Sparse Upcycling: Inference Inefficient Finetuning

TL;DR

Sparse upcycling transforms pretrained dense models into Mixture-of-Experts architectures, improving quality by over 20% compared to continued pretraining but at the cost of increased inference latency, highlighting a quality-efficiency trade-off.

Contribution

This work systematically compares sparse upcycling with continued pretraining across various settings, demonstrating its potential to enhance model quality despite higher inference costs.

Findings

Sparse upcycling achieves over 20% quality improvement in some scenarios.

It causes 40% slowdown in inference for larger models.

Trade-offs between model quality and inference efficiency are significant.

Abstract

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.