ollama37/tests/testcases/inference/TC-INFERENCE-005.yml

id: TC-INFERENCE-005
name: Large Model (27b) Dual-GPU Inference
suite: inference
priority: 5
timeout: 900000

dependencies:
  - TC-INFERENCE-004

steps:
  - name: Verify dual GPU availability
    command: |
      echo "=== GPU Configuration ==="
      GPU_COUNT=$(docker exec ollama37 nvidia-smi --query-gpu=name --format=csv,noheader | wc -l)
      echo "GPUs detected: $GPU_COUNT"

      if [ "$GPU_COUNT" -lt 2 ]; then
        echo "WARNING: Less than 2 GPUs detected. 27b model may not fit."
      fi

      docker exec ollama37 nvidia-smi --query-gpu=index,name,memory.total,memory.free --format=csv

  - name: Check if gemma3:27b model exists
    command: docker exec ollama37 ollama list | grep -q "gemma3:27b" && echo "Model exists" || echo "Model not found"

  - name: Pull gemma3:27b model if needed
    command: docker exec ollama37 ollama list | grep -q "gemma3:27b" || docker exec ollama37 ollama pull gemma3:27b
    timeout: 1200000

  - name: Verify model available
    command: docker exec ollama37 ollama list | grep gemma3:27b

  - name: Warmup model (preload into both GPUs)
    command: |
      curl -s http://localhost:11434/api/generate \
        -d '{"model":"gemma3:27b","prompt":"hi","stream":false}' \
        | jq -r '.response' | head -c 100
    timeout: 600000

  - name: Verify model loaded across GPUs
    command: |
      # Use log collector file if available, fallback to docker compose logs
      if [ -f "/tmp/test-${TEST_ID}-logs.txt" ]; then
        LOGS=$(cat /tmp/test-${TEST_ID}-logs.txt)
      else
        LOGS=$(cd docker && docker compose logs --since=10m 2>&1)
      fi

      echo "=== Model Loading Check for gemma3:27b ==="

      # 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

      # Check for multi-GPU allocation in memory logs
      echo ""
      echo "=== GPU Memory Allocation ==="
      echo "$LOGS" | grep -E "device=CUDA" | tail -10

  - name: Verify both GPUs have memory allocated
    command: |
      echo "=== GPU Memory Usage ==="
      docker exec ollama37 nvidia-smi --query-gpu=index,memory.used,memory.total --format=csv

      echo ""
      echo "=== Per-GPU Process Memory ==="
      docker exec ollama37 nvidia-smi pmon -c 1 2>/dev/null || docker exec ollama37 nvidia-smi

      # Check both GPUs are being used
      GPU0_MEM=$(docker exec ollama37 nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 0 | tr -d ' ')
      GPU1_MEM=$(docker exec ollama37 nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 1 | tr -d ' ')

      echo ""
      echo "GPU 0 memory used: ${GPU0_MEM} MiB"
      echo "GPU 1 memory used: ${GPU1_MEM} MiB"

      # Both GPUs should have significant memory usage for 27b model
      if [ "$GPU0_MEM" -gt 1000 ] && [ "$GPU1_MEM" -gt 1000 ]; then
        echo "SUCCESS: Both GPUs have significant memory allocation (dual-GPU split confirmed)"
      else
        echo "WARNING: One GPU may have low memory usage - model might not be split optimally"
      fi

  - name: Run inference test
    command: docker exec ollama37 ollama run gemma3:27b "Explain quantum entanglement in one sentence." 2>&1
    timeout: 300000

  - name: Check for inference errors
    command: |
      # Use log collector file if available, fallback to docker compose logs
      if [ -f "/tmp/test-${TEST_ID}-logs.txt" ]; then
        LOGS=$(cat /tmp/test-${TEST_ID}-logs.txt)
      else
        LOGS=$(cd docker && docker compose logs --since=10m 2>&1)
      fi

      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 gemma3:27b from VRAM..."
      curl -s http://localhost:11434/api/generate -d '{"model":"gemma3:27b","keep_alive":0}' || true
      sleep 3
      echo "Model unloaded"

  - name: Verify VRAM released
    command: |
      echo "=== Post-Unload GPU Memory ==="
      docker exec ollama37 nvidia-smi --query-gpu=index,memory.used,memory.total --format=csv

criteria: |
  The gemma3:27b model should run inference using both GPUs on Tesla K80.

  Expected:
  - Model downloads successfully (~17GB)
  - Model loads and splits across both K80 GPUs
  - Logs show "offloaded X/Y layers to GPU"
  - Logs show "llama runner started"
  - Both GPU 0 and GPU 1 show significant memory usage (>1GB each)
  - Inference returns a coherent response about quantum entanglement
  - NO CUBLAS_STATUS_ errors
  - NO CUDA errors
  - NO out of memory errors

  This is a large model that requires dual-GPU on K80 (11GB + 11GB = 22GB available).
  The model (~17GB) should split layers across both GPUs.
  Accept any reasonable explanation of quantum entanglement.
  Inference will be slower than smaller models due to cross-GPU communication.