TIDE: Efficient and Lossless MoE Diffusion LLM Inference with I/O-aware Expert Offload

Researchers introduced TIDE, a lossless inference system for diffusion MoE language models.

TIDE exploits the temporal stability of expert activations and refreshes expert placement at intervals, reducing I/O traffic and CPU work.

The method needs no retraining and improved throughput on single GPU-CPU setups in tests on LLaDA2.0-mini and LLaDA2.0-flash.

It offers a training-free way to make large diffusion models more efficient on limited hardware.