diff --git a/.github/actions/run-inference/action.yml b/.github/actions/run-inference/action.yml
new file mode 100644
index 00000000..314a8be5
--- /dev/null
+++ b/.github/actions/run-inference/action.yml
@@ -0,0 +1,62 @@
+name: Run Inference
+description: Run one llama-tornado inference pass and write the metrics + sidecar files.
+
+inputs:
+  backend:
+    description: 'GPU backend (opencl or ptx)'
+    required: true
+  model_file:
+    description: 'Model filename inside $MODELS_DIR (e.g. Llama-3.2-1B-Instruct-F16.gguf)'
+    required: true
+  model:
+    description: 'Human-readable model name for the sidecar (e.g. Llama-3.2-1B-Instruct)'
+    required: true
+  quantization:
+    description: 'Quantization type (e.g. F16, Q8_0)'
+    required: true
+  configuration:
+    description: 'Configuration key for the sidecar (e.g. standard, prefill-decode)'
+    required: true
+  flags:
+    description: 'Extra CLI flags passed to llama-tornado (omit for standard run)'
+    required: false
+    default: ''
+  metrics_file:
+    description: 'Absolute path for the output metrics JSON file'
+    required: true
+  prompt:
+    description: 'Prompt to pass to the model'
+    required: false
+    default: 'Say hello'
+
+runs:
+  using: composite
+  steps:
+    - name: Run inference
+      shell: bash
+      working-directory: ${{ github.workspace }}
+      env:
+        JAVA_TOOL_OPTIONS: >-
+          -Dllama.metrics.format=json
+          -Dllama.metrics.output=file
+          -Dllama.metrics.file=${{ inputs.metrics_file }}
+        METRICS_FILE: ${{ inputs.metrics_file }}
+      run: |
+        # Run inference and emit raw metrics JSON via JAVA_TOOL_OPTIONS
+        ./llama-tornado --gpu --${{ inputs.backend }} \
+          --model $MODELS_DIR/${{ inputs.model_file }} \
+          --prompt "${{ inputs.prompt }}" \
+          ${{ inputs.flags }}
+
+        # Write metadata sidecar so process_metrics.py can identify each metrics file
+        SIDECAR="${METRICS_FILE%.json}.meta.json"
+        python3 scripts/write_metrics_sidecar.py \
+          --out "$SIDECAR" \
+          backend="${{ inputs.backend }}" \
+          task=llama-inference \
+          model_file=${{ inputs.model_file }} \
+          model=${{ inputs.model }} \
+          quantization=${{ inputs.quantization }} \
+          configuration=${{ inputs.configuration }} \
+          "flags=${{ inputs.flags }}" \
+          prompt="${{ inputs.prompt }}"
diff --git a/.github/actions/setup-java/action.yml b/.github/actions/setup-java/action.yml
new file mode 100644
index 00000000..334b5432
--- /dev/null
+++ b/.github/actions/setup-java/action.yml
@@ -0,0 +1,22 @@
+name: Setup Java
+description: Install and activate a JDK via SDKMAN. Exports JAVA_HOME and updates PATH.
+
+inputs:
+  java_version:
+    description: 'SDKMAN Java version identifier (e.g. 21.0.2-open)'
+    required: true
+
+runs:
+  using: composite
+  steps:
+    - name: Set up Java with SDKMAN
+      shell: bash
+      run: |
+        source "$HOME/.sdkman/bin/sdkman-init.sh"
+        if ! sdk list java | grep -q "${{ inputs.java_version }}"; then
+          sdk install java "${{ inputs.java_version }}"
+        fi
+        sdk use java "${{ inputs.java_version }}"
+        echo "JAVA_HOME=$HOME/.sdkman/candidates/java/current" >> $GITHUB_ENV
+        echo "$HOME/.sdkman/candidates/java/current/bin" >> $GITHUB_PATH
+        java -version
diff --git a/.github/actions/setup-tornadovm/action.yml b/.github/actions/setup-tornadovm/action.yml
new file mode 100644
index 00000000..3b1c5070
--- /dev/null
+++ b/.github/actions/setup-tornadovm/action.yml
@@ -0,0 +1,81 @@
+name: Setup TornadoVM
+description: Build TornadoVM once per backend and reuse across runs via a local SHA sentinel. Exports TORNADOVM_HOME and updates PATH for all subsequent steps.
+
+inputs:
+  backend:
+    description: 'TornadoVM backend to build (opencl or ptx)'
+    required: true
+
+runs:
+  using: composite
+  steps:
+    - name: Get TornadoVM HEAD SHA
+      id: tornado_sha
+      shell: bash
+      run: |
+        SHA=$(git ls-remote https://github.com/beehive-lab/TornadoVM HEAD | cut -f1)
+        echo "sha=$SHA" >> $GITHUB_OUTPUT
+
+    - name: Check local build sentinel
+      id: sentinel
+      shell: bash
+      run: |
+        SENTINEL="$TORNADO_ROOT/.built-sha"
+        if [ -f "$SENTINEL" ] && [ "$(cat $SENTINEL)" = "${{ steps.tornado_sha.outputs.sha }}" ]; then
+          echo "up-to-date=true" >> $GITHUB_OUTPUT
+        else
+          echo "up-to-date=false" >> $GITHUB_OUTPUT
+        fi
+
+    - name: Clone TornadoVM master
+      if: steps.sentinel.outputs.up-to-date != 'true'
+      shell: bash
+      run: |
+        rm -rf $TORNADO_ROOT
+        git clone --depth 1 --branch master \
+          https://github.com/beehive-lab/TornadoVM.git \
+          $TORNADO_ROOT
+
+    - name: Set up Python venv for TornadoVM
+      if: steps.sentinel.outputs.up-to-date != 'true'
+      shell: bash
+      run: |
+        python3 -m venv $TORNADO_ROOT/venv
+        source $TORNADO_ROOT/venv/bin/activate
+        python --version
+
+    - name: Build TornadoVM
+      if: steps.sentinel.outputs.up-to-date != 'true'
+      shell: bash
+      run: |
+        cd $TORNADO_ROOT
+        mkdir -p graalJars && cp $GRAAL_JARS/* graalJars/
+        source venv/bin/activate
+        echo "=== Building TornadoVM ==="
+        make BACKEND=${{ inputs.backend }}
+        echo "${{ steps.tornado_sha.outputs.sha }}" > .built-sha
+
+    - name: Find TornadoVM SDK directory
+      id: find_sdk
+      shell: bash
+      run: |
+        SDK_DIR=$(find $TORNADO_ROOT/dist -type d -maxdepth 3 -path "*/tornadovm-*-${{ inputs.backend }}" | head -n 1)
+        if [ -z "$SDK_DIR" ]; then
+          echo "::error::Could not locate TornadoVM SDK directory!"
+          find $TORNADO_ROOT/dist -maxdepth 5 -type d
+          exit 1
+        fi
+        echo "sdk_dir=$SDK_DIR" >> $GITHUB_OUTPUT
+
+    # Runs on both fresh build and sentinel hit — sets TORNADOVM_HOME and PATH
+    # for all subsequent steps in the calling workflow.
+    - name: Configure TornadoVM environment
+      shell: bash
+      run: |
+        FULL_SDK="${{ steps.find_sdk.outputs.sdk_dir }}"
+        echo "Detected TornadoVM SDK: $FULL_SDK"
+        echo "TORNADOVM_HOME=$FULL_SDK" >> $GITHUB_ENV
+        echo "$FULL_SDK/bin" >> $GITHUB_PATH
+        echo "$JAVA_HOME/bin" >> $GITHUB_PATH
+        which tornado || { echo "::error::tornado not in PATH"; exit 1; }
+        tornado --devices
diff --git a/.github/workflows/build-and-run.yml b/.github/workflows/build-and-run.yml
index a17c9228..2a19099f 100644
--- a/.github/workflows/build-and-run.yml
+++ b/.github/workflows/build-and-run.yml
@@ -8,11 +8,10 @@ on:
     types: [opened, synchronize, reopened]
 
