nemo-automodel-model-onboarding

To onboard a new NeMo AutoModel architecture, you scaffold the model directory, implement model.py and state_dict_adapter.py, register the model in the registry, and add unit tests with tiny configs for validation.

Adding a Vision Language Model requires inspecting the target HF config, scaffolding the model components, implementing state_dict_adapter.py for weight conversion, updating the registry, and preparing minimal documentation and tiny-config tests.

Yes, the NeMo AutoModel onboarding process applies to dense LLMs, Mixture of Experts (MoE), and Vision Language Model (VLM) patterns by providing a structured workflow for discovery, implementation, and validation.

You need state_dict_adapter.py to adapt and convert external checkpoint weights into the NeMo AutoModel format, ensuring proper weight loading and reproducibility during the model onboarding process.

The standard workflow for registering a model in the NeMo AutoModel registry involves creating a standard Skill Unit, implementing the required model code, updating the registry, and adding SKILL.md frontmatter with unit-test scaffolding.