Least-Loaded Expert Parallelism: Load Balancing An Imbalanced Mixture-of-Experts

TL;DR

This paper introduces Least-Loaded Expert Parallelism (LLEP), a dynamic load balancing method for Mixture-of-Experts models that improves speed and memory efficiency during inference by rerouting tokens from overloaded to underutilized devices.

Contribution

LLEP is a novel expert parallelism algorithm that adaptively balances load across devices, addressing imbalance issues in MoE models during post-training and inference.

Findings

LLEP achieves up to 5x speedup over standard EP.

LLEP reduces peak memory usage by up to 4x.

LLEP enables faster inference, exemplified by 1.9x speedup on gpt-oss-120b.

Abstract

Mixture-of-Experts (MoE) models are typically pre-trained with explicit load-balancing constraints to ensure statistically balanced expert routing. Despite this, we observe that even well-trained MoE models exhibit significantly imbalanced routing. This behavior is arguably natural-and even desirable - as imbalanced routing allows models to concentrate domain-specific knowledge within a subset of experts. Expert parallelism (EP) is designed to scale MoE models by distributing experts across multiple devices, but with a less-discussed assumption of balanced routing. Under extreme imbalance, EP can funnel a disproportionate number of tokens to a small number of experts, leading to compute- and memory-bound failures on overloaded devices during post-training or inference, where explicit load balancing is often inapplicable. We propose Least-Loaded Expert Parallelism (LLEP), a novel EP algorithm that dynamically reroutes excess tokens and associated expert parameters from overloaded devices to underutilized ones. This ensures that all devices complete their workloads within the minimum collective latency while respecting memory constraints. Across different model scales, LLEP achieves up to 5x speedup and 4x reduction in peak memory usage compared to standard EP. This enables faster and higher-throughput post-training and inference, with ~1.9x faster for gpt-oss-120b. We support our method with extensive theoretical analysis and comprehensive empirical evaluations, including ablation studies. These results illuminate key trade-offs and enable a principled framework for hardware-specific hyper-parameter tuning to achieve optimal performance.