Proposal: GB10 / DGX Spark cluster support

[inputoutputcode]

Hi @ndipebot and @lalitadithya,

I love this project and have been experimenting with extending NVCRE to support my two-node NVIDIA GB10 / DGX Spark cluster.

The current prototype targets:

• Two GB10 nodes with one GPU per node
• k3s and Kubeflow Trainer
• Dual ConnectX-7 RoCE rails
• NCCL all-reduce and all-to-all validation
• CPU/GPU NVLink-C2C shared-memory checks
• A small distributed Llama 3.2 1B workload with goodput measurement
• Integration with nvcrectl certification run and reporting

I did something similar with Slurm recently and shared the work here:

• Sparky MLOps source code
• Building an LLM training pipeline for a DGX Spark cluster with Slurm

My experimental NVCRE branch is available here:

• GB10 support prototype

Would GB10 / DGX Spark support be a contribution you would be interested in?
If so, I’d be happy to invest a few days reviewing the ADRs and design in depth before opening a PR, making sure the contribution fits the project’s architecture and avoids any major design faux pas. I’d appreciate your guidance on whether GB10 should be a first-class GPU architecture and which validation categories belong in the initial contribution.

Thanks for making NVCRE available. It has been genuinely enjoyable to test and extend.

Best,
Christian