 env:
-  JAVA_HOME: /opt/jenkins/jdks/graal-23.1.0/jdk-21.0.3
-  TORNADO_ROOT: ${{ github.workspace }}/GPULlama3.java/external/tornadovm
-  LLAMA_ROOT: ${{ github.workspace }}
+  JAVA_VERSION: 21.0.2-open
   GRAAL_JARS: /opt/graalJars
   MODELS_DIR: /opt/models
+  QUARKUS_PORT: 8081
   # History file committed back to the repo on push to main
   PERF_HISTORY_FILE: docs/perf-history.jsonl
 
@@ -31,14 +30,17 @@ jobs:
           cd ${{ github.workspace }}
           # ./mvnw -T12C -Pspotless spotless:check
 
-  build-and-run:
+  # Build: TornadoVM → GPULlama3 → Quarkus LangChain4j
+  # max-parallel: 1 ensures the opencl and ptx variants run sequentially so
+  # there are no workspace conflicts between matrix jobs.
+  build:
     if: github.repository == 'beehive-lab/GPULlama3.java'
     runs-on: [self-hosted]
     needs: code-quality
     timeout-minutes: 30
-
     strategy:
       fail-fast: true
+      max-parallel: 1
       matrix:
         backend:
           - name: opencl
@@ -47,483 +49,265 @@ jobs:
     steps:
       - name: Checkout GPULlama3
         uses: actions/checkout@v4
+        with:
+          clean: false
 
-      - name: Clone TornadoVM master
-        run: |
-          git clone --depth 1 --branch master \
-            https://github.com/beehive-lab/TornadoVM.git \
-            $TORNADO_ROOT
-
-      - name: Set up Python venv for TornadoVM
-        run: |
-          python3 -m venv $TORNADO_ROOT/venv
-          source $TORNADO_ROOT/venv/bin/activate
-          python --version
+      - name: Set up Java
+        uses: ./.github/actions/setup-java
+        with:
+          java_version: ${{ env.JAVA_VERSION }}
 
-      - name: Build TornadoVM
-        run: |
-          cd $TORNADO_ROOT
-          mkdir -p graalJars && cp $GRAAL_JARS/* graalJars/
-          source venv/bin/activate
-          echo "=== Building TornadoVM ==="
-
-          make BACKEND=${{ matrix.backend.name }}
-
-          echo "=== Searching for TornadoVM SDK directory ==="
-          SDK_DIR=$(find dist -type d -maxdepth 3 -path "*/tornadovm-*-${{ matrix.backend.name }}" | head -n 1)
-          if [ -z "$SDK_DIR" ]; then
-            echo "::error::Could not locate TornadoVM SDK directory!"
-            find dist -maxdepth 5 -type d
-            exit 1
-          fi
-          FULL_SDK="${PWD}/${SDK_DIR}"
-          echo "Detected TornadoVM SDK: $FULL_SDK"
-
-          # Export for current shell session
-          export TORNADOVM_HOME="$FULL_SDK"
-          export PATH="$FULL_SDK/bin:$JAVA_HOME/bin:$PATH"
-
-          # Save for subsequent steps
-          echo "TORNADOVM_HOME=$FULL_SDK" >> $GITHUB_ENV
-          echo "PATH=$PATH" >> $GITHUB_ENV
-
-          echo "=== Checking tornado CLI ==="
-          which tornado || { echo "::error::tornado not in PATH"; exit 1; }
-          tornado --devices
+      - name: Setup TornadoVM
+        uses: ./.github/actions/setup-tornadovm
+        env:
+          TORNADO_ROOT: ${{ runner.tool_cache }}/tornadovm/tornadovm-${{ matrix.backend.name }}
+        with:
+          backend: ${{ matrix.backend.name }}
 
       - name: Build GPULlama3.java
         run: |
-          cd ${{ github.workspace }}
-          echo "Using TORNADOVM_HOME=$TORNADOVM_HOME"
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
           tornado --version
-          ./mvnw clean package -DskipTests
+          # Strip any pre-existing -SNAPSHOT suffix before appending, making this step idempotent
+          # across sequential matrix variants (ptx runs after opencl on the same workspace).
+          BASE_VERSION=$(./mvnw help:evaluate -Dexpression=project.version -q -DforceStdout | sed 's/-SNAPSHOT$//')
+          GPULLAMA3_VERSION="${BASE_VERSION}-SNAPSHOT"
+          echo "GPULlama3 version: $GPULLAMA3_VERSION"
+          ./mvnw versions:set -DnewVersion=$GPULLAMA3_VERSION
+          ./mvnw clean install -DskipTests
+          echo "GPULLAMA3_VERSION=$GPULLAMA3_VERSION" >> $GITHUB_ENV
+
+      - name: Clone Quarkus LangChain4j
+        run: |
+          rm -rf quarkus-langchain4j
+          git clone --depth 1 https://github.com/quarkiverse/quarkus-langchain4j.git
 
