Rollplex: Cross-Phase GPU Spatial Sharing for Vision Language Model Post-Training

Hanfeng Lu, Tianyu Feng, Suyi Li, Yuheng Zhao, Wei Gao et al.|August 14, 2026arXiv

Key Takeaway

By overlapping independent computation phases and sharing GPU memory intelligently, you can train vision-language models 1.2–2.2× faster without needing more hardware or changing your RL algorithm.

Summary

Rollplex is a GPU runtime that speeds up vision-language model training by overlapping different computational phases.

efficiency training architecture

Key Terms

tensor-parallel on-policy-rl phase-aware-memory-management weight-sharing