# LSH-MoE: Communication-efficient MoE Training via Locality-Sensitive Hashing

arXiv: https://arxiv.org/abs/2411.08446 · HTML: https://mlsanity.com/paper/2411.08446/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- Placed in Efficient MoE training and inference systems, one of 12 directions in Mixture-of-experts.
- Builds on 10 of its 41 references that are in the map.
- First posted on arXiv 13 Nov 2024.

## Placed in

- Mixture-of-experts / [Efficient MoE training and inference systems](https://mlsanity.com/moe/efficient-moe-training-and-inference-systems/)

## Builds on (most cited first, 10 of 10)

- [Scaling Laws for Neural Language Models](https://mlsanity.com/paper/2001.08361/) (2020)
- [ST-MoE: Designing Stable and Transferable Sparse Expert Models](https://mlsanity.com/paper/2202.08906/) (2022)
- [Efficient Large Scale Language Modeling with Mixtures of Experts](https://mlsanity.com/paper/2112.10684/) (2021)
- [Pre-gated MoE: An Algorithm-System Co-Design for Fast and Scalable Mixture-of-Expert Inference](https://mlsanity.com/paper/2308.12066/) (2023)
- [M6: A Chinese Multimodal Pretrainer](https://mlsanity.com/paper/2103.00823/) (2021)
- [Routing Networks and the Challenges of Modular and Compositional Computation](https://mlsanity.com/paper/1904.12774/) (2019)
- [HetuMoE: An Efficient Trillion-scale Mixture-of-Expert Distributed Training System](https://mlsanity.com/paper/2203.14685/) (2022)
- [EvoMoE: An Evolutional Mixture-of-Experts Training Framework via Dense-To-Sparse Gate](https://mlsanity.com/paper/2112.14397/) (2021)
- [M6-T: Exploring Sparse Expert Models and Beyond](https://mlsanity.com/paper/2105.15082/) (2021)
- [Angel-PTM: A Scalable and Economical Large-scale Pre-training System in Tencent](https://mlsanity.com/paper/2303.02868/) (2023)

## Built on by (most cited first, 5 of 5)

- [Brief analysis of DeepSeek R1 and its implications for Generative AI](https://mlsanity.com/paper/2502.02523/) (2025)
- [Fine-grained MoE Load Balancing with Linear Programming](https://mlsanity.com/paper/2511.16947/) (2025)
- [Expert Upcycling: Shifting the Compute-Efficient Frontier of Mixture-of-Experts](https://mlsanity.com/paper/2604.19835/) (2026)
- [UBEP: Re-architecting Expert Parallelism Communication Library for Production Superpods](https://mlsanity.com/paper/2607.06202/) (2026)
- [CIERA: Cross-Iteration Exponent Reuse for Lossless Allgather in Sharded MoE Training](https://mlsanity.com/paper/2609.04609/) (2026)

## Abstract (arXiv)

Larger transformer models always perform better on various tasks but require more costs to scale up the model size. To efficiently enlarge models, the mixture-of-experts (MoE) architecture is widely adopted, which consists of a gate network and a series of experts and keep the training cost constant by routing the input data to a fixed number of experts instead of all. In existing large-scale MoE training systems, experts would be distributed among different GPUs for parallelization, and thus input data requires additional all-to-all communications to access the target experts and conduct corresponding computations. However, upon evaluating the training process of three mainstream MoE models on commonly used GPU clusters, we found that the all-to-all communication ratio averaged around 45%, which significantly hinders the efficiency and scalability of training MoE models. In this paper, we propose LSH-MoE, a communication-efficient MoE training framework using locality-sensitive hashing (LSH). We first present the problems of scaling MoE training in existing systems and highlight the potential of exploiting token similarity to facilitate data compression. Then, we introduce an efficient LSH-based compression technique, which utilizes the cross-polytope hashing for rapid clustering and implements a residual-based error compensation scheme to alleviate the adverse impact of compression. To verify the effectiveness of our methods, we conduct experiments on both language models (e.g., RoBERTa, GPT, and T5) and vision models (e.g., Swin) for pre-training and fine-tuning tasks. The results demonstrate that our method substantially outperforms its counterparts across different tasks by 1.28x - 2.2x of speedup.