e2e-prefill-heavy-gke.yaml

name: GKE Prefill Heavy Test

on:
  # Runs with a PR comment /run-gke-prefill-heavy
  issue_comment:
    types: [created]
  workflow_dispatch:
    inputs:
      pr_or_branch:
        description: 'Pull-request number or branch name to test'
        required: true
        default: 'main'
        type: string

permissions:
  contents: read
  pull-requests: write

jobs:
  # Authorization Job: Ensures only authorized users can execute workflow
  # Note, even if user checks out branch to modify access, user will need to provide correct secret keys to deploy to GCP.
  check_access:
    runs-on: ubuntu-latest
    
    if: |
      (github.event_name == 'issue_comment' &&
      github.event.issue.pull_request &&
      contains(github.event.comment.body, '/run-gke-prefill-heavy')) || github.event_name == 'workflow_dispatch'
    
    outputs:
      authorized: ${{ steps.auth_logic.outputs.authorized }}

    steps:
      - name: Checkout Repository
        uses: actions/checkout@v4
        
      - name: Authorization Logic
        id: auth_logic
        shell: bash
        run: |
          authorized='false'
          auth_file=".github/authorized_workflow_users.txt"
          user=""
          role=""
          
          if [[ "${{ github.event_name }}" == "issue_comment" ]]; then
            user="${{ github.event.comment.user.login }}"
            role="${{ github.event.comment.author_association }}"
            
          elif [[ "${{ github.event_name }}" == "workflow_dispatch" ]]; then
            user="${{ github.actor }}"
          fi

          if [[ "$role" == "OWNER" || "$role" == "MAINTAINER" ]]; then
            echo "User authorized by role: $role"
            authorized='true'          
          
          elif grep -Fxq "$user" "$auth_file"; then
            echo "User authorized by file lookup: $auth_file"
            authorized='true'
          fi
          
          echo "authorized=$authorized" >> "$GITHUB_OUTPUT"

  deploy_and_validate:
    needs: [check_access]
    if: |
      (github.event_name == 'workflow_dispatch' || github.event_name == 'issue_comment') &&
        needs.check_access.outputs.authorized == 'true'
      
    name: Test on ${{ matrix.accelerator.name }}
    runs-on: ubuntu-latest

    strategy:
      fail-fast: false
      max-parallel: 1
      matrix:
        accelerator:
          - name: GPU

    env:
      GCP_PROJECT_ID: llm-d-scale
      GKE_CLUSTER_NAME: llm-d-e2e-us-east5
      GKE_CLUSTER_ZONE: us-east5
      NAMESPACE: igw-prefill-heavy
      GATEWAY: gke-l7-regional-external-managed
      GATEWAY_PROVIDER: gke
      PR_OR_BRANCH: ${{ github.event.inputs.pr_or_branch || github.event.issue.number || 'main' }}
      HF_TOKEN: ${{ secrets.HF_TOKEN }}
      DEPLOYMENT_NAME: vllm-qwen3-32b
      MODEL: Qwen/Qwen3-32B
      GSA_EMAIL: ${{ secrets.GCS_WORKLOAD_SA }}
      GCS_BUCKET: igw-e2e-benchmark-results
      KSA_NAME: igw-e2e-benchmark-sa

    steps:
      - name: Checkout
        uses: actions/checkout@v4
        with:
          persist-credentials: false

      - name: Determine if pr_or_branch is a PR number
        id: check_pr
        shell: bash
        run: |
          VAL="${{ env.PR_OR_BRANCH }}"
          echo "TARGET=$VAL" >> "$GITHUB_ENV"
          if [[ "$VAL" =~ ^[0-9]+$ ]]; then
            echo "is_pr=true" >> "$GITHUB_OUTPUT"
          else
            echo "is_pr=false" >> "$GITHUB_OUTPUT"
          fi

      - name: Fetch and checkout PR
        if: steps.check_pr.outputs.is_pr == 'true'
        run: |
          git fetch origin pull/"$TARGET"/head:pr-"$TARGET"
          git checkout pr-"$TARGET"

      - name: Checkout branch
        if: steps.check_pr.outputs.is_pr == 'false'
        run: |
          git fetch origin "$TARGET"
          git checkout "$TARGET"

