microsoft--agent-lightning/.github/workflows/examples-spider.yml

name: Examples - Spider
permissions:
  contents: read
on:
  schedule:
    # Every day at 4 AM UTC+8
    - cron: '0 20 * * *'

  workflow_dispatch:

  repository_dispatch:
    types: [ci-spider, ci-all]

run-name: >-
  ${{ github.event_name == 'repository_dispatch'
      && format(
        'Spider - PR #{0} - {1} - {2}',
        github.event.client_payload.pull_number,
        github.event.client_payload.ci_label,
        github.event.client_payload.correlation_id
      )
      || format('Spider - {0}', github.event_name) }}

jobs:
  spider:
    if: >
      github.event_name != 'repository_dispatch' ||
      github.event.action == 'ci-spider' ||
      github.event.action == 'ci-all'
    name: Spider (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
    timeout-minutes: 60
    strategy:
      matrix:
        include:
        # legacy is omitted because langchain doesn't work with legacy vllm versions
        - python-version: '3.12'
          setup-script: 'stable'
        - python-version: '3.13'
          setup-script: 'latest'
      fail-fast: false
    steps:
      - name: Check GPU status
        run: nvidia-smi
      - name: Check disk space
        run: df -h
      - uses: actions/checkout@v6
        with:
          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
      - uses: astral-sh/setup-uv@v7
        with:
          enable-cache: true
          python-version: ${{ matrix.python-version }}
      - name: Upgrade dependencies (latest)
        run: uv lock --upgrade
        if: matrix.setup-script == 'latest'
      - name: Sync dependencies (latest)
        run: |
          uv sync --frozen --no-default-groups --extra verl \
            --group dev --group experiment --group agents --group langchain --group torch-gpu-stable
        if: matrix.setup-script == 'latest'
      - name: Sync dependencies (stable)
        run: |
          uv sync --frozen --no-default-groups --extra verl \
            --group dev --group experiment --group agents --group langchain --group torch-gpu-${{ matrix.setup-script }}
        if: matrix.setup-script == 'stable'
      - name: Freeze dependencies
        run: |
          set -ex
          uv pip freeze | tee requirements-freeze.txt
          echo "UV_LOCKED=1" >> $GITHUB_ENV
          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
      - name: Upload dependencies artifact
        uses: actions/upload-artifact@v6
        with:
          name: dependencies-spider-${{ matrix.python-version }}-${{ matrix.setup-script }}
          path: requirements-freeze.txt
          compression-level: 0

      - name: Launch LiteLLM Proxy
        run: |
          ./scripts/litellm_run.sh
        env:
          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}

      - name: Prepare Spider dataset
        run: |
          set -ex
          cd examples/spider
          uv run gdown --fuzzy https://drive.google.com/file/d/1oi9J1jZP9TyM35L85CL3qeGWl2jqlnL6/view
          unzip -q spider-data.zip -d data
          rm spider-data.zip

      - name: Spider sanity check
        run: |
          set -ex
          cd examples/spider
          uv run sql_agent.py
        env:
          OPENAI_API_BASE: http://localhost:12306/
          OPENAI_API_KEY: dummy
        if: success() || failure()

      - name: Spider training
        run: |
          set -ex
          source .venv/bin/activate
          cd examples/spider
          ../../scripts/restart_ray.sh
          sleep 5
          PYTHONUNBUFFERED=1 python train_sql_agent.py fast
          sleep 10
        shell: bash
        env:
          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
        id: spider_train

      - name: Validate Spider training
        run: |
          set -ex
          uv run scripts/validate_example_wandb.py ${{ steps.spider_train.outputs.project_name }} ${{ steps.spider_train.outputs.run_name }} --reward-tolerance 5
        env:
          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}