MoNDE: Mixture of Near-Data Experts for Large-Scale Sparse Models

TL;DR

MoNDE introduces a near-data computing approach for large-scale MoE language models, significantly reducing memory transfer costs and improving inference efficiency by selectively transferring only hot experts.

Contribution

The paper proposes MoNDE, a novel near-data computing framework that minimizes expert parameter movement in MoE models, enabling more efficient large-scale language model inference.

Findings

MoNDE achieves substantial speedups over existing frameworks.

It reduces memory transfer by transferring only hot experts.

It improves inference efficiency for both encoder and decoder operations.

Abstract

Mixture-of-Experts (MoE) large language models (LLM) have memory requirements that often exceed the GPU memory capacity, requiring costly parameter movement from secondary memories to the GPU for expert computation. In this work, we present Mixture of Near-Data Experts (MoNDE), a near-data computing solution that efficiently enables MoE LLM inference. MoNDE reduces the volume of MoE parameter movement by transferring only the $\textit{hot}$ experts to the GPU, while computing the remaining $\textit{cold}$ experts inside the host memory device. By replacing the transfers of massive expert parameters with the ones of small activations, MoNDE enables far more communication-efficient MoE inference, thereby resulting in substantial speedups over the existing parameter offloading frameworks for both encoder and decoder operations.