[andyluo7]

Description

Mooncake's scheduled nightly workflow already builds ROCm wheels for Python 3.10 and 3.12, but those artifacts are neither stamped with the computed nightly version nor included in the TestPyPI upload.

This change:

• adds an optional version_override input to the reusable ROCm wheel workflow;
• passes the nightly .devYYYYMMDD version into the ROCm build;
• makes publish-testpypi wait for the ROCm wheel matrix and download its artifacts;
• preserves the existing mooncake-wheel-rocm-ubuntu-* artifact names used by the MI350X integration workflow; and
• adds a structural regression test for the version, dependency, artifact, and execution-order contracts.

This complements #4265, which rejects ROCm wheels built without multi-protocol routing. The PRs are independent: #4265 guards wheel correctness, while this PR makes validated development wheels available through the existing nightly TestPyPI channel.

Related roadmap: #3556.

Overlap: #3970 also edits ci_rocm.yml and nightly.yml as part of a broader packaging-backend migration, but it does not publish ROCm nightly wheels. If #3970 lands first, this small change will need to be rebased onto its root-package version override.

Deployment prerequisite: the existing TESTPYPI_API_TOKEN in the nightly environment must be authorized to create or upload the currently absent mooncake-transfer-engine-rocm TestPyPI project.

Module

• Transfer Engine (mooncake-transfer-engine)
• Mooncake Store (mooncake-store)
• Mooncake Conductor (mooncake-conductor)
• Reshard (mooncake-reshard)
• Mooncake EP (mooncake-ep)
• Mooncake PG (mooncake-pg)
• Integration (mooncake-integration)
• P2P Store (mooncake-p2p-store)
• Python Wheel (mooncake-wheel)
• Common (mooncake-common)
• Mooncake RL (mooncake-rl)
• CI/CD
• Docs
• Other

Type of Change

• Bug fix
• New feature
• Refactor
• Breaking change
• Documentation update
• Performance improvement
• Other

How Has This Been Tested?

Test commands:

python3 -m unittest discover -s scripts/ci/tests -v
python3 -m pytest -q mooncake-wheel/tests/test_testpypi_wheel_gate.py
pre-commit run --files \
  .github/workflows/ci.yml \
  .github/workflows/ci_rocm.yml \
  .github/workflows/nightly.yml \
  scripts/ci/tests/test_nightly_rocm_publish.py
actionlint -shellcheck= \
  .github/workflows/ci.yml \
  .github/workflows/ci_rocm.yml \
  .github/workflows/nightly.yml
git diff --check

Test results:

• Unit tests pass
• Integration tests pass (if applicable)
• Manual testing done (workflow YAML parsed successfully; the version-substitution command produced the expected PEP 440 nightly version)

Results: 7/7 CI-script tests and 6/6 TestPyPI gate tests passed. The scoped pre-commit suite and workflow syntax validation passed. Native ROCm wheel builds and TestPyPI publication require upstream CI and repository credentials and were not run locally.

Checklist

• I have performed a self-review of my own code
• I have formatted my code using the repository's scoped pre-commit hooks
• I have run pre-commit on the files changed in this PR and all hooks pass
• I have updated the documentation (not applicable; workflow-only change)
• I have added tests to prove my changes are effective
• For changes >500 LOC: not applicable (109 added lines)

AI Assistance Disclosure

• No AI tools were used
• AI tools were used (specified below)

OpenAI Codex inspected the existing nightly, ROCm build, integration, and publication workflows; implemented the scoped workflow changes and regression tests; and ran the validation commands listed above. The human submitter remains responsible for reviewing every changed line and for the TestPyPI credential configuration.

[stmatengss]

@Aionw Could you double-check?

[Aionw]

LGTM