ci: convert framework image pipelines to single multi-arch manifest builds (#7399)

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

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

@@ -8,7 +8,7 @@ inputs:
     description: 'Target stage for Docker rendering'
     required: true
   platform:
-    description: 'Platform to build (amd64 or arm64)'
+    description: 'Docker platform string (e.g. linux/amd64, linux/amd64,linux/arm64)'
     required: true
   cuda_version:
     description: 'CUDA version to build (e.g., 12.9, 13.0)'
@@ -68,15 +68,7 @@ inputs:
     required: false
     default: 'false'
   extra_tags:
-    description: 'Additional tags (newline-separated, -$platform suffix auto-appended)'
-    required: false
-    default: ''
-  build_only:
-    description: 'Build and push only — controls extra tag logic'
-    required: false
-    default: 'false'
-  sanitized_ref_name:
-    description: 'Sanitized git ref name for branch-tagged images'
+    description: 'Additional tags (newline-separated)'
     required: false
     default: ''
   show_summary:
@@ -121,8 +113,8 @@ runs:
         fi
         TARGET_TAG_PLAIN="${{ github.sha }}-${{ inputs.framework }}-${{ inputs.target }}${EFA_SUFFIX}"
         TEST_TAG_PLAIN="${{ github.sha }}-${{ inputs.framework }}-${{ inputs.target }}${EFA_SUFFIX}-test"
-        DEFAULT_TARGET_IMAGE_URI="${{ inputs.aws_account_id }}.dkr.ecr.${{ inputs.aws_default_region }}.amazonaws.com/ai-dynamo/dynamo:${TARGET_TAG_PLAIN}-cuda${CUDA_VERSION}-${{ inputs.platform }}"
-        TEST_IMAGE_URI="${{ inputs.aws_account_id }}.dkr.ecr.${{ inputs.aws_default_region }}.amazonaws.com/ai-dynamo/dynamo:${TEST_TAG_PLAIN}-cuda${CUDA_VERSION}-${{ inputs.platform }}"
+        DEFAULT_TARGET_IMAGE_URI="${{ inputs.aws_account_id }}.dkr.ecr.${{ inputs.aws_default_region }}.amazonaws.com/ai-dynamo/dynamo:${TARGET_TAG_PLAIN}-cuda${CUDA_VERSION}"
+        TEST_IMAGE_URI="${{ inputs.aws_account_id }}.dkr.ecr.${{ inputs.aws_default_region }}.amazonaws.com/ai-dynamo/dynamo:${TEST_TAG_PLAIN}-cuda${CUDA_VERSION}"
         echo "default_target_image_uri=${DEFAULT_TARGET_IMAGE_URI}" >> $GITHUB_OUTPUT
         echo "test_image_uri=${TEST_IMAGE_URI}" >> $GITHUB_OUTPUT
         echo "target_tag_plain=${TARGET_TAG_PLAIN}" >> $GITHUB_OUTPUT
@@ -135,7 +127,7 @@ runs:
         flavor: ${{ inputs.framework }}
         arch: ${{ inputs.platform }}
         cuda_version: ${{ inputs.cuda_version }}
-    - name: Calculate extra tags with platform suffix
+    - name: Calculate extra tags
       id: extra-tags
       shell: bash
       env:
@@ -149,20 +141,10 @@ runs:
         if [ -n "$EXTRA_TAGS" ]; then
           while IFS= read -r tag; do
             if [ -n "$tag" ]; then
-              RESULT+="${ECR_REGISTRY}/ai-dynamo/dynamo:${tag}-cuda${CUDA_VERSION_MAJOR}-${{ inputs.platform }}"$'\n'
+              RESULT+="${ECR_REGISTRY}/ai-dynamo/dynamo:${tag}-cuda${CUDA_VERSION_MAJOR}"$'\n'
             fi
           done <<< "$EXTRA_TAGS"
         fi
-        if [ "${{ inputs.build_only }}" == "true" ]; then
-          if [ -z "${{ inputs.sanitized_ref_name }}" ]; then
-            echo "::warning::sanitized_ref_name is empty but build_only is true; skipping branch tags"
-          else
-            BRANCH_TAG="${{ inputs.sanitized_ref_name }}-${{ inputs.framework }}"
-            RESULT+="${ECR_REGISTRY}/ai-dynamo/dynamo:${BRANCH_TAG}-cuda${CUDA_VERSION_MAJOR}-${{ inputs.platform }}"$'\n'
-            RESULT+="${ACR_REGISTRY}/ai-dynamo/dynamo:${BRANCH_TAG}-cuda${CUDA_VERSION_MAJOR}-${{ inputs.platform }}"$'\n'
-            RESULT+="${ACR_REGISTRY}/ai-dynamo/dynamo:${{ steps.calculate-target-tag.outputs.target_tag_plain }}-cuda${CUDA_VERSION_MAJOR}-${{ inputs.platform }}"$'\n'
-          fi
-        fi
         if [ -n "$RESULT" ]; then
           echo "tags<<EOF" >> $GITHUB_OUTPUT
           echo "$RESULT" >> $GITHUB_OUTPUT
@@ -236,7 +218,7 @@ runs:
         ECR_HOSTNAME: ${{ inputs.aws_account_id }}.dkr.ecr.${{ inputs.aws_default_region }}.amazonaws.com
       run: |
         CUDA_MAJOR=${{ steps.calculate-target-tag.outputs.cuda_version_plain }}
-        CACHE_TAG="test-${{ inputs.framework }}-cuda${CUDA_MAJOR}-${{ inputs.platform }}-cache"
+        CACHE_TAG="test-${{ inputs.framework }}-cuda${CUDA_MAJOR}-cache"
         CACHE_ARGS="--cache-from type=registry,ref=${ECR_HOSTNAME}/ai-dynamo/dynamo:main-${CACHE_TAG}"
         CACHE_ARGS+=" --cache-from type=registry,ref=${ECR_HOSTNAME}/ai-dynamo/dynamo:release-${CACHE_TAG}"
         if [[ "$GITHUB_REF_NAME" == "main" ]]; then
@@ -261,7 +243,7 @@ runs:
           --progress=plain \
           ${PUSH_ARGS} \
           ${NO_CACHE_ARG} \
-          --platform linux/${{ inputs.platform }} \
+          --platform ${{ inputs.platform }} \
           -f container/Dockerfile.test \
           --build-arg BASE_IMAGE=${{ steps.calculate-target-tag.outputs.default_target_image_uri }} \
           ${CACHE_ARGS} \
@@ -270,7 +252,7 @@ runs:
       shell: bash
       if: ${{ inputs.push_image == 'true' && inputs.show_summary == 'true' }}
       run: |
-        echo "### 🐳 ${{ inputs.framework }}-cuda${{ inputs.cuda_version }}-${{ inputs.platform }} Default Image" >> $GITHUB_STEP_SUMMARY
+        echo "### 🐳 ${{ inputs.framework }}-cuda${{ inputs.cuda_version }} Default Image" >> $GITHUB_STEP_SUMMARY
         echo "" >> $GITHUB_STEP_SUMMARY
         echo "| Image URI |" >> $GITHUB_STEP_SUMMARY
         echo "|-----|" >> $GITHUB_STEP_SUMMARY