      - name: Install Python Dependencies
        run: |
          cd benchmarking
          python3 -m venv .venv
          pip install inference-perf pandas numpy matplotlib

      - name: Authenticate to Google Cloud
        id: auth
        uses: google-github-actions/auth@b7593ed2efd1c1617e1b0254da33b86225adb2a5
        with:
          credentials_json: ${{ secrets.GCP_SA_KEY }}

      - name: Set up gcloud CLI and kubectl
        uses: google-github-actions/setup-gcloud@cb1e50a9932213ecece00a606661ae9ca44f3397
        with:
          project_id: ${{ env.GCP_PROJECT_ID }}
          install_components: 'kubectl,gke-gcloud-auth-plugin'

      - name: Get GKE credentials
        run: |
          gcloud container clusters get-credentials "${{ env.GKE_CLUSTER_NAME }}" --zone "${{ env.GKE_CLUSTER_ZONE }}"

      - name: Create namespace
        run: |
          kubectl create namespace "${NAMESPACE}" || echo "Namespace already exists"

      - name: Create hf-token secret
        run: |
          kubectl create secret generic hf-token \
            --from-literal="token=${{ secrets.HF_TOKEN }}" \
            --namespace "${NAMESPACE}" \
            --dry-run=client -o yaml | kubectl apply -f -

      - name: Create and Annotate KSA for Workload Identity
        run: |     
          kubectl create serviceaccount $KSA_NAME --namespace "${NAMESPACE}" --dry-run=client -o yaml | kubectl apply -f -
          kubectl annotate serviceaccount $KSA_NAME \
            iam.gke.io/gcp-service-account=$GSA_EMAIL \
            --overwrite \
            --namespace "${NAMESPACE}"

      - name: Fetch Latest IGW Release
        run: |
          IGW_LATEST_RELEASE=$(curl -s https://api.github.com/repos/kubernetes-sigs/gateway-api-inference-extension/releases \
            | jq -r '.[] | select(.prerelease == false) | .tag_name' \
            | sort -V \
            | tail -n1)
                    
          echo "IGW_CHART_VERSION=$IGW_LATEST_RELEASE" >> $GITHUB_ENV

      - name: Deploy Model Server and CRDs
        run: |
          cd config/manifests/vllm
          echo "Deploying Model Server..."
          kubectl apply -f gpu-deployment.yaml -n ${NAMESPACE} | tee ~/igw-prefill-heavy-deployment.log
          kubectl scale deployment ${DEPLOYMENT_NAME} -n ${NAMESPACE} --replicas=8
          echo "Installing CRDs"
          kubectl apply -f https://github.com/kubernetes-sigs/gateway-api-inference-extension/releases/download/${IGW_CHART_VERSION}/manifests.yaml
          echo "---------------------------------------" >> ~/igw-prefill-heavy-deployment.log

      - name: Wait for all pods to be ready
        run: |
          kubectl wait pod \
            --for=condition=Ready \
            --all \
            -n "${NAMESPACE}" \
            --timeout=25m
          echo "✅ All pods are ready."
          kubectl get pods -n "${NAMESPACE}"

      - name: Deploy Gateway
        run: |
          GATEWAY_NAME=inference-gateway
          echo "Deploying Gateway..."
          kubectl apply -f https://raw.githubusercontent.com/kubernetes-sigs/gateway-api-inference-extension/refs/tags/${IGW_CHART_VERSION}/config/manifests/gateway/gke/gateway.yaml -n ${NAMESPACE} | tee ~/igw-prefill-heavy-deployment.log 
          echo "---------------------------------------" >> ~/igw-prefill-heavy-deployment.log

      - name: Wait for gateway to be ready
        run: |
          GATEWAY_NAME=inference-gateway
          kubectl wait gateway/${GATEWAY_NAME} \
            --for=condition=Programmed=True \
            -n "${NAMESPACE}" \
            --timeout=500s
          echo "✅ Gateway is ready."
          kubectl get gateway -n "${NAMESPACE}"

