Introduction
AReaL (Asynchronous Reinforcement Learning) is a highly scalable asynchronous RL training system designed for large-scale LLM inference and agent training. It supports multiple training backends, distributed configurations, and modular algorithm components.
Key Features
- Fully asynchronous training pipeline to improve resource utilization and throughput.
- Rich integration with training/inference backends (vLLM, Megatron, FSDP, Ray, etc.).
- Configurable and reproducible toolchain for both research and engineering.
Use Cases
- Large-scale RLHF / agent training.
- Algorithm research and rapid prototyping.
Technical Highlights
- Primarily implemented in Python, with documentation and examples for easy onboarding.