TileQ: Efficient Low-Rank Quantization of Mixture-of-Experts with 2D Tiling

Hongyaoxing Gu, Xinzhe Chen, Lijuan Hu, Liu Fangfang
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:36981-37010, 2026.

Abstract

Mixture-of-Experts (MoE) models achieve remarkable performance by sparsely activating specialized experts, yet their massive parameters in experts pose significant challenges for deployment. While low-rank quantization offers a promising route to compress MoE models, existing methods still incur nonnegligible memory overhead and inference latency. To address these limitations, we propose TileQ, a fine-tuning-free post-training quantization (PTQ) method that employs 2D-tiling structured low-rank quantization to share low-rank factors across both input and output dimensions of MoE experts. Furthermore, we introduce an efficient inference technique for TileQ that fuses multiple low-rank expert computations into a single-pass operation, significantly improving hardware utilization. Experiments show that TileQ cuts down additional memory usage up to 10x and reduces inference latency to 5% while preserving state-of-the-art accuracy.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-gu26c, title = {{T}ile{Q}: Efficient Low-Rank Quantization of Mixture-of-Experts with 2{D} Tiling}, author = {Gu, Hongyaoxing and Chen, Xinzhe and Hu, Lijuan and Fangfang, Liu}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {36981--37010}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/gu26c/gu26c.pdf}, url = {https://proceedings.mlr.press/v306/gu26c.html}, abstract = {Mixture-of-Experts (MoE) models achieve remarkable performance by sparsely activating specialized experts, yet their massive parameters in experts pose significant challenges for deployment. While low-rank quantization offers a promising route to compress MoE models, existing methods still incur nonnegligible memory overhead and inference latency. To address these limitations, we propose TileQ, a fine-tuning-free post-training quantization (PTQ) method that employs 2D-tiling structured low-rank quantization to share low-rank factors across both input and output dimensions of MoE experts. Furthermore, we introduce an efficient inference technique for TileQ that fuses multiple low-rank expert computations into a single-pass operation, significantly improving hardware utilization. Experiments show that TileQ cuts down additional memory usage up to 10x and reduces inference latency to 5% while preserving state-of-the-art accuracy.} }
Endnote
%0 Conference Paper %T TileQ: Efficient Low-Rank Quantization of Mixture-of-Experts with 2D Tiling %A Hongyaoxing Gu %A Xinzhe Chen %A Lijuan Hu %A Liu Fangfang %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-gu26c %I PMLR %P 36981--37010 %U https://proceedings.mlr.press/v306/gu26c.html %V 306 %X Mixture-of-Experts (MoE) models achieve remarkable performance by sparsely activating specialized experts, yet their massive parameters in experts pose significant challenges for deployment. While low-rank quantization offers a promising route to compress MoE models, existing methods still incur nonnegligible memory overhead and inference latency. To address these limitations, we propose TileQ, a fine-tuning-free post-training quantization (PTQ) method that employs 2D-tiling structured low-rank quantization to share low-rank factors across both input and output dimensions of MoE experts. Furthermore, we introduce an efficient inference technique for TileQ that fuses multiple low-rank expert computations into a single-pass operation, significantly improving hardware utilization. Experiments show that TileQ cuts down additional memory usage up to 10x and reduces inference latency to 5% while preserving state-of-the-art accuracy.
APA
Gu, H., Chen, X., Hu, L. & Fangfang, L.. (2026). TileQ: Efficient Low-Rank Quantization of Mixture-of-Experts with 2D Tiling. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:36981-37010 Available from https://proceedings.mlr.press/v306/gu26c.html.

Related Material