.github/workflows/ssh-runner.yml

name: SSH into our runners

on:
  workflow_dispatch:
    inputs:
      runner_type:
        description: 'Type of runner to test (a10)'
        required: true
      docker_image:
        description: 'Name of the Docker image'
        required: true
      num_gpus:
        description: 'Type of the number of gpus to use (`single` or `multi`)'
        required: true

env:
  HF_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
  HF_HOME: /mnt/cache
  TRANSFORMERS_IS_CI: yes
  OMP_NUM_THREADS: 8
  MKL_NUM_THREADS: 8
  RUN_SLOW: yes # For gated repositories, we still need to agree to share information on the Hub repo. page in order to get access. # This token is created under the bot `hf-transformers-bot`.
  TF_FORCE_GPU_ALLOW_GROWTH: true
  CUDA_VISIBLE_DEVICES: 0,1

jobs:
  get_runner:
    name: "Get runner to use"
    runs-on: ubuntu-22.04
    outputs:
      RUNNER: ${{ steps.set_runner.outputs.RUNNER }}
    steps:
      - name: Get runner to use
        shell: bash
        env:
          NUM_GPUS: ${{ github.event.inputs.num_gpus }}
          RUNNER_TYPE: ${{ github.event.inputs.runner_type }}
        run: |
          if [[ "$NUM_GPUS" == "single" && "$RUNNER_TYPE" == "a10" ]]; then
            echo "RUNNER=aws-g5-4xlarge-cache-ssh" >> $GITHUB_ENV
          elif [[ "$NUM_GPUS" == "multi" && "$RUNNER_TYPE" == "a10" ]]; then
            echo "RUNNER=aws-g5-12xlarge-cache-ssh" >> $GITHUB_ENV
          else
            echo "RUNNER=" >> $GITHUB_ENV
          fi

      - name: Set runner to use
        id: set_runner
        run: |
          echo "$RUNNER"
          echo "RUNNER=$RUNNER" >> $GITHUB_OUTPUT

  ssh_runner:
    name: "SSH"
    needs: get_runner
    runs-on:
      group: ${{ needs.get_runner.outputs.RUNNER }}
    container:
      image: ${{ github.event.inputs.docker_image }}
    steps:
      - name: Update clone
        working-directory: /transformers
        env:
          commit_sha: ${{ github.sha }}
        run: |
          git fetch && git checkout "$commit_sha"

      - name: Cleanup
        working-directory: /transformers
        run: |
          rm -rf tests/__pycache__
          rm -rf tests/models/__pycache__
          rm -rf reports

      - name: Show installed libraries and their versions
        working-directory: /transformers
        run: pip freeze

      - name: NVIDIA-SMI
        run: |
          nvidia-smi

      - name: Create python alias
        run: |
          ln -sf $(which python3) /usr/local/bin/python
          ln -sf $(which pip3) /usr/local/bin/pip
          echo "✅ python -> python3 symlink created"

      - name: Install psutil for memory monitor
        run: |
          pip install psutil --break-system-packages

      - name: Download memory monitor script
        working-directory: /transformers
        run: |
          apt-get update && apt-get install -y curl
          curl -o memory_monitor.py https://raw.githubusercontent.com/huggingface/transformers/refs/heads/utility_scripts/utils/memory_monitor.py

      - name: Start memory monitor
        working-directory: /transformers
        continue-on-error: true  # Don't fail workflow if monitor has issues
        run: |
          python3 memory_monitor.py --threshold 90 --interval 1 > memory_monitor.log 2>&1 &
          echo $! > memory_monitor.pid
          echo "Memory monitor started with PID $(cat memory_monitor.pid)"
          # Give it a moment to start
          sleep 2
          # Verify it's running
          ps aux | grep memory_monitor | grep -v grep || echo "Warning: memory monitor may not be running"

      - name: Install utilities
        run: |
          apt-get install -y nano

      - name: Store Slack infos
        #because the SSH can be enabled dynamically if the workflow failed, so we need to store slack infos to be able to retrieve them during the waitforssh step
        shell: bash
        env:
          GITHUB_ACTOR: ${{ github.actor }}
        run: |
          echo "$GITHUB_ACTOR"
          github_actor=$GITHUB_ACTOR
          github_actor=${github_actor/'-'/'_'}
          echo "$github_actor"
          echo "github_actor=$github_actor" >> $GITHUB_ENV

      - name: Setup automatic environment for SSH login
        run: |
          # Create shared environment setup
          cat > /root/.env_setup << 'EOF'
          # Auto-setup (non-sensitive vars)
          export HF_HOME=/mnt/cache
          export TRANSFORMERS_IS_CI=yes
          export OMP_NUM_THREADS=8
          export MKL_NUM_THREADS=8
          export RUN_SLOW=yes
          export TF_FORCE_GPU_ALLOW_GROWTH=true
          export CUDA_VISIBLE_DEVICES=0,1
          
          cd /transformers 2>/dev/null || true
          
          # Remind user to set token if needed
          if [ -z "$HF_TOKEN" ]; then
              echo "⚠️  HF_TOKEN not set. Set it with:"
              echo "    export HF_TOKEN=hf_xxxxx"
          else
              echo "✅ HF_TOKEN is set"
          fi
          
          echo "📁 Working directory: $(pwd)"
          EOF
          
          # Source from both .bash_profile and .bashrc
          echo 'source /root/.env_setup' >> /root/.bash_profile
          echo 'source /root/.env_setup' >> /root/.bashrc

      - name: Store Slack infos
        #because the SSH can be enabled dynamically if the workflow failed, so we need to store slack infos to be able to retrieve them during the waitforssh step
        shell: bash
        env:
          user_slack_id: ${{ secrets[format('{0}_{1}', env.github_actor, 'SLACK_ID')] }}
          default_slack_channel: ${{ secrets.SLACK_CIFEEDBACK_CHANNEL }}
        run: |
          echo "$github_actor"
          if [ "$user_slack_id" != "" ]; then
            echo "SLACKCHANNEL=$user_slack_id" >> $GITHUB_ENV
          else
            echo "SLACKCHANNEL=$default_slack_channel" >> $GITHUB_ENV
          fi
        
      - name: Tailscale # In order to be able to SSH when a test fails
        uses: huggingface/tailscale-action@main
        with:
          authkey: ${{ secrets.TAILSCALE_SSH_AUTHKEY }}
          slackChannel: ${{ env.SLACKCHANNEL }}
          slackToken: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
          waitForSSH: true
          sshTimeout: 15m