Sparse Upcycling: Inference Inefficient Finetuning

Small, highly trained, open-source LLMs are widely used due to their inference efficiency, but further improving their quality remains a challenge. Sparse upcycling is a promising approach that transforms a pretrained dense model into a Mixture-of-Experts (MoE) architecture, increasing the model’s parameter count and potential quality. In this work, we compare the effectiveness of sparse upcycling against continued pretraining (CPT) across different model sizes, FLOP budgets, and pretraining durations. Our experiments show that sparse upcycling can achieve better quality, with improvements of over 20% relative to CPT in certain scenarios. However, this comes with a significant inference cost, leading to 40% slowdowns in high-demand inference settings for larger models. These results highlight the trade-off between model quality and inference efficiency, offering insights for practitioners seeking to balance performance with practical deployment costs.