adding-model-support

To add a new HuggingFace model to Megatron-Bridge, you analyze the model configs, create custom weight mapping classes, and implement configuration translation. This Skill guides you through establishing the bridge, mapping, and test scripts for seamless integration.

Multimodal model integration with Megatron conversion tooling involves analyzing vision-language configs, identifying custom layers, and handling quantization. This Skill provides implementation patterns to ensure compatibility and efficient deployment of diverse architectures.

Yes, you need familiarity with HuggingFace configs and Megatron conversion tooling to use this model integration framework. The Skill requires this prerequisite knowledge to facilitate config analysis, custom weight mapping, and architecture validation.

The best way to handle custom weight mapping when onboarding a new LLM is to develop model-specific code for configuration translation. This Skill assists in creating mapping classes and establishing testing pipelines for accurate weight conversion.

Yes, this approach supports both language and vision-language model architectures. It provides comprehensive guidance for incorporating diverse architectures into Megatron-Bridge, ensuring compatibility and efficient validation across multimodal configurations.