SkyRL Documentation

Welcome to SkyRL's Documentation!

SkyRL is a full-stack reinforcement learning library for training LLMs, designed for modularity and extensibility.

Getting Started

• Installation - Set up SkyRL with Docker or from source
• Quick Start - Run your first GRPO training on GSM8K
• System Overview - Understand SkyRL's architecture

Tutorials

• Creating a New Environment - Build custom environments and tasks
• Async Training - Implement asynchronous training pipelines
• Using Tools - Add tool-use capabilities to your agents

Examples

• Megatron Backend - 5D parallelism for large-scale training
• Multi-Turn Text2SQL - Train agents for database queries
• Search with SkyRL - Multi-turn RL for search tasks

Reference

• Configuration - Training configuration options
• Algorithms - DAPO and custom algorithm implementation
• Recipes - End-to-end training recipes