Skip to content

adding-an-engine.md

Latest commit

 

History

History
41 lines (37 loc) · 2.27 KB

adding-an-engine.md

File metadata and controls

41 lines (37 loc) · 2.27 KB
Raw

Add an engine to the catalog

  1. Copy and adapt the InferRSI engine delivery template. Keep the implementation in its real owner repository, retain upstream notices and use the common organization attribution/asset provenance. Existing engines adapt only relevant files; they do not need the InferRSI Python runtime.

  2. Document architecture, environment, installation, startup, minimal request, reproduction, tests and limitations. Verify each claimed hardware independently. Keep campaign gates and comparison pins in acceptance issues and evidence (plan/, results/, patch/, validator/). Missing hardware leaves the entry unverified and the acceptance INCONCLUSIVE.

  3. Copy InferRSI's recipe YAML into recipes/<id>.yaml in the engine repository. Follow schema v1. Use exact deployment tags/commits and maintainers, one complete configuration per ID, with reviewed commands. Commit that source before catalog registration; source URLs must be reachable by the sync identity.

  4. Add an explicit declaration to the catalog's sources.yaml:

    schema_version: 1
sources:
  - id: your-engine
    repository: https://github.com/your-org/your-engine
    ref: main
    paths:
      - recipes/your-model-your-device.yaml

    The names above illustrate syntax and are not publishable content. Do not copy a second owned YAML into this catalog. If no engine-owned source can be maintained, use one local YAML here with upstream repository attribution.

  5. Run pnpm sync:sources, review the exact lock/cache diff, then pnpm test, pnpm validate and pnpm build. Do not execute commands during content sync. Review cards, model associations, mobile layout, copy and original edit links.

  6. Submit the collection change with source ownership, actual checks, hardware validation gaps and evidence scope. Engine maintenance, catalog collection, experiment acceptance and website publication are distinct responsibilities.

For changes to an existing engine recipe, update the owner file, synchronize and review the new locked snapshot. A failed sync preserves the previous snapshot but exits nonzero; resolve the reported source/path before publishing.