Hecate: Unlocking Efficient Sparse Model Training via Fully Sharded Sparse Data Parallelism

TL;DR

Hecate introduces Fully Sharded Sparse Data Parallelism (FSSDP) to improve MoE training efficiency by fully sharding parameters and using sparse collectives, significantly reducing stragglers and boosting training speed.

Contribution

The paper presents FSSDP, a novel parallelization approach for MoE models that enhances memory efficiency and reduces stragglers, implemented in the Hecate system.

Findings

Achieves up to 3.54x speedup over existing systems

Reduces memory and communication overhead

Demonstrates consistent improvements across architectures

Abstract

Mixture-of-Experts (MoE) has emerged as a promising sparse paradigm for scaling up pre-trained models (PTMs) with remarkable cost-effectiveness. However, the dynamic nature of MoE leads to rapid fluctuations and imbalances in expert loads during training, resulting in significant straggler effects that hinder training performance when using expert parallelism (EP). Existing MoE training systems attempt to mitigate these effects through expert rearrangement strategies, but they face challenges in terms of memory efficiency and timeliness of rearrangement. This paper proposes Fully Sharded Sparse Data Parallelism (FSSDP), an innovative approach that tackles the parallelization of MoE layers and potential straggler effects caused by imbalanced expert loads from a new perspective. FSSDP fully shards the parameters and optimizer states of MoE layers across devices and sparsely materializes MoE parameters from scratch in each iteration with two sparse collectives SparseAllGather and SparseReduceScatter. We build Hecate, a high-performance MoE training system that incorporates FSSDP to fully unlock its potential. Hecate introduces heterogeneous sharding, sparse materialization, and re-materialization techniques to construct flexible and efficient expert placements with low memory and communication overhead. Our evaluation reveals that Hecate achieves up to 3.54x speedup compared over state-of-the-art MoE training systems and consistently demonstrates improvements across model architectures and hardware environments.