Dev & Engineering ✓ NVIDIA · Official nemo-automodelmodel-onboardinghugging-facestate-dict-adaptersmixture-of-expertsvision-language-modelsunit-testing

NeMo AutoModel Model Onboarding

A structured guide for adding new LLM, MoE, and VLM architectures to NeMo AutoModel.

FollowSkills review · FSRS-2.0
Use with care
60/ 100 5-point scale 3.0 / 5
Trust18 / 25 · 3.6/5

The skill declares Apache-2.0, scopes its guidance to model onboarding, and includes explicit warnings around weight mapping and validation. No malware, credential theft, covert exfiltration, or destructive default was found. However, it directs users to fetch Hugging Face configuration and rely on external model repositories without explicit confirmation, data-flow disclosure, source validation, dependency security, rollback, or least-privilege guidance, so points are deducted.

Reliability9 / 20 · 2.3/5

The workflow, file layout, registry steps, adapter guidance, and testing requirements are detailed, with some abnormal-case warnings. However, this review is static: key paths were not executed, examples and external dependencies were not verified, and the document says five phases while specifying six. The score therefore remains conservative under the static cap.

Adaptability12 / 15 · 4.0/5

Audience, use cases, routing boundary, positive examples, negative examples, and Dense/MoE/VLM classification signals are relatively clear. Missing are stronger handling rules for incomplete inputs, Chinese-language support, and mainland-China network reachability; the discovery path also depends on Hugging Face, so points are deducted.

Convention9 / 15 · 3.0/5

The main guide, pattern references, precision guidance, checklist, skill card, evaluations, and benchmark report provide useful progressive disclosure and identify the license, author, and some update context. Versioning, changelog, maintenance ownership, dependency installation, FAQs, known limitations, and stable citation practices remain incomplete. The benchmark also records schema and author-format findings.

Effectiveness7 / 15 · 2.3/5

The skill provides actionable guidance for discovery, implementation files, registration, state-dict conversion, tests, and parity checks. Its supplied benchmark reports strong results on three positive tasks. Nevertheless, no independently reproduced output is available in this static review, and the guidance still requires engineering implementation and review, so the score is limited by the static cap.

Verifiability5 / 10 · 2.5/5

The revision includes detailed primary skill material, checklists, evaluation tasks, and a benchmark report, giving the claims some auditability. It does not provide independently reproducible CI evidence, committed key-path test results, or corroboration from multiple sources within the supplied material, so the score is limited to the static-review ceiling.

Evidence confidence:Low Reviewed Jul 20, 2026 Reviewed revision 55f18499943e
The upstream repository has new commits since this review. The score still applies to the reviewed revision shown and may not cover the latest changes.
Before you use it
  • Do not treat the benchmark PASS or high scores as independent verification for this review; run adapter round-trip, layer-equivalence, image-text, and end-to-end parity tests against the target NeMo AutoModel version.
  • Before fetching config.json or checkpoints from Hugging Face, verify network reachability, model provenance, licensing, dependency versions, and user authorization; no mainland-China fallback procedure is documented.
  • The guide references several companion files and repository paths; if they are absent from an installation, those references may be unusable.
See the full review method →

What does this skill do, and when should you use it?

This skill helps NeMo AutoModel developers add or modify model architecture support. It starts with Hugging Face config.json discovery and proceeds through implementation, weight adaptation, registry updates, example configuration, testing, documentation, and parity validation. It covers dense LLMs, mixture-of-experts models, and vision-language models, with explicit safeguards for expert mappings and numerical equivalence. It is one skill in NVIDIA/skills, a 324-skill monorepo installed through the repository’s CLI flow.

Instructs an agent to inspect a Hugging Face config.json and classify the architecture as dense LLM, MoE, or VLM; identify model.py, state_dict_adapter.py, and optional config, layer, and RoPE files under components/models/<name>/; update model and custom-config registrations in _transformers/registry.py; map router, expert, text, and vision weights; create an example YAML and tiny-config, forward-shape, adapter round-trip, layer-equivalence, loss, and parity tests; and update model-coverage documentation.

  1. A NeMo AutoModel developer adding a new Hugging Face causal-LM architecture can use it to determine the required files, registry entries, and validation gates.
  2. A model engineer onboarding an MoE checkpoint with routers, shared experts, or combined projections can use it to audit expert ordering and state-dict mappings.
  3. A VLM developer adding a model with vision_config, text_config, and a conditional-generation architecture can use it to check the vision tower, projector, processor assumptions, and image-text tests.
  4. A maintainer changing custom attention, RoPE, normalization, MLP, or precision-sensitive parameters can use it to identify capability declarations and equivalence tests.

What are this skill's strengths and limitations?

Pros
  • Covers the workflow from config.json discovery through end-to-end parity testing.
  • Provides concrete classification signals and file locations for dense LLM, MoE, and VLM onboarding.
  • Highlights commonly missed checks including expert ordering, tied weights, tiny configurations, and numerical equivalence.
  • Defines a clear boundary that excludes standalone recipe, launcher, and distributed-strategy questions.
Limitations
  • It is an implementation guide, not a runnable onboarding generator or complete model implementation.
  • It assumes access to Hugging Face configuration and NeMo AutoModel source code, plus a PyTorch test environment.
  • The source provides no platform test matrix, performance measurements, or complete list of already supported models.
  • The supplied repository metadata says NOASSERTION for the repository license, while this skill declares Apache-2.0; verify the applicable license before adoption.

