ipex-llm/.github/workflows/llm-harness-evaluation.yml

name: LLM Harness Evaluation

# Cancel previous runs in the PR when you push new commits
concurrency:
  group: ${{ github.workflow }}-llm-nightly-test-${{ github.event.pull_request.number || github.run_id }}
  cancel-in-progress: true

permissions:
  contents: read

# Controls when the action will run.
on:
  schedule:
    - cron: "30 12 * * *" # GMT time, 12:30 GMT == 20:30 China
  pull_request:
    branches: [main]
    paths:
      - ".github/workflows/llm-harness-evaluation.yml"
  # Allows you to run this workflow manually from the Actions tab
  workflow_dispatch:
    inputs:
      model_name:
        description: 'Model names, separated by comma and must be quoted.'
        required: true
        type: string
      precision:
        description: 'Precisions, separated by comma and must be quoted.'
        required: true
        type: string
      task:
        description: 'Tasks, separated by comma and must be quoted.'
        required: true
        type: string
      runs-on:
        description: 'Labels to filter the runners, separated by comma and must be quoted.'
        default: "accuracy"
        required: false
        type: string


# A workflow run is made up of one or more jobs that can run sequentially or in parallel
jobs:
  llm-cpp-build:
    uses: ./.github/workflows/llm-binary-build.yml
    with:
      platform: 'Linux'
  # Set the testing matrix based on the event (schedule, PR, or manual dispatch)
  set-matrix:
    runs-on: ubuntu-latest
    outputs:
      model_name: ${{ steps.set-matrix.outputs.model_name }}
      precision: ${{ steps.set-matrix.outputs.precision }}
      task: ${{ steps.set-matrix.outputs.task }}
      runner: ${{ steps.set-matrix.outputs.runner }}
    steps:
      - name: set-nightly-env
        if: ${{github.event_name == 'schedule'}}
        env:
          NIGHTLY_MATRIX_MODEL_NAME: '["Llama2-7b-guanaco-dolphin-500", "falcon-7b-instruct-with-patch", 
                        "Mistral-7B-v0.1", "mpt-7b-chat", "Baichuan2-7B-Chat-LLaMAfied"]'
          NIGHTLY_MATRIX_TASK: '["arc", "truthfulqa", "winogrande"]'
          NIGHTLY_MATRIX_PRECISION: '["sym_int4", "fp8"]'
          NIGHTLY_LABELS: '["self-hosted", "llm", "accuracy-nightly"]'
        run: |
            echo "model_name=$NIGHTLY_MATRIX_MODEL_NAME" >> $GITHUB_ENV
            echo "precision=$NIGHTLY_MATRIX_PRECISION" >> $GITHUB_ENV
            echo "task=$NIGHTLY_MATRIX_TASK" >> $GITHUB_ENV
            echo "runner=$NIGHTLY_LABELS" >> $GITHUB_ENV            

      - name: set-pr-env
        if: ${{github.event_name == 'pull_request'}}
        env:
          PR_MATRIX_MODEL_NAME: '["stablelm-3b-4e1t"]'
          PR_MATRIX_TASK: '["winogrande"]'
          PR_MATRIX_PRECISION: '["sym_int4"]'
          PR_LABELS: '["self-hosted", "llm", "temp-arc01"]'

