Lory: Fully Differentiable Mixture-of-Experts for Autoregressive Language Model Pre-training

TL;DR

Lory introduces a fully-differentiable MoE architecture for autoregressive language model pre-training, employing novel routing and batching techniques to improve efficiency, expert specialization, and downstream performance.

Contribution

It is the first to scale fully-differentiable MoE architectures to autoregressive language model pre-training, demonstrating significant performance improvements.

Findings

Achieved +13.9% perplexity improvement over dense models

Models with segment routing perform competitively with token-level routing MoEs

Experts capture domain-level specialization without supervision

Abstract

Mixture-of-experts (MoE) models facilitate efficient scaling; however, training the router network introduces the challenge of optimizing a non-differentiable, discrete objective. Recently, a fully-differentiable MoE architecture, SMEAR, was proposed (Muqeeth et al., 2023), which softly merges experts in the parameter space; nevertheless, its effectiveness was only demonstrated in downstream fine-tuning on classification tasks. In this paper, we present Lory, the first approach that scales such architectures to autoregressive language model pre-training. Lory introduces two key techniques: (1) a causal segment routing strategy that achieves high efficiency for expert merging operations while preserving the autoregressive nature of language models; (2) a similarity-based data batching method that encourages expert specialization by grouping similar documents in training instances. We pre-train a series of Lory models on 150B tokens from scratch, with up to 32 experts and 30B (1.5B active) parameters. Experimental results show significant performance gains over parameter-matched dense models on both perplexity (+13.9%) and a variety of downstream tasks (+1.5%-11.1%). Despite segment-level routing, Lory models achieve competitive performance compared to state-of-the-art MoE models with token-level routing. We further demonstrate that the trained experts in Lory capture domain-level specialization without supervision. Our work highlights the potential of fully-differentiable MoE architectures for language model pre-training and advocates future research in this area.