feat(infra): split Docker image into sglang and vllm variants

.github/workflows/build-docker-image.yml

@@ -14,7 +14,7 @@ permissions:
 env:
   GCP_PROJECT_ID: ${{ secrets.GCP_PROJECT_ID }}
   VALIDATOR_LABELS: gcp-docker-validator
-  RUNNER_VERSION: '2.317.0'
+  RUNNER_VERSION: '2.332.0'
   IMAGE_NAME: ghcr.io/inclusionai/areal-runtime
   IMAGE_TAG: test
 
@@ -111,12 +111,12 @@ jobs:
 
             throw new Error(`Timed out waiting for builder runner ${instanceName} to come online.`);
 
-  build-and-push-image:
+  build-and-push-images:
     needs:
       - start-builder
-    name: Build and push Docker image
+    name: Build and push Docker images
     runs-on: [self-hosted, areal-docker-builder]
-    timeout-minutes: 180
+    timeout-minutes: 240  # ~90 min/image x 2 + buffer for cache misses
     steps:
       - uses: actions/checkout@v6
 
@@ -132,35 +132,65 @@ jobs:
           username: inclusionai
           password: ${{ secrets.GHCR_TOKEN }}
 
-      - name: Build and push Docker image
+      - name: Build and push sglang image
         uses: docker/build-push-action@v7
         with:
           context: .
           file: ./Dockerfile
           push: true
-          tags: ${{ env.IMAGE_NAME }}:${{ env.IMAGE_TAG }}
+          tags: ${{ env.IMAGE_NAME }}:${{ env.IMAGE_TAG }}-sglang
+          build-args: |
+            VARIANT=sglang
+
+      - name: Build and push vllm image
+        uses: docker/build-push-action@v7
+        with:
+          context: .
+          file: ./Dockerfile
+          push: true
+          tags: ${{ env.IMAGE_NAME }}:${{ env.IMAGE_TAG }}-vllm
+          build-args: |
+            VARIANT=vllm
 
       - name: Image details
         run: |
-          echo "✅ Docker image built and pushed successfully!"
-          echo "Image: ${{ env.IMAGE_NAME }}:${{ env.IMAGE_TAG }}"
+          echo "✅ Docker images built and pushed successfully!"
+          echo "Images:"
+          echo "  - ${{ env.IMAGE_NAME }}:${{ env.IMAGE_TAG }}-sglang"
+          echo "  - ${{ env.IMAGE_NAME }}:${{ env.IMAGE_TAG }}-vllm"
           echo "Commit: ${{ github.sha }}"
           echo "Branch: ${{ github.head_ref || github.ref_name }}"
 
-  run-tests:
-    name: Run tests with test image
+  run-tests-sglang:
+    name: Run tests with sglang image
+    needs:
+      - build-and-push-images
+    uses: ./.github/workflows/test-areal.yml
+    with:
+      image_tag: test
+      variant: sglang
+    secrets: inherit
+
+  run-tests-vllm:
+    name: Run tests with vllm image
     needs:
-      - build-and-push-image
+      - build-and-push-images
     uses: ./.github/workflows/test-areal.yml
     with:
       image_tag: test
+      variant: vllm
     secrets: inherit
 
-  promote-image:
-    name: Promote test image to dev
+  promote-images:
+    name: Promote ${{ matrix.variant }} test image to dev
     needs:
-      - run-tests
+      - run-tests-sglang
+      - run-tests-vllm
     runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        variant: [sglang, vllm]
     steps:
       - name: Log in to GitHub Container Registry
         uses: docker/login-action@v4
@@ -173,45 +203,53 @@ jobs:
         env:
           IMAGE_NAME: ghcr.io/inclusionai/areal-runtime
         run: |
-          docker pull $IMAGE_NAME:test
-          docker tag $IMAGE_NAME:test $IMAGE_NAME:dev
-          docker push $IMAGE_NAME:dev
-          echo "✅ Image promoted from :test to :dev"
+          docker pull $IMAGE_NAME:test-${{ matrix.variant }}
+          docker tag $IMAGE_NAME:test-${{ matrix.variant }} $IMAGE_NAME:dev-${{ matrix.variant }}
+          docker push $IMAGE_NAME:dev-${{ matrix.variant }}
+          echo "✅ ${{ matrix.variant }} image promoted from :test-${{ matrix.variant }} to :dev-${{ matrix.variant }}"
 
