[NemoClaw][spark 1.120.38] [doc] spark-install .md command fails because ubuntu image lacks nvidia-smi

[eko-liu]

Description

[Description]
The documented GPU smoke test command uses the ubuntu base image:bash
docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi

However, the official Ubuntu image on Docker Hub does not include the nvidia-smi binary, so this command fails with nvidia-smi: not found even when GPUs are correctly configured. The example is therefore misleading as a validation step for GPU availability.

[Environment]
Device: spark (1.120.38)
Node.js: v20.20.1
npm: 10.8.2
Docker: Docker Engine 29.3.0
OpenShell CLI: 0.0.16
NemoClaw: v0.1.0 (cec1e4)
OpenClaw: 2026.3.11 (29dc654)
Note: Behavior reproducible with the official ubuntu image from Docker Hub.

[Steps to Reproduce]

1. On a DGX Spark host with NVIDIA runtime configured, run:bash docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi
2. Observe that the container starts but the command fails because nvidia-smi is not found in the Ubuntu image.
3. Confirm that a CUDA-based image works, for example:bash docker run --rm --gpus all nvidia/cuda:12.2.0-base-ubuntu22.04 nvidia-smi

[Expected Result]
The documentation’s GPU smoke test should either:

• Use an image that actually includes nvidia-smi (for example nvidia/cuda:*‑base‑*), so the command can successfully validate GPU access, or
• Clearly state that nvidia-smi is expected to run on the host, not inside a minimal Ubuntu container.

[Actual Result]
The documented command suggests that nvidia-smi will run inside the ubuntu container, but the binary is missing in that image, causing users to see nvidia-smi: command not found or similar error even when GPU setup is correct. This can lead to confusion and misdiagnosis of GPU configuration issues.

Steps to reproduce

No steps provided.

---

[NVB# 6033961]

[NVB#6033961]

[paritoshd-nv]

Hi @eko-liu , the nvidia-container-toolkit injects/mounts nvidia-smi binary into the ubuntu container:

nvidia@spark:~$ nvidia@spark:~$ docker run --rm --runtime=nvidia --gpus all ubuntu mount | grep nvidia-smi
/dev/nvme0n1p2 on /usr/bin/nvidia-smi type ext4 (ro,nosuid,nodev,relatime,errors=remount-ro)^C

nvidia@spark:~$ nvidia-container-cli list | grep nvidia-smi
/usr/bin/nvidia-smi

It should work if your nvidia-contaner-toolkit is configured correctly.

[latenighthackathon]

Our PR #1226 switches the smoke test image to nvidia/cuda:12.8.0-base-ubuntu24.04 which bundles nvidia-smi.

However, @paritoshd-nv notes above that nvidia-container-toolkit should inject nvidia-smi into any container (including ubuntu) via bind mount when --runtime=nvidia --gpus all is used. If the toolkit is configured correctly, the plain ubuntu image should work. Our PR may be a docs improvement (more explicit about GPU requirements) rather than a strict fix — happy to adjust the approach if the maintainers prefer keeping the ubuntu image and documenting the toolkit prerequisite instead.

[paritoshd-nv]

Please see the official Spark doc: https://build.nvidia.com/spark/nemoclaw/instructions
It is also using the Ubuntu image.

We already mention here: https://github.com/NVIDIA/NemoClaw/blob/main/spark-install.md#1-verify-the-nvidia-container-runtime

If this fails, configure the NVIDIA runtime and restart Docker:

sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

Closing the issue. Please let us know if you still see it.