Repository Issue Activity (beta)

#1966 preempt: JobOrderFn tiebreak has no resource-awareness, falls back to creation-time order

#1944 Knative scale-to-zero causes PodGroup creation failure when Knative gang scheduling is enabled

#1913 feat(scheduler): allow jobs to override stale gang eviction grace period

#1968 stalegangeviction evicts pods of successfully completing jobs

#1967 [Bug] CUDA_DEVICE_MEMORY_LIMIT is not set by HAMI Core plugin on UMA nodes where nvidia.com/gpu.memory label is missing

#1827 perf(scheduler): bound per-task per-node fit-error retention during allocation

#1964 Segmented PyTorch grouper ignores elastic minReplicas and requires all worker segments

#1927 podgrouper: segmented PyTorch and LWS emit invalid minMember on parent SubGroups

#1756 [Numa awareness] Support best effort topology manager

#1945 status-updater floods logs and API server retrying status/patch updates for deleted PodGroups

#1948 Pipeline-only allocation leaks speculative fit errors

#1888 PodGroup schedulingConditions not cleared after the workload schedules

#1929 docs/metrics: inconsistent naming between queue_deserved_gpus and queue_quota_* metrics

#1933 How to schedule a pod onto a cordoned node?

#1936 Set annotation to ignore quota occupied by workload

#1939 reclaim: FeasibleNodesForJob reuses the reclaimer's node set to re-home victims, needlessly killing relocatable CPU-only victims

#1873 DRA GPU count overflow can understate queue demand and prevent eligible GPU reclaim

#1584 Design kai behavior for priority class preemptionPolicy

#848 Scheduler Assigns Multiple Workloads to Fully-Allocated GPU Node Causing Infinite Retry Loop

#1821 feat: Container level VRAM metrics - integration with HAMi

#1934 Can reclaim enforce fair-share (not just deserved quota) between sibling queues?

#1930 shared DRA ResourceClaim double-counted per pod makes node unschedulable

#1856 NUMA scale tests

#1912 Cannot find parameter scheduler.gpuSharing.hamicoreEnabled in values.yaml file

#1923 queue validator: validateParentChildQuota only sums CPU across siblings