-  cleanup-test-image:
-    name: Delete test image from registry
+  cleanup-test-images:
+    name: Delete ${{ matrix.variant }} test image from registry
     needs:
-      - build-and-push-image
-      - run-tests
-      - promote-image
-    if: always() && needs.build-and-push-image.result == 'success'
+      - build-and-push-images
+      - run-tests-sglang
+      - run-tests-vllm
+      - promote-images
+    if: always() && needs.build-and-push-images.result == 'success'
     runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        variant: [sglang, vllm]
     steps:
       - name: Delete test image from GHCR
         env:
           GH_TOKEN: ${{ secrets.GHCR_TOKEN }}
         run: |
+          PACKAGE_NAME="areal-runtime"
+          TAG="test-${{ matrix.variant }}"
           # Get the package version ID for the test tag
           PACKAGE_VERSION_ID=$(curl -s -H "Authorization: Bearer $GH_TOKEN" \
-            "https://api.github.com/orgs/inclusionai/packages/container/areal-runtime/versions" \
-            | jq -r '.[] | select(.metadata.container.tags[] == "test") | .id')
+            "https://api.github.com/orgs/inclusionai/packages/container/$PACKAGE_NAME/versions?per_page=100" \
+            | jq -r ".[] | select(.metadata.container.tags[] == \"$TAG\") | .id")
 
           if [ -n "$PACKAGE_VERSION_ID" ] && [ "$PACKAGE_VERSION_ID" != "null" ]; then
             curl -X DELETE -H "Authorization: Bearer $GH_TOKEN" \
-              "https://api.github.com/orgs/inclusionai/packages/container/areal-runtime/versions/$PACKAGE_VERSION_ID"
-            echo "✅ Deleted test image from registry"
+              "https://api.github.com/orgs/inclusionai/packages/container/$PACKAGE_NAME/versions/$PACKAGE_VERSION_ID"
+            echo "✅ Deleted ${{ matrix.variant }} test image from registry"
           else
-            echo "⚠️ Test image not found or already deleted"
+            echo "⚠️ ${{ matrix.variant }} test image not found or already deleted"
           fi
 
   stop-builder:
     name: Stop areal-docker-builder instance
     needs:
       - start-builder
-      - build-and-push-image
-      - run-tests
-      - promote-image
-      - cleanup-test-image
+      - build-and-push-images
+      - run-tests-sglang
+      - run-tests-vllm
+      - promote-images
+      - cleanup-test-images
     if: always() && needs.start-builder.outputs.was_running != 'true'
     runs-on: ubuntu-latest
     env:

.github/workflows/install-test.yml

@@ -77,8 +77,13 @@ jobs:
     # Test CUDA extras installation on Linux only (CUDA packages have Linux-only wheels)
     # Note: flash-attn is excluded because it requires CUDA compilation infrastructure
     # which is not available on standard GitHub runners
+    # sglang and vllm are conflicting extras, so we test each separately
     runs-on: ubuntu-latest
-    name: Install test (Linux, CUDA extras)
+    strategy:
+      fail-fast: false
+      matrix:
+        variant: [sglang, vllm]
+    name: Install test (Linux, CUDA extras, ${{ matrix.variant }})
 
     steps:
       - name: Checkout repository
@@ -96,17 +101,20 @@ jobs:
       - name: Install package with CUDA extras (excluding flash-attn)
         # flash-attn requires CUDA toolkit for compilation, skip it in CI
         # Test individual extras that have pre-built wheels
-        run: uv sync --extra vllm --extra sglang --extra megatron --extra tms
+        run: uv sync --extra ${{ matrix.variant }} --extra megatron --extra tms
 
       - name: Verify package import with CUDA extras
         run: |
           uv run python -c "import areal; print(f'areal version: {areal.__version__}')"
 
       - name: Verify CUDA-dependent packages are installed
+        env:
+          VARIANT: ${{ matrix.variant }}
         run: |
           uv run python -c "
-          import importlib.util
-          packages = ['vllm', 'sglang', 'megatron.core']
+          import os, importlib.util
+          variant = os.environ['VARIANT']
+          packages = [variant, 'megatron.core']
           for pkg in packages:
               spec = importlib.util.find_spec(pkg)
               status = 'installed' if spec else 'not found'
@@ -120,9 +128,13 @@ jobs:
     # and `no space on device left` errors on busy runners
     if: github.event_name == 'workflow_dispatch'
     runs-on: ubuntu-latest
-    name: Install test (Docker runtime image)
+    strategy:
+      fail-fast: false
+      matrix:
+        variant: [sglang, vllm]
+    name: Install test (Docker ${{ matrix.variant }} image)
     container:
-      image: ghcr.io/inclusionai/areal-runtime:v1.0.1
+      image: ghcr.io/inclusionai/areal-runtime:dev-${{ matrix.variant }}
 