.github/actions/builder-refresher/action.yml

@@ -22,17 +22,13 @@ inputs:
     description: 'Buildkit flavor used to route workers on re-init (vllm, sglang, trtllm, general)'
     required: true
   arch:
-    description: 'Target architecture used on re-init (amd64, arm64)'
+    description: 'Docker platform string used on re-init (e.g. linux/amd64, linux/arm64, linux/amd64,linux/arm64)'
     required: false
-    default: 'amd64'
+    default: 'linux/amd64'
   cuda_version:
     description: 'CUDA version used on re-init (e.g. 12.9, 13.0). Leave empty for general flavor.'
     required: false
     default: ''
-  all_arch:
-    description: 'If true, re-initialize builder for both amd64 and arm64 architectures'
-    required: false
-    default: 'false'
 
   # Kubernetes fallback passthrough inputs (forwarded to init-dynamo-builder)
   ephemeral_storage:
@@ -90,7 +86,6 @@ runs:
         builder_name: ${{ inputs.builder_name }}
         flavor: ${{ inputs.flavor }}
         arch: ${{ inputs.arch }}
-        all_arch: ${{ inputs.all_arch }}
         cuda_version: ${{ inputs.cuda_version }}
         ephemeral_storage: ${{ inputs.ephemeral_storage }}
         namespace: ${{ inputs.namespace }}

.github/actions/compliance-scan/action.yml

@@ -12,9 +12,9 @@ inputs:
     description: 'Name for the uploaded artifact (e.g., compliance-vllm-cuda12-amd64)'
     required: true
   arch:
-    description: 'Target architecture (amd64, arm64)'
+    description: 'Docker platform string (e.g. linux/amd64, linux/arm64)'
     required: false
-    default: 'amd64'
+    default: 'linux/amd64'
   framework:
     description: 'Framework name for base image resolution (vllm, sglang, trtllm, dynamo)'
     required: false
@@ -78,7 +78,7 @@ runs:
         mkdir -p /tmp/compliance-target
         docker buildx build \
           --builder ${{ steps.builder.outputs.name }} \
-          --platform linux/${{ inputs.arch }} \
+          --platform ${{ inputs.arch }} \
           --build-arg TARGET_IMAGE="${{ inputs.image }}" \
           --output "type=local,dest=/tmp/compliance-target" \
           --pull \
@@ -102,7 +102,7 @@ runs:
         mkdir -p /tmp/compliance-base
         docker buildx build \
           --builder ${{ steps.builder.outputs.name }} \
-          --platform linux/${{ inputs.arch }} \
+          --platform ${{ inputs.arch }} \
           --build-arg TARGET_IMAGE="${{ steps.resolve-base.outputs.base_image }}" \
           --output "type=local,dest=/tmp/compliance-base" \
           --pull \

.github/actions/init-dynamo-builder/action.yml

@@ -11,8 +11,8 @@ description: 'Route buildkit workers and bootstrap buildx builder for dynamo bui
 #   4. Falls back to Kubernetes driver if no remote workers are available
 #
 # Architecture modes:
-#   - Single arch: Set arch to 'amd64' or 'arm64' to initialize for one architecture
-#   - Multi arch: Set all_arch to 'true' to initialize for both amd64 and arm64
+#   - Single arch: Set arch to 'linux/amd64' or 'linux/arm64'
+#   - Multi arch:  Set arch to 'linux/amd64,linux/arm64'
 #
 # Flavor routing:
 #   BuildKit pods are assigned to flavors based on pod index modulo 3:
@@ -26,14 +26,14 @@ description: 'Route buildkit workers and bootstrap buildx builder for dynamo bui
 #     with:
 #       builder_name: my-builder
 #       flavor: general
-#       all_arch: 'true'
+#       arch: 'linux/amd64,linux/arm64'
 #
 #   # Initialize for single architecture with specific flavor and CUDA version:
 #   - uses: ./.github/actions/init-dynamo-builder
 #     with:
 #       builder_name: my-builder
 #       flavor: vllm
-#       arch: amd64
+#       arch: 'linux/amd64'
 #       cuda_version: '12.9'
 
 inputs:
@@ -45,13 +45,9 @@ inputs:
     required: false
     default: 'general'
   arch:
-    description: 'Target architecture (amd64, arm64). Ignored if all_arch is true.'
+    description: 'Docker platform string: linux/amd64, linux/arm64, or linux/amd64,linux/arm64'
     required: false
-    default: 'amd64'
-  all_arch:
-    description: 'If true, initialize builder for both amd64 and arm64 architectures'
-    required: false
-    default: 'false'
+    default: 'linux/amd64'
   cuda_version:
     description: 'CUDA version (12.9, 13.0). Optional for general flavor.'
     required: false
@@ -99,22 +95,29 @@ runs:
         if [[ -n "${{ inputs.cuda_version }}" ]]; then
           CUDA_ARG="--cuda ${{ inputs.cuda_version }}"
         fi
-        if [[ "${{ inputs.all_arch }}" == "true" ]]; then
-           echo "running with --arch all --flavor ${{ inputs.flavor }} $CUDA_ARG"
-          .github/scripts/route_buildkit.sh --arch all --flavor ${{ inputs.flavor }} $CUDA_ARG
+        # Strip linux/ prefix (e.g. linux/amd64,linux/arm64 → amd64,arm64)
+        ARCH="${{ inputs.arch }}"
+        ARCH="${ARCH//linux\//}"
+        if [[ "$ARCH" == *","* ]]; then
+          ROUTE_ARCH="all"
         else
-          echo "running with --arch ${{ inputs.arch }} --flavor ${{ inputs.flavor }} $CUDA_ARG"
-          .github/scripts/route_buildkit.sh --arch ${{ inputs.arch }} --flavor ${{ inputs.flavor }} $CUDA_ARG
+          ROUTE_ARCH="$ARCH"
         fi
+        echo "running with --arch ${ROUTE_ARCH} --flavor ${{ inputs.flavor }} $CUDA_ARG"
+        .github/scripts/route_buildkit.sh --arch ${ROUTE_ARCH} --flavor ${{ inputs.flavor }} $CUDA_ARG
 
     - name: Prepare worker addresses and platform
       id: prepare
       shell: bash
+      env:
+        AMD64_ADDRS: ${{ steps.route-buildkit.outputs[format('{0}_amd64', inputs.flavor)] }}
+        ARM64_ADDRS: ${{ steps.route-buildkit.outputs[format('{0}_arm64', inputs.flavor)] }}
       run: |
-        if [[ "${{ inputs.all_arch }}" == "true" ]]; then
-          # Combine both architecture outputs for multi-arch builds
-          AMD64_ADDRS="${{ steps.route-buildkit.outputs[format('{0}_amd64', inputs.flavor)] }}"
-          ARM64_ADDRS="${{ steps.route-buildkit.outputs[format('{0}_arm64', inputs.flavor)] }}"
+        # Strip linux/ prefix (e.g. linux/amd64,linux/arm64 → amd64,arm64)
+        ARCH="${{ inputs.arch }}"
+        ARCH="${ARCH//linux\//}"
+        if [[ "$ARCH" == *","* ]]; then
+          # Multi-arch: combine both worker pools
           if [[ -n "$AMD64_ADDRS" && -n "$ARM64_ADDRS" ]]; then
             echo "worker_addresses=${AMD64_ADDRS},${ARM64_ADDRS}" >> "$GITHUB_OUTPUT"
           elif [[ -n "$AMD64_ADDRS" ]]; then
@@ -124,9 +127,10 @@ runs:
           else
             echo "worker_addresses=" >> "$GITHUB_OUTPUT"
           fi
+        elif [[ "$ARCH" == "arm64" ]]; then
+          echo "worker_addresses=${ARM64_ADDRS}" >> "$GITHUB_OUTPUT"
         else
-          # Single architecture build
-          echo "worker_addresses=${{ steps.route-buildkit.outputs[format('{0}_{1}', inputs.flavor, inputs.arch)] }}" >> "$GITHUB_OUTPUT"
+          echo "worker_addresses=${AMD64_ADDRS}" >> "$GITHUB_OUTPUT"
         fi
 
