb13b2

📥 Commits

Reviewing files that changed from the base of the PR and between 81068be and 8727ad4.

📒 Files selected for processing (2)

• .github/workflows/pr.yaml
• .github/workflows/test.yaml

Included review availability: Your plan provides up to 12 included reviews per hour; 11 remain after this review.

---

📝 Walkthrough

Summary by CodeRabbit

• Chores
  • Updated documentation, build, and validation workflows to run on RTX PRO 6000 GPU runners.
• Tests
  • Migrated Java, C++, notebook, pandas, cuML compatibility, Narwhals, sanitizer, benchmark, and memcheck jobs to the updated GPU runner environment.
  • Packaged Java tests now use RTX PRO 6000 runners on AMD64 while retaining existing runners for other architectures.
  • Maintains automated coverage across supported test and verification workflows with updated hardware.

Walkthrough

Changes

GPU-based jobs in the build, sanitizer, Pandas, pull request, and test workflows now select RTX PRO 6000 runners. Packaged Java tests select RTX PRO 6000 for amd64 and L4 for other architectures.

GPU runner migration

Layer / File(s)	Summary
Supporting workflow runner updates .github/workflows/build.yaml, .github/workflows/compute-sanitizer-run.yaml, .github/workflows/pandas-tests.yaml	Documentation build, sanitizer, and Pandas test jobs now select RTX PRO 6000 runners.
CI workflow runner updates .github/workflows/pr.yaml, .github/workflows/test.yaml	Pull request and test jobs now select RTX PRO 6000 runners. Packaged Java tests retain L4 runners for non-amd64 architectures.

Estimated code review effort: 1 (Trivial) | ~5 minutes

Merge Risk: ⚪ Minimal · up to 8727a

This PR updates CI jobs to use RTX PRO 6000 runners while retaining L4 for ARM64; no actionable merge-blocking risk remains beyond normal checks and review.

Suggested reviewers: vyasr, galipremsagar, jameslamb

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0…
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.
Title check	✅ Passed	The title clearly describes the main change: migrating custom CI jobs from L4 runners to RTX PRO 6000 runners.
Description check	✅ Passed	The description directly explains the CI runner migration and the ARM64 exception for packaged Java jobs.

Full details: Docstring Coverage

Explanation

No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0 files. (2 skipped: 2 unsupported.)

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[bdice]

CI runtime comparison: L4 vs. RTX PRO 6000

Updated baseline: the earlier version of this report used PR #23869, which ran before dependency caching was enabled. This comparison instead uses the cached L4 run from PR #23770 (run 33193794748). It started only 35 minutes before this PR's RTX PRO 6000 run and overlapped it in wall-clock time. Both runs include the Setup proxy cache step.

I compared that L4 run with the completed RTX PRO 6000 jobs from this PR's previous commit, 81068be1f0 (run 33196496195). The table includes only name-matched AMD64 jobs whose explicit runner selector this PR changed. It excludes the two ARM64 Java jobs and unrelated shared-workflow matrix jobs that happened to use either GPU. Durations are job wall-clock time from GitHub's started_at and completed_at timestamps.

Job	L4 with cache	RTX PRO 6000 with cache	Change
conda-java-tests	10m 27s	10m 15s	-1.9%
pandas-tests (0)	29m 56s	27m 18s	-8.8%
pandas-tests (1)	24m 14s	26m 46s	+10.5%
third-party-integration-tests-cudf-pandas	17m 34s	13m 05s	-25.5%
cuml-compat-tests	11m 52s	10m 28s	-11.8%
docs-build	32m 24s	6m 40s	-79.4%
conda-notebook-tests	29m 47s	4m 10s	-86.0%
narwhals-tests	4m 41s	6m 19s	+34.9%
java-tests (CUDA 12, AMD64)	27m 15s	6m 43s	-75.4%
java-tests (CUDA 13, AMD64)	18m 11s	7m 12s	-60.4%

