Reproduce “X-Cache: Cross-Chunk Block Caching for Few-Step Autoregressive World Models Inference” 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
Yixiao Zeng, Jianlei Zheng, Chaoda Zheng, Shijia Chen, Mingdian Liu, Tongping Liu, Tengwei Luo, Yu Zhang and 5 more
Researchers at XPeng Inc. developed X-Cache, a training-free method to accelerate few-step autoregressive video diffusion models for autonomous driving. It exploits cross-chunk redundancy by caching and reusing Diffusion Transformer (DiT) block residuals across consecutive video chunks, achieving a 2.6-2.7x speedup with imperceptible quality degradation.
Key insights
- In physically continuous environments like autonomous driving, a unique form of "cross-chunk" redundancy persists in DiT block inputs across sequential generations, even when few-step distillation eliminates "cross-step" redundancy.
- Effective caching for interactive world models requires a similarity assessment that is both structure-aware (via spatio-temporal subsampling) and action-aware (via explicit conditioning channels).
- Protecting critical operations, particularly the KV cache update frames, is essential to prevent long-term error propagation and maintain the fidelity of autoregressive generative models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.