Skip to content

ci_dependency_tiers.md

Latest commit

 

History

History
59 lines (43 loc) · 3.28 KB

ci_dependency_tiers.md

File metadata and controls

59 lines (43 loc) · 3.28 KB
Raw

CI dependency tiers

YOLOZU uses a tiered CI dependency model to keep signal high while avoiding optional-extras combination explosion.

Default cost policy:

  • Pull requests run lightweight Ubuntu checks: docs/metadata gates, runtime smoke, and focused quality tests.
  • Pushes to main run the full CPU evaluation gate, including Python matrix and full unittest discovery.
  • GPU checks stay on Ubuntu/Linux GPU runners and run only by manual dispatch.
  • macOS is reserved for release-time wheel build validation only.
  • Expensive full/fuzz/regression sweeps are manual-only unless there is an explicit release or incident reason to re-enable a schedule.

Tiers

  1. core
  • Install: python3 tools/ci/install_with_hashes.py --requirements requirements-locks/requirements-runtime.lock --install-local-wheel
  • Purpose: packaging and CLI/runtime smoke only.
  • Jobs: smoke_gate; pip_smoke is main-push only.
  1. docs_mcp
  • Install: python3 tools/ci/install_with_hashes.py --requirements requirements-locks/requirements-docs-actions.lock, then python3 tools/ci/install_with_hashes.py --requirements requirements-locks/requirements-web-docs.lock
  • The web-docs candidate gate sets YOLOZU_REQUIRE_REAL_COCO=1; a missing pycocotools installation is a failure, not a dry-run fallback.
  • Purpose: docs/manual/README and MCP/Actions surface checks without running full runtime regression gates.
  • Jobs: docs_mcp_gate.
  1. workflows_meta
  • Install: none beyond stock Python on the runner.
  • Purpose: release/security workflow regression checks for .github/workflows/**-only changes.
  • Jobs: workflows_meta.
  1. recommended
  • Install: python3 tools/ci/install_with_hashes.py --requirements requirements-locks/requirements-ci.lock
  • Purpose: pinned interface contract/behavior gates (schema, manifest, reference regression, deep smoke walkthrough, focused tests).
  • Jobs: quality_gate; test is main-push only.
  1. full
  • Purpose: GPU/backend matrix (TensorRT/CUDA/provider parity, full reference regression profile).
  • Trigger: manual dispatch only.
  • Workflows: gpu_smoke_machine.yml, gpu_practical_suite_machine.yml, gpu_zisn_pipeline.yml, reference_adapter_full.yml, cflite_batch.yml.
  1. release
  • Purpose: publish-time packaging confidence, including macOS wheel build validation before the Ubuntu publish job.
  • Workflows: publish.yml, container.yml, manual_doi.yml, announce_release.yml.

Why this split

  • core catches packaging/runtime breakages cheaply.
  • docs_mcp validates docs/MCP changes quickly without spending runtime-heavy CI budget.
  • recommended gives stable regression signals with an exact-version lock that is installed via a generated --require-hashes wheelhouse.
  • full is intentionally separated because GPU/provider stacks are expensive and noisy for every PR or nightly schedule. The GPU shell helpers and container images now also layer exact-version extras on top of requirements-locks/requirements-runtime.lock via tools/ci/install_with_hashes.py.

Optional extras policy

Optional extras are defined in pyproject.toml and their rationale is recorded under:

  • [tool.yolozu.optional_extras_rationale]

CI does not depend on full extras as a single install target. Instead, it uses the tiered installs above so failures are easier to localize.