REDDIT · REDDIT// 4d agoRESEARCH PAPER

DFlash speeds lossless speculative decoding

DFlash is a research project from Z Lab that applies a lightweight block diffusion model as the drafter in speculative decoding. By conditioning the draft model on target-model features and generating token blocks in parallel, it reports up to 6x lossless speedup overall and roughly 2.5x better speedup than EAGLE-3 on Qwen3-8B. The project ships a paper, GitHub repo, and Hugging Face model collection, with SGLang support for serving.

// ANALYSIS

This is a systems-first paper that makes diffusion useful by narrowing its job: not to replace the base LLM, but to draft blocks quickly while the verifier preserves exactness.

–The key idea is practical: parallel block drafting matters more than chasing standalone generation quality.
–Conditioning the drafter on target-model hidden features is the real unlock; it raises acceptance length without making the drafter huge.
–The reported gains are strong for an inference optimization paper, especially because they stay lossless.
–If the SGLang path is stable, this has a clearer route to real deployments than many speculative-decoding experiments.
–The main question is generality: how much of the speedup survives across more models, longer contexts, and production traffic patterns.

// TAGS

llminferencespeculative decodingdiffusionopen sourcesglanghugging face

DISCOVERED

4d ago

2026-04-07

PUBLISHED

4d ago

2026-04-07

RELEVANCE

9/ 10

AUTHOR

Total-Resort-3120