     steps:
       - name: Checkout repository

.github/workflows/tag-release-image.yml

@@ -26,6 +26,8 @@ jobs:
   start-builder:
     name: Start areal-docker-builder instance
     runs-on: ubuntu-latest
+    outputs:
+      was_running: ${{ steps.start-instance.outputs.was_running }}
     env:
       INSTANCE_NAME: areal-docker-builder
       INSTANCE_ZONE: us-central1-f
@@ -39,6 +41,7 @@ jobs:
         uses: google-github-actions/setup-gcloud@v3
 
       - name: Start builder instance if stopped
+        id: start-instance
         run: |
           set -euo pipefail
 
@@ -55,12 +58,14 @@ jobs:
 
           if [ "$status" = "RUNNING" ]; then
             echo "Instance $INSTANCE_NAME is already running."
+            echo "was_running=true" >> $GITHUB_OUTPUT
           elif [ "$status" = "TERMINATED" ] || [ "$status" = "SUSPENDED" ]; then
             echo "Instance $INSTANCE_NAME is $status. Starting it..."
             gcloud compute instances start "$INSTANCE_NAME" \
               --project "$GCP_PROJECT_ID" \
               --zone "$INSTANCE_ZONE"
             echo "Instance started successfully."
+            echo "was_running=false" >> $GITHUB_OUTPUT
           else
             echo "Instance $INSTANCE_NAME has unexpected status: $status" >&2
             exit 1
@@ -113,9 +118,9 @@ jobs:
   build-and-push-release:
     needs:
       - start-builder
-    name: Build and push release Docker image
+    name: Build and push release Docker images
     runs-on: [self-hosted, areal-docker-builder]
-    timeout-minutes: 180
+    timeout-minutes: 300
     outputs:
       package_version: ${{ steps.get-version.outputs.version }}
       release_tag: ${{ steps.get-version.outputs.release_tag }}
@@ -150,25 +155,41 @@ jobs:
           username: inclusionai
           password: ${{ secrets.GHCR_TOKEN }}
 
-      - name: Build and push Docker image
+      - name: Build and push sglang image
         uses: docker/build-push-action@v7
         with:
           context: .
           file: ./Dockerfile
           push: true
           tags: |
-            ${{ env.IMAGE_NAME }}:${{ steps.get-version.outputs.release_tag }}
+            ${{ env.IMAGE_NAME }}:${{ steps.get-version.outputs.release_tag }}-sglang
             ${{ env.IMAGE_NAME }}:latest
-            ${{ env.IMAGE_NAME }}:dev
+            ${{ env.IMAGE_NAME }}:dev-sglang
+          build-args: |
+            VARIANT=sglang
+
+      - name: Build and push vllm image
+        uses: docker/build-push-action@v7
+        with:
+          context: .
+          file: ./Dockerfile
+          push: true
+          tags: |
+            ${{ env.IMAGE_NAME }}:${{ steps.get-version.outputs.release_tag }}-vllm
+            ${{ env.IMAGE_NAME }}:dev-vllm
+          build-args: |
+            VARIANT=vllm
 
       - name: Image details
         run: |
-          echo "Docker image built and pushed successfully!"
+          echo "Docker images built and pushed successfully!"
           echo "Image: ${{ env.IMAGE_NAME }}"
           echo "Tags:"
-          echo "  - ${{ steps.get-version.outputs.release_tag }}"
-          echo "  - latest"
-          echo "  - dev"
+          echo "  - ${{ steps.get-version.outputs.release_tag }}-sglang"
+          echo "  - ${{ steps.get-version.outputs.release_tag }}-vllm"
+          echo "  - latest (sglang)"
+          echo "  - dev-sglang"
+          echo "  - dev-vllm"
           echo "Release: ${{ github.event.release.name || github.event.inputs.tag }}"
           echo "Commit: ${{ github.sha }}"
 
@@ -177,7 +198,7 @@ jobs:
     needs:
       - build-and-push-release
       - start-builder
-    if: always()
+    if: always() && needs.start-builder.outputs.was_running != 'true'
     runs-on: ubuntu-latest
     env:
       INSTANCE_NAME: areal-docker-builder