Reproduce “Adapting Tree-Structured Speculative Decoding to DeepSeek-V4 for Efficient 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
Changxu Liu, Zhaogeng Li
The paper adapts tree-based speculative decoding to DeepSeek-V4’s compressed-attention states, keeping divergent branch histories separate during drafting and verification. It evaluates DeepSeek-V4-Flash on three tasks across budgets and batch sizes. Tree speculation raises accepted length, including 3.387 versus 2.842 tokens on GSM8K at D=8, and reaches 18.5% higher throughput, but gains are near break-even at some D=5 settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.