-      - name: FP16 - Run Llama-3.2-1B-Instruct-F16.gguf - Standard
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-f16-standard.json
+      - name: Build Quarkus LangChain4j
         run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/Llama-3.2-1B-Instruct-F16.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-f16-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=Llama-3.2-1B-Instruct-F16.gguf \
-            model=Llama-3.2-1B-Instruct \
-            quantization=F16 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+          cd ${{ github.workspace }}/quarkus-langchain4j
+          sed -i 's/<gpu-llama3\.version>.*<\/gpu-llama3\.version>/<gpu-llama3.version>'$GPULLAMA3_VERSION'<\/gpu-llama3.version>/' pom.xml
+          # -Dtornado activates the TornadoVM profile; -am builds only the gpu-llama3 module + deps
+          mvn clean install -pl integration-tests/gpu-llama3 -am -DskipTests -Dtornado
 
-      - name: FP16 - Run Llama-3.2-1B-Instruct-F16.gguf - Prefill-Decode
+  standalone-inference:
+    if: github.repository == 'beehive-lab/GPULlama3.java'
+    runs-on: [self-hosted]
+    needs: build
+    timeout-minutes: 30
+    strategy:
+      fail-fast: true
+      matrix:
+        backend:
+          - name: opencl
+          - name: ptx
+
+    steps:
+      - name: Checkout GPULlama3
+        uses: actions/checkout@v4
+        with:
+          clean: false
+
+      - name: Set up Java
+        uses: ./.github/actions/setup-java
+        with:
+          java_version: ${{ env.JAVA_VERSION }}
+
+      - name: Setup TornadoVM
+        uses: ./.github/actions/setup-tornadovm
         env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-f16-prefill-decode.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/Llama-3.2-1B-Instruct-F16.gguf \
-            --prompt "Say hello" \
-            --with-prefill-decode
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-f16-prefill-decode.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=Llama-3.2-1B-Instruct-F16.gguf \
-            model=Llama-3.2-1B-Instruct \
-            quantization=F16 \
-            configuration=prefill-decode \
-            "flags=--with-prefill-decode" \
-            prompt="Say hello"
+          TORNADO_ROOT: ${{ runner.tool_cache }}/tornadovm/tornadovm-${{ matrix.backend.name }}
+        with:
+          backend: ${{ matrix.backend.name }}
+
+      # Test standalone mode per model family and quantization
+      # Note: variants can be represented with matrices
+      - name: FP16 - Run Llama-3.2-1B-Instruct-F16.gguf - Standard
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Llama-3.2-1B-Instruct-F16.gguf
+          model: Llama-3.2-1B-Instruct
+          quantization: F16
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-f16-standard.json
+
+      - name: FP16 - Run Llama-3.2-1B-Instruct-F16.gguf - Prefill-Decode
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Llama-3.2-1B-Instruct-F16.gguf
+          model: Llama-3.2-1B-Instruct
+          quantization: F16
+          configuration: prefill-decode
+          flags: --with-prefill-decode
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-f16-prefill-decode.json
 
       - name: FP16 - Run Llama-3.2-1B-Instruct-F16.gguf - Batch-Prefill-Decode
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-f16-batch-prefill-decode.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/Llama-3.2-1B-Instruct-F16.gguf \
-            --prompt "Say hello" \
-            --with-prefill-decode --batch-prefill-size 32
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-f16-batch-prefill-decode.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=Llama-3.2-1B-Instruct-F16.gguf \
-            model=Llama-3.2-1B-Instruct \
-            quantization=F16 \
-            configuration=batch-prefill-decode \
-            "flags=--with-prefill-decode --batch-prefill-size 32" \
-            prompt="Say hello"
-
-      # ── PTX-only: CUDA-graph variants ────────────────────────────────────────
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Llama-3.2-1B-Instruct-F16.gguf
+          model: Llama-3.2-1B-Instruct
+          quantization: F16
+          configuration: batch-prefill-decode
+          flags: --with-prefill-decode --batch-prefill-size 32
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-f16-batch-prefill-decode.json
+
+      # PTX-only: CUDA-graph variants
       - name: PTX - FP16 - Run Llama-3.2-1B-Instruct-F16.gguf - Prefill-Decode-CUDA-Graphs
         if: matrix.backend.name == 'ptx'
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-ptx-llama-1b-f16-prefill-decode-cuda-graphs.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --ptx \
-            --model $MODELS_DIR/Llama-3.2-1B-Instruct-F16.gguf \
-            --prompt "Say hello" \
-            --with-prefill-decode \
-            --cuda-graphs
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-ptx-llama-1b-f16-prefill-decode-cuda-graphs.meta.json" \
-            backend=ptx \
-            task=llama-inference \
-            model_file=Llama-3.2-1B-Instruct-F16.gguf \
-            model=Llama-3.2-1B-Instruct \
-            quantization=F16 \
-            configuration=prefill-decode-cuda-graphs \
-            "flags=--with-prefill-decode --cuda-graphs" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Llama-3.2-1B-Instruct-F16.gguf
+          model: Llama-3.2-1B-Instruct
+          quantization: F16
+          configuration: prefill-decode-cuda-graphs
+          flags: --with-prefill-decode --cuda-graphs
+          metrics_file: ${{ runner.temp }}/metrics-ptx-llama-1b-f16-prefill-decode-cuda-graphs.json
 
