reference/code/mbd-lms/docs/guidelines/inference_setup.md

Inference Setup

MBD-LMs are trained and defined in this repository, but the production inference engine is Diffulex.

Use the Diffulex mbd-lms branch when you want to reproduce the reported MBD-LMs experiment behavior:

git clone https://github.com/SJTU-DENG-Lab/Diffulex.git
cd Diffulex
git checkout mbd-lms

Use Diffulex main when you want to develop the engine, contribute new runtime features, or explore new decoding algorithms:

git checkout main

The method name in the paper is Multi-Block Diffusion (MultiBD). In Diffulex configs, the corresponding runtime strategy is:

decoding_strategy: multi_bd

Diffulex main may contain newer performance optimizations and model-specific runtime changes, so use mbd-lms for experiment reproduction and main for future development.