几何引导的视觉-语言混合专家负载均衡方法 ReBA
原标题：Relax Within, Balance Across: Geometry-Guided Load Balancing for Vision-Language Mixture-of-Experts
AI 摘要
Hugging Face 每日论文发布了一篇关于视觉-语言混合专家模型负载均衡的研究。论文提出 ReBA 方法，通过分别平衡图像和文本的负载、以图像为单位进行路由决策，解决了标准辅助损失在混合负载下可能隐藏模态特定不平衡的问题。实验表明，ReBA 在四个视觉-语言 MoE 骨干网络上降低了负载不均衡，同时保持了任务准确率。
正文节选
Relax Within, Balance Across: Geometry-Guided Load Balancing for Vision-Language Mixture-of-Experts Abstract Vision-language MoE batches contain different numbers of image and text tokens. Image resolution, image count, tiling, and prompt length all change this token mix. We call the standard token-level Switch auxiliary loss Std-Aux. Std-Aux balances only the mixed load, so large image and text load errors can cancel at one mix. On our main model, the same trained router shows more than a fivef