      - name: Deploy InferencePool and Endpoint Picker Extension
        run: |
          helm upgrade --install ${DEPLOYMENT_NAME} \
          --dependency-update \
          --namespace $NAMESPACE \
          --set inferencePool.modelServers.matchLabels.app=${DEPLOYMENT_NAME} \
          --set provider.name=$GATEWAY_PROVIDER \
          --set experimentalHttpRoute.enabled=true \
          --version $IGW_CHART_VERSION \
          oci://registry.k8s.io/gateway-api-inference-extension/charts/inferencepool | tee ~/igw-prefill-heavy-deployment.log
          echo "---------------------------------------" >> ~/igw-prefill-heavy-deployment.log

      - name: Wait for all resources to be ready
        run: |
          kubectl rollout status deployment/$DEPLOYMENT_NAME-epp -n $NAMESPACE --timeout=120s

          echo "Allow extra time to initialize resources..."
          sleep 30

          kubectl wait httproute/${DEPLOYMENT_NAME} \
            --for=jsonpath='{.status.parents[0].conditions[?(@.type=="ResolvedRefs")].status}'=True \
            -n ${NAMESPACE} \
            --timeout=5m

          kubectl wait inferencepool/${DEPLOYMENT_NAME} \
            --for=jsonpath='{.status.parents[0].conditions[?(@.type=="ResolvedRefs")].status}'=True \
            -n ${NAMESPACE} \
            --timeout=5m

      - name: Show deployment status
        run: |
          echo "=== Deployments ==="
          kubectl get deployments -n "${NAMESPACE}"
          echo ""
          echo "=== Pods ==="
          kubectl get pods -n "${NAMESPACE}"
          echo ""
          echo "=== Services ==="
          kubectl get svc -n "${NAMESPACE}"
          echo ""
          echo "=== Helm releases ==="
          helm list -n "${NAMESPACE}" || true
          echo ""
          echo "=== Inference Pools ==="
          kubectl get inferencepools -n "${NAMESPACE}" -o yaml || true
          echo ""
          echo "=== HTTPRoutes ==="
          kubectl get httproutes -n "${NAMESPACE}" -o yaml || true
          echo ""
          echo "=== Gateway ==="
          kubectl get Gateway -n "${NAMESPACE}" || true
          echo ""

      - name: Verify installation and run validation test
        run: |
          cd .github/scripts/e2e
          ./e2e-validate.sh -n "${NAMESPACE}" -v -m ${MODEL}

      - name: Run benchmarking test
        run: |
          TIMESTAMP=$(date +"%Y-%m-%d-%H-%M-%S")
          echo "TIMESTAMP=$TIMESTAMP" >> "$GITHUB_ENV"
          cd benchmarking/single-workload
          host="${GATEWAY_HOST:-$(kubectl get gateway -n "$NAMESPACE" \
          -o jsonpath='{.items[0].status.addresses[0].value}' 2>/dev/null || true)}"
          if [[ -z "$host" ]]; then
            echo "Error: could not discover a Gateway address in namespace '$NAMESPACE'." >&2
            exit 1
          fi
          port=80
          svc_host="${host}:${port}"
          helm upgrade --install prefill-heavy-benchmark ../inference-perf/ -f prefill-heavy-values.yaml \
            --namespace "${NAMESPACE}" \
            --create-namespace \
            --set token.hfToken="${HF_TOKEN}" \
            --set "config.server.base_url=http://${svc_host}" \
            --set "job.serviceAccountName=$KSA_NAME" \
            --set "job.image.tag=latest" \
            --set "config.storage.google_cloud_storage.bucket_name=${GCS_BUCKET}" \
            --set "config.storage.google_cloud_storage.path=${NAMESPACE}/${TIMESTAMP}" \
            --set "gcsPath=${GCS_BUCKET}/datasets/billsum_conversations.json" \
            --set "config.data.path=/gcsDataset/gcs-dataset.json" \
            --set-string 'job.resources.limits.nvidia\.com/gpu=1'

      - name: Wait for benchmarking job to finish
        run: |
          job_name=prefill-heavy-benchmark-inference-perf-job
          TIMEOUT_DURATION="7200s"
          if ! kubectl wait --for=condition=complete job/"$job_name" -n "$NAMESPACE" --timeout="$TIMEOUT_DURATION"; then
            echo "Error: Benchmark job $job_name did not complete successfully within $TIMEOUT_DURATION." >&2
            echo "--- Job Description ---" >&2
            kubectl describe job "$job_name" -n "$NAMESPACE" >&2
            echo "--- Pod Logs (Last 50 lines) ---" >&2
            kubectl logs -l job-name="$job_name" -n "$NAMESPACE" --all-containers=true --tail 50 >&2
            exit 1
          fi
          echo "✅ Benchmarking Job Completed."

