Reproduce “Step-dLLM: Adaptive Step-aware Sparse Attention for Efficient Diffusion LLM 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
Zhichen Zeng, Xichong Zhang, Junpan Wu, Yifei Zuo, Chi-Chih Chang, Jiayi Wang, Maohua Nie, Ji Liu and 2 more
Step-dLLM reduces attention work in diffusion language models by periodically refreshing block-sparse attention masks and sharing a sparsity budget across heads. It reuses masks between anchor steps while recomputing attention. On LLaDA-1.5, it reports 61.79% GSM8K accuracy versus SparseD’s 29.64% at 90% nominal sparsity, and 2.78× end-to-end speedup at 8K context and 256 steps, relative to dense FlashAttention.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.