Volcano

主な変更 (7)

• Network-Topology-Aware scheduling soft mode corrected for jobs and subjobs
• Scalar in-queue resource accounting now uses milli-units, fixing precision errors in resource calculations
• HAMi vGPU scheduling failures in medium and large clusters resolved
• Ascend vNPU health check switched to Allocatable resources instead of the previous method
• Preemption now reprievess higher-priority pods first, restoring correct ordering
• job.Status.Conditions unbounded growth prevented, reducing memory and etcd pressure on long-running jobs
• Several smaller scheduler fixes: device annotation cleanup on pod release, minAvailable fallback to replicas when minPartitions is omitted, ImageStates restored in node snapshots, and maxFloat/maxInt display corrections