dLLM Notes

Learning notes · diffusion language models

Diffusion LLMs, unmasked.

Autoregressive models write left to right, one token per forward pass. Diffusion language models start from a sequence of blanks and fill them in, several at a time, in whatever order they are most sure of. These notes follow two tracks: how the models work, and how to serve them.

Start reading →
1 / 13 unmasked

Track M · 1 / 9 published

Inside the Model

How a diffusion LLM is built, trained, and sampled: the objective, the architecture choices, and the algorithms that turn noise into text.

Foundations

  1. M01Why not left to right?Autoregression commits one token at a time. What if we did not?Sep 2026
  2. M02Diffusion on discrete tokensThere is no Gaussian noise for a vocabulary.planned
  3. M03Masked diffusion modelsOne kind of noise: [MASK]. One loss: a weighted cross-entropy.planned
  4. M04Sampling and remaskingWhich blanks do we fill, and when do we change our mind?planned

Scaling up

  1. M05LLaDADoes masked diffusion scale like autoregression?planned
  2. M06Starting from an AR checkpointDream and DiffuLLaMA: reuse the weights, change the objective.planned
  3. M07Block diffusionAutoregressive across blocks, diffusion within them.planned

Steering

  1. M08Constrained decodingGrammars, when the order is not left to right.planned
  2. M09Reinforcement learning for dLLMsPolicy gradients without a tractable likelihood.planned

Track P · 0 / 4 published

Into Production

What it takes to serve a diffusion LLM fast and cheaply: caching, parallel decoding, batching, and the shape of an inference engine.

Making one request fast

  1. P01Caching without causalityBidirectional attention breaks the KV cache. Mostly.planned
  2. P02Parallel decodingHow many tokens can we commit in one step?planned

Serving many requests

  1. P03Batching denoising stepsEvery request is at a different noise level. Can they share a step?planned
  2. P04A serving engine for dLLMsScheduler, cache manager, sampler: what changes from an AR engine.planned