id: TC-INFERENCE-004
name: Medium Model (12b) Inference
suite: inference
priority: 4
timeout: 600000

dependencies:
  - TC-INFERENCE-003

steps:
  - name: Check if gemma3:12b model exists
    command: docker exec ollama37 ollama list | grep -q "gemma3:12b" && echo "Model exists" || echo "Model not found"

  - name: Pull gemma3:12b model if needed
    command: docker exec ollama37 ollama list | grep -q "gemma3:12b" || docker exec ollama37 ollama pull gemma3:12b
    timeout: 900000

  - name: Verify model available
    command: docker exec ollama37 ollama list | grep gemma3:12b

  - name: Warmup model (preload into GPU)
    command: |
      curl -s http://localhost:11434/api/generate \
        -d '{"model":"gemma3:12b","prompt":"hi","stream":false}' \
        | jq -r '.response' | head -c 100
    timeout: 300000

  - name: Verify model loaded to GPU
    command: |
      cd docker
      LOGS=$(docker compose logs --since=5m 2>&1)

      echo "=== Model Loading Check for gemma3:12b ==="

      # Check for layer offloading to GPU
      if echo "$LOGS" | grep -q "offloaded.*layers to GPU"; then
        echo "SUCCESS: Model layers offloaded to GPU"
        echo "$LOGS" | grep "offloaded.*layers to GPU" | tail -1
      else
        echo "ERROR: Model layers not offloaded to GPU"
        exit 1
      fi

      # Check llama runner started
      if echo "$LOGS" | grep -q "llama runner started"; then
        echo "SUCCESS: Llama runner started"
      else
        echo "ERROR: Llama runner not started"
        exit 1
      fi

  - name: Run inference test
    command: docker exec ollama37 ollama run gemma3:12b "What is the capital of France? Answer in one word." 2>&1
    timeout: 180000

  - name: Check GPU memory usage
    command: |
      echo "=== GPU Memory Usage ==="
      docker exec ollama37 nvidia-smi --query-gpu=index,memory.used,memory.total --format=csv
      echo ""
      echo "=== GPU Processes ==="
      docker exec ollama37 nvidia-smi --query-compute-apps=pid,used_memory --format=csv 2>/dev/null || echo "No GPU processes listed"

  - name: Check for inference errors
    command: |
      cd docker
      LOGS=$(docker compose logs --since=5m 2>&1)

      echo "=== Inference Error Check ==="

      if echo "$LOGS" | grep -qE "CUBLAS_STATUS_"; then
        echo "CRITICAL: CUBLAS error during inference:"
        echo "$LOGS" | grep -E "CUBLAS_STATUS_"
        exit 1
      fi

      if echo "$LOGS" | grep -qE "CUDA error"; then
        echo "CRITICAL: CUDA error during inference:"
        echo "$LOGS" | grep -E "CUDA error"
        exit 1
      fi

      if echo "$LOGS" | grep -qi "out of memory"; then
        echo "ERROR: Out of memory"
        echo "$LOGS" | grep -i "out of memory"
        exit 1
      fi

      echo "SUCCESS: No inference errors"

  - name: Unload model after test
    command: |
      echo "Unloading gemma3:12b from VRAM..."
      curl -s http://localhost:11434/api/generate -d '{"model":"gemma3:12b","keep_alive":0}' || true
      sleep 2
      echo "Model unloaded"

criteria: |
  The gemma3:12b model should run inference on Tesla K80.

  Expected:
  - Model downloads successfully (~8GB)
  - Model loads into GPU (single GPU should be sufficient)
  - Logs show "offloaded X/Y layers to GPU"
  - Logs show "llama runner started"
  - Inference returns a response mentioning "Paris"
  - NO CUBLAS_STATUS_ errors
  - NO CUDA errors
  - NO out of memory errors
  - GPU memory shows allocation (~10GB)

  This is a medium-sized model that should fit in a single K80 GPU.
  Accept any reasonable answer about France's capital.