Two-Node Deployment (Remote GPU)
Kamiwaza supports a two-node topology where the control plane runs on one host and model serving containers run on a paired GPU host. This is useful for DGX Spark-style deployments and other environments where the GPU node is separate from the management node.
It is explicitly and only tested for DGX Spark or AMD Strix Halo.
How it works
- The control plane launches and manages model serving containers on the remote GPU host over SSH.
- Model files are synchronized to the remote host using
rsync. - GPU inventory is collected from the paired node and used for scheduling.
Requirements
- SSH access from the control-plane host to the GPU host
- Docker installed on the GPU host
- Network connectivity between hosts for model sync and container lifecycle
Configuration
Set the following environment variables on the control-plane host in the env.sh file post-installation and restart services:
| Variable | Description | Example |
|---|---|---|
KAMIWAZA_PAIRED_NODE | Hostname or IP of the GPU node | 10.0.0.25 |
KAMIWAZA_PAIRED_USER | SSH username | kamiwaza |
KAMIWAZA_PAIRED_KEY | SSH private key path | /etc/kamiwaza/ssl/cluster.key |
KAMIWAZA_PAIRED_MODELS_ROOT | Model storage path on GPU host | /opt/kamiwaza/models |
KAMIWAZA_PAIRED_STRICT_KNOWN_HOSTS | Enforce SSH host key checking | true or false |
Operational notes
- Use a dedicated SSH key with limited access to the GPU host.
- Ensure the GPU host can reach any upstream registries or artifact stores required for model deployment.
- If you rotate the SSH key or change the remote host, restart the control-plane services so the new settings take effect.