fix(ci): centralize GPU CI runtime pins

.github/actions/aicr-build/action.yml

@@ -32,6 +32,10 @@ inputs:
     description: 'Comma-separated validator phases to build (e.g., "conformance,deployment"), or "none" to skip all. Takes precedence over build_validators.'
     required: false
     default: ''
+  snapshot_agent_cuda_image:
+    description: 'CUDA base image for the snapshot agent image; defaults to .settings.yaml via load-versions'
+    required: false
+    default: ''
 
 runs:
   using: 'composite'
@@ -50,9 +54,16 @@ runs:
         GOFLAGS: -mod=vendor
       run: bash "${{ github.action_path }}/build-cli.sh"
 
+    - name: Load snapshot agent image version
+      if: inputs.build_snapshot_agent == 'true' && inputs.snapshot_agent_cuda_image == ''
+      id: snapshot-versions
+      uses: ./.github/actions/load-versions
+
     - name: Build snapshot agent image and load into kind
       if: inputs.build_snapshot_agent == 'true'
       shell: bash
+      env:
+        SNAPSHOT_AGENT_CUDA_IMAGE: ${{ inputs.snapshot_agent_cuda_image || steps.snapshot-versions.outputs.snapshot_agent_cuda_image }}
       run: bash "${{ github.action_path }}/build-snapshot-agent.sh"
 
     - name: Build validator images and load into kind

.github/actions/aicr-build/build-snapshot-agent.sh

@@ -15,14 +15,8 @@
 
 set -euo pipefail
 
-if ! command -v yq >/dev/null 2>&1; then
-  echo "::error::yq is required to read testing.snapshot_agent_cuda_image from .settings.yaml"
-  exit 1
-fi
-
-SNAPSHOT_AGENT_CUDA_IMAGE="$(yq eval '.testing.snapshot_agent_cuda_image // ""' .settings.yaml)"
-if [[ -z "${SNAPSHOT_AGENT_CUDA_IMAGE}" || "${SNAPSHOT_AGENT_CUDA_IMAGE}" == "null" ]]; then
-  echo "::error::testing.snapshot_agent_cuda_image must be set in .settings.yaml"
+if [[ -z "${SNAPSHOT_AGENT_CUDA_IMAGE:-}" || "${SNAPSHOT_AGENT_CUDA_IMAGE}" == "null" ]]; then
+  echo "::error::SNAPSHOT_AGENT_CUDA_IMAGE must be provided by the aicr-build action"
   exit 1
 fi
 

.github/actions/load-versions/action.yml

@@ -97,6 +97,12 @@ outputs:
   h100_kind_node_image:
     description: 'Kind node image for H100 GPU tests'
     value: ${{ steps.versions.outputs.h100_kind_node_image }}
+  gpu_operator_chart_version:
+    description: 'GPU Operator Helm chart version for GPU smoke tests'
+    value: ${{ steps.versions.outputs.gpu_operator_chart_version }}
+  snapshot_agent_cuda_image:
+    description: 'CUDA base image for the snapshot agent image'
+    value: ${{ steps.versions.outputs.snapshot_agent_cuda_image }}
 
 runs:
   using: 'composite'
@@ -149,6 +155,8 @@ runs:
         # Testing configuration
         echo "kind_node_image=$(yq eval '.testing.kind_node_image' .settings.yaml)" >> $GITHUB_OUTPUT
         echo "h100_kind_node_image=$(yq eval '.testing.h100_kind_node_image' .settings.yaml)" >> $GITHUB_OUTPUT
+        echo "gpu_operator_chart_version=$(yq eval '.testing.gpu_operator_chart_version' .settings.yaml)" >> $GITHUB_OUTPUT
+        echo "snapshot_agent_cuda_image=$(yq eval '.testing.snapshot_agent_cuda_image' .settings.yaml)" >> $GITHUB_OUTPUT
 
     - name: Display loaded versions
       shell: bash
@@ -182,3 +190,5 @@ runs:
         echo "  test_timeout: ${{ steps.versions.outputs.test_timeout }}"
         echo "  kind_node_image: ${{ steps.versions.outputs.kind_node_image }}"
         echo "  h100_kind_node_image: ${{ steps.versions.outputs.h100_kind_node_image }}"
+        echo "  gpu_operator_chart_version: ${{ steps.versions.outputs.gpu_operator_chart_version }}"
+        echo "  snapshot_agent_cuda_image: ${{ steps.versions.outputs.snapshot_agent_cuda_image }}"

.github/actions/runtime-install/action.yml

@@ -39,6 +39,10 @@ inputs:
     description: 'Continue deploying remaining bundle components after a component failure'
     required: false
     default: 'true'
+  gpu_operator_chart_version:
+    description: 'GPU Operator Helm chart version for helm mode; defaults to .settings.yaml via load-versions'
+    required: false
+    default: ''
 
 runs:
   using: 'composite'
@@ -58,9 +62,15 @@ runs:
 
     # --- Helm mode: standalone GPU operator chart ---
 
+    - name: Load GPU Operator chart version
+      if: inputs.method == 'helm' && inputs.gpu_operator_chart_version == ''
+      id: helm-versions
+      uses: ./.github/actions/load-versions
     - name: Install GPU Operator (helm)
       if: inputs.method == 'helm'
       shell: bash
+      env:
+        GPU_OPERATOR_CHART_VERSION: ${{ inputs.gpu_operator_chart_version || steps.helm-versions.outputs.gpu_operator_chart_version }}
       run: bash "${{ github.action_path }}/install-gpu-operator-helm.sh"
     - name: Wait for GPU operands (helm)
       if: inputs.method == 'helm'

.github/actions/runtime-install/install-gpu-operator-helm.sh

@@ -15,14 +15,8 @@
 
 set -euo pipefail
 
-if ! command -v yq >/dev/null 2>&1; then
-  echo "::error::yq is required to read testing.gpu_operator_chart_version from .settings.yaml"
-  exit 1
-fi
-
-GPU_OPERATOR_CHART_VERSION="$(yq eval '.testing.gpu_operator_chart_version // ""' .settings.yaml)"
-if [[ -z "${GPU_OPERATOR_CHART_VERSION}" || "${GPU_OPERATOR_CHART_VERSION}" == "null" ]]; then
-  echo "::error::testing.gpu_operator_chart_version must be set in .settings.yaml"
+if [[ -z "${GPU_OPERATOR_CHART_VERSION:-}" || "${GPU_OPERATOR_CHART_VERSION}" == "null" ]]; then
+  echo "::error::GPU_OPERATOR_CHART_VERSION must be provided by the runtime-install action"
   exit 1
 fi
 

.settings.yaml

@@ -73,6 +73,8 @@ docs_tools:
 testing:
   kind_node_image: 'kindest/node:v1.32.0'
   h100_kind_node_image: 'kindest/node:v1.35.0'
+
+  # GPU CI runtime pins consumed through .github/actions/load-versions.
   gpu_operator_chart_version: 'v25.10.1'
   snapshot_agent_cuda_image: 'nvcr.io/nvidia/cuda:13.1.0-base-ubuntu24.04'