kai-scheduler · davidLif · Aug 10, 2026
diff --git a/.changes/unreleased/added-20260810-112731.yaml b/.changes/unreleased/added-20260810-112731.yaml
@@ -0,0 +1,3 @@
+kind: Added
+body: |-
+  Install gpu-sharing as an OCI subchart
diff --git a/.github/workflows/on-pr.yaml b/.github/workflows/on-pr.yaml
@@ -233,6 +233,7 @@ jobs:
 
       - name: Build helm chart
         run: |
+          helm dependency build ./deployments/kai-scheduler
           helm package ./deployments/kai-scheduler -d ./charts --app-version $PACKAGE_VERSION --version $PACKAGE_VERSION
           cp charts/kai-scheduler-$PACKAGE_VERSION.tgz /mnt/images/
 

diff --git a/.github/workflows/on-release.yaml b/.github/workflows/on-release.yaml
@@ -63,6 +63,7 @@ jobs:
       - name: Build helm chart
         run: |
           sed -i 's#registry/local/kai-scheduler#${{ env.DOCKER_REGISTRY }}#' deployments/kai-scheduler/values.yaml
+          helm dependency build ./deployments/kai-scheduler
           helm package ./deployments/kai-scheduler -d ./charts --app-version $PACKAGE_VERSION --version $PACKAGE_VERSION
           cp charts/kai-scheduler-$PACKAGE_VERSION.tgz /mnt/images/
 

diff --git a/.github/workflows/push-artifacts.yaml b/.github/workflows/push-artifacts.yaml
@@ -59,6 +59,7 @@ jobs:
       - name: Build helm chart
         run: |
           sed -i 's#registry/local/kai-scheduler#${{ env.DOCKER_REGISTRY }}#' deployments/kai-scheduler/values.yaml
+          helm dependency build ./deployments/kai-scheduler
           helm package ./deployments/kai-scheduler -d ./charts --app-version $PACKAGE_VERSION --version $PACKAGE_VERSION
 
       - name: Push Helm Chart

diff --git a/Makefile b/Makefile
@@ -27,8 +27,13 @@ K8S_COPYRIGHTED_MANIFEST_FILES := deployments/kai-scheduler/crds/kai.scheduler_t
 lint: fmt-go vet-go lint-go
 .PHONY: lint
 
+.PHONY: chart-deps
+chart-deps:
+	@echo "Fetching Helm chart dependencies: kai-scheduler"
+	helm dependency build ./deployments/kai-scheduler
+
 .PHONY: test-chart
-test-chart:
+test-chart: chart-deps
 	@echo "Running tests for Helm chart: kai-scheduler"
 	docker run -t --rm -v ./deployments/kai-scheduler:/apps helmunittest/helm-unittest:3.17.2-0.8.1 . -f 'tests/**/*_test.yaml'
 

diff --git a/deployments/kai-scheduler/Chart.lock b/deployments/kai-scheduler/Chart.lock
@@ -0,0 +1,6 @@
+dependencies:
+- name: gpu-sharing
+  repository: oci://ghcr.io/kai-scheduler/gpu-sharing
+  version: v0.1.0-alpha.1
+digest: sha256:c3d7f76c6d2b2fd7b256b7450fb435a82693e8acef1265d10a8d5270fcb028db
+generated: "2026-08-02T20:53:54.374149+03:00"
diff --git a/deployments/kai-scheduler/Chart.yaml b/deployments/kai-scheduler/Chart.yaml
@@ -6,4 +6,10 @@ name: kai-scheduler
 description: KAI Scheduler by NVIDIA
 type: application
 version: 0.0.0
-appVersion: 0.0.0
+appVersion: 0.0.0
+
+dependencies:
+  - name: gpu-sharing
+    version: v0.1.0-alpha.1
+    repository: oci://ghcr.io/kai-scheduler/gpu-sharing
+    condition: global.nvFractions.set
diff --git a/deployments/kai-scheduler/templates/_helpers.tpl b/deployments/kai-scheduler/templates/_helpers.tpl
@@ -112,10 +112,14 @@ spec:
     {{- if .Values.podgrouper.queueLabelKey }}
     queueLabelKey: {{ .Values.podgrouper.queueLabelKey | quote }}
     {{- end }}
-    {{- /* GpuSharingMode precedence: an explicit (non-empty) mode wins; otherwise a
-           legacy gpuSharing value is passed through (operator derives the mode); when
-           neither is set, default to NonMemoryEnforced. */}}
-    {{- if .Values.global.gpuSharingMode }}
+    {{- /* GpuSharingMode precedence: nvFractions.set forces NvFractions (validate.yaml has
+           already rejected a conflicting explicit mode, so this only guards against a legacy
+           gpuSharing value downgrading it); then an explicit (non-empty) mode; then a legacy
+           gpuSharing value is passed through (operator derives the mode); when none is set,
+           default to NonMemoryEnforced. */}}
+    {{- if .Values.global.nvFractions.set }}
+    gpuSharingMode: NvFractions
+    {{- else if .Values.global.gpuSharingMode }}
     gpuSharingMode: {{ .Values.global.gpuSharingMode }}
     {{- else if not (hasKey .Values.global "gpuSharing") }}
     gpuSharingMode: NonMemoryEnforced
@@ -264,7 +268,7 @@ spec:
         maxUnavailable: {{ .Values.admission.podDisruptionBudget.maxUnavailable }}
         {{- end }}
       {{- end }}
-    {{- if and (not .Values.global.gpuSharingMode) (hasKey .Values.global "gpuSharing") }}
+    {{- if and (not .Values.global.nvFractions.set) (not .Values.global.gpuSharingMode) (hasKey .Values.global "gpuSharing") }}
     gpuSharing: {{ .Values.global.gpuSharing }}
     {{- end }}
     blockNvidiaVisibleDevices: {{ .Values.global.blockNvidiaVisibleDevices | default false }}

