Tag
Ai2 分享其 GPU 调度器的工程实践，通过 GPU 时间预算、分层公平共享分配和时间片机制取代优先级调度器，将其最大 H100 集群的中位排队等待时间从 5 分钟降至 24 秒。
Ai2's AI Infrastructure team details replacing their priority-based GPU scheduler with a system using GPU time budgets, hierarchical fair-share allocation, and time-slicing contracts to prioritize high-impact research while maintaining full cluster occupancy across thousands of H100, B200, and B300 GPUs.
AI21 published a case study on how it replaced manual GPU allocation in internal chat channels with Kueue on GKE to manage a shared fleet of ~10,000 GPUs across teams. The collaboration with Google's Kueue team drove new features such as Admission Fair Sharing (AFS), delivering an 83% reduction in time-to-start for high-priority workloads.
This tweet comments on the limitations of multi-agent systems, emphasizing that using them in unsuitable scenarios can lead to negative effects, and points out that agent clusters require governance and design.