# ZHZisZZ/dllm

dLLM: Simple Diffusion Language Modeling

Repository: https://github.com/ZHZisZZ/dllm
Canonical: https://ross.abutalabs.com/products/dllm
Homepage: https://arxiv.org/pdf/2602.22661
Language: Python
License: Apache-2.0
License Family: permissive
Topics: llm, discrete-diffusion-models, nlp
Last push: 2026-07-17T21:15:25+00:00

## Health v2 (maintenance only)
Score: 59/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 93, release rhythm 35, longevity 25
- inputs: {"age_days": 353, "days_push": 47, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2672, forks 278 (observed 2026-08-28T04:07:09.580519+00:00)

## What it is
dLLM is a Python library that unifies training, inference, and evaluation of diffusion language models such as LLaDA and Dream. It builds on Hugging Face transformers Trainer and lm-evaluation-harness, with support for LoRA, DeepSpeed, FSDP, and algorithms like MDLM, BD3LM, and diffu-GRPO.

## Use cases
- train a masked diffusion language model
- fine-tune LLaDA or Dream with LoRA
- evaluate diffusion LMs on benchmarks
- run GRPO reinforcement learning on diffusion language models
- implement block diffusion or edit flows training algorithms
- accelerate diffusion LM inference with caching

## When to choose
- you want a reproducible, unified pipeline for diffusion language model training and evaluation
- you need recipes for open-weight diffusion models like LLaDA and Dream
- you want to experiment with masked diffusion or block diffusion training algorithms

## When to avoid
- you only need standard autoregressive LLM training, which existing frameworks handle better
- you need production inference serving rather than research training and evaluation

## Facets
- artifact type: library
- maturity: active
- function: llm-training, machine-learning, llm-inference, rag
- domain: large-language-models, machine-learning, deep-learning
- platform: python
- tags: diffusion-language-models, masked-diffusion, llada, dream, grpo, lora, deepspeed, fsdp, lm-evaluation-harness, natural-language-processing, gpu

## Member repositories
- ZHZisZZ/dllm (main) score 59

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:09.580519+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T02:18:06.265016+00:00, confidence not recorded.
  - readme: https://github.com/ZHZisZZ/dllm (fetched 2026-08-28T04:07:09.580519+00:00, sha 58bb8b133034)
- Data as of 2026-08-30T08:39:29.467469+00:00.