Sparse Upcycling: Inference Inefficient Finetuning
This work addresses the trade-off between model quality and inference efficiency for practitioners deploying language models, though it is incremental as it builds on existing sparse upcycling methods.
The paper tackled the challenge of improving the quality of small, open-source large language models by comparing sparse upcycling, which transforms dense models into Mixture-of-Experts architectures, against continued pretraining. It found that sparse upcycling can achieve over 20% better quality in some scenarios but incurs a 40% inference slowdown for larger models.
Small, highly trained, open-source large language models 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 quality. In this work, we compare the effectiveness of sparse upcycling against continued pretraining (CPT) across different model sizes, compute 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. Our findings highlight the trade-off between model quality and inference efficiency, offering insights for practitioners seeking to balance model quality and deployment constraints.