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

Setup

pip install -r requirements.txt

The code expects a CUDA GPU.

Run

Run a minimal generation example:

python main.py

Run MBPP evaluation:

./eval.sh

eval.sh evaluates mbpp_sanitized_llada_moe for multiple jump_steps ($\tau$) values and writes outputs under ./out.

Acknowledgements

This project builds on inclusionAI/LLaDA2.0.

The evaluation code is adapted from dInfer.

Citation

If you find our work useful, please consider citing our paper or giving it a star! ⭐️

@article{chen2026tide,
      title={TIDE: Efficient and Lossless MoE Diffusion LLM Inference with I/O-aware Expert Offload}, 
      author={Zhiben Chen and Youpeng Zhao and Yang Sui and Jun Wang and Yuzhang Shang},
      year={2026},
      eprint={2605.20179},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.20179}, 
}