Repository navigation
Conversation
Signed-off-by: oliver könig <okoenig@nvidia.com>
|
🌿 Preview your docs: https://nvidia-preview-pr-5610-aa7fe4cd3124.docs.buildwithfern.com/nemo/gym |
svcnvidia-nemo-ci
left a comment
There was a problem hiding this comment.
SHIP
LGTM — no concerns found.
The runner selection moves the default AWS container build to CPU capacity while retaining nondefault routes, attempt-qualified ephemeral labels, Azure environment selection, and upstream/cancellation gates. The updated contract test now checks the container job specifically. I also traced the Dockerfile/prefetch path and downstream GPU job; I found no GPU requirement during image construction or changed publication contract.
Static review only: I didn't run tests or a container build, or independently verify the live runner configuration and registry permissions.
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.
Background
The protected-main
Build Gym containerjob is queued behind H100 workloads onnemo-ci-aws-gpu-x2(run 41861701599). The pool was at its 24-runner limit while the existing CPU pool had available capacity. Image construction and dependency prefetch require no GPU.This leaves downstream CI using an older admissible image: CI-Events-Workflows run 41857672334, attempt 2 fails because that image has uv 0.11.29 while current Gym requires uv >=0.13.4. This focused infrastructure repair is tracked by those failing runs; no separate issue is needed.
What changed
Route the default AWS container build to the existing
nemo-ci-aws-cpu-4-16pool. Preserve nondefault preflight routes, Azure environment selection, attempt-qualified ephemeral labels, scheduled/manual triggers, upstream unit-test checks, and cancellation behavior.Details
The existing CI contract test now checks the parsed
container_buildrunner selection rather than matching a runner expression elsewhere in the workflow. The trigger policy from #2810 and ephemeral attempt routing from #3163 remain in force.flowchart LR P[Preflight and unit checks] --> C[CPU image builder] C --> E[ECR image and source revision] E --> R[CI Events protected-main image resolver] R --> T[Gym cluster build test]Compatibility and benchmark impact: none; the Dockerfile, image contents, runtime GPU jobs, and image admission rules are unchanged.
Tested
aa7fe4cd3124d9ce94e77fe1a90bbab5d3fad042, signed and DCO signed off..venv/bin/python -m pytest tests/unit_tests/test_ci_environment.py -q: 81 passed..venv/bin/pre-commit run --all-files: passed, including guardwords and import layering. Staged checks and signed-commit hooks also passed;git diff --checkpassed.bash /tmp/gym-cpu-image-pr-local-build.sh: built the canonicaldocker/Dockerfilerelease target from a clean archive of this exact head, using the official namednemo-gymsource context,NEMO_GYM_PREFETCH_CONFIGS=tests/e2e/gpu_e2e.yaml, and a BuildKit container limited to 4 CPUs and 16 GiB. Dependency prefetch, Python 3.13 setup, and the nonroot CLI check completed.gym:ci-events-local-aa7fe4cd3124d9ce94e77fe1a90bbab5d3fad042; ECR independently confirms indexsha256:16a316341daacad6e88dfb043fa2d788d1beb59bee2fb43a4dbd200f4791aa68and image size 14,654,181,462 bytes. Local validation tags do not satisfy protected-main admission.