     - name: Bootstrap buildkit

.github/actions/skopeo-copy/action.yml

@@ -53,6 +53,10 @@ inputs:
   target_azure_acr_password:
     description: 'Azure ACR password for target registry'
     required: false
+  override_arch:
+    description: 'Copy only the specified architecture (e.g. amd64). If empty, copies all platforms (manifest list).'
+    required: false
+    default: ''
 
 outputs:
   target_image_ref:
@@ -109,7 +113,13 @@ runs:
         RETRY_DELAY=10
         for attempt in $(seq 1 $MAX_RETRIES); do
           echo "Attempt ${attempt}/${MAX_RETRIES}..."
-          if skopeo copy --all --retry-times 4 "${SOURCE_REF}" "${TARGET_REF}"; then
+          ARCH_FLAG=""
+          if [ -n "${{ inputs.override_arch }}" ]; then
+            ARCH_FLAG="--override-arch ${{ inputs.override_arch }}"
+          else
+            ARCH_FLAG="--all"
+          fi
+          if skopeo copy ${ARCH_FLAG} --retry-times 4 "${SOURCE_REF}" "${TARGET_REF}"; then
             echo "target_image_ref=${{ inputs.target_registry }}/${TARGET_IMAGE}:${TARGET_TAG}" >> $GITHUB_OUTPUT
             echo "✅ Image copied successfully"
             exit 0

.github/workflows/build-flavor-matrix.yml → .github/workflows/build-flavor.yml

 # SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-name: Build Framework Image Matrix
+name: Build Framework Image
 
 on:
   workflow_call:
@@ -14,10 +14,11 @@ on:
         description: 'Target stage for Docker rendering'
         required: true
         type: string
-      platforms:
-        description: 'Platforms to build (JSON array, e.g., ["amd64", "arm64"])'
-        required: true
+      platform:
+        description: 'Docker platform(s) to build (e.g. linux/amd64,linux/arm64)'
+        required: false
         type: string
+        default: 'linux/amd64,linux/arm64'
       cuda_versions:
         description: 'CUDA versions to build (JSON array, e.g., ["12.9", "13.0"])'
         required: true
@@ -27,7 +28,7 @@ on:
         required: true
         type: string
       build_timeout_minutes:
-        description: 'Timeout in minutes for the build step'
+        description: 'Timeout in minutes for the build job'
         required: false
         type: number
         default: 60
@@ -36,13 +37,8 @@ on:
         required: false
         type: boolean
         default: false
-      extra_tags:
-        description: 'Additional tags (newline-separated, -$platform suffix auto-appended)'
-        required: false
-        type: string
-        default: ''
-      make_efa:
-        description: 'Enable AWS EFA support in the build'
+      copy_to_acr:
+        description: 'Also push the image to ACR'
         required: false
         type: boolean
         default: false
@@ -51,31 +47,26 @@ on:
         required: false
         type: boolean
         default: false
-      sanitized_ref_name:
-        description: 'Sanitized git ref name for branch-tagged images'
+      extra_tags:
+        description: 'Additional tags (newline-separated)'
         required: false
         type: string
         default: ''
-      build_only:
-        description: 'Build and push only — enables branch-tagged images'
-        required: false
-        type: boolean
-        default: false
       run_compliance_scan:
         description: 'Run compliance scan after build'
         required: false
         type: boolean
         default: false
-      copy_to_acr:
-        description: 'Copy the built image from ECR to ACR using skopeo after the build'
-        required: false
-        type: boolean
-        default: false
       copy_timeout_minutes:
         description: 'Timeout in minutes for the copy to ACR step'
         required: false
         type: number
         default: 10
+      make_efa:
+        description: 'Enable AWS EFA support in the build'
+        required: false
+        type: boolean
+        default: false
     secrets:
       AWS_DEFAULT_REGION:
         required: true
@@ -98,26 +89,25 @@ on:
 
 jobs:
   build:
-    name: Build cuda${{ matrix.cuda_version }}-${{ matrix.platform }}
+    name: Build cuda${{ matrix.cuda_version }}
     runs-on: prod-builder-v3
     timeout-minutes: ${{ inputs.build_timeout_minutes }}
     strategy:
       fail-fast: false
       matrix:
-        platform: ${{ fromJson(inputs.platforms) }}
         cuda_version: ${{ fromJson(inputs.cuda_versions) }}
     steps:
       - name: Checkout repository
         uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955  # v4.3.0
         with:
           lfs: true
-      - name: Build
+      - name: Build Flavor
         id: build
         uses: ./.github/actions/build-flavor
         with:
           framework: ${{ inputs.framework }}
           target: ${{ inputs.target }}
-          platform: ${{ matrix.platform }}
+          platform: ${{ inputs.platform }}
           cuda_version: ${{ matrix.cuda_version }}
           builder_name: ${{ inputs.builder_name }}
           aws_default_region: ${{ secrets.AWS_DEFAULT_REGION }}
@@ -132,19 +122,9 @@ jobs:
           build_timeout_minutes: ${{ inputs.build_timeout_minutes }}
           push_image: ${{ inputs.push_image }}
           no_cache: ${{ inputs.no_cache }}
-          make_efa: ${{ inputs.make_efa }}
           extra_tags: ${{ inputs.extra_tags }}
-          sanitized_ref_name: ${{ inputs.sanitized_ref_name }}
-          build_only: ${{ inputs.build_only }}
           show_summary: ${{ inputs.push_image }}
-      - name: Refresh BuildKit builder
-        if: ${{ inputs.target != 'dev' }}
-        uses: ./.github/actions/builder-refresher
-        with:
-          builder_name: ${{ inputs.builder_name }}
-          flavor: ${{ inputs.framework }}
-          arch: ${{ matrix.platform }}
-          cuda_version: ${{ matrix.cuda_version }}
+          make_efa: ${{ inputs.make_efa }}
       - name: Copy image to ACR
         if: inputs.copy_to_acr
         timeout-minutes: ${{ inputs.copy_timeout_minutes }}
@@ -152,15 +132,16 @@ jobs:
         with:
           source_registry: ${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com
           source_image: ai-dynamo/dynamo
-          source_tag: ${{ steps.build.outputs.target_tag_plain }}-cuda${{ steps.build.outputs.cuda_version_plain }}-${{ matrix.platform }}
+          source_tag: ${{ steps.build.outputs.target_tag_plain }}-cuda${{ steps.build.outputs.cuda_version_plain }}
           target_registry: ${{ secrets.AZURE_ACR_HOSTNAME }}
           target_image: ai-dynamo/dynamo
-          target_tag: ${{ steps.build.outputs.target_tag_plain }}-cuda${{ steps.build.outputs.cuda_version_plain }}-${{ matrix.platform }}
+          target_tag: ${{ steps.build.outputs.target_tag_plain }}-cuda${{ steps.build.outputs.cuda_version_plain }}
           source_aws_default_region: ${{ secrets.AWS_DEFAULT_REGION }}
           source_aws_account_id: ${{ secrets.AWS_ACCOUNT_ID }}
           target_azure_acr_hostname: ${{ secrets.AZURE_ACR_HOSTNAME }}
           target_azure_acr_user: ${{ secrets.AZURE_ACR_USER }}
           target_azure_acr_password: ${{ secrets.AZURE_ACR_PASSWORD }}
+          override_arch: amd64 # We are using AMD64 images only on the rest of the clusters.
       - name: Calculate compliance image URI
         id: compliance-image
         if: inputs.run_compliance_scan
@@ -168,20 +149,25 @@ jobs:
         run: |
           CUDA_MAJOR="${{ matrix.cuda_version }}"
           CUDA_MAJOR="${CUDA_MAJOR%%.*}"
-          EFA_SUFFIX=""
-          if [ "${{ inputs.make_efa }}" == "true" ]; then
-            EFA_SUFFIX="-efa"
-          fi
-          TARGET_TAG="${{ github.sha }}-${{ inputs.framework }}-${{ inputs.target }}${EFA_SUFFIX}"
-          IMAGE="${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${TARGET_TAG}-cuda${CUDA_MAJOR}-${{ matrix.platform }}"
+          TARGET_TAG="${{ github.sha }}-${{ inputs.framework }}-${{ inputs.target }}"
+          IMAGE="${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${TARGET_TAG}-cuda${CUDA_MAJOR}"
           echo "runtime_image=${IMAGE}" >> $GITHUB_OUTPUT
           echo "cuda_major=${CUDA_MAJOR}" >> $GITHUB_OUTPUT
-      - name: Compliance scan
-        if: inputs.run_compliance_scan
+      - name: Compliance scan (amd64)
+        if: inputs.run_compliance_scan && contains(inputs.platform, 'amd64')
+        uses: ./.github/actions/compliance-scan
+        with:
+          image: ${{ steps.compliance-image.outputs.runtime_image }}
+          artifact_name: compliance-${{ inputs.framework }}-${{ inputs.target }}-cuda${{ steps.compliance-image.outputs.cuda_major }}-amd64
+          arch: linux/amd64
+          framework: ${{ inputs.framework }}
+          cuda_version: ${{ matrix.cuda_version }}
+      - name: Compliance scan (arm64)
+        if: inputs.run_compliance_scan && contains(inputs.platform, 'arm64')
         uses: ./.github/actions/compliance-scan
         with:
           image: ${{ steps.compliance-image.outputs.runtime_image }}
-          artifact_name: compliance-${{ inputs.framework }}-${{ inputs.target }}${{ inputs.make_efa && '-efa' || '' }}-cuda${{ steps.compliance-image.outputs.cuda_major }}-${{ matrix.platform }}
-          arch: ${{ matrix.platform }}
+          artifact_name: compliance-${{ inputs.framework }}-${{ inputs.target }}-cuda${{ steps.compliance-image.outputs.cuda_major }}-arm64
+          arch: linux/arm64
           framework: ${{ inputs.framework }}
           cuda_version: ${{ matrix.cuda_version }}

