|
| 1 | +# Test workflow for EKS Auto Mode GPU runners with ARC |
| 2 | +# This workflow validates the GPU runner infrastructure |
| 3 | + |
| 4 | +name: Test ARC GPU Runners |
| 5 | + |
| 6 | +on: |
| 7 | + workflow_dispatch: |
| 8 | + inputs: |
| 9 | + test_light_runners: |
| 10 | + description: 'Test light-gpu-runners (1 GPU)' |
| 11 | + required: false |
| 12 | + default: true |
| 13 | + type: boolean |
| 14 | + test_heavy_runners: |
| 15 | + description: 'Test heavy-gpu-runners (2 GPUs)' |
| 16 | + required: false |
| 17 | + default: true |
| 18 | + type: boolean |
| 19 | + |
| 20 | +jobs: |
| 21 | + test-light-gpu-runner: |
| 22 | + if: ${{ inputs.test_light_runners }} |
| 23 | + runs-on: light-gpu-runners |
| 24 | + timeout-minutes: 30 |
| 25 | + steps: |
| 26 | + - name: Checkout repository |
| 27 | + uses: actions/checkout@v4 |
| 28 | + |
| 29 | + - name: System Information |
| 30 | + run: | |
| 31 | + echo "=== System Information ===" |
| 32 | + echo "Hostname: $(hostname)" |
| 33 | + echo "Kernel: $(uname -r)" |
| 34 | + echo "CPU Info:" |
| 35 | + lscpu | grep -E "^(Architecture|CPU\(s\)|Model name)" |
| 36 | + echo "" |
| 37 | + echo "Memory Info:" |
| 38 | + free -h |
| 39 | + echo "" |
| 40 | + echo "Disk Info:" |
| 41 | + df -h / |
| 42 | +
|
| 43 | + - name: Verify GPU Access |
| 44 | + run: | |
| 45 | + echo "=== GPU Verification ===" |
| 46 | + if command -v nvidia-smi &> /dev/null; then |
| 47 | + echo "nvidia-smi found, checking GPU..." |
| 48 | + nvidia-smi |
| 49 | + echo "" |
| 50 | + echo "GPU Count: $(nvidia-smi -L | wc -l)" |
| 51 | + nvidia-smi -L |
| 52 | + else |
| 53 | + echo "nvidia-smi not found in runner, checking via Docker..." |
| 54 | + fi |
| 55 | +
|
| 56 | + - name: Test GPU via Docker |
| 57 | + run: | |
| 58 | + echo "=== Docker GPU Test ===" |
| 59 | + docker run --rm --gpus all nvidia/cuda:12.2.0-runtime-ubuntu22.04 nvidia-smi |
| 60 | +
|
| 61 | + - name: Verify Single GPU Allocation |
| 62 | + run: | |
| 63 | + echo "=== Verifying 1 GPU Allocation ===" |
| 64 | + GPU_COUNT=$(docker run --rm --gpus all nvidia/cuda:12.2.0-runtime-ubuntu22.04 nvidia-smi -L | wc -l) |
| 65 | + echo "GPUs available: $GPU_COUNT" |
| 66 | + if [ "$GPU_COUNT" -ge 1 ]; then |
| 67 | + echo "✅ GPU allocation verified" |
| 68 | + else |
| 69 | + echo "❌ Expected at least 1 GPU, got $GPU_COUNT" |
| 70 | + exit 1 |
| 71 | + fi |
| 72 | +
|
| 73 | + - name: Test CUDA Compute |
| 74 | + run: | |
| 75 | + echo "=== CUDA Compute Test ===" |
| 76 | + docker run --rm --gpus all nvidia/cuda:12.2.0-devel-ubuntu22.04 bash -c ' |
| 77 | + cat > /tmp/test.cu << EOF |
| 78 | + #include <stdio.h> |
| 79 | + __global__ void hello() { |
| 80 | + printf("Hello from GPU thread %d\\n", threadIdx.x); |
| 81 | + } |
| 82 | + int main() { |
| 83 | + hello<<<1, 5>>>(); |
| 84 | + cudaDeviceSynchronize(); |
| 85 | + return 0; |
| 86 | + } |
| 87 | + EOF |
| 88 | + nvcc /tmp/test.cu -o /tmp/test && /tmp/test |
| 89 | + ' |
| 90 | +
|
| 91 | + test-heavy-gpu-runner: |
| 92 | + if: ${{ inputs.test_heavy_runners }} |
| 93 | + runs-on: heavy-gpu-runners |
| 94 | + timeout-minutes: 30 |
| 95 | + steps: |
| 96 | + - name: Checkout repository |
| 97 | + uses: actions/checkout@v4 |
| 98 | + |
| 99 | + - name: System Information |
| 100 | + run: | |
| 101 | + echo "=== System Information ===" |
| 102 | + echo "Hostname: $(hostname)" |
| 103 | + echo "Kernel: $(uname -r)" |
| 104 | + echo "CPU Info:" |
| 105 | + lscpu | grep -E "^(Architecture|CPU\(s\)|Model name)" |
| 106 | + echo "" |
| 107 | + echo "Memory Info:" |
