skills.rest
NVIDIA-NeMo
NVIDIA-NeMoOFFICIAL
Explore all skills by this author →
WebsiteGitHub

create-model-verification-card

Create and validate agent-readable Megatron Bridge model verification cards in YAML.

896|481|Updated May 21, 2025
Search Tags:#llm training#model verification#megatron bridge#checkpoint conversion#deterministic inference#Slurm#YAML validation
One-click install
npx skills add https://github.com/NVIDIA-NeMo/Megatron-Bridge --skill create-model-verification-card
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: create-model-verification-card
Source: https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/main/skills/create-model-verification-card
Command: npx skills add https://github.com/NVIDIA-NeMo/Megatron-Bridge --skill create-model-verification-card

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires pyyaml, torch, transformers, and includes scripts (resource) components.

What problem does it solve?

Adding support for a new model in Megatron Bridge requires recording conversion, inference, training, and performance verification results in a consistent, auditable format. This Skill produces concise, agent-readable YAML verification cards that follow a strict schema, enforce privacy boundaries, and stay comparable across models.

Core Features & Use Cases

  • Structured card authoring: Generates card.yaml files with a required twelve-item core inventory covering conversion, forward-pass correlation, inference, pretraining, SFT, PEFT, and checkpoint resume, plus optional performance, FSDP, and weak-scaling items.
  • Schema and privacy validation: Uses scripts/validate_card.py to enforce status values, hardware allowlists, metric fields, verification-index consistency, and to reject private cluster details, credentials, paths, and job metadata.
  • Deterministic HF inference verification: Uses scripts/verify_hf_inference.py to run greedy generation from an exported Hugging Face checkpoint and confirm strict checkpoint reload.
  • Use Case: When preparing a model-support PR for a new LLM or VLM, use this Skill to draft the verification card, record public Slurm launcher commands and training metrics, and validate the card before submission.

Quick Start

Create a model verification card for the new model I am adding to Megatron Bridge and validate it with the card validator script.

Frequently Asked Questions about create-model-verification-card

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I create a Megatron Bridge model verification card?▼

Create a card.yaml under examples/model_verification_cards/<model-slug>/ with the twelve required inventory items, a verification_index, and public launcher commands. Then run scripts/validate_card.py to check schema, status values, and privacy rules before submitting the PR.

How do I verify deterministic Hugging Face inference after SFT export?▼

Run scripts/verify_hf_inference.py with the exported HF model directory, a prompt, and a max token count. It performs greedy generation, requires a strict checkpoint reload with no missing or unexpected keys, and can enforce GPU-only placement.

What items are required in a model verification card?▼

Twelve items are mandatory: four CPU/GPU conversion directions, manual forward-pass correlation, deterministic inference, pretraining, SFT, SFT export with inference, long-context SFT, PEFT, and checkpoint resume. Performance, FSDP, and weak-scaling items are optional.

What information is prohibited in a verification card?▼

Cards must never contain hostnames, IPs, usernames, accounts, partitions, private image paths, mount sources, tokens, job IDs, or environment-specific launcher overlays. The validator rejects these patterns and supports custom deny terms for private codenames.

Which hardware keys and statuses does the card validator accept?▼

Hardware keys come from a public allowlist including H100, H200, B200, GB200, and GB300, with 'all' reserved for global unsupported or not_applicable limitations. Statuses are limited to unverified, verified, unsupported, and not_applicable.

Related Skills

Explore More Software Engineering Skills→
sayalinvidiasayalinvidia

nemo-mbridge-resiliency

Configures Megatron Bridge fault tolerance, straggler detection, and restart for distributed GPU training.

Community
Advanced
wnz99wnz99
1

cross-review-pr

Coordinate independent cross-model pull request reviews with one-way validation.

Community
Advanced
luokai25luokai25

hugging-face-tool-builder

Creates and executes scripts that interact with the Hugging Face API.

Community
Advanced
randoneeringrandoneering

adding-dbt-unit-test

Generate YAML unit test definitions for dbt models with mocked inputs and outputs.

Community
Intermediate