.github/workflows/build-frontend-image.yaml

@@ -91,7 +91,7 @@ jobs:
         with:
           builder_name: ${{ needs.changed-files.outputs.builder_name }}
           flavor: general
-          all_arch: 'true'
+          arch: 'linux/amd64,linux/arm64'
       - name: Docker Login
         uses: ./.github/actions/docker-login
         with:
@@ -157,6 +157,7 @@ jobs:
           echo "epp_image_uri: ${{ steps.calculate-target-tag.outputs.epp_image_uri }}"
           echo "push_image: true"
 
+
       - name: Build Frontend Container
         id: build-image
         timeout-minutes: 30

.github/workflows/build-on-demand.yml

@@ -52,46 +52,49 @@ jobs:
   vllm-pipeline:
     needs: [init]
     if: inputs.build_vllm
-    uses: ./.github/workflows/build-test-distribute-flavor-matrix.yml
+    uses: ./.github/workflows/build-flavor.yml
     with:
       framework: vllm
       target: runtime
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       builder_name: ${{ needs.init.outputs.builder_name }}
       build_timeout_minutes: 60
-      build_only: true
-      sanitized_ref_name: ${{ needs.init.outputs.sanitized_ref_name }}
+      push_image: true
+      run_compliance_scan: false
+      copy_to_acr: true
     secrets: inherit
 
   sglang-pipeline:
     needs: [init]
     if: inputs.build_sglang
-    uses: ./.github/workflows/build-test-distribute-flavor-matrix.yml
+    uses: ./.github/workflows/build-flavor.yml
     with:
       framework: sglang
       target: runtime
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       builder_name: ${{ needs.init.outputs.builder_name }}
       build_timeout_minutes: 60
-      build_only: true
-      sanitized_ref_name: ${{ needs.init.outputs.sanitized_ref_name }}
+      push_image: true
+      run_compliance_scan: false
+      copy_to_acr: true
     secrets: inherit
 
   trtllm-pipeline:
     needs: [init]
     if: inputs.build_trtllm
-    uses: ./.github/workflows/build-test-distribute-flavor-matrix.yml
+    uses: ./.github/workflows/build-flavor.yml
     with:
       framework: trtllm
       target: runtime
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["13.1"]'
       builder_name: ${{ needs.init.outputs.builder_name }}
       build_timeout_minutes: 60
-      build_only: true
-      sanitized_ref_name: ${{ needs.init.outputs.sanitized_ref_name }}
+      push_image: true
+      run_compliance_scan: false
+      copy_to_acr: true
     secrets: inherit
 
   # ============================================================================
@@ -115,7 +118,7 @@ jobs:
         with:
           builder_name: ${{ needs.init.outputs.builder_name }}
           flavor: general
-          all_arch: 'true'
+          arch: 'linux/amd64,linux/arm64'
       - name: Docker Login
         uses: ./.github/actions/docker-login
         with:

.github/workflows/build-test-distribute-flavor-matrix.yml

@@ -14,10 +14,11 @@ on:
         description: 'Target stage for Docker rendering'
         required: true
         type: string
-      platforms:
-        description: 'Platforms to build (JSON array, e.g., ["amd64", "arm64"])'
-        required: true
+      platform:
+        description: 'Docker platform(s) to build (e.g. linux/amd64,linux/arm64)'
+        required: false
         type: string
+        default: 'linux/amd64,linux/arm64'
       cuda_versions:
         description: 'CUDA versions to build (JSON array, e.g., ["12.9", "13.0"])'
         required: true
@@ -84,7 +85,7 @@ on:
         required: true
         type: string
       extra_tags:
-        description: 'Additional tags (newline-separated, -$platform suffix auto-appended)'
+        description: 'Additional tags (newline-separated)'
         required: false
         type: string
         default: ''
@@ -113,16 +114,6 @@ on:
         required: false
         type: boolean
         default: false
-      build_only:
-        description: 'Build and push only — skip all tests, show summary'
-        required: false
-        type: boolean
-        default: false
-      sanitized_ref_name:
-        description: 'Sanitized git ref name for branch-tagged images (used with build_only)'
-        required: false
-        type: string
-        default: ''
     secrets:
       AWS_DEFAULT_REGION:
         required: true
@@ -148,7 +139,6 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        platform: ${{ fromJson(inputs.platforms) }}
         cuda_version: ${{ fromJson(inputs.cuda_versions) }}
     # This name weirdly shows in the checks overview, but not once you dive into
     # a specific workflow. Keeping it as a short placeholder
@@ -156,13 +146,13 @@ jobs:
     uses: ./.github/workflows/build-test-distribute-flavor.yml
     with:
       framework: ${{ inputs.framework }}
-      platform: ${{ matrix.platform }}
+      platform: ${{ inputs.platform }}
       target: ${{ inputs.target }}
       cuda_version: ${{ matrix.cuda_version }}
       extra_tags: ${{ inputs.extra_tags }}
       no_cache: ${{ inputs.no_cache }}
       builder_name: ${{ inputs.builder_name }}
-      build_image: ${{ inputs.build_only || inputs.build_image }}
+      build_image: ${{ inputs.build_image }}
       build_timeout_minutes: ${{ inputs.build_timeout_minutes }}
       push_image: ${{ inputs.push_image }}
       run_cpu_only_tests: ${{ inputs.run_cpu_only_tests }}
@@ -174,10 +164,8 @@ jobs:
       run_multi_gpu_tests: ${{ inputs.run_multi_gpu_tests }}
       multi_gpu_test_markers: ${{ inputs.multi_gpu_test_markers }}
       multi_gpu_test_timeout_minutes: ${{ inputs.multi_gpu_test_timeout_minutes }}
-      copy_to_acr: ${{ inputs.copy_to_acr && matrix.platform == 'amd64' }} # no reason to copy ARM images to ACR
+      copy_to_acr: ${{ inputs.copy_to_acr }}
       copy_timeout_minutes: ${{ inputs.copy_timeout_minutes }}
       make_efa: ${{ inputs.make_efa }}
-      show_summary: ${{ inputs.build_only || inputs.show_summary }}
-      build_only: ${{ inputs.build_only }}
-      sanitized_ref_name: ${{ inputs.sanitized_ref_name }}
+      show_summary: ${{ inputs.show_summary }}
     secrets: inherit

