AstraFlow v0.1.1 Documentation

AstraFlow is an asynchronous RL training system for large reasoning and agentic models on distributed GPU clusters.

Get Started

• Installation
• Quick Start

System Components

• System Overview
• Dataflow
• RaaS
• Trainer
• WeightManager
• Megatron Weight Synchronization

Recipes

• Math
• Math (Offline)
• Code
• Multi-Agent (Math)
• AgentBench
• Search
• TextCraft (Recursive Agent)