       - name: PTX - FP16 - Run Llama-3.2-1B-Instruct-F16.gguf - Batch-Prefill-Decode-CUDA-Graphs
         if: matrix.backend.name == 'ptx'
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-ptx-llama-1b-f16-batch-prefill-decode-cuda-graphs.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --ptx \
-            --model $MODELS_DIR/Llama-3.2-1B-Instruct-F16.gguf \
-            --prompt "Say hello" \
-            --with-prefill-decode --batch-prefill-size 32 \
-            --cuda-graphs
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-ptx-llama-1b-f16-batch-prefill-decode-cuda-graphs.meta.json" \
-            backend=ptx \
-            task=llama-inference \
-            model_file=Llama-3.2-1B-Instruct-F16.gguf \
-            model=Llama-3.2-1B-Instruct \
-            quantization=F16 \
-            configuration=batch-prefill-decode-cuda-graphs \
-            "flags=--with-prefill-decode --batch-prefill-size 32 --cuda-graphs" \
-            prompt="Say hello"
-
-      # ── Additional models — standard inference, all backends ─────────────────
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Llama-3.2-1B-Instruct-F16.gguf
+          model: Llama-3.2-1B-Instruct
+          quantization: F16
+          configuration: batch-prefill-decode-cuda-graphs
+          flags: --with-prefill-decode --batch-prefill-size 32 --cuda-graphs
+          metrics_file: ${{ runner.temp }}/metrics-ptx-llama-1b-f16-batch-prefill-decode-cuda-graphs.json
+
       - name: FP16 - Run Qwen3-4B-f16.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen3-4b-f16-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/Qwen3-4B-f16.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen3-4b-f16-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=Qwen3-4B-f16.gguf \
-            model=Qwen3-4B \
-            quantization=F16 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Qwen3-4B-f16.gguf
+          model: Qwen3-4B
+          quantization: F16
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen3-4b-f16-standard.json
 
       - name: FP16 - Run Mistral-7B-Instruct-v0.3.fp16.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-mistral-7b-fp16-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/Mistral-7B-Instruct-v0.3.fp16.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-mistral-7b-fp16-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=Mistral-7B-Instruct-v0.3.fp16.gguf \
-            model=Mistral-7B-Instruct-v0.3 \
-            quantization=F16 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Mistral-7B-Instruct-v0.3.fp16.gguf
+          model: Mistral-7B-Instruct-v0.3
+          quantization: F16
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-mistral-7b-fp16-standard.json
 
       - name: FP16 - Run Qwen2.5-1.5b-instruct-fp16.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen2-5-1-5b-fp16-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/qwen2.5-1.5b-instruct-fp16.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen2-5-1-5b-fp16-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=qwen2.5-1.5b-instruct-fp16.gguf \
-            model=Qwen2.5-1.5B-Instruct \
-            quantization=F16 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: qwen2.5-1.5b-instruct-fp16.gguf
+          model: Qwen2.5-1.5B-Instruct
+          quantization: F16
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen2-5-1-5b-fp16-standard.json
 
       - name: FP16 - Run Phi-3-mini-4k-instruct-fp16.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-phi3-mini-fp16-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/Phi-3-mini-4k-instruct-fp16.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-phi3-mini-fp16-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=Phi-3-mini-4k-instruct-fp16.gguf \
-            model=Phi-3-mini-4k-instruct \
-            quantization=F16 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Phi-3-mini-4k-instruct-fp16.gguf
+          model: Phi-3-mini-4k-instruct
+          quantization: F16
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-phi3-mini-fp16-standard.json
 
       - name: FP16 - Run Granite-3.2-2b-instruct-f16.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-3-2-2b-f16-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/granite-3.2-2b-instruct-f16.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-3-2-2b-f16-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=granite-3.2-2b-instruct-f16.gguf \
-            model=Granite-3.2-2B-Instruct \
-            quantization=F16 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: granite-3.2-2b-instruct-f16.gguf
+          model: Granite-3.2-2B-Instruct
+          quantization: F16
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-3-2-2b-f16-standard.json
 
       - name: FP16 - Run Granite-4.0-1b-F16.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-4-0-1b-f16-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/granite-4.0-1b-F16.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-4-0-1b-f16-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=granite-4.0-1b-F16.gguf \
-            model=Granite-4.0-1B \
-            quantization=F16 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: granite-4.0-1b-F16.gguf
+          model: Granite-4.0-1B
+          quantization: F16
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-4-0-1b-f16-standard.json
 
       - name: Q8 - Run Llama-3.2-1B-Instruct-Q8_0.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-q8-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/Llama-3.2-1B-Instruct-Q8_0.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-q8-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=Llama-3.2-1B-Instruct-Q8_0.gguf \
-            model=Llama-3.2-1B-Instruct \
-            quantization=Q8_0 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Llama-3.2-1B-Instruct-Q8_0.gguf
+          model: Llama-3.2-1B-Instruct
+          quantization: Q8_0
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-llama-1b-q8-standard.json
 
       - name: Q8 - Run Qwen3-0.6B-Q8_0.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen3-0-6b-q8-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/Qwen3-0.6B-Q8_0.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen3-0-6b-q8-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=Qwen3-0.6B-Q8_0.gguf \
-            model=Qwen3-0.6B \
-            quantization=Q8_0 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Qwen3-0.6B-Q8_0.gguf
+          model: Qwen3-0.6B
+          quantization: Q8_0
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen3-0-6b-q8-standard.json
 
       - name: Q8 - Run Phi-3-mini-4k-instruct-Q8_0.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-phi3-mini-q8-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/Phi-3-mini-4k-instruct-Q8_0.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-phi3-mini-q8-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=Phi-3-mini-4k-instruct-Q8_0.gguf \
-            model=Phi-3-mini-4k-instruct \
-            quantization=Q8_0 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Phi-3-mini-4k-instruct-Q8_0.gguf
+          model: Phi-3-mini-4k-instruct
+          quantization: Q8_0
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-phi3-mini-q8-standard.json
 
       - name: Q8 - Run Qwen2.5-1.5b-instruct-q8_0.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen2-5-1-5b-q8-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/qwen2.5-1.5b-instruct-q8_0.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen2-5-1-5b-q8-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=qwen2.5-1.5b-instruct-q8_0.gguf \
-            model=Qwen2.5-1.5B-Instruct \
-            quantization=Q8_0 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
-
-      - name: Q8 - Mistral-7B-Instruct-v0.3.Q8_0.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-mistral-7b-q8-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/Mistral-7B-Instruct-v0.3.Q8_0.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-mistral-7b-q8-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=Mistral-7B-Instruct-v0.3.Q8_0.gguf \
-            model=Mistral-7B-Instruct-v0.3 \
-            quantization=Q8_0 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: qwen2.5-1.5b-instruct-q8_0.gguf
+          model: Qwen2.5-1.5B-Instruct
+          quantization: Q8_0
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-qwen2-5-1-5b-q8-standard.json
+
+      - name: Q8 - Run Mistral-7B-Instruct-v0.3.Q8_0.gguf
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: Mistral-7B-Instruct-v0.3.Q8_0.gguf
+          model: Mistral-7B-Instruct-v0.3
+          quantization: Q8_0
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-mistral-7b-q8-standard.json
 