diff --git a/deployments/kai-scheduler/templates/validate.yaml b/deployments/kai-scheduler/templates/validate.yaml
@@ -0,0 +1,9 @@
+{{/*
+Copyright 2025 NVIDIA CORPORATION
+SPDX-License-Identifier: Apache-2.0
+*/}}
+{{- /* Renders no resource. Aborts the render when nvFractions.set asks for the gpu-sharing
+       operator while gpuSharingMode explicitly selects a different mode. */}}
+{{- if and .Values.global.nvFractions.set .Values.global.gpuSharingMode (ne .Values.global.gpuSharingMode "NvFractions") }}
+{{- fail (printf "global.nvFractions.set=true requires global.gpuSharingMode to be empty or \"NvFractions\", got %q" .Values.global.gpuSharingMode) }}
+{{- end }}
diff --git a/deployments/kai-scheduler/tests/nv_fractions_test.yaml b/deployments/kai-scheduler/tests/nv_fractions_test.yaml
@@ -0,0 +1,108 @@
+# Copyright 2025 NVIDIA CORPORATION
+# SPDX-License-Identifier: Apache-2.0
+
+suite: test nvFractions subchart gating and validation
+
+tests:
+  - it: rejects nvFractions.set=true with a conflicting gpuSharingMode
+    templates:
+      - templates/validate.yaml
+    set:
+      global.nvFractions.set: true
+      global.gpuSharingMode: HamiCore
+    asserts:
+      - failedTemplate:
+          errorMessage: 'global.nvFractions.set=true requires global.gpuSharingMode to be empty or "NvFractions", got "HamiCore"'
+
+  - it: allows nvFractions.set=true with an explicit NvFractions mode
+    templates:
+      - templates/validate.yaml
+    set:
+      global.nvFractions.set: true
+      global.gpuSharingMode: NvFractions
+    asserts:
+      - notFailedTemplate: {}
+
+  - it: allows nvFractions.set=true with an empty gpuSharingMode
+    templates:
+      - templates/validate.yaml
+    set:
+      global.nvFractions.set: true
+    asserts:
+      - notFailedTemplate: {}
+
+  - it: allows gpuSharingMode NvFractions while the subchart is disabled
+    templates:
+      - templates/validate.yaml
+    set:
+      global.gpuSharingMode: NvFractions
+    asserts:
+      - notFailedTemplate: {}
+
+  - it: nvFractions.set=true selects NvFractions and omits legacy gpuSharing
+    templates:
+      - kai-config.yaml
+    set:
+      kaiConfigDeployer.enabled: false
+      kaiConfig.render: true
+      global.nvFractions.set: true
+    asserts:
+      - equal:
+          path: spec.global.gpuSharingMode
+          value: NvFractions
+      - notExists:
+          path: spec.admission.gpuSharing
+
+  - it: nvFractions.set=true is not downgraded by a legacy gpuSharing value
+    templates:
+      - kai-config.yaml
+    set:
+      kaiConfigDeployer.enabled: false
+      kaiConfig.render: true
+      global.nvFractions.set: true
+      global.gpuSharing: false
+    asserts:
+      - equal:
+          path: spec.global.gpuSharingMode
+          value: NvFractions
+      - notExists:
+          path: spec.admission.gpuSharing
+
+  - it: nvFractions.set=true with an explicit NvFractions mode agrees
+    templates:
+      - kai-config.yaml
+    set:
+      kaiConfigDeployer.enabled: false
+      kaiConfig.render: true
+      global.nvFractions.set: true
+      global.gpuSharingMode: NvFractions
+    asserts:
+      - equal:
+          path: spec.global.gpuSharingMode
+          value: NvFractions
+
+  - it: default is unchanged when nvFractions.set is false
+    templates:
+      - kai-config.yaml
+    set:
+      kaiConfigDeployer.enabled: false
+      kaiConfig.render: true
+    asserts:
+      - equal:
+          path: spec.global.gpuSharingMode
+          value: NonMemoryEnforced
+
+  # No mirror-image test asserting the deployment is absent when nvFractions.set is false:
+  # helm-unittest treats a conditioned-out subchart template as "not exists or not selected"
+  # and errors instead of rendering zero documents. The off state is covered by Helm's
+  # condition: semantics and by the NonMemoryEnforced default asserted above.
+  - it: renders the gpu-sharing operator deployment when nvFractions.set is true
+    templates:
+      - charts/gpu-sharing/templates/deployment.yaml
+    set:
+      global.nvFractions.set: true
+    asserts:
+      - hasDocuments:
+          count: 1
+      - isKind:
+          of: Deployment
diff --git a/deployments/kai-scheduler/values.yaml b/deployments/kai-scheduler/values.yaml
@@ -25,6 +25,14 @@ global:
   # Deprecated: gpuSharing is honored only for upgrades when gpuSharingMode is empty
   # (true -> NonMemoryEnforced, false -> Disabled). Prefer gpuSharingMode.
   # gpuSharing: false
+  # nvFractions.set installs the gpu-sharing operator
+  # (oci://ghcr.io/kai-scheduler/gpu-sharing) as a subchart and selects the NvFractions
+  # GPU sharing mode. Setting it together with a global.gpuSharingMode other than
+  # NvFractions is rejected at render time.
+  # KAI's global.imagePullSecrets/nodeSelector/affinity/tolerations are NOT propagated to
+  # the subchart; override them under the top-level `gpu-sharing:` key.
+  nvFractions:
+    set: false
   blockNvidiaVisibleDevices: false
   clusterAutoscaling: false
   nodeSelector: {}

