Xiaohongshu Teams Up with Peking University to Launch UltraEP: An Open-Source Load Balancing Solution for Large-Scale Mixture-of-Experts (MoE) Training and Inference
1 day ago / Read about 0 minute
Author:小编   

On July 20th, Xiaohongshu revealed in a blog post that it had joined forces with Peking University to develop UltraEP. This innovative solution marks the first instance of integrating a real-time load balancing technology, grounded in precise routing information, into production systems. UltraEP dynamically replicates hotspot experts—key components in the MoE architecture—at each microbatch and across every layer, thereby driving actual training and inference performance to near-optimal levels.

According to reports, UltraEP delivers an average performance that reaches 94.3% of the theoretical ideal, marking a 1.49-fold improvement over the current state-of-the-art (SOTA) frameworks in the industry. Moreover, it has already been successfully deployed in real-world production environments for large-scale pre-training tasks.