Reproduce “Tsubame: Tree Replay for Diffusion-Based Speculative Decoding” 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
Yepeng Weng, Qiao Hu, Takehisa Yairi
Tsubame separates tree planning from candidate sampling in speculative decoding. It first fixes a context-aware tree topology using draft-path scores, then fills its slots with freshly sampled tokens for target-model verification; diffusion-based drafters make replay practical. With replay, it outperforms both first-pass trees and sampled-chain Block Verification on six datasets for Domino and DFlash2, in the tested settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.