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

Benchmark Guide

For reported MBD-LMs results, benchmark through the Diffulex mbd-lms reproduction branch:

https://github.com/SJTU-DENG-Lab/Diffulex/tree/mbd-lms

That branch is the stable reference for reproducing the paper experiments. It keeps the benchmark configs, model choices, and runtime assumptions aligned with the MBD-LMs training and evaluation setup.

Diffulex main is the active engine branch:

https://github.com/SJTU-DENG-Lab/Diffulex/tree/main

Use main for continued engine development, open-source contributions, new algorithm exploration, and deployment-oriented optimization. Because main evolves with runtime changes, its exact metrics may differ from the reproduction branch.

When comparing results, keep the terminology consistent:

Term	Meaning
SingleBD	Native one-block-at-a-time block diffusion decoding.
MultiBD	Multi-Block Diffusion over a bounded running-set of blocks.
MultiTF	The post-training recipe used to align MBD-LMs with practical MultiBD inference states.
Block Buffer	The fixed-size inference mechanism that makes MultiBD static-shape friendly.