Reproduce “One Block Ahead: Promoting Sampler State in Diffusion Language Models” 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
Yangyi Shen, Pin-Jui Ku, Nicolai Oswald, Jean-Marie Lemercier, Yonggan Fu, Stefano Ermon, Jure Leskovec, Boris Ginsburg and 1 more
OBA overlaps denoising the next block with current-block refinement, carrying forward tokens, self-conditioning, and denoising time. On DiffusionGemma, this sampler-state promotion raises tokens per forward pass from 15.64 to 22.25 and denoising throughput from 2,375 to 2,761 tokens per second across ten benchmarks, while accuracy falls 1.37 points. Results are for batch one on a GB300 GPU and uniform-state diffusion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.