redai-infraOfficial·16 Agent Skills Included

Relax

Train and debug large multimodal AI models with reinforcement learning

Runs large-scale reinforcement learning post-training for text, vision, and audio models on Ray, Megatron, and SGLang clusters. Guides AI agents through code reviews, git commits, model integration, hang debugging, and performance tuning of training scripts. Eliminates manual cluster inspection, misconfigured GPU settings, and risky trial-and-error when migrating recipes from verl or RedAccel.

585|144

↗

npx skills add redai-infra/Relax --all -g -y

Available:

Tells the AI agent how the Relax training framework is structured, which coding and distributed-computing rules it must follow, and which expert subagent or skill to invoke for each type of task.

All Skills in This Repository (16)

Pure Emerald Level Indicators

📦 In Repo

redai-infra

model-integration

Integrate custom model architectures into the Relax training system.

git-commit

Create standardized git commit messages with markdown-formatted bodies.

creating-skills

Provides step-by-step instructions for developing Claude Code skills with file organization and metadata configuration.

redaccel-to-relax

Converts RedAccel RL training components and scripts to Relax framework equivalents.

doc-writer

Verify source code and write bilingual documentation for the Relax framework.

verl-to-relax

Convert verl reinforcement learning recipes and code to Relax.

code-review

Analyze Python codebases for quality, security, and architectural integrity.

debug-hang

Diagnose distributed training hangs in Ray clusters by analyzing call stacks and task states.

relax-dev-debug

Submit and monitor reinforcement learning training jobs on remote Ray clusters.

sync-github

Synchronizes code between internal GitLab and external GitHub via gated cherry-pick workflows.

agentic-rollout

Assess, integrate, and validate external agents for Relax resident Agentic rollout.

nemo-gym-recipe-integration

Integrates NVIDIA NeMo Gym environments into Relax as three-step RL training recipes.

Official

Advanced

Frequently Asked Questions

FAQPage Schema

How to install Relax skills?▼

Run `npx skills add redai-infra/Relax --all -g -y` in your terminal to install all skills in this suite globally.

What is the Relax framework used for?▼

Relax is an open-source engine for reinforcement learning post-training of large multimodal models, supporting algorithms like PPO, GRPO, and GSPO on Ray, Megatron, and SGLang.

How to debug a hanging distributed training job?▼

The debug-hang skill automatically collects Ray cluster status, task call stacks, and actor states to trace the blocking chain and locate the root cause.

Can I migrate training recipes from verl to Relax?▼

Yes. The verl-to-relax skill converts reward functions, tool environments, datasets, and launch scripts from verl into Relax's format step by step.

Do these skills work with Claude Code and other AI agents?▼

Yes. All skills follow the standard SKILL.md format and work in Claude Code, OpenCode, and other compatible AI coding agents.