[CI] Four verified workflow config bugs: pr-gates miss the reusable runner, dead uv.lock change detection, push-event lanes silently skipped, shared non-PR concurrency group

[arham766]

Describe the bug

Four verified defects in the CI workflow configs. Impact: should have — none block development, but two cause silent test-coverage gaps on merges and one fails a scheduled job. Found while auditing the workflows; each is confirmed either from this repo's own run history (via the GitHub API) or by reproducing the workflow steps locally. Sending one surgical PR per item.

1. example_tests / gpu_tests pr-gate does not watch the reusable runner or the cache action.
example_tests.yml's pr-gate files: list includes example_tests.yml itself but not .github/workflows/_example_tests_runner.yml (the reusable workflow every example-test job executes through) or .github/actions/cache-extensions/** (used by the runner and by gpu_tests.yml). Reusable workflows run from the PR's ref, so a PR that changes only the runner merges with the very jobs it rewires never executed. Observed on PR #1651 (touched _example_tests_runner.yml + cache-extensions/action.yml + gpu_tests.yml): its example-tests run shows pr-gate success, then torch/trtllm-pr/megatron/onnx all skipped, and the required check green.

2. bump_uv_lock.yml change detection is dead code; a no-update week fails the scheduled run.
The torch-override step rewrites pyproject.toml with toml.load/toml.dump, which drops comments and reformats the whole file, so the later unscoped git diff --quiet is always dirty and changed=false is unreachable. When uv lock --upgrade changes nothing, the create-pull-request step stages nothing (git add uv.lock) and git commit exits 1 under bash -e, failing the run instead of no-oping. Reproduced locally by executing the workflow's steps against this repo's files: post-mutation git diff --quiet reports dirty with uv.lock byte-identical, and the commit step exits 1. Latent so far only because the lock has had upgrades every week.

3. unit_tests.yml: on push events the file-change check outputs empty, silently skipping six lanes.
In check-file-changes, the non-pr step runs only for schedule/workflow_dispatch and the changed step only for pull_request. For push neither runs, any_changed is empty, and the windows, multi-version, partial-install, launcher, mcp, and skills lanes all skip — despite the workflow-level push paths: filter listing tools/launcher/**, tools/mcp/**, .agents/skills/**, etc. Observed on the latest push run on main: check-file-changes success, linux success, all other lanes skipped. A push to main touching only tools/launcher/** runs the irrelevant modelopt suite and skips the launcher tests; release/*/feature/* branches get no coverage of these lanes at all (nightly only covers main).

4. code_quality.yml: non-PR runs share one concurrency group and cancel each other.
The group is ${{ github.workflow }}-${{ github.event.pull_request.number }} with no || github.sha fallback (unlike unit_tests.yml), so every nightly and manual dispatch shares the literal group Code Quality- with cancel-in-progress: true — a manual dispatch cancels an in-flight nightly and vice versa.

Steps/Code to reproduce bug

• Item 1: compare example_tests.yml pr-gate files: with the uses: ./.github/workflows/_example_tests_runner.yml job wiring; see PR ci: cache JIT-compiled CUDA torch extensions in GPU/example tests #1651's example-tests run for the observed skip.
• Item 2: run the workflow's steps locally in a clean checkout: apply the toml override snippet, then git diff --quiet; echo $? → 1 with uv.lock unchanged; git add uv.lock && git commit -s -m x → exit 1.
• Item 3: inspect any push-event run of Unit tests on main: check-file-changes succeeds with empty output and the gated lanes report skipped.
• Item 4: trigger workflow_dispatch of Code Quality while a nightly is in flight.

Expected behavior

1. Changes to the test runner/cache action execute the example/GPU test jobs on the PR that changes them.
2. A no-update week no-ops (changed=false) instead of failing.
3. Push events run the lanes whose paths triggered the workflow.
4. Nightly and manual code-quality runs do not cancel each other.

Who can help?

@kevalmorabia97

System information

• Container used (if applicable): N/A (CI workflow configs)
• OS: N/A — reproductions done on Windows 11 (git steps) and via GitHub API run history
• CPU architecture / GPU: N/A
• Library versions: ModelOpt commit 4b9225b (current main)
• Any other details that may help: sending four surgical PRs, one per item, referencing this issue

[h-guo18]

Hi @kevalmorabia97 , seems like these bugs and PRs are related to CI/setup, could you take a look please. Thanks!