adding-model-support

To add a HuggingFace model to Megatron-Bridge, you extract HF config details, determine if it is an LLM or VLM, map components into bridge and provider layers, and add recipes with tests. This ensures correct checkpoint conversion and generation verification.

For FP8 or FP4 weight conversion, you must detect quantized safetensors and apply the required dequantization approaches. This prevents silently broken models during the bridge and provider implementation by ensuring correct checkpoint import-export roundtrips.

LLM and VLM onboarding requires extracting HF config fields including model_type, architectures, dimensions, and RoPE or MoE specifics. Analyzing these fields determines the architecture type and guides the correct model-specific module scoping and embedding tying for VLMs.

Yes, adding VLM support requires special embedding tying conventions. The implementation must satisfy required conventions for config field locations and local model-specific module scoping to ensure successful checkpoint conversion and generation tests.

Onboarding a new model family requires unit and functional testing alongside runnable examples. These tests verify successful checkpoint conversion and generation workflows, ensuring the bridge registration and provider configuration are maintainable and correct.