diff --git a/docs/operator/README.md b/docs/operator/README.md
@@ -82,6 +82,32 @@ The KAI operator reports installation health on `Config.status.conditions`. The
 
 When `global.gpuSharingMode` is `NvFractions`, the KAI operator also checks the cluster-scoped `GpuSharingConfig` resource from the GPU-sharing operator, as running fractional GPU pods depends on it.
 
+### Installing the GPU-sharing operator
+
+The GPU-sharing operator is required only for `NvFractions`; every other `gpuSharingMode` runs without it. It also raises the cluster's requirements: [NVIDIA GPU Operator](https://github.com/NVIDIA/gpu-operator) **v26.7.1 or newer** and an **`r615` or newer driver (CUDA 13.4)** on the GPU nodes. `r615` is not the GPU Operator default, so it must be requested explicitly via `driver.version`; on an older driver the `GpuSharingConfig` `Ready` condition reports `GPUDriverVersionUnsupported` and the node-level daemons are not rolled out. See the [gpu-sharing prerequisites](https://github.com/kai-scheduler/gpu-sharing#prerequisites) for the full list, including containerd 2.0+ with NRI enabled.
+
+Set `global.nvFractions.set=true` to install the [gpu-sharing](https://github.com/kai-scheduler/gpu-sharing) operator as a subchart and select the `NvFractions` mode in one step:
+
+```bash
+helm upgrade -i kai-scheduler oci://ghcr.io/kai-scheduler/kai-scheduler/kai-scheduler \
+  -n kai-scheduler --create-namespace --set global.nvFractions.set=true
+```
+
+Leaving `global.gpuSharingMode` empty is enough — `nvFractions.set` selects `NvFractions`. Setting it to any other mode at the same time fails the install at render time. To run `NvFractions` against a GPU-sharing operator installed out-of-band, leave `global.nvFractions.set` at its `false` default and set `global.gpuSharingMode: NvFractions` instead.
+
+KAI's `global.imagePullSecrets`, `global.nodeSelector`, `global.affinity` and `global.tolerations` are **not** propagated to the subchart, because the gpu-sharing chart reads those keys at its own top level rather than under `global`. Set them explicitly under the `gpu-sharing:` key — relevant for private-registry and air-gapped installs:
+
+```yaml
+global:
+  nvFractions:
+    set: true
+gpu-sharing:
+  imagePullSecrets:
+    - name: my-registry-secret
+  nodeSelector:
+    nvidia.com/gpu.present: "true"
+```
+
 If all KAI services are deployed and available, but the GPU-sharing operator reports a dependency failure, the KAI `Config` can look like this:
 
 ```yaml

diff --git a/hack/run-e2e-upgrade-kind.sh b/hack/run-e2e-upgrade-kind.sh
@@ -138,6 +138,7 @@ if [ "$LOCAL_IMAGES_BUILD" = "true" ]; then
 fi
 
 # Package the new helm chart
+helm dependency build ./deployments/kai-scheduler
 helm package ./deployments/kai-scheduler -d ./charts --app-version $PACKAGE_VERSION --version $PACKAGE_VERSION
 export UPGRADE_CHART_PATH=${REPO_ROOT}/charts/kai-scheduler-$PACKAGE_VERSION.tgz
 

diff --git a/hack/setup-e2e-cluster.sh b/hack/setup-e2e-cluster.sh
@@ -198,6 +198,7 @@ if [ "$LOCAL_IMAGES_BUILD" = "true" ]; then
     fi
 
     # Package helm chart
+    helm dependency build ./deployments/kai-scheduler
     helm package ./deployments/kai-scheduler -d ./charts --app-version $PACKAGE_VERSION --version $PACKAGE_VERSION
     if [ "$SKIP_KAI_INSTALL" = "true" ]; then
         echo "Skipping KAI install; packaged chart kept at ./charts/kai-scheduler-$PACKAGE_VERSION.tgz"