Reproduce “Accelerating Parallel Diffusion Model Serving with Residual Compression” 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
Jiajun Luo, Yicheng Xiao, Jianru Xu, Yangxiu You, Rongwei Lu, Chen Tang, Jingyan Jiang, Zhi Wang
CompactFusion, a residual compression framework, accelerates parallel diffusion model serving by reducing inter-device communication. It achieves this by transmitting only compressed differences between successive activations, resulting in up to a 3.0x speedup over single-device inference and superior generation quality (e.g., FID of 3.263 with 2-bit quantization) compared to prior overlap-based methods, while also enabling communication-intensive strategies on low-bandwidth networks.
Key insights
- Diffusion model activations exhibit strong temporal redundancy across steps, making the smaller-magnitude 'residuals' a significantly more efficient target for compression than full activations.
- Directly compressing these residuals with error feedback is more effective for maintaining high generation fidelity and reducing communication volume than merely overlapping the transfer of redundant full activations.
- For extreme compression ratios using low-rank approximation, prioritizing 'directional coverage' (approximating more dimensions of change with quantization) proves more critical for preserving generation quality than maximizing per-step approximation accuracy in fewer dimensions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.