Repository navigation
Support DGX Spark Nodes #1794
Description
Describe the bug
Joining my DGX Spark to an existing k3s cluster. The error kicked back is the device plugin reports back saying "error getting device memory: Not Supported". I assume this is because of the unified RAM on the GB10.
To Reproduce
Add a Spark to a cluster
Expected behavior
I would like to be able to add the Spark as a member of my cluster and consume the GPU resources via Time Slicing at a minimum.
Environment (please provide the following information):
- GPU Operator Version: v25.3.4
- OS: DGXOS 7
- Kubernetes Distro and Version: k3s v1.33.5
Reacted by Vladimir Karnaukhov, Geoff, Benjamin Etheredge, YangShihao-Twave, Sam B and Ray Jasson
Activity
devgig commentedon Oct 22, 2025
I need the same capability
elezar commentedon Oct 22, 2025
Although not officially supported, we updated the implementation of the device plugin as of v0.17.4 to work on DGX Spark Nodes. Please set your
devicePlugin.versiontov0.17.4when deploying the GPU Operator. Alternatively, the upcoming release of the GPU Operator will include this by default.aldmbmtl commentedon Oct 28, 2025
Upgraded to v25.10.0 and the Spark now shows up properly. Thank you!
miharothl commentedon Dec 4, 2025
steps that worked for me
Fix for NVIDIA GB10 (DGX Spark) on Kubernetes with GPU Operator v25.3.2
Problem
The nvidia-device-plugin was crashing with:
error getting device memory: Not Supported
This happens because the GB10 GPU doesn't support the NVML memory query - nvidia-smi shows Memory-Usage: Not Supported.
Solution
Upgrade the device plugin to v0.17.4 (or later). This version gracefully ignores the unsupported memory query instead of
failing:
W1204 18:29:54.725748 Ignoring error getting device memory: Not Supported
I1204 18:29:54.733542 Registered device plugin for 'nvidia.com/gpu' with Kubelet
Helm Command
helm upgrade gpu-operator nvidia/gpu-operator
-n gpu-operator
--version v25.3.2
--set driver.enabled=false
--set toolkit.enabled=false
--set devicePlugin.version=v0.17.4
Note: Adjust driver.enabled and toolkit.enabled based on your setup. We had these disabled because the Spark runs the
host-installed driver.
Environment
fix: Upgrade GPU Operator to v25.3.0 for DGX Spark support
Sam-Bieberich commentedon Dec 16, 2025
Does this error have relation to:
NVMLCHK: rank 0 node [hostname] gpu 000f:01:00.0 Not Supportedfix: Upgrade GPU Operator to v25.10.1 for DGX Spark/GB10 support
fix: Upgrade GPU Operator to v25.3.0 for DGX Spark support
fix: Upgrade GPU Operator to v25.10.1 for DGX Spark/GB10 support