Repository navigation
fix(env): address managed lockfile review follow-ups - #5604
Conversation
Signed-off-by: oliver könig <okoenig@nvidia.com>
svcnvidia-nemo-ci
left a comment
There was a problem hiding this comment.
SHIP — LGTM — no concerns found.
I inspected all 15 changed files and traced the affected runtime installers, wheel lock restoration, dependency-policy merge, COMET callers, and CI result gates. The packaging hook preserves the two directly consumed runtime locks while retaining the archive path for other components. The required CI gate rejects non-successful lock results before the docs-only and nightly branches.
I couldn't retrieve one delegated report, so I completed its packaging and setup-documentation inspection directly. This was a read-only source review; I didn't execute tests or independently reproduce the reported build and CPU inference results.
Note: formal approval is not available for this request, so this review is posted as a comment only.
Review generated by codex. This was a full review. Request a full review with /review reset=true.
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/41872099263 |
Background
Follow up on the four remaining requests in Kajal's review of #4837. The merged packaging hook replaced directly consumed runtime locks with placeholders, the translation locks forced COMET outside its declared dependency limits, lock freshness was outside the required CI gate, and contributor instructions still advertised an unsupported uv version. This review supplies the issue context; a separate issue is unnecessary.
What changed
uv sync --lockedinstallers can read them. Other component locks continue to use the compressed archive. Add packaging regression coverage.<5exception in the component policy and regenerate those two locks. Test effective component constraints and COMET's supported version ranges.Nemo_CICD_Testdepends on it and rejects failure, cancellation, missing results, and unexpected skips, including docs-only runs. Include it in failure summaries. Remove its duplicate standalonemainpush trigger.Details
The old Transformers 5.22.1 override reproduces a COMET inference failure:
ValueError: not enough values to unpack (expected 3, got 2)inXLMREncoder.forward. Both new CPU inference tests reproduce this failure under the previous override and pass with each corrected lock.wmt_translationlongmt_evalThe shared Transformers 5 security floor is explicitly waived for these two isolated COMET 2.2.7 environments to restore its declared API contract. This exception is recorded in
scripts/component-lock-policy.jsonand the development guide. Existing uv installations must be upgraded before syncing; upgrading uv itself requires no lock regeneration.Tested
Signed commit:
af0196eca96c51984480f8e42b399e47eb86b727, based on mergedmainat1203b0be210f8fd19ab3fc4cdc5262feb73ffed9. Signature and DCO sign-off verified.pre-commit run --all-files: passed.git diff --cached --check: passed.python scripts/component_locks.py --jobs 4: 267 projects, 0 failed.Root regression suite: 127 passed, 1 skipped. The skip is the optional COMET module in the root environment; its two tests pass separately in each component environment.
pytest -q -o addopts='' --tb=short \ tests/unit_tests/test_wheel_locks.py \ tests/unit_tests/test_ci_environment.py \ tests/unit_tests/test_component_dependencies.py \ tests/unit_tests/test_locked_env.py \ tests/unit_tests/test_comet_compatibility.pyHF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 OMP_NUM_THREADS=1 python -m pytest -q -o addopts='' --tb=short tests: WMT 118 passed; LongMT 45 passed.torch==2.12.0+cpuandtorchvision==0.27.0+cpufor local validation. Run the following from the repo root with each component's Python:2 passed in each environment. These construct small local XLM-R encoders and run actual referenceless and XCOMET prediction on CPU without downloading pretrained weights. Both tests fail under the previous Transformers 5.22.1 override.
uv build --wheel --out-dir /tmp/gym-4837-followup-dist: passed. Extracted the two runtime projects from that wheel and ranuv sync --locked --dry-run --inexact --no-dev --project <extracted-project> --python <version>: both passed, using Python 3.11 for MolOpt and 3.13.14 for OSWorld. This validates lock acceptance without installing optional runtime packages.actionlint -shellcheck='' -ignore 'label ".+" is unknown' .github/workflows/component-locks.yml .github/workflows/cicd-main.yml .github/workflows/full-test-suite.yml: passed. Shell diagnostics and existing custom runner-label warnings are excluded from this syntax/dependency validation.fern/,npm run check -- --warnings: 0 errors; the authenticated redirects check is skipped because no Fern token is configured locally.CI triage
af0196eca96c51984480f8e42b399e47eb86b727: Both Linux/macOS component-lock checks and sandbox tests passed. All four core and all eight server test steps completed successfully. GitHub still reports several finished jobs as in progress, while combined coverage, the unit/server aggregates, andNemo_CICD_Testremain queued without assigned runners. Those aggregate verdicts remain pending.model_errorscenario. The same failure occurs on currentmainat1203b0be210f8fd19ab3fc4cdc5262feb73ffed9in job 126670755114; its publish/check aggregates reflect that failure. The main baseline's required CI gate and all unit/server jobs passed.orchestration/executors/judge_profile.mdxlibrary page.libraries.nemo-gym.input.gitreads the library from the upstream default branch. The preview workflow, library-input configuration, andjudge_profile.pyare unchanged frommain(git diff origin/main HEAD -- .github/workflows/fern-docs-preview.yml fern/docs.yml nemo_gym/orchestration/executors/judge_profile.pyis empty). This is the existing pinned preview generator's failure on main-derived source; main publication uses a different CLI path and succeeds.