What problem does it solve? Adding support for a new Hugging Face model architecture in NeMo AutoModel involves many error-prone steps: classifying the architecture, implementing model files, writing state-dict adapters, registering classes, and validating numerical parity. This Skill provides a structured five-phase workflow with checklists and reference patterns so nothing is missed. ## Core Features & Use Cases - Architecture Discovery: Classify dense LLM, MoE, or VLM models from config.json fields such as architectures, expert counts, and vision_config/text_config. - Implementation Patterns: Reference guides for dense LLMs, MoE state-dict adapters (router, expert, and shared-expert weight mapping), and VLM wrappers (vision tower, projector, language model). - Registration & Validation: Register classes in MODEL_ARCH_MAPPING, declare ModelCapabilities and weight-tying policies, and run unit, layer-equivalence, and end-to-end parity tests. - Use Case: When adding support for a new Hugging Face causal LM, follow the checklist to create components/models/<name>/, register the class, add a tiny-config test, and verify checkpoint round-trip before full training. ## Quick Start Ask the agent to onboard a new Hugging Face model architecture into NeMo AutoModel, for example by saying: add support for the new model in my config.json as a NeMo AutoModel architecture.
