AIP-001: Plugin Architecture for AIPerf

[ajcasagrande]

AIP-001: Plugin Architecture for AIPerf

Summary

Adopt a production-grade plugin architecture for AIPerf based on Python's entry points ecosystem, lazy loading, the pluggy hook system, and dependency injector. This enables third-party developers to add new endpoint types, data processors, metrics, exporters, timing strategies, and other functionality as standalone packages with independent release cycles, while maintaining zero-boilerplate plugin authorship and minimal startup overhead for the application.

Read the full DEP in user-friendly format:
https://github.com/ai-dynamo/enhancements/blob/ajc/aip-001-plugin-architecture/deps/AIP-001-Plugin-Architecture.md

Note

• The use of pluggy hook system and dependency-injector are tbd. The entire architecture works without them as well
• We have a very sophisticated hook system already (albeit not yet documented), and the pyproject.toml entry-points work just as well without pluggy using native libraries.
• Dependency injection is a thoughtful nice to have, but not necessary either.