       - name: Q8 - Run Granite-3.2-2b-instruct-Q8_0.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-3-2-2b-q8-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/granite-3.2-2b-instruct-Q8_0.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-3-2-2b-q8-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=granite-3.2-2b-instruct-Q8_0.gguf \
-            model=Granite-3.2-2B-Instruct \
-            quantization=Q8_0 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: granite-3.2-2b-instruct-Q8_0.gguf
+          model: Granite-3.2-2B-Instruct
+          quantization: Q8_0
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-3-2-2b-q8-standard.json
 
       - name: Q8 - Run Granite-4.0-1b-Q8_0.gguf
-        env:
-          JAVA_TOOL_OPTIONS: >-
-            -Dllama.metrics.format=json
-            -Dllama.metrics.output=file
-            -Dllama.metrics.file=${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-4-0-1b-q8-standard.json
-        run: |
-          cd ${{ github.workspace }}
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          ./llama-tornado --gpu --${{ matrix.backend.name }} \
-            --model $MODELS_DIR/granite-4.0-1b-Q8_0.gguf \
-            --prompt "Say hello"
-          python3 scripts/write_metrics_sidecar.py \
-            --out "${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-4-0-1b-q8-standard.meta.json" \
-            backend="${{ matrix.backend.name }}" \
-            task=llama-inference \
-            model_file=granite-4.0-1b-Q8_0.gguf \
-            model=Granite-4.0-1B \
-            quantization=Q8_0 \
-            configuration=standard \
-            flags="" \
-            prompt="Say hello"
-
-      # ── Upload metrics for the publish job ────────────────────────────────────
+        uses: ./.github/actions/run-inference
+        with:
+          backend: ${{ matrix.backend.name }}
+          model_file: granite-4.0-1b-Q8_0.gguf
+          model: Granite-4.0-1B
+          quantization: Q8_0
+          configuration: standard
+          metrics_file: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-granite-4-0-1b-q8-standard.json
+
+      # Upload metrics for the publish job
       - name: Upload metrics artifacts
         if: always()
         uses: actions/upload-artifact@v4
@@ -532,7 +316,102 @@ jobs:
           path: ${{ runner.temp }}/metrics-${{ matrix.backend.name }}-*.json
           if-no-files-found: warn
 
-  # ── Separate job: collect all matrix metrics and update history ───────────────
+  # Test integration with Quarkus-langchain4j
+  quarkus-langchain4j-integration:
+    if: github.repository == 'beehive-lab/GPULlama3.java'
+    runs-on: [self-hosted]
+    needs: build
+    timeout-minutes: 10
+    strategy:
+      fail-fast: true
+      matrix:
+        backend:
+          - name: opencl
+          - name: ptx
+
+    steps:
+      - name: Checkout GPULlama3
+        uses: actions/checkout@v4
+        with:
+          clean: false
+
+      - name: Set up Java
+        uses: ./.github/actions/setup-java
+        with:
+          java_version: ${{ env.JAVA_VERSION }}
+
+      - name: Setup TornadoVM
+        uses: ./.github/actions/setup-tornadovm
+        env:
+          TORNADO_ROOT: ${{ runner.tool_cache }}/tornadovm/tornadovm-${{ matrix.backend.name }}
+        with:
+          backend: ${{ matrix.backend.name }}
+
+      - name: Verify GPULlama3 Dependency
+        run: |
+          cd ${{ github.workspace }}/quarkus-langchain4j/integration-tests/gpu-llama3
+          mvn dependency:tree | grep "io.github.beehive-lab:gpu-llama3"
+
+      - name: Start Quarkus Application
+        run: |
+          cd ${{ github.workspace }}/quarkus-langchain4j/integration-tests/gpu-llama3
+          java  @"$TORNADOVM_HOME/tornado-argfile" \
+                -Dtornado.device.memory=8GB \
+                -Dquarkus.http.port=$QUARKUS_PORT \
+                -jar target/quarkus-app/quarkus-run.jar &
+          APP_PID=$!
+          echo "APP_PID=$APP_PID" >> $GITHUB_ENV
+          for i in {1..30}; do
+            if curl -s http://localhost:$QUARKUS_PORT/q/health > /dev/null 2>&1; then
+              echo "Application ready after ${i} seconds"
+              break
+            elif [ $i -eq 30 ]; then
+              echo "::error::Application failed to start within 30 seconds"
+              kill $APP_PID || true
+              exit 1
+            else
+              [ $((i % 5)) -eq 0 ] && echo "Still waiting... (${i}s)"
+              sleep 1
+            fi
+          done
+
+      - name: Trigger Blocking Endpoint
+        run: |
+          for attempt in 1 2 3; do
+            echo "Attempt $attempt of 3 for blocking endpoint..."
+            HTTP_RESPONSE=$(curl -s -w "%{http_code}" http://localhost:$QUARKUS_PORT/chat/blocking)
+            HTTP_CODE="${HTTP_RESPONSE: -3}"
+            if [ "$HTTP_CODE" = "200" ]; then
+              echo "SUCCESS: HTTP $HTTP_CODE"
+              echo "Response body: ${HTTP_RESPONSE%???}"
+              break
+            fi
+            echo "Failed: HTTP $HTTP_CODE"
+            [ $attempt -lt 3 ] && sleep 2
+            [ $attempt -eq 3 ] && { echo "::error::Blocking endpoint failed after 3 attempts"; exit 1; }
+          done
+
+      - name: Trigger Streaming Endpoint
+        run: |
+          for attempt in 1 2 3; do
+            echo "Attempt $attempt of 3 for streaming endpoint..."
+            HTTP_CODE=$(timeout 10s curl -s -o /dev/null -w "%{http_code}" http://localhost:$QUARKUS_PORT/chat/streaming)
+            if [ "$HTTP_CODE" = "200" ]; then
+              echo "SUCCESS: HTTP $HTTP_CODE"
+              break
+            fi
+            echo "Failed: HTTP $HTTP_CODE"
+            [ $attempt -lt 3 ] && sleep 2
+            [ $attempt -eq 3 ] && { echo "::error::Streaming endpoint failed after 3 attempts"; exit 1; }
+          done
+
+      - name: Cleanup & Shutdown
+        if: always()
+        run: |
+          kill $APP_PID || true
+          wait $APP_PID 2>/dev/null || true
+
+  # Collect all matrix metrics and update history
   publish-performance-history:
     # Guard: only commit history on real pushes to main, not on PRs or forks.
     # Prevents duplicate entries from PR runs and avoids push-permission errors on forks.
@@ -540,8 +419,9 @@ jobs:
       github.repository == 'beehive-lab/GPULlama3.java' &&
       github.event_name == 'push' &&
       github.ref == 'refs/heads/main'
+
     runs-on: [self-hosted]
-    needs: build-and-run
+    needs: standalone-inference
     timeout-minutes: 15
 
