skills.rest
DeepExperience
DeepExperienceOFFICIAL
Explore all skills by this author →
GitHub

model-integration

Guide new model architecture integration into the Relax training pipeline.

71|Updated May 9, 2026
Search Tags:#reinforcement learning#model integration#huggingface#fsdp#megatron#tensor parallel#weight conversion
One-click install
npx skills add https://github.com/DeepExperience/HyperEyes --skill model-integration-deepexperience
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: model-integration
Source: https://github.com/DeepExperience/HyperEyes/tree/main/RL/relax-hypereyes/Relax/skills/model-integration
Command: npx skills add https://github.com/DeepExperience/HyperEyes --skill model-integration-deepexperience

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps engineers add new model architectures to the Relax training pipeline by guiding weight conversion, synchronization, and backend adaptation steps that are otherwise complex and error-prone.

Core Features & Use Cases

  • Model Integration Guidance: Provides workflows for connecting models across Megatron and FSDP backends, including bridge and raw conversion paths.
  • Weight Conversion Support: Explains how to implement Megatron-to-HuggingFace converters, register mappings, and handle tensor parallel transformations.
  • Use Case: A machine learning engineer adding a new multimodal or MoE model can follow this Skill to configure training scripts, implement required converters, and validate weight synchronization.

Quick Start

Ask the model-integration skill to guide the steps required to add a new model architecture to the Relax training pipeline.

Frequently Asked Questions about model-integration

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

FAQPage Schema
How do I integrate a new model architecture into the Relax training pipeline?▼

To integrate a new model architecture into the Relax training pipeline, you need to define weight conversion and synchronization workflows. This involves configuring model providers, implementing backend-specific integration steps, and validating rollout weight synchronization.

How do I implement Megatron to HuggingFace weight conversion?▼

Megatron to HuggingFace weight conversion requires implementing specific weight converters and registering correct model mapping logic. You must handle tensor parallel transformations to ensure reliable rollout weight synchronization across the integrated backends.

Does the Relax pipeline support both FSDP and Megatron backends for model integration?▼

Yes, the Relax training pipeline supports both FSDP and Megatron backend adaptation. It provides dedicated workflows for connecting models across these backends, utilizing both bridge and raw conversion paths for flexible integration.

What is tensor parallel handling and when is it needed during model integration?▼

Tensor parallel handling is the process of managing weight transformations across distributed training backends. It is needed when adapting new model architectures to Megatron or FSDP to ensure correct weight synchronization and parallel execution.

Why does weight synchronization fail when adding new models to distributed training backends?▼

Weight synchronization fails when model mapping logic or tensor parallel transformations are incorrectly configured. Validating rollout weight synchronization and implementing proper backend-specific weight converters are required to prevent these integration errors.

What's the best way to configure training launch scripts for a new multimodal model?▼

The best way to configure training launch scripts for a new multimodal model is to follow guided integration workflows. This ensures you correctly set up model providers, implement necessary weight converters, and validate the synchronization process.

Related Skills

Explore More Software Engineering Skills→
BirgerMoellBirgerMoell

megatron-hf-converter — Claude Code skill guide

Convert HuggingFace and Megatron-LM checkpoints with numerical equivalence validation.

Community
Intermediate
heysuhasheysuhas
2

slime-rl-training

Integrate Megatron-LM and SGLang for large-scale LLM reinforcement learning training.

Community
Advanced
DeepExperienceDeepExperience
71

redaccel-to-relax

Migrates RedAccel reinforcement learning algorithms to Relax-compatible modules and interfaces.

Official
Advanced
EnkeZEnkeZ
1

model-integration

Integrate model architectures into the Relax training framework with Megatron and FSDP backends.

Community
Advanced