
Unlocking Lossless Speedups in LLMs via Discrete Diffusion
Generating text one token at a time creates a bottleneck for LLM inference. Can models generate multiple tokens in parallel without changing their output distribution?
In this Nebius Science Paper Club session, Subham Sahoo, who leads the Diffusion LLM team at the Institute of Foundation Models, presents Unlocking Lossless Speedups in LLMs via Discrete Diffusion.
The paper introduces Uno, a family of models that combines autoregressive language modeling with lightweight diffusion weights for parallel token generation. We’ll examine how diffusion distillation and Ψ-Spec sampling enable lossless acceleration without a separate draft model, what the reported results show, and what they mean for efficient LLM inference.
What you’ll learn
- How Uno combines autoregressive modeling with discrete diffusion
- How diffusion distillation enables parallel token generation
- What “lossless” acceleration means and how Ψ-Spec preserves the model’s distribution
- How the approach compares with speculative decoding and diffusion LLMs across the evaluated workloads and batch sizes
Who should attend
ML researchers, research engineers and PhD students working on language modeling, generative models or inference efficiency. Familiarity with autoregressive LLMs and basic diffusion concepts will help you follow the discussion.
Recommended reading: Unlocking Lossless Speedups in LLMs via Discrete Diffusion
About Nebius Science Paper Club
Nebius Science Paper Club is a webinar series led by Nebius researchers. Each session brings together paper authors and practitioners to discuss new ideas, research, and discoveries in AI.