Open-source alternatives to Dflash

Constrained Decoding of Diffusion LLMs with Context-Free Grammars

dVoting: Fast Voting for dLLMs

The pipeline of dKV-Cache

https://github.com/user-attachments/assets/09c67a58-b425-463d-a998-c1a6049bc171

dInfer is an efficient and extensible inference framework for dLLMs. As illustrated in the following architecture, it modularizes inference into four components: model, diffusion iteration manager, decoder and KV-cache manager. It provides well-designed APIs for flexible algorithms combinations…

MaskKV: Fine-Grained Cache Eviction for Efficient dLLM Inference

Time-Annealed Perturbation Sampling (TAPS) is an inference-time method for improving diversity in diffusion language models without sacrificing generation quality.

Jihoon Lee 1 , Hoyeon Moon 1 , Kevin Zhai 5 , Arun Kumar Chithanar, Anit Kumar Sahu 2 , Soummya Kar 3 , Chul Lee, Souradip Chakraborty 4 , Amrit Singh Bedi 5

Research dLLM Codebase for Diffusion Language Models Research

We introduce ReMDM, a simple and general framework to design remasking samplers for masked discrete diffusion models. In this repo, we provide our implementation of different ReMDM strategies for unconditional text generation on OpenWebText. We also provide a demo in this notebook showing how to…

Beyond Fixed: Training-Free Variable-Length Denoising for Diffusion Large Language Models

Official PyTorch implementation of the paper "Accelerating Diffusion Large Language Models with SlowFast Sampling: The Three Golden Principles" (Slow Fast Sampling).

This repository contains the code for Optimizing Decoding Paths in Masked Diffusion Models by Quantifying Uncertainty.

Code repository for the paper Think While You Generate: Discrete Diffusion with Planned Denoising, by Sulin Liu, Juno Nam, Andrew Campbell, Hannes Stärk, Yilun Xu, Tommi Jaakkola, Rafael Gómez-Bombarelli. Tweet and video for the main idea.

DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (Paper Coming Soon)

DSB: Dynamic Sliding Block Scheduling for Diffusion LLMs

Lingkun Long 1 , Yushi Huang 2,3 , Shihao Bai 3 , Ruihao Gong 1,3 , Jun Zhang 2 , Ao Zhou 1 , Jianlei Yang 1,*

Official PyTorch implementation of the paper "dLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive Caching" (dLLM-Cache) in ICML 2026.

Yicun Yang 1 , Cong Wang 1 , Shaobo Wang 1 , Zichen Wen 1 , Biqing Qi 2 , Hanlin Xu 3 , Linfeng Zhang 1 †

2026-1-28

We would like to express our gratitude to the previous studies on DLLM KV caching that inspired our work. Special thanks to the authors of d2Cache for providing excellent open-source code, which served as a valuable foundation for our experimental framework. We also acknowledge the developers of…

ACL '26 Source code for paper "Empirical Analysis of Decoding Biases in Masked Diffusion Models"

Performance: Picture on the left shows BLEU Scores of different models for the paraphrase task on the QQP dataset. Our FMSeq beats all the models when using a single sampling step and achieves comparable performance to DiffuSeq (2000 steps) with only 10 steps. Workflow: Picture on the right…