Skip to content

v0.18.0

Choose a tag to compare

View all tags
@KaiPilotBot KaiPilotBot released this 23 Sep 13:14
· 60 commits to main since this release
v0.18.0
2df9a1f
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Added

  • Improve simulation performance by caching equivalent node-affinity constraints
  • Persist fractional GPU compute mode #1316 davidLif
  • In-place pod resize accounting (KEP-1287) and queue admission webhook
  • Validate NvFractions GPU memory annotations #1316 davidLif
  • Add gpujoborder plugin for configurable GPU-based JobOrderFn tiebreak on priority ties
  • Configurable admission Service name preserves webhook DNS and certificate identity during migrations
  • Default memory limits and CPU/memory requests for the operator and chart hook jobs
  • Background pods plugin: ignore maintenance pods when scheduling, evict them when displaced
  • Semi-preemptible mode (alpha): minimal shape is non-preemptible and in-quota, surplus is elastic; gang staleness now respects minSubGroup
  • Schedule NvFractions GPU memory requests #1316 davidLif
  • Bind NvFractions GPU memory devices #1316 davidLif
  • Configure NvFractions GPU sharing mode #1316 davidLif
  • Gate NvFractions on GPU sharing readiness #1316 davidLif
  • Add typed fractional GPU group API #1316 davidLif
  • Enforce fractional GPU compute modes #1316 davidLif
  • Config to group all pods of a deployment under a single podgroup. From now on, this will be the grouping behavior for deployments by default. #2109 davidLif
  • Install gpu-sharing as an OCI subchart #1316 davidLif
  • Support kai.scheduler gpu-memory.portion.limit annotation for per-container GPU memory limits #1316 davidLif
  • Add fipsMode=only to allow the enforcement of GODEBUG=fips140=only at runtime on all KAI containers
  • Add queuePriorityInQuotaReclaim strategy for reclaim. If it's enabled (off by default) higher priority queues can reclaim in-quota resources from lower priority queues. This is alpha functionality, and the exact behavior of this feature might change in future releases.
  • Publish images.yaml manifest with per-platform image digests as a release asset
  • Skill to diagnose why a pod is pending #1731 somanythingstodo
  • Added support for configuring binder Pod Disruption Budget via Helm values (binder.podDisruptionBudget) when running multiple replicas #1477 dttung2905
  • Added support for configuring queue-controller Pod Disruption Budget via Helm values (queuecontroller.podDisruptionBudget) when running multiple replicas #1477 dttung2905

Changed

  • Images use scratch as base instead of distroless; resource-reservation stays on distroless for NVML
  • Renamed Helm value global.fips to global.fipsMode with on/off options #1906
  • Queue allocated status now includes init-container peak and sidecar requests
  • Renamed gpu-compute-sharing-mode annotation to per-container nvidia.com/container..gpu-compute.mode with validation #1316 davidLif
  • Hook image crd-upgrader replaced by helm-hooks: a Go binary on distroless, no shell or kubectl

Fixed

  • Allow the post-delete cleanup job to run on OpenShift
  • Allow PodGroup minMember and minSubGroup of 0 for workloads with no gang requirement
  • Propagate Kubernetes client QPS and burst settings to operator-managed controllers
  • Preserve reservation pods during BindRequest cache lag #1316 davidLif
  • Scheduler and binder now fail fast when Kubernetes API discovery fails at startup, instead of silently disabling dynamic resource allocation and node resource topology for the lifetime of the process. #2038 hcnguyen-ai
  • Podgrouper skips WorkloadRunner wrapper so wrapped workloads keep their gang grouping
  • Respect hierarchical gang floors during allocation and stale-gang eviction
  • Isolate same-named PodGroups across Kubernetes namespaces
  • DRA allocator now honors PartitionableDevices/ConsumableCapacity feature gates instead of hardcoded empty Features
  • Add resource-reservation service account to OpenShift SecurityContextConstraints
  • Disable TLS ML-KEM hybrid curve so fipsMode=only containers can complete TLS handshakes. See golang/go#78298, kubernetes/kubernetes#133743
  • Releasing pods no longer block required inter-pod anti-affinity, so reclaim and preempt can place such pods
  • Prevent the scheduler from getting stuck when a pod cannot be marked as pipelined.
  • Bug fix: consistent resourceList parsing for resourceInfo and node vector
  • Do not grant the GPU-sharing node score to pods that request no GPU #2154 universome
  • Binder hamicore plugin fails closed instead of admitting pods uncapped when GPU memory limit calc fails
  • Topology migration hook now creates KAI Topologies with the served v1alpha1 API version
  • Queue quota validation sums sibling GPU and memory against the parent and treats -1 as unlimited
  • Reduce scheduler allocations from disabled verbose logs
  • Scheduler status-updater no longer keeps re-applying a cached PodGroup status update after another writer (e.g. pod-group-assigner clearing scheduling conditions) modified the PodGroup. Previously the stale condition was overlaid on every snapshot, so the PodGroup was skipped as "not assigned to current scheduler" until the scheduler restarted.
  • Improve greedy GPU matching performance by avoiding repeated scans
  • Apply fractional GPU compute-mode compatibility consistently during fit and ordering
  • Count a GPU shared by multiple pods through one DRA ResourceClaim once per node, preventing negative idle GPUs. #1930 TensorRaya
  • stalegangeviction no longer evicts remaining pods of a gang whose pods complete successfully #1968 Thezone-1
  • tighten configmap verb for admission, nodescaleadjuster, podgroupcontroller, podgrouper, queuecontroller, scheduler ClusterRole #2060 dttung2905
  • Render global.daemonsetsTolerations into kai-config dttung2905
  • Honor postCleanup.serviceAccountName for the post-delete ServiceAccount and ClusterRoleBinding subject dttung2905
Assets 4
Loading