Several fixes and improvements for CI

.github/actions/run-inference/action.yml

@@ -0,0 +1,61 @@
+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 }}
+      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 }}" \
+          ${{ inputs.flags }}
+
+        # Write metadata sidecar so process_metrics.py can identify each metrics file
+        SIDECAR="${{ inputs.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 }}"

.github/actions/setup-tornadovm/action.yml

@@ -0,0 +1,93 @@
+name: Setup TornadoVM
+description: Clone (or restore from cache), build, and configure TornadoVM. 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
+
+    # actions/cache `path:` is an expression — it cannot access the calling
+    # workflow's env vars via ${{ env.* }}, so we resolve TORNADO_ROOT here.
+    - name: Resolve TORNADO_ROOT for cache path
+      id: paths
+      shell: bash
+      run: echo "tornado_root=$TORNADO_ROOT" >> $GITHUB_OUTPUT
+
+    - name: Restore TornadoVM cache
+      id: cache
+      uses: actions/cache@v4
+      with:
+        path: ${{ steps.paths.outputs.tornado_root }}
+        key: tornadovm-${{ inputs.backend }}-${{ steps.tornado_sha.outputs.sha }}
+
+    - name: Clone TornadoVM master
+      if: steps.cache.outputs.cache-hit != 'true'
+      shell: bash
+      run: |
+        git clone --depth 1 --branch master \
+          https://github.com/beehive-lab/TornadoVM.git \
+          $TORNADO_ROOT
+
+    - name: Set up Python venv for TornadoVM
+      if: steps.cache.outputs.cache-hit != 'true'
+      shell: bash
+      run: |
+        python3 -m venv $TORNADO_ROOT/venv
+        source $TORNADO_ROOT/venv/bin/activate
+        python --version
+
+    - name: Build TornadoVM
+      if: steps.cache.outputs.cache-hit != '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 "=== Verifying TornadoVM SDK directory ==="
+        SDK_DIR=$(find dist -type d -maxdepth 3 -path "*/tornadovm-*-${{ inputs.backend }}" | head -n 1)
+        if [ -z "$SDK_DIR" ]; then
+          echo "::error::Could not locate TornadoVM SDK directory!"
+          find dist -maxdepth 5 -type d
+          exit 1
+        fi
+
+    # Runs on both cache hit and miss — sets TORNADOVM_HOME and PATH for all
+    # subsequent steps in the calling workflow.
+    - name: Configure TornadoVM environment
+      shell: bash
+      run: |
+        cd $TORNADO_ROOT
+        SDK_DIR=$(find 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 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" >> $GITHUB_ENV
+        echo "$FULL_SDK/bin" >> $GITHUB_PATH
+        echo "$JAVA_HOME/bin" >> $GITHUB_PATH
+
+        echo "=== Checking tornado CLI ==="
+        which tornado || { echo "::error::tornado not in PATH"; exit 1; }
+        tornado --devices