TL;DR
A framework for LLM post-training with reinforcement learning at scale, connecting Megatron and SGLang.
Key features
High-Performance Training: Supports efficient training in various modes by connecting Megatron with SGLang.
Flexible Data Generation: Enables arbitrary training data generation workflows through custom data generation interfaces and server-based engines.
Native Engine Pass-Through: Keeps Megatron and SGLang control surfaces close to the upstream engines while adding the RL dataflow.
Production Validation: Validated through complete post-training workflows for large-scale models like the GLM, Qwen, and DeepSeek series.
When to use it
When you need a unified framework for large-scale reinforcement learning-based LLM post-training (e.g., RLHF, RLAIF) that integrates high-performance training and flexible data generation.