.github/workflows/build-test-distribute-flavor.yml

@@ -15,7 +15,7 @@ on:
         required: true
         type: string
       platform:
-        description: 'Platform to build (amd64 or arm64)'
+        description: 'Docker platform(s) to build (e.g. linux/amd64,linux/arm64)'
         required: true
         type: string
       cuda_version:
@@ -84,7 +84,7 @@ on:
         required: true
         type: string
       extra_tags:
-        description: 'Additional tags (newline-separated, -$platform suffix auto-appended)'
+        description: 'Additional tags (newline-separated)'
         required: false
         type: string
         default: ''
@@ -118,16 +118,6 @@ on:
         required: false
         type: boolean
         default: false
-      build_only:
-        description: 'Build and push only — skip all tests, show summary'
-        required: false
-        type: boolean
-        default: false
-      sanitized_ref_name:
-        description: 'Sanitized git ref name for branch-tagged images (used with build_only)'
-        required: false
-        type: string
-        default: ''
     secrets:
       AWS_DEFAULT_REGION:
         required: true
@@ -153,12 +143,14 @@ jobs:
   # ============================================================================
   build:
     if: inputs.build_image
-    name: Build cuda${{ inputs.cuda_version }}-${{ inputs.platform }}
+    name: Build cuda${{ inputs.cuda_version }}
     runs-on: prod-builder-v3
     timeout-minutes: ${{ inputs.build_timeout_minutes }}
     outputs:
       target_tag_plain: ${{ steps.build.outputs.target_tag_plain }}
       test_tag_plain: ${{ steps.build.outputs.test_tag_plain }}
+      compliance_arches: ${{ steps.compliance-arches.outputs.arches }}
+      test_runners: ${{ steps.test-runners.outputs.runners }}
     env:
       FRAMEWORK: ${{ inputs.framework }}
     steps:
@@ -166,6 +158,32 @@ jobs:
         uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955  # v4.3.0
         with:
           lfs: true
+      - name: Compute compliance arches
+        id: compliance-arches
+        shell: bash
+        run: |
+          PLATFORM="${{ inputs.platform }}"
+          # Convert comma-separated Docker platform string to JSON array
+          # "linux/amd64,linux/arm64" -> ["linux/amd64","linux/arm64"]
+          # "linux/amd64"             -> ["linux/amd64"]
+          JSON=$(printf '%s\n' ${PLATFORM//,/ } | jq -Rsc '[splits("\n") | select(length>0)]')
+          echo "arches=${JSON}" >> $GITHUB_OUTPUT
+      - name: Compute test runners
+        id: test-runners
+        shell: bash
+        run: |
+          PLATFORM="${{ inputs.platform }}"
+          # Emit a JSON array of {arch, runner} objects for platforms actually being built
+          # "linux/amd64,linux/arm64" -> both runners
+          # "linux/amd64"             -> amd64 runner only (covers EFA and other single-arch)
+          # "linux/arm64"             -> arm64 runner only
+          if [[ "$PLATFORM" == *"amd64"* && "$PLATFORM" == *"arm64"* ]]; then
+            echo 'runners=[{"arch":"amd64","runner":"prod-tester-amd-gpu-v1"},{"arch":"arm64","runner":"prod-tester-arm-v1"}]' >> $GITHUB_OUTPUT
+          elif [[ "$PLATFORM" == *"arm64"* ]]; then
+            echo 'runners=[{"arch":"arm64","runner":"prod-tester-arm-v1"}]' >> $GITHUB_OUTPUT
+          else
+            echo 'runners=[{"arch":"amd64","runner":"prod-tester-amd-gpu-v1"}]' >> $GITHUB_OUTPUT
+          fi
       - name: Build
         id: build
         uses: ./.github/actions/build-flavor
@@ -190,23 +208,23 @@ jobs:
           no_cache: ${{ inputs.no_cache }}
           make_efa: ${{ inputs.make_efa }}
           extra_tags: ${{ inputs.extra_tags }}
-          build_only: ${{ inputs.build_only }}
-          sanitized_ref_name: ${{ inputs.sanitized_ref_name }}
           show_summary: ${{ inputs.push_image && inputs.show_summary }}
 