        run: |
            echo "model_name=$PR_MATRIX_MODEL_NAME" >> $GITHUB_ENV
            echo "precision=$PR_MATRIX_PRECISION" >> $GITHUB_ENV
            echo "task=$PR_MATRIX_TASK" >> $GITHUB_ENV
            echo "runner=$PR_LABELS" >> $GITHUB_ENV            
      - name: set-manual-env
        if: ${{github.event_name == 'workflow_dispatch'}}
        env:
          MANUAL_MATRIX_MODEL_NAME: ${{format('[ {0} ]', inputs.model_name)}}
          MANUAL_MATRIX_TASK: ${{format('[ {0} ]', inputs.task)}}
          MANUAL_MATRIX_PRECISION: ${{format('[ {0} ]', inputs.precision)}}
          MANUAL_LABELS: ${{format('["self-hosted", "llm", {0}]', inputs.runs-on)}}
        run: |
            echo "model_name=$MANUAL_MATRIX_MODEL_NAME" >> $GITHUB_ENV
            echo "precision=$MANUAL_MATRIX_PRECISION" >> $GITHUB_ENV
            echo "task=$MANUAL_MATRIX_TASK" >> $GITHUB_ENV
            echo "runner=$MANUAL_LABELS" >> $GITHUB_ENV            
      - name: set-matrix
        id: set-matrix
        run: |
            echo "model_name=$model_name" >> $GITHUB_OUTPUT
            echo "precision=$precision" >> $GITHUB_OUTPUT
            echo "task=$task" >> $GITHUB_OUTPUT
            echo "runner=$runner" >> $GITHUB_OUTPUT            
  llm-harness-evaluation:
    timeout-minutes: 1000
    needs: [llm-cpp-build, set-matrix]
    strategy:
      fail-fast: false
      matrix:
        python-version: ["3.9"]
        model_name: ${{ fromJson(needs.set-matrix.outputs.model_name) }}
        task: ${{ fromJson(needs.set-matrix.outputs.task) }}
        precision: ${{ fromJson(needs.set-matrix.outputs.precision) }}
        device: [xpu]
        
    runs-on: ${{ fromJson(needs.set-matrix.outputs.runner) }}
    env:
      ANALYTICS_ZOO_ROOT: ${{ github.workspace }}
      ORIGIN_DIR: /mnt/disk1/models
      HARNESS_HF_HOME: /mnt/disk1/harness_home
    steps:
      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # actions/checkout@v3
      - name: Set up Python ${{ matrix.python-version }}
        uses: actions/setup-python@v4
        with:
          python-version: ${{ matrix.python-version }}
      - name: Install dependencies
        shell: bash
        run: |
          set -e
          python -m pip install --upgrade pip
          python -m pip install --upgrade setuptools==58.0.4
          python -m pip install --upgrade wheel          

      - name: Download llm binary
        uses: ./.github/actions/llm/download-llm-binary
        with:
          platform: 'Linux'

      - name: Run LLM install (all) test
        uses: ./.github/actions/llm/setup-llm-env
        with:
          extra-dependency: "xpu_2.1"

      - name: Install harness
        working-directory: ${{ github.workspace }}/python/llm/dev/benchmark/harness/
        shell: bash
        run: |
          git clone https://github.com/EleutherAI/lm-evaluation-harness.git
          cd lm-evaluation-harness
          git checkout b281b09
          pip install -e .          
      
      - name: Download models and datasets
        shell: bash
        run: |
          echo "MODEL_PATH=${ORIGIN_DIR}/${{ matrix.model_name }}/" >> "$GITHUB_ENV"
          MODEL_PATH=${ORIGIN_DIR}/${{ matrix.model_name }}/
          if [ ! -d $HARNESS_HF_HOME ]; then
            mkdir -p $HARNESS_HF_HOME
          fi
          wget -r -nH -nc -l inf --no-verbose --cut-dirs=2 ${LLM_FTP_URL}/llm/LeaderBoard_Datasets/ -P $HARNESS_HF_HOME/
          wget -r -nH -nc --no-verbose --cut-dirs=1 ${LLM_FTP_URL}/llm/${{ matrix.model_name }} -P ${ORIGIN_DIR}
                    
      - name: Upgrade packages
        shell: bash
        run: |
          pip install --upgrade datasets==2.14.6 
          if [ "${{ matrix.model_name }}" = "Mistral-7B-v0.1" ]; then
            pip install --upgrade transformers==4.36
          else
            pip install --upgrade transformers==4.31
          fi          
      

      - name: Run harness
        shell: bash
        working-directory: ${{ github.workspace }}/python/llm/dev/benchmark/harness
        env:
          USE_XETLA: OFF
          # SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS: 1
        run: |
          export HF_HOME=${HARNESS_HF_HOME}
          export HF_DATASETS=$HARNESS_HF_HOME/datasets
          export HF_DATASETS_CACHE=$HARNESS_HF_HOME/datasets
          source /opt/intel/oneapi/setvars.sh
          