| 108 | + free -h |
| 109 | + echo "" |
| 110 | + echo "Disk Info:" |
| 111 | + df -h / |
| 112 | +
|
| 113 | + - name: Verify GPU Access |
| 114 | + run: | |
| 115 | + echo "=== GPU Verification ===" |
| 116 | + if command -v nvidia-smi &> /dev/null; then |
| 117 | + echo "nvidia-smi found, checking GPU..." |
| 118 | + nvidia-smi |
| 119 | + echo "" |
| 120 | + echo "GPU Count: $(nvidia-smi -L | wc -l)" |
| 121 | + nvidia-smi -L |
| 122 | + else |
| 123 | + echo "nvidia-smi not found in runner, checking via Docker..." |
| 124 | + fi |
| 125 | +
|
| 126 | + - name: Test GPU via Docker |
| 127 | + run: | |
| 128 | + echo "=== Docker GPU Test ===" |
| 129 | + docker run --rm --gpus all nvidia/cuda:12.2.0-runtime-ubuntu22.04 nvidia-smi |
| 130 | +
|
| 131 | + - name: Verify Multi-GPU Allocation |
| 132 | + run: | |
| 133 | + echo "=== Verifying 2 GPU Allocation ===" |
| 134 | + GPU_COUNT=$(docker run --rm --gpus all nvidia/cuda:12.2.0-runtime-ubuntu22.04 nvidia-smi -L | wc -l) |
| 135 | + echo "GPUs available: $GPU_COUNT" |
| 136 | + if [ "$GPU_COUNT" -ge 2 ]; then |
| 137 | + echo "✅ Multi-GPU allocation verified" |
| 138 | + else |
| 139 | + echo "❌ Expected at least 2 GPUs, got $GPU_COUNT" |
| 140 | + exit 1 |
| 141 | + fi |
| 142 | +
|
| 143 | + - name: Test Multi-GPU NCCL Communication |
| 144 | + run: | |
| 145 | + echo "=== Multi-GPU NCCL Test ===" |
| 146 | + docker run --rm --gpus all \ |
| 147 | + -e NCCL_DEBUG=INFO \ |
| 148 | + nvidia/cuda:12.2.0-devel-ubuntu22.04 bash -c ' |
| 149 | + apt-get update && apt-get install -y libnccl2 libnccl-dev > /dev/null 2>&1 |
| 150 | + cat > /tmp/nccl_test.cu << EOF |
| 151 | + #include <stdio.h> |
| 152 | + #include <cuda_runtime.h> |
| 153 | + int main() { |
| 154 | + int deviceCount; |
| 155 | + cudaGetDeviceCount(&deviceCount); |
| 156 | + printf("CUDA Device Count: %d\\n", deviceCount); |
| 157 | + for (int i = 0; i < deviceCount; i++) { |
| 158 | + cudaDeviceProp prop; |
| 159 | + cudaGetDeviceProperties(&prop, i); |
| 160 | + printf("Device %d: %s (Compute %d.%d)\\n", i, prop.name, prop.major, prop.minor); |
| 161 | + } |
| 162 | + // Test P2P access |
| 163 | + if (deviceCount >= 2) { |
| 164 | + int canAccess; |
| 165 | + cudaDeviceCanAccessPeer(&canAccess, 0, 1); |
| 166 | + printf("P2P Access 0->1: %s\\n", canAccess ? "Yes" : "No"); |
| 167 | + cudaDeviceCanAccessPeer(&canAccess, 1, 0); |
| 168 | + printf("P2P Access 1->0: %s\\n", canAccess ? "Yes" : "No"); |
| 169 | + } |
| 170 | + return 0; |
| 171 | + } |
| 172 | + EOF |
| 173 | + nvcc /tmp/nccl_test.cu -o /tmp/nccl_test && /tmp/nccl_test |
| 174 | + ' |
| 175 | +
|
| 176 | + summary: |
| 177 | + needs: [test-light-gpu-runner, test-heavy-gpu-runner] |
| 178 | + if: always() |
| 179 | + runs-on: ubuntu-latest |
| 180 | + steps: |
| 181 | + - name: Test Summary |
| 182 | + run: | |
| 183 | + echo "=== ARC GPU Runner Test Summary ===" |
| 184 | + echo "" |
| 185 | + echo "Light GPU Runner (1 GPU): ${{ needs.test-light-gpu-runner.result }}" |
| 186 | + echo "Heavy GPU Runner (2 GPUs): ${{ needs.test-heavy-gpu-runner.result }}" |
| 187 | + echo "" |
| 188 | + if [ "${{ needs.test-light-gpu-runner.result }}" == "success" ] || [ "${{ needs.test-light-gpu-runner.result }}" == "skipped" ]; then |
| 189 | + if [ "${{ needs.test-heavy-gpu-runner.result }}" == "success" ] || [ "${{ needs.test-heavy-gpu-runner.result }}" == "skipped" ]; then |
| 190 | + echo "✅ All requested tests passed!" |
| 191 | + exit 0 |
| 192 | + fi |
| 193 | + fi |
| 194 | + echo "❌ Some tests failed" |
| 195 | + exit 1 |
0 commit comments