 
-
   # ============================================================================
   # TEST
   # ============================================================================
   test:
     if: |
-        !inputs.build_only &&
         ( inputs.run_cpu_only_tests || inputs.run_single_gpu_tests ) &&
         inputs.build_image
     needs: [build]
-    name: Test cuda${{ inputs.cuda_version }}-${{ inputs.platform }}
-    runs-on: ${{ inputs.platform == 'amd64' && 'prod-tester-amd-gpu-v1' || 'prod-tester-arm-v1' }}
+    name: Test cuda${{ inputs.cuda_version }} (${{ matrix.arch }})
+    strategy:
+      fail-fast: false
+      matrix:
+        include: ${{ fromJson(needs.build.outputs.test_runners) }}
+    runs-on: ${{ matrix.runner }}
     env:
       FRAMEWORK: ${{ inputs.framework }}
     steps:
@@ -219,9 +237,9 @@ jobs:
           CUDA_VERSION_RAW=${{ inputs.cuda_version }}
           CUDA_VERSION=${CUDA_VERSION_RAW%%.*}
           echo "cuda_version_plain=${CUDA_VERSION}" >> $GITHUB_OUTPUT
-          RUNTIME_IMAGE=${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${{ needs.build.outputs.target_tag_plain }}-cuda${CUDA_VERSION}-${{ inputs.platform }}
+          RUNTIME_IMAGE=${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${{ needs.build.outputs.target_tag_plain }}-cuda${CUDA_VERSION}
           echo "runtime_image=${RUNTIME_IMAGE}" >> $GITHUB_OUTPUT
-          TEST_IMAGE=${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${{ needs.build.outputs.test_tag_plain }}-cuda${CUDA_VERSION}-${{ inputs.platform }}
+          TEST_IMAGE=${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${{ needs.build.outputs.test_tag_plain }}-cuda${CUDA_VERSION}
           echo "test_image=${TEST_IMAGE}" >> $GITHUB_OUTPUT
       - name: Docker Login
         uses: ./.github/actions/docker-login
@@ -273,7 +291,7 @@ jobs:
           pytest_marks: ${{ inputs.cpu_only_test_markers }}
           framework: ${{ inputs.framework }}
           test_type: "pre_merge_cpu"
-          platform_arch: ${{ inputs.platform }}
+          platform_arch: ${{ matrix.arch }}
           hf_token: ${{ secrets.HF_TOKEN }}
           parallel_mode: 'auto'
           dind_as_sidecar: 'true'
@@ -282,14 +300,14 @@ jobs:
       # These are e2e tests marked with gpu_1 that require GPU hardware
       - name: Run GPU tests (sequential)
         timeout-minutes: ${{ inputs.single_gpu_test_timeout_minutes }}
-        if: ( inputs.platform == 'amd64' && inputs.run_single_gpu_tests == true ) # We only run GPU tests on amd64
+        if: inputs.run_single_gpu_tests && matrix.arch == 'amd64'
         uses: ./.github/actions/pytest
         with:
           image_tag: ${{ steps.calculate-target-tag.outputs.test_image }}
           pytest_marks: ${{ inputs.single_gpu_test_markers }}
           framework: ${{ inputs.framework }}
           test_type: "pre_merge_gpu"
-          platform_arch: ${{ inputs.platform }}
+          platform_arch: ${{ matrix.arch }}
           hf_token: ${{ secrets.HF_TOKEN }}
           parallel_mode: 'none'
           dind_as_sidecar: 'true'
@@ -301,12 +319,10 @@ jobs:
   multi-gpu-test:
     # Multi-GPU support limited to AMD64 only
     if: |
-        !inputs.build_only &&
         inputs.run_multi_gpu_tests &&
-        inputs.build_image &&
-        ( inputs.platform != 'arm64' )
+        inputs.build_image
     needs: [build]
-    name: Multi-gpu test cuda${{ inputs.cuda_version }}-${{ inputs.platform }}
+    name: Multi-gpu test cuda${{ inputs.cuda_version }}
     runs-on: prod-tester-amd-gpu-4-v1
     env:
       FRAMEWORK: ${{ inputs.framework }}
@@ -320,9 +336,9 @@ jobs:
           CUDA_VERSION_RAW=${{ inputs.cuda_version }}
           CUDA_VERSION=${CUDA_VERSION_RAW%%.*}
           echo "cuda_version_plain=${CUDA_VERSION}" >> $GITHUB_OUTPUT
-          RUNTIME_IMAGE=${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${{ needs.build.outputs.target_tag_plain }}-cuda${CUDA_VERSION}-${{ inputs.platform }}
+          RUNTIME_IMAGE=${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${{ needs.build.outputs.target_tag_plain }}-cuda${CUDA_VERSION}
           echo "runtime_image=${RUNTIME_IMAGE}" >> $GITHUB_OUTPUT
-          TEST_IMAGE=${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${{ needs.build.outputs.test_tag_plain }}-cuda${CUDA_VERSION}-${{ inputs.platform }}
+          TEST_IMAGE=${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${{ needs.build.outputs.test_tag_plain }}-cuda${CUDA_VERSION}
           echo "test_image=${TEST_IMAGE}" >> $GITHUB_OUTPUT
       - name: Docker Login
         uses: ./.github/actions/docker-login
@@ -352,7 +368,7 @@ jobs:
           pytest_marks: ${{ inputs.multi_gpu_test_markers }}
           framework: ${{ inputs.framework }}
           test_type: "pre_merge_gpu"
-          platform_arch: ${{ inputs.platform }}
+          platform_arch: amd64
           hf_token: ${{ secrets.HF_TOKEN }}
           parallel_mode: 'none'
           dind_as_sidecar: 'true'
@@ -364,7 +380,11 @@ jobs:
   compliance:
     if: inputs.build_image && inputs.push_image
     needs: [build]
-    name: Compliance cuda${{ inputs.cuda_version }}-${{ inputs.platform }}
+    strategy:
+      fail-fast: false
+      matrix:
+        arch: ${{ fromJson(needs.build.outputs.compliance_arches) }}
+    name: Compliance cuda${{ inputs.cuda_version }}-${{ matrix.arch }}
     runs-on: prod-builder-v3
     steps:
       - name: Checkout repository
@@ -377,21 +397,24 @@ jobs:
           azure_acr_hostname: ${{ secrets.AZURE_ACR_HOSTNAME }}
           azure_acr_user: ${{ secrets.AZURE_ACR_USER }}
           azure_acr_password: ${{ secrets.AZURE_ACR_PASSWORD }}
-      - name: Calculate image URI
+      - name: Calculate image URI and arch suffix
         id: images
         shell: bash
         run: |
           CUDA_VERSION_RAW=${{ inputs.cuda_version }}
           CUDA_VERSION=${CUDA_VERSION_RAW%%.*}
           echo "cuda_major=${CUDA_VERSION}" >> $GITHUB_OUTPUT
-          RUNTIME_IMAGE=${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${{ needs.build.outputs.target_tag_plain }}-cuda${CUDA_VERSION}-${{ inputs.platform }}
+          RUNTIME_IMAGE=${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com/ai-dynamo/dynamo:${{ needs.build.outputs.target_tag_plain }}-cuda${CUDA_VERSION}
           echo "runtime_image=${RUNTIME_IMAGE}" >> $GITHUB_OUTPUT
+          # Sanitize arch for artifact name: linux/amd64 -> amd64 (artifact names can't contain /)
+          ARCH="${{ matrix.arch }}"
+          echo "arch_suffix=${ARCH#linux/}" >> $GITHUB_OUTPUT
       - name: Compliance scan
         uses: ./.github/actions/compliance-scan
         with:
           image: ${{ steps.images.outputs.runtime_image }}
-          artifact_name: compliance-${{ inputs.framework }}-${{ inputs.target }}${{ inputs.make_efa && '-efa' || '' }}-cuda${{ steps.images.outputs.cuda_major }}-${{ inputs.platform }}
-          arch: ${{ inputs.platform }}
+          artifact_name: compliance-${{ inputs.framework }}-${{ inputs.target }}${{ inputs.make_efa && '-efa' || '' }}-cuda${{ steps.images.outputs.cuda_major }}-${{ steps.images.outputs.arch_suffix }}
+          arch: ${{ matrix.arch }}
           framework: ${{ inputs.framework }}
           cuda_version: ${{ inputs.cuda_version }}
 
@@ -404,11 +427,10 @@ jobs:
     # Run if copy_to_acr is true AND build succeeded AND (test succeeded OR test was skipped)
     if: |
       always() &&
-      !inputs.build_only &&
       inputs.copy_to_acr &&
       needs.build.result == 'success' &&
       (needs.test.result == 'success' || needs.test.result == 'skipped')
-    name: copy-to-acr cuda${{ inputs.cuda_version }}-${{ inputs.platform }}
+    name: copy-to-acr cuda${{ inputs.cuda_version }}
     runs-on: prod-default-small-v2
     outputs:
       target_tag_plain: ${{ needs.build.outputs.target_tag_plain }}
@@ -430,12 +452,13 @@ jobs:
         with:
           source_registry: ${{ secrets.AWS_ACCOUNT_ID }}.dkr.ecr.${{ secrets.AWS_DEFAULT_REGION }}.amazonaws.com
           source_image: ai-dynamo/dynamo
-          source_tag: ${{ needs.build.outputs.target_tag_plain }}-cuda${{ steps.calculate-target-tag.outputs.cuda_version_plain }}-${{ inputs.platform }}
+          source_tag: ${{ needs.build.outputs.target_tag_plain }}-cuda${{ steps.calculate-target-tag.outputs.cuda_version_plain }}
           target_registry: ${{ secrets.AZURE_ACR_HOSTNAME }}
           target_image: ai-dynamo/dynamo
-          target_tag: ${{ needs.build.outputs.target_tag_plain }}-cuda${{ steps.calculate-target-tag.outputs.cuda_version_plain }}-${{ inputs.platform }}
+          target_tag: ${{ needs.build.outputs.target_tag_plain }}-cuda${{ steps.calculate-target-tag.outputs.cuda_version_plain }}
           source_aws_default_region: ${{ secrets.AWS_DEFAULT_REGION }}
           source_aws_account_id: ${{ secrets.AWS_ACCOUNT_ID }}
           target_azure_acr_hostname: ${{ secrets.AZURE_ACR_HOSTNAME }}
           target_azure_acr_user: ${{ secrets.AZURE_ACR_USER }}
           target_azure_acr_password: ${{ secrets.AZURE_ACR_PASSWORD }}
+          override_arch: amd64 # We are using AMD64 images only on the rest of the clusters.

