GreenRay AIO

Hands-on lab simulator and mock exam for the NVIDIA-Certified Professional: AI Operations (NCP-AIO) exam.

GreenRay AIO drops you into a broken or half-configured 4-node DGX H100 cluster managed by Base Command Manager (BCM). You investigate and fix it with the same commands you'd use on real systems:

• nvidia-smi and dcgmi
• Slurm
• kubectl and helm with the GPU Operator
• cmsh
• InfiniBand tools and NCCL tests
• Docker and NGC containers

check grades the end state of the cluster, not whether you typed a particular command.

• 39 incident and task scenarios covering every topic of the exam blueprint (docs/COVERAGE.md)
• Incident mode: only the incident report is shown, and the root cause stays hidden until the debrief
• explain <command> for learning tools as you need them, plus progressive hints and a debrief after every lab
• Feels like a real node: --help and man pages, nano/vim (your real editor), cd with tab completion of paths, command history, watch, live dmon samples, a coloured prompt and real-tool colours
• Full mock exam in the current format: 30 questions plus 3 hands-on labs in 120 minutes (per NVIDIA's exam page; re-check before booking)
• Runs anywhere Python runs: Linux, WSL, macOS or Windows. No GPU needed.

root@bcm-head:~# ssh dgx-03
root@dgx-03:~# dmesg | grep -i xid
[ 8812.402113] NVRM: Xid (PCI:0000:5d:00): 79, pid=0, GPU has fallen off the bus.
root@dgx-03:~# nvidia-smi
Unable to determine the device handle for GPU00000000:5D:00.0: Unknown Error
...
root@dgx-03:~# check
  PASS  you read the kernel log on dgx-03
  ....  2 more success criteria not met yet

GreenRay AIO is an independent study tool. It is not affiliated with, endorsed by or sponsored by NVIDIA. NVIDIA, DGX, NCP-AIO and related names are trademarks of NVIDIA Corporation, used here only to describe the exam this tool helps you prepare for. Every node, GPU and log is simulated.

Install

Pick one. A GPU is not required.

With uv or pipx (any OS with Python 3.10+):

uv tool install greenray-aio      # or: pipx install greenray-aio
greenray

With Docker (nothing to install):

docker run -it --rm -v greenray:/home/lab/.greenray ghcr.io/rehmanayounis/greenray-aio

The named volume keeps your progress between runs.

From source (to contribute):

git clone https://github.com/RehmanaYounis/greenray-aio.git
cd greenray-aio && uv sync && uv run greenray

Use WSL rather than PowerShell on Windows for the best terminal experience. greenray doctor prints your environment when something looks wrong.

Developer commands (see CONTRIBUTING.md):

uv run pytest            # test suite (add --cov for coverage)
uv run ruff check .      # lint
uv run ruff format .     # format
uv run mypy              # type checks
uv build                 # wheel + sdist
docker build -t greenray-aio .

Use

Installed with uv tool, pipx or Docker, the command is greenray …. From a source checkout use uv run greenray ….

Inside a lab, these meta-commands are available alongside the simulated cluster commands:

When you pass, you get the debrief: the reference solution, why it works, the edge cases and exam traps, your own commands for comparison, and a related scenario to practise next. Read it even when you passed. Ctrl-C cancels the current line, as in bash.

Guided or independent

greenray progress shows three counts per domain:

• passed
• without hints
• as an incident, unguided: passed in incident mode with no hints and no revealed objectives

A good routine is to learn a topic in guided mode first, then prove it with greenray incident -t <topic>.

What the grader can and cannot judge

• It grades the end state of the cluster, so any valid fix passes. For example, nvidia-smi -pm 1 and systemctl enable --now nvidia-persistenced are both accepted for persistence mode.
• Evidence checks confirm that you ran the diagnostic commands. Except where a check says otherwise, they do not check that you ran them before changing anything. Diagnose first anyway; the debrief reminds you.
• It does not generally detect unnecessary changes. Only some scenarios guard against specific harmful ones, such as raising power limits to hide a thermal fault, or killing a user's job.
• If a command ever crashes inside the simulator (a bug), the cluster is restored to its state before that command and you are told so.

Progress is stored in ~/.greenray/progress.json on your machine (set GREENRAY_HOME to move it). Progress from the older ~/.aiolab folder is copied over automatically the first time you run it.

The simulated cluster

Simulated tools:

• GPU: nvidia-smi (queries, -q -d, power, persistence, ECC, reset, MIG, topology, NVLink)
• Health: dcgmi, dmesg, journalctl, systemctl, ipmitool
• Slurm: sinfo, squeue, scontrol, sbatch, srun, salloc (with pyxis), sacct, sacctmgr, sprio, sshare
• Kubernetes: kubectl, helm (GPU Operator, Network Operator), Run:ai CLI
• Containers: docker, enroot, ngc, nvidia-ctk, crictl
• BCM: cmsh (device, category, softwareimage, user, power, imageupdate, grabimage)
• Fabric and Magnum IO: ibstat, ibdiagnet, mlxlink, perfquery, NCCL tests (all_reduce_perf, mpirun), GPUDirect RDMA/Storage checks
• DPU: BlueField tools (mlxconfig, bfb-install, DOCA services)
• Shell basics: cd, pwd, ls, cat, grep, sed -i, heredocs, pipes and redirects; nano, vi and vim open your real editor; --help and man work as on a real node

commands inside a lab prints the full list.

Limits: outputs are realistic but simplified, and only the options that matter for the exam are implemented. If something doesn't parse, try the canonical form of the command. If a real command behaves differently from the simulator, trust the real one, and please open an issue.

Licence

Apache-2.0. See LICENSE and NOTICE.

Trust model

Scenario files are code: checks are Python expressions and setup may contain Python, both run on your machine. They come only from this package and, if you use one, your own solutions directory. Commands you type in a lab never reach eval/exec; they are parsed by the simulated tools. Only install scenario files you trust.

Code layout

src/greenray/
  cli.py            command-line entry point (argparse)
  core/             cluster state model, bash-like shell, errors
  commands/         one module per tool family (nvidia_smi, slurm, kubectl, cmsh, ib, ...)
  content/          loaders: scenarios, questions, solutions, command guides, validation, domains
  app/              interactive layer: lab runner, question drill, mock exam, progress, UI
  data/scenarios/   scenario files (problem only)
  data/questions/   question bank (questions only)
  data/solutions/   reference solutions, answer keys and lab solutions
  data/command_guides.yaml   scenario-agnostic guides for `explain`
tests/              pytest suite, including a replay of every reference solution
docs/               SCENARIO_FORMAT.md, COVERAGE.md

Adding scenarios and questions

• Scenarios: the problem goes in src/greenray/data/scenarios/ and the solution in src/greenray/data/solutions/scenarios/. See docs/SCENARIO_FORMAT.md. Add every new scenario to docs/COVERAGE.md.
• Questions: these go in src/greenray/data/questions/, with the answer key in src/greenray/data/solutions/questions/<same file>. The format is documented in src/greenray/content/questions.py.
• Run uv run greenray validate and uv run pytest before committing. CI runs lint and tests on every push.

All content is original and written from public documentation concepts. It is not copied from the exam.

Related

• Real-hardware practice that complements the simulator: the RTX 4060 / WSL2 labs and the rented-GPU cloud session kit live in the companion study repository, NVIDIA-NCP-AIO-Prep (labs/gpu-local, labs/cloud-gpu).