Paper

Train High, Ship Low: Efficient LoRA Synchronization for Distributed On-Policy Post-Training · ByteTrail