SGLang implements fully deterministic inference with only 34.35% performance overhead and enables 100% reproducible RL training in collaboration with slime, providing reliable solutions for rigorous scientific experiments.
We implemented an end-to-end FP8 sampling and training pipeline for RL. Experiments show that for MoE models, using BF16 training with FP8 rollout leads to severe train-inference inconsistency as model size increases. Unified FP8 for both training and rollout effectively eliminates quantization-induced inconsistency, improving RL training speed and stability.
The SGLang RL team achieves major breakthroughs in RL training stability and efficiency, implementing end-to-end INT4 QAT that enables ~1TB model deployment on a single H200 GPU while maintaining training-inference consistency.
