Just found the dLLM library to create Diffusion Language Models It's still early but it's insanely fun to experiment with diffusion (training, inference, eval) dLLM has the potential of becoming the main library for diffusion LLMs

💻 Library: Source: by Individual-Ninja-141

Diffusion LLMs are like teaching language models to dream before they talk. Instead of guessing the next word, they refine whole thoughts from noise, piece by piece. dLLM might just make text generation feel more like painting sentences than typing them. It’s chaos turning into poetry, one token at a time.

liquid training a ?

Can these models be distilled or merged?

Definitely! Especially merging

I suspect there will be some fantastic results, especially with reasoning isolation, distillation and transfer.

Dang, UI is well put together in the terminal and it looks so fun to try this model

We’ve been experimenting with it for speech understanding too; the idea of generating or reasoning over continuous representations instead of discrete tokens feels like the next frontier.

Have to try it out soon.

neat

Sounds like a wild ride!

I did something like this with a critique that self corrects it self also. This one is equivalent to autoregression

So cool

A model can think in Diffusion and talk in Autoregression.

Nice!

curious to see what kind of unique applications this enables compared to standard transformer architectures

@rdiandaru

That's a cool find, Maxime! The potential of dLLM sounds quite exciting, especially for diffusion LLMs, don't you think?

Very cool

The idea of combining diffusion models with LLMs seems like it could open up new research directions. How beginner-friendly is dLLM for someone new to this?

This is new to me. Can this generate tokens in parallel or is it still linear?

I suppose dLLM is similar to BERT