.github/workflows/container-validation-dynamo.yml

@@ -74,7 +74,7 @@ jobs:
         with:
           builder_name: ${{ needs.changed-files.outputs.builder_name }}
           flavor: general
-          arch: amd64
+          arch: linux/amd64
       - name: Docker Login
         if: github.event.pull_request.head.repo.full_name == github.repository || github.event_name == 'push'
         uses: ./.github/actions/docker-login

.github/workflows/nightly-ci.yml

@@ -24,7 +24,7 @@ jobs:
       framework: vllm
       target: runtime
       no_cache: true
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       extra_tags: |
         ${{ github.ref_name == 'main' && 'main-vllm' || '' }}
@@ -47,7 +47,7 @@ jobs:
       framework: sglang
       target: runtime
       no_cache: true
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       extra_tags: |
         ${{ github.ref_name == 'main' && 'main-sglang' || '' }}
@@ -70,7 +70,7 @@ jobs:
       framework: trtllm
       target: runtime
       no_cache: true
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["13.1"]'
       extra_tags: |
         ${{ github.ref_name == 'main' && 'main-trtllm' || '' }}

.github/workflows/post-merge-ci.yml

@@ -25,7 +25,7 @@ jobs:
     with:
       framework: vllm
       target: runtime
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       extra_tags: |
         ${{ github.ref_name == 'main' && 'main-vllm' || '' }}
@@ -50,7 +50,7 @@ jobs:
     with:
       framework: sglang
       target: runtime
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       extra_tags: |
         ${{ github.ref_name == 'main' && 'main-sglang' || '' }}
@@ -75,7 +75,7 @@ jobs:
     with:
       framework: trtllm
       target: runtime
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["13.1"]'
       extra_tags: |
         ${{ github.ref_name == 'main' && 'main-trtllm' || '' }}
@@ -96,11 +96,11 @@ jobs:
   # ============================================================================
   vllm-dev-pipeline:
     name: vllm-dev
-    uses: ./.github/workflows/build-flavor-matrix.yml
+    uses: ./.github/workflows/build-flavor.yml
     with:
       framework: vllm
       target: dev
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       builder_name: b-${{ github.run_id }}-${{ github.run_attempt }}
       build_timeout_minutes: 60
@@ -111,11 +111,11 @@ jobs:
 
   sglang-dev-pipeline:
     name: sglang-dev
-    uses: ./.github/workflows/build-flavor-matrix.yml
+    uses: ./.github/workflows/build-flavor.yml
     with:
       framework: sglang
       target: dev
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       builder_name: b-${{ github.run_id }}-${{ github.run_attempt }}
       build_timeout_minutes: 60
@@ -126,11 +126,11 @@ jobs:
 
   trtllm-dev-pipeline:
     name: trtllm-dev
-    uses: ./.github/workflows/build-flavor-matrix.yml
+    uses: ./.github/workflows/build-flavor.yml
     with:
       framework: trtllm
       target: dev
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["13.1"]'
       builder_name: b-${{ github.run_id }}-${{ github.run_attempt }}
       build_timeout_minutes: 60
@@ -150,7 +150,7 @@ jobs:
     with:
       framework: vllm
       target: runtime
-      platforms: '["amd64"]'
+      platform: 'linux/amd64'
       cuda_versions: '["12.9"]'
       make_efa: true
       extra_tags: |
@@ -174,7 +174,7 @@ jobs:
     with:
       framework: trtllm
       target: runtime
-      platforms: '["amd64"]'
+      platform: 'linux/amd64'
       cuda_versions: '["13.1"]'
       make_efa: true
       extra_tags: |
@@ -211,7 +211,7 @@ jobs:
         with:
           builder_name: b-${{ github.run_id }}-${{ github.run_attempt }}
           flavor: general
-          all_arch: 'true'
+          arch: 'linux/amd64,linux/arm64'
       - name: Docker Login
         uses: ./.github/actions/docker-login
         with:
@@ -322,7 +322,7 @@ jobs:
     with:
       framework: vllm
       profiles: '["agg", "agg_router", "disagg", "disagg_router"]'
-      image_suffix: vllm-runtime-cuda12-amd64
+      image_suffix: vllm-runtime-cuda12
       namespace: ${{ needs.deploy-operator.outputs.namespace }}
       vcluster_name: ${{ needs.deploy-operator.outputs.vcluster_name }}
       operator_tag: ${{ needs.deploy-operator.outputs.operator_tag }}
@@ -334,7 +334,7 @@ jobs:
     with:
       framework: sglang
       profiles: '["agg", "agg_router"]'
-      image_suffix: sglang-runtime-cuda12-amd64
+      image_suffix: sglang-runtime-cuda12
       namespace: ${{ needs.deploy-operator.outputs.namespace }}
       vcluster_name: ${{ needs.deploy-operator.outputs.vcluster_name }}
       operator_tag: ${{ needs.deploy-operator.outputs.operator_tag }}
@@ -346,7 +346,7 @@ jobs:
     with:
       framework: trtllm
       profiles: '["agg", "agg_router"]'
-      image_suffix: trtllm-runtime-cuda13-amd64
+      image_suffix: trtllm-runtime-cuda13
       namespace: ${{ needs.deploy-operator.outputs.namespace }}
       vcluster_name: ${{ needs.deploy-operator.outputs.vcluster_name }}
       operator_tag: ${{ needs.deploy-operator.outputs.operator_tag }}