How do you install this skill?

Install the skill using the command documented in the repository README: npx skills add nvidia/skills --skill nemo-automodel-model-onboarding --yes. You can target an agent explicitly, for example: npx skills add nvidia/skills --skill nemo-automodel-model-onboarding --agent codex.

How do you use this skill?

After installation, invoke it for a model-architecture onboarding task, such as: “Add support for a new Hugging Face MoE architecture in NeMo AutoModel, including state-dict adaptation, registration, and tests.” The source does not document a standalone executable or dedicated runtime command; code changes and test execution depend on the target repository environment.

FAQ

Does installation require payment?
The supplied material mentions no fee. The README only documents installation through the npx skills CLI flow.
Can it handle ordinary training recipes or launcher configuration?
Not by default. The skill explicitly places optimizer, dataset, scheduler, trainer wiring, distributed-strategy, and launcher questions out of scope unless they are directly part of onboarding a new architecture.
What are the main MoE failure modes it guards against?
Do not validate only with from_pretrained(). Map router and expert weights separately, preserve expert index order, handle gate/up/down projections and shared experts explicitly, and run key-mapping and tiny-config numerical-equivalence tests.
What external access does it require?
The guide calls for obtaining Hugging Face config.json and reading or modifying NeMo AutoModel model, registry, example, test, and documentation files, so network, shell, and filesystem access are normally required.

More skills from this repository

All from NVIDIA/skills

Dev & Engineering ✓ NVIDIA · Official

MoE Dispatcher Selection Guide

Choose alltoall, DeepEP, or HybridEP from the hardware, expert-parallel degree, and optimization stage.

Dev & Engineering ✓ NVIDIA · Official

Earth2Studio Diagnostic Builder

Build Earth2Studio wrappers for single-step diagnostic data transformations.

Dev & Engineering ✓ NVIDIA · Official

NeMo MBridge MoE Communication Overlap

Tune expert-parallel communication and computation overlap in Megatron Bridge.

Dev & Engineering ✓ NVIDIA · Official

NeMo MBridge CUDA Graphs

Configure, validate, and benchmark CUDA Graph capture in Megatron Bridge to reduce host-driver overhead.

Dev & Engineering ✓ NVIDIA · Official

NeMo MBridge MoE Performance Workflow

Systematically diagnose and improve MoE training throughput in Megatron Bridge.

Dev & Engineering ✓ NVIDIA · Official

NV-Reason-CXR Chest X-ray Reasoning Test

Runs reproducible command-shape and live inference smoke tests for chest X-ray reasoning.

Data & Analysis ✓ NVIDIA · Official

TAO Image Grounding

Convert image-caption pairs into phrase-grounded annotations with pixel-space bounding boxes.

Dev & Engineering ✓ NVIDIA · Official

Earth2Studio Forecast Wrapper Builder

Build and validate Earth2Studio time-stepping forecast wrappers.

Automation & Ops ✓ NVIDIA · Official

NeMo-RL Brev Runbook

Helps NeMo-RL agents manage storage, caches, logs, and authentication safely on Brev instances.

Dev & Engineering ✓ NVIDIA · Official

NeMo AutoModel Distributed Training

Choose and configure multi-GPU and multi-node training strategies for NeMo AutoModel.

Dev & Engineering ✓ NVIDIA · Official

NeMo AutoModel Recipe Development

Build, modify, and validate NeMo AutoModel training and evaluation recipes.

Automation & Ops ✓ NVIDIA · Official

Jetson LLM Serving

Deploy OpenAI-compatible LLM and VLM servers on Jetson with vLLM or SGLang.

Automation & Ops ✓ NVIDIA · Official

NeMo AutoModel Launcher Configuration

Configure NeMo AutoModel launches across local, Slurm, and cloud runs.

Dev & Engineering ✓ NVIDIA · Official

Megatron-Core Testing Guide

Design, run, and reproduce distributed Megatron-Core tests with CI-aligned workflows.

Dev & Engineering ✓ NVIDIA · Official

Catheter Navigation Setup Verification

Checks host, GPU, and Python path readiness for catheter navigation.

Dev & Engineering ✓ NVIDIA · Official

Catheter Navigation Fluoroscopy Smoke Tests

Validate core catheter-navigation workflow behavior on CPU-only CI environments.

Dev & Engineering ✓ NVIDIA · Official

Catheter Navigation E2E Smoke Test

Runs the Isaac for Healthcare v0.7 catheter workflow from setup verification through digital-twin DRR rendering and unit tests.

Data & Analysis ✓ NVIDIA · Official

Nemotron Speech ASR Customization Orchestrator

Chooses and coordinates the lowest-cost path for adapting speech recognition to a domain or language.

Dev & Engineering ✓ NVIDIA · Official

Earth2Studio Discover

Find weather and climate models, data sources, and examples that fit a task.

Data & Analysis ✓ NVIDIA · Official

TAO CLIP Fine-Tuning and Deployment

Train and deploy NVIDIA TAO CLIP for image-text retrieval, zero-shot classification, and embedding extraction.

Related skills