[S29] Sep 18, Friday: Unlocking Lossless Speedups in LLMs via Discrete Diffusion

[Diffusion LLM]

Hello folks, 

Diffusion LLMs have two limitations relative to AR models:
(1) Lower quality, and
(2) Slower inference at large batch sizes.

In this talk, Subham will introduce a new class of models called Diffusion-augmented LLMs address these issues. They

> Retain the AR architecture of LLMs
> Each layer has  two sets of weights:  AR weights and Diffusion weights
> Diffusion weights enable parallel sampling from the AR distribution losslessly

The resulting model Uno, is
💥 Faster than all speculative decoding methods: DFlash and EAGLE-3 

💥 Speeds up RL-postraining unlike speculative decoding methods
🔥 Beats ALL diffusion LLMs: Mercury 2, Diffusion Gemma, Llada

Meeting Link: click here

Time: Sep 17 (Friday) 1pm ET / 10am PT / 7pm CET / 10:30pm IST

Paper: https://arxiv.org/abs/2609.04010

[Diffusion LLM]

Starting in 30 mins!

[Diffusion LLM]

Hi folks, we just uploaded the recording of today's session, make sure to check it out: https://www.youtube.com/watch?v=IqbltB4ppqs