     steps:
@@ -559,24 +439,36 @@ jobs:
         run: |
           python3 scripts/process_metrics.py \
             --metrics-dir "${{ runner.temp }}/metrics-artifacts" \
-            --commit       "${{ github.sha }}" \
-            --branch       "${{ github.ref_name }}" \
-            --run-id       "${{ github.run_id }}" \
-            --run-number   "${{ github.run_number }}" \
-            --run-attempt  "${{ github.run_attempt }}" \
-            --workflow     "${{ github.workflow }}" \
-            --history      "$PERF_HISTORY_FILE"
+            --commit "${{ github.sha }}" \
+            --branch "${{ github.ref_name }}" \
+            --run-id "${{ github.run_id }}" \
+            --run-number "${{ github.run_number }}" \
+            --run-attempt "${{ github.run_attempt }}" \
+            --workflow "${{ github.workflow }}" \
+            --history "$PERF_HISTORY_FILE"
 
       - name: Commit performance history
         run: |
-          git config user.name  "github-actions[bot]"
+          SHORT_SHA=$(echo "${GITHUB_SHA}" | cut -c1-8)
+
+          git config user.name "github-actions[bot]"
           git config user.email "github-actions[bot]@users.noreply.github.com"
+
           git add "$PERF_HISTORY_FILE"
-          git diff --cached --quiet && echo "No history changes to commit" && exit 0
-          git commit -m "perf: record run #${{ github.run_number }} @ ${GITHUB_SHA::8}"
+
+          git diff --cached --quiet && \
+            echo "No history changes to commit" && exit 0
+
+          git commit -m "perf: record run #${{ github.run_number }} @ ${SHORT_SHA}"
+
           for attempt in 1 2 3; do
             git pull --rebase origin main && git push && break || {
-              [ $attempt -lt 3 ] && { echo "Attempt $attempt failed, retrying in $((attempt * 5))s..."; sleep $((attempt * 5)); } \
-                                 || { echo "::error::Failed to push after 3 attempts"; exit 1; }
+              if [ $attempt -lt 3 ]; then
+                echo "Attempt $attempt failed, retrying in $((attempt * 5))s..."
+                sleep $((attempt * 5))
+              else
+                echo "::error::Failed to push after 3 attempts"
+                exit 1
+              fi
             }
           done
diff --git a/.github/workflows/deploy-maven-central.yml b/.github/workflows/deploy-maven-central.yml
index 88bcee8e..53f62226 100644
--- a/.github/workflows/deploy-maven-central.yml
+++ b/.github/workflows/deploy-maven-central.yml
@@ -31,19 +31,18 @@ jobs:
       matrix:
         jdk:
           - name: jdk21
-            java_home: /opt/jenkins/jdks/graal-23.1.0/jdk-21.0.3
+            java_version: 21.0.2-open
           - name: jdk25
-            java_home: /opt/jenkins/jdks/jdk-25.0.2
-    env:
-      JAVA_HOME: ${{ matrix.jdk.java_home }}
+            java_version: 25.0.2-open
 
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Setup environment
-        run: |
-          echo "$JAVA_HOME/bin" >> $GITHUB_PATH
+      - name: Set up Java
+        uses: ./.github/actions/setup-java
+        with:
+          java_version: ${{ matrix.jdk.java_version }}
 
