artifacts/check2/amd_developer_cloud_runbook.md

AMD Developer Cloud Runbook

This runbook validates cuda_first_repo on AMD Developer Cloud without executing untrusted code inside the ROCmPort AI Space.

1. Create an AMD GPU VM

Use an AMD Developer Cloud VM with an AMD Instinct GPU and ROCm-ready Docker support.

2. Build the ROCm container

docker build -f Dockerfile.rocm -t rocmport-cuda_first_repo .

3. Run a smoke check

docker run --rm -it \
  --device /dev/kfd \
  --device /dev/dri \
  --group-add video \
  --ipc=host \
  --network=host \
  --security-opt seccomp=unconfined \
  rocmport-cuda_first_repo

4. Run vLLM on ROCm

docker run --rm -it \
  --device /dev/kfd \
  --device /dev/dri \
  --group-add video \
  --ipc=host \
  --network=host \
  --security-opt seccomp=unconfined \
  -v "$PWD:/workspace/cuda_first_repo" \
  vllm/vllm-openai-rocm:latest \
  vllm serve Qwen/Qwen3-Coder-Next-FP8 --tensor-parallel-size 1

5. Capture benchmark metadata

rocm-smi --showproductname --showmeminfo vram --showuse
python scripts/collect_benchmark_result.py --output benchmark_result.json

Replace data/benchmark_result.json with the captured result before final submission.