adding-model-support

To add support for a new HF model family in Megatron-Bridge, follow the structured phase-based workflow covering discovery, bridge wiring, mapping, provider creation, recipe scaffolding, tests, and documentation.

Yes, the Megatron-Bridge extension pattern is model-agnostic and applicable to both LLMs and VLMs, providing recommended strategies for top-level versus sub-config mappings and optional VLM providers.

The recommended workflow is a five-phase process: Phase 1 discovery, Phase 2 bridge and provider wiring, Phase 3 recipe scaffolding, Phase 4 tests, and Phase 5 docs, ensuring consistent mappings and validation checks.

Yes, you can use private HF configs when mapping new models. The bridge extension pattern includes specific guidance on handling local mappings and private configurations for new model families.

Scaffolding recipes and tests requires following Phase 3 and Phase 4 of the workflow, which involve creating recipe structures and implementing test coverage to minimize integration risk and validate hooks.

You should implement provider hooks during Phase 2 bridge and provider wiring if the new model requires custom integration logic, ensuring reliable mappings and proper validation checks throughout the process.