.github/workflows/pr.yaml

@@ -99,7 +99,7 @@ jobs:
         with:
           builder_name: ${{ needs.changed-files.outputs.builder_name }}
           flavor: general
-          all_arch: 'true'
+          arch: 'linux/amd64,linux/arm64'
       - name: Docker Login
         uses: ./.github/actions/docker-login
         with:
@@ -187,7 +187,7 @@ jobs:
     with:
       framework: vllm
       target: runtime
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       builder_name: ${{ needs.changed-files.outputs.builder_name }}
       build_timeout_minutes: 60
@@ -213,7 +213,7 @@ jobs:
     with:
       framework: sglang
       target: runtime
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       builder_name: ${{ needs.changed-files.outputs.builder_name }}
       build_timeout_minutes: 60
@@ -238,7 +238,7 @@ jobs:
     with:
       framework: trtllm
       target: runtime
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["13.1"]'
       builder_name: ${{ needs.changed-files.outputs.builder_name }}
       build_timeout_minutes: 60
@@ -259,11 +259,11 @@ jobs:
     name: vllm-dev
     needs: [changed-files]
     if: needs.changed-files.outputs.core == 'true' || needs.changed-files.outputs.vllm == 'true'
-    uses: ./.github/workflows/build-flavor-matrix.yml
+    uses: ./.github/workflows/build-flavor.yml
     with:
       framework: vllm
       target: dev
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       builder_name: ${{ needs.changed-files.outputs.builder_name }}
       build_timeout_minutes: 60
@@ -274,11 +274,11 @@ jobs:
     name: sglang-dev
     needs: [changed-files]
     if: needs.changed-files.outputs.core == 'true' || needs.changed-files.outputs.sglang == 'true'
-    uses: ./.github/workflows/build-flavor-matrix.yml
+    uses: ./.github/workflows/build-flavor.yml
     with:
       framework: sglang
       target: dev
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["12.9", "13.0"]'
       builder_name: ${{ needs.changed-files.outputs.builder_name }}
       build_timeout_minutes: 60
@@ -289,11 +289,11 @@ jobs:
     name: trtllm-dev
     needs: [changed-files]
     if: needs.changed-files.outputs.core == 'true' || needs.changed-files.outputs.trtllm == 'true'
-    uses: ./.github/workflows/build-flavor-matrix.yml
+    uses: ./.github/workflows/build-flavor.yml
     with:
       framework: trtllm
       target: dev
-      platforms: '["amd64", "arm64"]'
+      platform: 'linux/amd64,linux/arm64'
       cuda_versions: '["13.1"]'
       builder_name: ${{ needs.changed-files.outputs.builder_name }}
       build_timeout_minutes: 60
@@ -343,7 +343,7 @@ jobs:
     with:
       framework: vllm
       profiles: '["agg", "agg_router", "disagg", "disagg_router"]'
-      image_suffix: vllm-runtime-cuda12-amd64
+      image_suffix: vllm-runtime-cuda12
       namespace: ${{ needs.deploy-operator.outputs.namespace }}
       vcluster_name: ${{ needs.deploy-operator.outputs.vcluster_name }}
       operator_tag: ${{ needs.deploy-operator.outputs.operator_tag }}
@@ -360,7 +360,7 @@ jobs:
     with:
       framework: sglang
       profiles: '["agg", "agg_router"]'
-      image_suffix: sglang-runtime-cuda12-amd64
+      image_suffix: sglang-runtime-cuda12
       namespace: ${{ needs.deploy-operator.outputs.namespace }}
       vcluster_name: ${{ needs.deploy-operator.outputs.vcluster_name }}
       operator_tag: ${{ needs.deploy-operator.outputs.operator_tag }}
@@ -377,7 +377,7 @@ jobs:
     with:
       framework: trtllm
       profiles: '["agg", "agg_router"]'
-      image_suffix: trtllm-runtime-cuda13-amd64
+      image_suffix: trtllm-runtime-cuda13
       namespace: ${{ needs.deploy-operator.outputs.namespace }}
       vcluster_name: ${{ needs.deploy-operator.outputs.vcluster_name }}
       operator_tag: ${{ needs.deploy-operator.outputs.operator_tag }}

.github/workflows/release.yml

@@ -189,10 +189,14 @@ jobs:
           echo "========================================"
 
           copy_image() {
-            local SRC="$1" DST="$2" LABEL="$3"
+            local SRC="$1" DST="$2" LABEL="$3" PLATFORM="${4:-}"
+            local PLATFORM_ARG=""
+            if [ -n "${PLATFORM}" ]; then
+              PLATFORM_ARG="--platform ${PLATFORM}"
+            fi
             echo "----------------------------------------"
             echo "Copying: ${LABEL}"
-            if crane copy "${SRC}" "${DST}"; then
+            if crane copy ${PLATFORM_ARG} "${SRC}" "${DST}"; then
               echo "  Copied: ${LABEL}"
               SUCCESSFUL_COPIES+=("${LABEL}")
               return 0
@@ -223,9 +227,9 @@ jobs:
           for FRAMEWORK in "${CUDA12_FRAMEWORKS[@]}"; do
             NGC_NAME="${FRAMEWORK}-runtime"
             for ARCH in "${ARCHITECTURES[@]}"; do
-              SOURCE="${ECR_HOSTNAME}/${REGISTRY_IMAGE}:${COMMIT_SHA}-${FRAMEWORK}-runtime-cuda12-${ARCH}"
+              SOURCE="${ECR_HOSTNAME}/${REGISTRY_IMAGE}:${COMMIT_SHA}-${FRAMEWORK}-runtime-cuda12"
               TARGET="${NGC_REGISTRY}/${NGC_ORG}/ai-dynamo/${NGC_NAME}:${NGC_VERSION_TAG}-${ARCH}"
-              copy_image "${SOURCE}" "${TARGET}" "${NGC_NAME}:${NGC_VERSION_TAG}-${ARCH}"
+              copy_image "${SOURCE}" "${TARGET}" "${NGC_NAME}:${NGC_VERSION_TAG}-${ARCH}" "linux/${ARCH}"
             done
 
             create_manifest \
@@ -247,9 +251,9 @@ jobs:
             fi
 
             for ARCH in "${ARCHITECTURES[@]}"; do
-              SOURCE="${ECR_HOSTNAME}/${REGISTRY_IMAGE}:${COMMIT_SHA}-${FRAMEWORK}-runtime-cuda13-${ARCH}"
+              SOURCE="${ECR_HOSTNAME}/${REGISTRY_IMAGE}:${COMMIT_SHA}-${FRAMEWORK}-runtime-cuda13"
               TARGET="${NGC_REGISTRY}/${NGC_ORG}/ai-dynamo/${NGC_NAME}:${NGC_VERSION_TAG}-cuda13-${ARCH}"
-              copy_image "${SOURCE}" "${TARGET}" "${NGC_NAME}:${NGC_VERSION_TAG}-cuda13-${ARCH}"
+              copy_image "${SOURCE}" "${TARGET}" "${NGC_NAME}:${NGC_VERSION_TAG}-cuda13-${ARCH}" "linux/${ARCH}"
             done
 
             create_manifest \
@@ -264,12 +268,12 @@ jobs:
           echo "=== EFA Runtime Images ==="
 
           # vllm EFA (CUDA 12, amd64 only)
-          SOURCE="${ECR_HOSTNAME}/${REGISTRY_IMAGE}:${COMMIT_SHA}-vllm-runtime-efa-cuda12-amd64"
+          SOURCE="${ECR_HOSTNAME}/${REGISTRY_IMAGE}:${COMMIT_SHA}-vllm-runtime-efa-cuda12"
           TARGET="${NGC_REGISTRY}/${NGC_ORG}/ai-dynamo/vllm-runtime:${NGC_VERSION_TAG}-efa"
           copy_image "${SOURCE}" "${TARGET}" "vllm-runtime:${NGC_VERSION_TAG}-efa"
 
           # trtllm EFA (CUDA 13, amd64 only)
-          SOURCE="${ECR_HOSTNAME}/${REGISTRY_IMAGE}:${COMMIT_SHA}-trtllm-runtime-efa-cuda13-amd64"
+          SOURCE="${ECR_HOSTNAME}/${REGISTRY_IMAGE}:${COMMIT_SHA}-trtllm-runtime-efa-cuda13"
           TARGET="${NGC_REGISTRY}/${NGC_ORG}/ai-dynamo/tensorrtllm-runtime:${NGC_VERSION_TAG}-efa"
           copy_image "${SOURCE}" "${TARGET}" "tensorrtllm-runtime:${NGC_VERSION_TAG}-efa"