parity-testing

Validate numerical parity between HF and M-Core weight conversions for Megatron Bridge models.

2.8k|332|Updated Feb 25, 2026

Search Tags:#debugging #verification #bridge #parity #megatron #weight-conversion

Ready to run

npx skills add https://github.com/NVIDIA/skills --skill parity-testing

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: parity-testing
Source: https://github.com/NVIDIA/skills/tree/main/skills/Megatron-Bridge/parity-testing
Command: npx skills add https://github.com/NVIDIA/skills --skill parity-testing

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Structured framework for verifying numerical parity of HF<->MCore weight conversions. References existing tools and the add-model-support skill. Use when debugging weight mismatches, verifying checkpoint round-trips, or choosing which verification tool to run.

Core Features & Use Cases

Defines a decision framework to guide tool selection and interpretation of results across single-GPU and multi-GPU parity tests.
References existing verification utilities under examples/conversion and supports debugging workflows for Megatron Bridge model onboarding.
Enables end-to-end parity validation from conversion to forward-pass checks and VLM-grade checks when needed.

Quick Start

Run the parity testing workflow for Megatron bridge models to validate weight parity.

Frequently Asked Questions about parity-testing

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

FAQPage Schema

How do I verify weight parity between HuggingFace and Megatron-Core checkpoints?▼

To verify weight parity between HuggingFace and Megatron-Core checkpoints, use the structured decision framework to select appropriate verification utilities and validate numerical consistency across single or multi-GPU configurations.

What is the best way to debug weight mismatches during Megatron Bridge model onboarding?▼

Debugging weight mismatches during Megatron Bridge model onboarding involves applying a parity testing workflow that validates numerical consistency from initial weight conversion through forward-pass checks.

How do I validate checkpoint round-trips for Megatron Bridge models across multi-GPU configurations?▼

Validating checkpoint round-trips for Megatron Bridge models across multi-GPU configurations requires applying the defined decision framework to select and run appropriate verification tools for numerical parity checks.

Does the parity testing workflow support VLM-grade forward-pass checks?▼

Yes, the parity testing workflow supports VLM-grade checks, enabling end-to-end validation from initial weight conversion to forward-pass checks when verifying numerical parity for Megatron Bridge models.

When should I use the add-model-support skill versus existing conversion verification utilities?▼

The decision framework guides tool selection between existing verification utilities under examples/conversion and the add-model-support skill, determining the appropriate tool based on your specific weight parity debugging requirements.