GRIN: GRadient-INformed MoE
This addresses the problem of scaling MoE models for researchers and practitioners in deep learning by improving training efficiency and performance, though it is incremental as it builds on existing MoE frameworks.
The paper tackles the challenge of training Mixture-of-Experts (MoE) models by introducing GRIN, a method that uses sparse gradient estimation for expert routing to enable effective gradient-based optimization, resulting in a model with 6.6B activated parameters that outperforms a 7B dense model and matches a 14B dense model on language tasks.
Mixture-of-Experts (MoE) models scale more effectively than dense models due to sparse computation through expert routing, selectively activating only a small subset of expert modules. However, sparse computation challenges traditional training practices, as discrete expert routing hinders standard backpropagation and thus gradient-based optimization, which are the cornerstone of deep learning. To better pursue the scaling power of MoE, we introduce GRIN (GRadient-INformed MoE training), which incorporates sparse gradient estimation for expert routing and configures model parallelism to avoid token dropping. Applying GRIN to autoregressive language modeling, we develop a top-2 16$\times$3.8B MoE model. Our model, with only 6.6B activated parameters, outperforms a 7B dense model and matches the performance of a 14B dense model trained on the same data. Extensive evaluations across diverse tasks demonstrate the potential of GRIN to significantly enhance MoE efficacy, achieving 79.4 on MMLU, 83.7 on HellaSwag, 74.4 on HumanEval, and 58.9 on MATH.