Search papers, labs, and topics across Lattice.
The paper introduces "expert upcycling," a method for efficiently scaling Mixture-of-Experts (MoE) models by progressively increasing the number of experts during continued pre-training through expert duplication and router extension. By duplicating experts and using gradient-based importance scores for non-uniform duplication, the method provides a warm initialization that significantly reduces training costs compared to training from scratch. Experiments on 7B-13B parameter models demonstrate a 32% reduction in GPU hours while matching the performance of fixed-size baselines.
Expert upcycling lets you scale MoEs for 32% less compute by intelligently duplicating and specializing existing experts, challenging the need to train massive MoEs from scratch.
Mixture-of-Experts (MoE) has become the dominant architecture for scaling large language models: frontier models routinely decouple total parameters from per-token computation through sparse expert routing. Scaling laws show that under fixed active computation, model quality scales predictably with total parameters, and MoEs realize this by increasing expert count. However, training large MoEs is expensive, as memory requirements and inter-device communication both scale with total parameter count. We propose expert upcycling, a method for progressively expanding MoE capacity by increasing the number of experts during continued pre-training (CPT). Given a trained E-expert model, the upcycling operator constructs an mE-expert model through expert duplication and router extension while holding top-K routing fixed, preserving per-token inference cost. Duplication provides a warm initialization: the expanded model inherits the source checkpoint's learned representations, starting from a substantially lower loss than random initialization. Subsequent CPT then breaks the symmetry among duplicated experts to drive specialization. We formalize the upcycling operator and develop a theoretical framework decomposing the quality gap into a capacity term and an initialization term. We further introduce utility-based expert selection, which uses gradient-based importance scores to guide non-uniform duplication, more than tripling gap closure when CPT is limited. In our 7B-13B total parameter experiments, the upcycled model matches the fixed-size baseline on validation loss while saving 32% of GPU hours. Comprehensive ablations across model scales, activation ratios, MoE architectures, and training budgets yield a practical recipe for deploying expert upcycling, establishing it as a principled, compute-efficient alternative to training large MoE models from scratch.
Amazon Science