       - name: Configure Maven settings
         run: |
diff --git a/.github/workflows/integration-quarkus-langchain4j.yml b/.github/workflows/integration-quarkus-langchain4j.yml
deleted file mode 100644
index 8ee6e900..00000000
--- a/.github/workflows/integration-quarkus-langchain4j.yml
+++ /dev/null
@@ -1,260 +0,0 @@
-name: Integration Quarkus-LangChain4j
-
-on:
-  push:
-    branches: [ main ]
-  pull_request:
-     branches: [ main ]
-     types: [opened, synchronize, reopened]
-  schedule:
-     # Run every Saturday at 02:30 UTC to catch dependency breakages
-     - cron: '30 2 * * 6'
-  workflow_dispatch:
-    inputs:
-      quarkus_langchain4j_version:
-        description: 'Quarkus LangChain4j version to test against'
-        required: false
-        type: string
-      quarkus_version:
-        description: 'Quarkus platform version to test against'
-        required: false
-        type: string
-
-env:
-  JAVA_HOME: /opt/jenkins/jdks/graal-23.1.0/jdk-21.0.3
-  TORNADO_ROOT: ${{ github.workspace }}/GPULlama3.java/external/tornadovm
-  GRAAL_JARS: /opt/graalJars
-  QUARKUS_PORT: 8081
-
-jobs:
-  quarkus-integration-test:
-    if: github.repository == 'beehive-lab/GPULlama3.java'
-    runs-on: [self-hosted]
-    timeout-minutes: 30
-    strategy:
-      fail-fast: true
-      matrix:
-        backend:
-          - name: opencl
-          - name: ptx
-
-    steps:
-      - name: Checkout GPULlama3
-        uses: actions/checkout@v4
-
-      # Step 1: Clone and build TornadoVM
-      - name: Clone TornadoVM master
-        run: |
-          git clone --depth 1 --branch master \
-            https://github.com/beehive-lab/TornadoVM.git \
-            $TORNADO_ROOT
-      - name: Set up Python venv for TornadoVM
-        run: |
-          python3 -m venv $TORNADO_ROOT/venv
-          source $TORNADO_ROOT/venv/bin/activate
-          python --version
-      - name: Build TornadoVM
-        run: |
-          cd $TORNADO_ROOT
-          mkdir -p graalJars && cp $GRAAL_JARS/* graalJars/
-          source venv/bin/activate
-          echo "=== Building TornadoVM ==="
-
-          make BACKEND=${{ matrix.backend.name }}
-
-          echo "=== Searching for TornadoVM SDK directory ==="
-          SDK_DIR=$(find dist -type d -maxdepth 3 -path "*/tornadovm-*-${{ matrix.backend.name }}" | head -n 1)
-          if [ -z "$SDK_DIR" ]; then
-            echo "::error::Could not locate TornadoVM SDK directory!"
-            find dist -maxdepth 5 -type d
-            exit 1
-          fi
-          FULL_SDK="${PWD}/${SDK_DIR}"
-          echo "Detected TornadoVM SDK: $FULL_SDK"
-
-          # Export for current shell session
-          export TORNADOVM_HOME="$FULL_SDK"
-          export PATH="$FULL_SDK/bin:$JAVA_HOME/bin:$PATH"
-
-          # Save for subsequent steps
-          echo "TORNADOVM_HOME=$FULL_SDK" >> $GITHUB_ENV
-          echo "PATH=$PATH" >> $GITHUB_ENV
-
-          echo "=== Checking tornado CLI ==="
-          which tornado || { echo "::error::tornado not in PATH"; exit 1; }
-          tornado --devices
-
-      # Step 2: Build GPULlama3.java
-      - name: Build GPULlama3.java
-        run: |
-          cd ${{ github.workspace }}
-          echo "Using TORNADOVM_HOME=$TORNADOVM_HOME"
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          tornado --version
-          
-          # Append SNAPSHOT to GPULlama3 version
-          GPULLAMA3_VERSION=$(./mvnw help:evaluate -Dexpression=project.version -q -DforceStdout)
-          GPULLAMA3_VERSION="${GPULLAMA3_VERSION}-SNAPSHOT"
-          echo "GPULlama3 version: $GPULLAMA3_VERSION"
-          ./mvnw versions:set -DnewVersion=$GPULLAMA3_VERSION
-          
-          # Build
-          ./mvnw clean install -DskipTests
-    
-          # Save GPULlama3.java version for subsequent steps
-          echo "GPULLAMA3_VERSION=$GPULLAMA3_VERSION" >> $GITHUB_ENV
-
-      # Step 3: Clone Quarkus LangChain4j
-      - name: Clone Quarkus LangChain4j
-        run: |
-          cd ${{ github.workspace }}
-          git clone https://github.com/quarkiverse/quarkus-langchain4j.git
-
-      # Step 4: Build Quarkus LangChain4j with current GPULlama3.java
-      - name: Build Quarkus LangChain4j
-        run: |
-          cd ${{ github.workspace }}/quarkus-langchain4j
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          
-          # Update the GPULlama3 version used by quarkus-langchain4j
-          POM="pom.xml"
-          sed -i 's/<gpu-llama3\.version>.*<\/gpu-llama3\.version>/<gpu-llama3.version>'$GPULLAMA3_VERSION'<\/gpu-llama3.version>/' "$POM"
-          
-          # Use reactor to build *only *GPULlama3 integration test + dependencies
-          # This recompiles everything with the same Java version, avoiding compatibility issues
-          # The -Dtornado flag activates the TornadoVM profile which includes gpu-llama3 module
-          mvn clean install -pl integration-tests/gpu-llama3 -am -DskipTests -Dtornado
-
-      # Step 4.5: Verify GPULlama3 JAR
-      - name: Verify GPULlama3 Dependency
-        run: |
-          cd ${{ github.workspace }}/quarkus-langchain4j/integration-tests/gpu-llama3
-          echo "GPULlama3 dependency used by Quarkus-LangChain4j:"
-          mvn dependency:tree | grep "io.github.beehive-lab:gpu-llama3"
-
-      # Step 5: Start Quarkus Application and Wait for Startup
-      - name: Start Quarkus Application and Wait for Startup
-        run: |
-          cd ${{ github.workspace }}/quarkus-langchain4j/integration-tests/gpu-llama3
-          export PATH="$TORNADOVM_HOME/bin:$JAVA_HOME/bin:$PATH"
-          
-          echo "Starting Quarkus application on port $QUARKUS_PORT..."
-          
-          # Start the Quarkus application in the background
-          java  @"$TORNADOVM_HOME/tornado-argfile" \
-                -Dtornado.device.memory=8GB \
-                -Dquarkus.http.port=$QUARKUS_PORT \
-                -jar target/quarkus-app/quarkus-run.jar &
-          APP_PID=$!
-          
-          if [ -z "$APP_PID" ]; then
-            echo "ERROR: Failed to start Quarkus application"
-            exit 1
-          fi
-
-          echo "Waiting for Quarkus application to start..."
-          
-          # Wait for application to be ready
-          for i in {1..30}; do
-            if curl -s http://localhost:$QUARKUS_PORT/q/health > /dev/null 2>&1; then
-              echo "Application ready after ${i} seconds"
-              echo "Health endpoint: http://localhost:$QUARKUS_PORT/q/health"
-              break
-            elif [ $i -eq 30 ]; then
-              echo "ERROR: Application failed to start within 30 seconds"
-              echo "Debugging info:"
-              echo "- Port: $QUARKUS_PORT"
-              echo "- Process ID: $APP_PID"
-              echo "- Health URL: http://localhost:$QUARKUS_PORT/q/health"
-              kill $APP_PID || true
-              exit 1
-            else
-              [ $((i % 5)) -eq 0 ] && echo "Still waiting... (${i}s)"
-              sleep 1
-            fi
-          done
-
-      # Step 6: Run test 1
-      - name: Trigger Blocking Endpoint
-        run: |
-          MAX_ATTEMPTS=3
-          ATTEMPT=1
-          SUCCESS=false
-    
-          while [ $ATTEMPT -le $MAX_ATTEMPTS ]; do
-            echo "Attempt $ATTEMPT of $MAX_ATTEMPTS for blocking endpoint..."
-    
-            # Trigger endpoint
-            HTTP_RESPONSE=$(curl -s -w "%{http_code}" http://localhost:$QUARKUS_PORT/chat/blocking)
-            HTTP_RESPONSE_CODE="${HTTP_RESPONSE: -3}"
-            HTTP_RESPONSE_BODY="${HTTP_RESPONSE%???}"
-    
-            # Check response code
-            if [ "$HTTP_RESPONSE_CODE" = "200" ]; then
-              echo "SUCCESS: Blocking endpoint returned HTTP code: $HTTP_RESPONSE_CODE"
-              echo "HTTP Response body: $HTTP_RESPONSE_BODY"
-              SUCCESS=true
-              break
-            else
-              echo "Attempt $ATTEMPT failed: Blocking endpoint returned HTTP code $HTTP_RESPONSE_CODE"
-              echo "Response body: $HTTP_RESPONSE_BODY"
-    
-              if [ $ATTEMPT -lt $MAX_ATTEMPTS ]; then
-                echo "Retrying in 2 seconds..."
-                sleep 2
-              fi
-            fi
-    
-            ATTEMPT=$((ATTEMPT + 1))
-          done
-    
-          if [ "$SUCCESS" = false ]; then
-            echo "ERROR: Blocking endpoint failed after $MAX_ATTEMPTS attempts"
-            exit 1
-          fi
-
-      # Step 7: Run test 2
-      - name: Trigger Streaming Endpoint
-        run: |
-          MAX_ATTEMPTS=3
-          ATTEMPT=1
-          SUCCESS=false
-          
-          while [ $ATTEMPT -le $MAX_ATTEMPTS ]; do
-            echo "Attempt $ATTEMPT of $MAX_ATTEMPTS for streaming endpoint..."
-
-            # Trigger endpoint
-            HTTP_RESPONSE=$(timeout 10s curl -s -w "%{http_code}" http://localhost:$QUARKUS_PORT/chat/streaming)
-            HTTP_RESPONSE_CODE="${HTTP_RESPONSE: -3}"
-            #HTTP_RESPONSE_BODY="$HTTP_RESPONSE%???}"
-          
-            # Check response code
-            if [ "$HTTP_RESPONSE_CODE" == "200" ]; then
-              echo "SUCCESS: Streaming endpoint returned HTTP code: ${HTTP_RESPONSE: -3}"
-              # do not show ugly streaming response body
-              #echo "HTTP Response body: $HTTP_RESPONSE_BODY"
-              SUCCESS=true
-              break
-            else
-              echo "Attempt $ATTEMPT failed: Streaming endpoint returned HTTP code $HTTP_RESPONSE_CODE"
-          
-              if [ $ATTEMPT -lt $MAX_ATTEMPTS ]; then
-                echo "Retrying in 2 seconds..."
-                sleep 2
-              fi
-            fi
-          
-              ATTEMPT=$((ATTEMPT + 1))
-          done
-          
-          if [ "$SUCCESS" = false ]; then
-            echo "ERROR: Streaming endpoint failed after $MAX_ATTEMPTS attempts"
-            exit 1
-          fi
-
-      # Step 8: Cleanup & Shutdown
-      - name: Cleanup & Shutdown
-        run: |
-          # Clean shutdown
-          kill $APP_PID || true
-          wait $APP_PID 2>/dev/null || true
diff --git a/.github/workflows/prepare-release.yml b/.github/workflows/prepare-release.yml
index 3ff71f90..00d6c63b 100644
--- a/.github/workflows/prepare-release.yml
+++ b/.github/workflows/prepare-release.yml
@@ -20,6 +20,7 @@ on:
 env:
   VERSION: ${{ inputs.version }}
   PREV_VERSION: ${{ inputs.previous_version }}
+  JAVA_VERSION: 21.0.2-open
 
