Load balancing is a core bottleneck in MoE training, and DeepSeek's aux-loss-free method has been the de facto reference since it shipped. A technique claiming better results at the same throughput could change which balancing default new MoE implementations adopt. Researchers and engineers training sparse models should review the paper's routing metrics before their next run.
ID Balancing provides a mathematically elegant solution to a critical bottleneck in MoE scaling, offering stronger error correction without auxiliary losses. Researchers and engineers working on large-scale model training can implement this controller to improve stability and efficiency in increasingly sparse MoE architectures, particularly as models grow beyond 70B parameters.
微软要在SIGCOMM上晒网络研究，有负载均衡新招和AI模拟网络，搞网络的快去看看。
如果你在做MoE推理，TEMPO这个调度器在内存/算力瓶颈混合时能比常规方案快15.5%；不过DeepSeek-V3那种通信瓶颈场景它不灵。
Anthropic 更新了 MCP 协议，去掉了会话绑定，现在 MCP 服务器能上云水平扩展了，serverless 也能跑，做工具链的可以试。
MCP 终于去掉会话 ID 了，部署起来像普通 HTTP 一样灵活。还有 MRTR 解决中途确认问题，很实用。如果你在生产环境用 MCP，得仔细看迁移指南。
SGLang给DeepSeek模型加了两个新负载均衡器，跑DeepSeek V3/R1速度能快最多7%，而且不改精度，想加速推理的可以试试。
MoE推理的通信瓶颈终于有了任务感知的解法——做多任务MoE部署的团队可以直接参考TACG的静态分组+GESR动态复制策略，实测通信成本降三成，公平性几乎无损。
想搞懂 MoE 为什么又快又强，这篇视觉解释把路由和负载均衡的坑讲透了，做模型训练或推理优化的开发者值得一看。
做MoE模型训练或路由优化的研究者，这篇论文解释了路由崩溃和负载均衡损失的底层机制，看完会对如何设计更有效的路由策略有直接启发。