          # set --limit if it's pr-triggered to accelerate pr action
          if ${{github.event_name == 'pull_request'}}; then
            export LIMIT="--limit 6"
          fi

          python run_llb.py \
            --model ipex-llm \
            --pretrained ${MODEL_PATH} \
            --precision ${{ matrix.precision }} \
            --device ${{ matrix.device }} \
            --tasks ${{ matrix.task }} \
            --batch_size 1 --no_cache --output_path results \
            $LIMIT          

      - uses: actions/upload-artifact@v3
        with:
          name: harness_results
          path:
            ${{ github.workspace }}/python/llm/dev/benchmark/harness/results/**

      - name: echo single result
        shell: bash
        
        working-directory: ${{ github.workspace }}/python/llm/dev/benchmark/harness/results/
        run: |
          cat ${{ matrix.model_name }}/${{ matrix.device }}/${{ matrix.precision }}/${{ matrix.task }}/result.json          

  llm-harness-summary:
    if: ${{ always() }}
    needs: llm-harness-evaluation
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # actions/checkout@v3
      - name: Set up Python 3.9
        uses: actions/setup-python@v4
        with:
          python-version: 3.9
      - name: Install dependencies
        shell: bash
        run: |
          pip install --upgrade pip
          pip install jsonlines  pytablewriter regex          
      - name: Download all results
        uses: actions/download-artifact@v3
        with:
          name: harness_results
          path: results        
      - name: Summarize the results
        shell: bash
        run: |
          ls results
          python ${{ github.workspace }}/python/llm/dev/benchmark/harness/make_table.py results          

  # TODO: change machine to store the results later        
  llm-harness-html:
    if: ${{github.event_name == 'schedule' || github.event_name == 'pull_request'}}
    needs: [llm-harness-evaluation]
    runs-on: ["self-hosted", "llm", "accuracy1", "accuracy-nightly"]
    steps:
      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # actions/checkout@v3
      - name: Set up Python 3.9
        uses: actions/setup-python@v4
        with:
          python-version: 3.9
      - name: Install dependencies
        shell: bash
        run: |
          pip install --upgrade pip
          pip install jsonlines  pytablewriter regex
          pip install pandas==1.5.3          
      
      - name: Set output path
        shell: bash
        run: |    
          echo "DATE=$(date +%Y-%m-%d)" >> $GITHUB_ENV
          if ${{github.event_name == 'pull_request'}}; then
            echo 'ACC_FOLDER=/home/arda/action-runners/harness/pr-accuracy-data' >> $GITHUB_ENV
          fi
          if ${{github.event_name == 'schedule'}}; then
            echo 'ACC_FOLDER=/home/arda/action-runners/harness/nightly-accuracy-data' >> $GITHUB_ENV
          fi

      - name: Download harness results
        uses: actions/download-artifact@v3
        with:
          name: harness_results
          path: ${{ env.ACC_FOLDER}}/${{ env.DATE }}


      # Save fp16.csv in the parent folder of env.nightly_folder
      - name: Download FP16 results
        shell: bash
        run: |
          wget https://raw.githubusercontent.com/intel-analytics/ipex-llm/main/python/llm/test/benchmark/harness/fp16.csv -O $ACC_FOLDER/../fp16.csv
          ls $ACC_FOLDER/..          

      - name: Write to CSV
        working-directory: ${{ github.workspace }}/python/llm/dev/benchmark/harness
        shell: bash
        run: |
          ls $ACC_FOLDER/$DATE
          python make_csv.py $ACC_FOLDER/$DATE $ACC_FOLDER          

      - name: Update HTML
        working-directory: ${{ github.workspace }}/python/llm/test/benchmark/harness
        shell: bash
        run: |
          python harness_csv_to_html.py -f $ACC_FOLDER
          if ${{github.event_name == 'schedule'}}; then
            python update_html_in_parent_folder.py -f $ACC_FOLDER
          fi