Summary

• 8 of 10 matched jobs were faster on RTX PRO 6000.
• Summed job time fell from 3h 26m 21s to 1h 58m 56s, saving 1h 27m 25s (-42.4%) of runner time.
• Median job time fell from 21m 12s to 8m 44s (-58.9%).
• Summed Run script step time fell from 3h 05m 48s to 1h 45m 53s (-43.0%).
• The two pandas shards were effectively unchanged in aggregate: 54m 10s on L4 versus 54m 04s on RTX PRO 6000 (-0.2%). conda-java-tests was also effectively flat (-1.9%). This confirms that the large differences for those jobs in the original report were primarily a pre-cache/post-cache comparison, not an RTX PRO 6000 effect.
• third-party-integration-tests-cudf-pandas improved 25.5% and cuml-compat-tests improved 11.8%; narwhals-tests regressed 34.9%.

Same-run matrix comparison

As a tighter cross-check, run 33196496195 contains five shared-workflow matrix jobs with both AMD64 L4 and RTX PRO 6000 entries. This holds the commit, cache configuration, and run timing constant. All entries use the latest driver; CUDA, Python, OS, and dependency variants differ as shown.

Matrix job	L4 entry	RTX PRO 6000 entries	RTX mean vs. L4
conda-cpp-tests	39m 08s (CUDA 13.0, Python 3.12, Ubuntu 24.04)	8m 44s (CUDA 13.3, Python 3.13, Ubuntu 26.04)	-77.7%
conda-python-cudf-tests	47m 04s (CUDA 12.2, Python 3.12, Ubuntu 22.04)	14m 40s (CUDA 13.0, Python 3.12, Ubuntu 24.04); 13m 37s (CUDA 13.3, Python 3.13, Ubuntu 26.04)	-70.0%
conda-python-other-tests	48m 22s (CUDA 12.2, Python 3.12, Ubuntu 22.04)	11m 32s (CUDA 13.0, Python 3.12, Ubuntu 24.04); 12m 40s (CUDA 13.3, Python 3.13, Ubuntu 26.04)	-75.0%
wheel-tests-cudf	16m 40s (CUDA 12.9, Python 3.11, Ubuntu 22.04, oldest deps)	12m 38s (CUDA 13.0, Python 3.12, Ubuntu 24.04); 13m 14s (CUDA 13.3, Python 3.13, Rocky 8); 13m 00s (CUDA 13.3, Python 3.14, Ubuntu 26.04)	-22.3%
wheel-tests-cudf-streaming	21m 17s (CUDA 12.9, Python 3.11, Ubuntu 22.04, oldest deps)	2m 57s (CUDA 13.0, Python 3.12, Ubuntu 24.04); 3m 05s (CUDA 13.3, Python 3.13, Rocky 8); 2m 59s (CUDA 13.3, Python 3.14, Ubuntu 26.04)	-85.9%

RTX PRO 6000 was faster in all five matrix groups. Giving each matrix equal weight, mean job time was 34m 30s on L4 versus 10m 11s on RTX PRO 6000, a 70.5% reduction (3.39× speedup). Per-matrix reductions ranged from 22.3% to 85.9%.

This same-run result avoids the cache and temporal confound in the original cross-PR comparison, but it still does not isolate GPU hardware: the matrix intentionally varies CUDA, Python, OS, and, for the wheel jobs, dependency age. In particular, the L4 wheel entries use oldest dependencies while the RTX PRO 6000 entries use latest dependencies.

This is still a single observational before/after sample, not an isolated hardware benchmark. The aggregate is dominated by unusually long L4 docs, notebook, and Java jobs, so the 42.4% aggregate reduction should not be interpreted as a general hardware speedup. Although the runs overlapped and used the same cache setup, PR source differences, cache-hit state, network behavior, and runner contention remain uncontrolled. Multiple paired runs would be needed to estimate stable hardware-specific speedups.

[vyasr]

/merge