Reproduce “WorldPager: Zero-Stall Tiered Memory for Long-Horizon World Models on a Single Consumer GPU” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bingnan Liu, Ruini Xue
WorldPager extends interactive world models by storing compressed attention state across GPU, host memory, and NVMe, and prefetching needed chunks during generation. It predicts retrieval from camera pose and reuses the set after a miss. On an RTX 4090, it generated 3,000 chunks, or 37.5 minutes of video, within host-memory limits without stalls above 0.5 ms; the run took 63.2 minutes, not real-time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.