 jobs:
   prepare-release:
@@ -29,8 +30,6 @@ jobs:
       contents: write
       pull-requests: write
     timeout-minutes: 15
-    env:
-      JAVA_HOME: /opt/jenkins/jdks/graal-23.1.0/jdk-21.0.3
 
     steps:
       - name: Validate version format
@@ -48,9 +47,10 @@ jobs:
           fetch-depth: 0
           token: ${{ secrets.GITHUB_TOKEN }}
 
-      - name: Setup environment
-        run: |
-          echo "$JAVA_HOME/bin" >> $GITHUB_PATH
+      - name: Set up Java
+        uses: ./.github/actions/setup-java
+        with:
+          java_version: ${{ env.JAVA_VERSION }}
 
       - name: Configure Git
         run: |
diff --git a/.github/workflows/rerun-workflow.yml b/.github/workflows/rerun-workflow.yml
index e4b14e75..62eec7d3 100644
--- a/.github/workflows/rerun-workflow.yml
+++ b/.github/workflows/rerun-workflow.yml
@@ -47,8 +47,30 @@ jobs:
               core.setOutput('is_help', 'false');
               }
 
-      - name: Get PR SHA
+      - name: Check commenter permissions
+        id: check_permission
         if: steps.help.outputs.is_help != 'true'
+        uses: actions/github-script@v7
+        with:
+          script: |
+            const { data: perm } = await github.rest.repos.getCollaboratorPermissionLevel({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              username: context.payload.comment.user.login
+            });
+            const authorized = ['write', 'admin'].includes(perm.permission);
+            if (!authorized) {
+              await github.rest.issues.createComment({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                issue_number: context.issue.number,
+                body: `@${context.payload.comment.user.login} You need write permission to trigger workflow reruns.`
+              });
+            }
+            core.setOutput('authorized', authorized ? 'true' : 'false');
+
+      - name: Get PR SHA
+        if: steps.help.outputs.is_help != 'true' && steps.check_permission.outputs.authorized == 'true'
         id: pr
         uses: actions/github-script@v7
         with:
@@ -64,7 +86,7 @@ jobs:
             console.log(`PR head ref: ${pr.head.ref}`);
 
       - name: Add reaction to comment
-        if: steps.help.outputs.is_help != 'true'
+        if: steps.help.outputs.is_help != 'true' && steps.check_permission.outputs.authorized == 'true'
         uses: actions/github-script@v7
         with:
           script: |
@@ -76,7 +98,7 @@ jobs:
             });
 
       - name: Post start comment
-        if: steps.help.outputs.is_help != 'true'
+        if: steps.help.outputs.is_help != 'true' && steps.check_permission.outputs.authorized == 'true'
         uses: actions/github-script@v7
         with:
           script: |
@@ -92,7 +114,7 @@ jobs:
             });
 
       - name: Rerun failed workflows
-        if: steps.help.outputs.is_help != 'true'
+        if: steps.help.outputs.is_help != 'true' && steps.check_permission.outputs.authorized == 'true'
         uses: actions/github-script@v7
         with:
           script: |
@@ -167,7 +189,7 @@ jobs:
             console.log(`Reran ${rerunCount} workflow(s)`);
 
       - name: Post completion comment
-        if: always() && steps.help.outputs.is_help != 'true'
+        if: always() && steps.help.outputs.is_help != 'true' && steps.check_permission.outputs.authorized == 'true'
         uses: actions/github-script@v7
         with:
           script: |