      - name: Analyze Benchmarking Results
        run: |
          cd benchmarking
          python3 -m venv .venv
          source .venv/bin/activate

          benchmark_id='prefill-heavy' ./download-results.bash gcs $GCS_BUCKET ${NAMESPACE}/${TIMESTAMP}
          inference-perf --analyze output/default-run/prefill-heavy/results/json/${TIMESTAMP}

      - name: Upload Benchmark Results
        uses: actions/upload-artifact@v4
        with:
          name: benchmarking-artifacts
          path: benchmarking/output/default-run/prefill-heavy/results/json/**/*

      - name: Comment on PR
        if: steps.check_pr.outputs.is_pr == 'true'
        env:
          GH_TOKEN: ${{ github.token }}
          PR_NUMBER: ${{ github.event.pull_request.number }}
          RUN_ID: ${{ github.run_id }}
        run: |
          cd benchmarking/output/default-run/prefill-heavy/results/json/${TIMESTAMP}
          FILE_LIST=$(ls | sed 's/^/- /')

          COMMENT_BODY=$(cat <<EOF
          ✅ **Benchmarking Job Completed!**
          
          You can find the benchmarking artifacts in the [Workflow Summary](https://github.com{{ github.repository }}/actions/runs/$RUN_ID).
          
          **Files generated:**
          $FILE_LIST
          EOF
          )

          gh pr comment $PR_NUMBER --body "$COMMENT_BODY"

      - name: Collect and upload Kubernetes pod logs
        if: always()
        run: |
            mkdir -p pod-logs-inference-prefill-heavy
            cd pod-logs-inference-prefill-heavy
            echo "Fetching ${NAMESPACE} pods log..."
            kubectl get pods -n "${NAMESPACE}" --no-headers -o custom-columns=":metadata.name" \
            | xargs -I{} sh -c 'kubectl logs --all-containers=true -n "${NAMESPACE}" {} > "{}.log" 2>&1'
            echo "Fetching ${NAMESPACE} pods descriptions..."
            kubectl get pods -n "${NAMESPACE}" --no-headers -o custom-columns=":metadata.name" \
            | xargs -I{} sh -c 'kubectl describe pod -n "${NAMESPACE}" {} > "{}-describe.log" 2>&1'
            mv ~/igw-prefill-heavy-deployment.log . || true
            mv ~/install-deps.log . || true

      - name: Upload pod logs as artifact
        uses: actions/upload-artifact@v4
        if: always()
        with:
          name: igw-pod-logs-inference-prefill-heavy-${{ matrix.accelerator.name }}
          path: pod-logs-inference-prefill-heavy

      - name: Send Google Chat notification on failure
        if: failure()
        uses: SimonScholz/google-chat-action@3b3519e5102dba8aa5046fd711c4b553586409bb
        with:
          webhookUrl: ${{ secrets.GOOGLE_CHAT_WEBHOOK }}
          jobStatus: ${{ job.status }}
          title: '${{ github.workflow }} - ${{ matrix.accelerator.name }}'

      - name: Cleanup deployment
        if: always()
        run: |
          GATEWAY_NAME=inference-gateway
          helm uninstall ${DEPLOYMENT_NAME} -n ${NAMESPACE} --ignore-not-found
          helm uninstall prefill-heavy-benchmark -n ${NAMESPACE} --ignore-not-found
          kubectl delete httproute llm-route -n ${NAMESPACE} --ignore-not-found
          kubectl delete gateway ${GATEWAY_NAME} -n ${NAMESPACE} --ignore-not-found
          kubectl delete deployment ${DEPLOYMENT_NAME} -n ${NAMESPACE} --ignore-not-found