Skip to main content
Enterprise

HAMi Enterprise Air-Gap Deployment Guide

This document is intended for SRE / platform engineers and describes how to deploy HAMi Enterprise in a Kubernetes cluster using the All-in-One air-gap package, complete license activation, enable GPU nodes, and verify with sample workloads.

This delivery package uses Zarf to perform image import, Helm chart installation, and subsequent upgrades in environments without external network access or with restricted networks, reducing the manual effort of syncing images and maintaining installation order.

Zarf is an application packaging and deployment tool for Kubernetes air-gap / semi-air-gap environments; it can bundle images, Helm charts, scripts, and deployment actions into a portable package.

The installation itself does not depend on a licence; you can complete the deployment first, then apply for and import the licence in subsequent steps.

In short: install the software first, then obtain the licence; if not activated, vGPU partitioning and scheduling functionality will be unavailable, and function verification will also fail.

Air-Gap Package Contents

The outer delivery package is named as follows:

hami-enterprise-v<VERSION>-airgap-<ARCH>.tar.gz
hami-enterprise-v<VERSION>-airgap-<ARCH>.tar.gz.sha256

The current hami-enterprise-v0.0.2-airgap-amd64.tar.gz contains the following key files:

FilePurpose
zarf-linux-amd64Linux amd64 Zarf CLI
zarf-init-amd64-v0.76.0.tar.zstZarf init air-gap package
hami-enterprise-v0.0.2-airgap-amd64.tar.zstHAMi Enterprise main deployment package
zarf-package-hami-example-gpu-burn-amd64-v0.0.1.tar.zstGPU burn sample verification package
zarf-package-hami-example-vllm-qwen-amd64-v0.0.2.tar.zstvLLM + Qwen sample verification package
collect-hami-license-info.shLicense request information collection script
collect-cluster-info.shCluster diagnostic information collection script
hami/README.mdhami-enterprise complete values documentation

You can use the following command to inspect the contents of the outer package and confirm the delivery is complete:

tar -tvf hami-enterprise-v0.0.2-airgap-amd64.tar.gz

Prerequisites Checklist

TypeRequirementVerification Command
Kubernetes≥ 1.24kubectl version --short
Container Runtimecontainerd or Dockerkubectl get nodes -o wide
Helm≥ 3.14helm version --short
GPU DriverNVIDIA driver ≥ 470 (recommended ≥ 550)nvidia-smi
Prometheus CRDPrometheus monitoring CRD must be installed to be compatible with different monitoring metric collection systems: Prometheus, VictoriaMetrics, etc.kubectl api-resources | grep monitoring.coreos.com/v1
GPU OperatorInstalled and devicePlugin.enabled = false, recommended version: v25.3.2)helm list -A | grep gpu-operator
Storage SpaceRecommended greater than 30 GBdf -h

Key Constraint: HAMi includes its own device-plugin, which conflicts with the built-in device-plugin of the NVIDIA GPU Operator. If GPU Operator is already installed, be sure to disable its built-in plugin with --set devicePlugin.enabled=false.

Extract, Verify, and Install Zarf

# Download the delivery package and checksum file
curl -L -O <URL>
curl -L -O <SHA256_URL>

# Verify integrity
shasum -a 256 -c hami-enterprise-v0.0.2-airgap-amd64.tar.gz.sha256

# Extract the outer tar.gz
tar -xzf hami-enterprise-v0.0.2-airgap-amd64.tar.gz

# Enter the extracted directory
cd hami-enterprise-v0.0.2-airgap-amd64

The delivery package already includes the Linux amd64 version of the Zarf CLI. After entering the extracted directory, install the bundled Zarf first:

chmod +x ./zarf-linux-amd64
sudo install -m 0755 ./zarf-linux-amd64 /usr/local/bin/zarf
zarf version

Zarf includes the Helm tool. For subsequent troubleshooting of Helm release, values, and chart status, use zarf tools helm to avoid relying on a separately installed Helm in the target environment:

zarf tools helm version
zarf tools helm list -A

Initialize Zarf

Before deploying a Zarf package to the target cluster for the first time, run zarf init.

Use the Zarf built-in registry:

zarf init zarf-init-amd64-v0.76.0.tar.zst --confirm

Use an external registry:

zarf init zarf-init-amd64-v0.76.0.tar.zst \
--registry-url=<registry.example.com> \
--registry-push-username=<username> \
--registry-push-password=<password> \
--confirm

External registry parameter descriptions:

ParameterDescription
--registry-urlExternal image registry address
--registry-push-usernameUsername used to push images
--registry-push-passwordPassword used to push images

After initialization, subsequent zarf package deploy operations will automatically import images, "rewrite" (actually relying on an admission webhook to redirect image names) the container image field, and deploy Helm charts.

Deploy HAMi Enterprise

All components in the HAMi Enterprise main package are set as optional. Select --components based on your deployment scenario. Please keep the component order as documented.

The component list is as follows:

Component NameDescriptionRequiredRecommended
toolsOperations toolkit: jq, nerdctl, etc.NoAs needed
hami-deploy-scriptsHAMi deployment and pre-check scriptsYesYes
hamihami-enterprise Helm ChartNoYes
prometheus-crdsPrometheus monitoring CRDNoYes
prometheusKube-prometheus-stack Helm ChartNoAs needed
gpu-operatorNVIDIA GPU OperatorNoAs needed

Enterprise deployment requires a custom Helm Chart values file to override cluster, image, scheduling, and monitoring-related configurations. You must pass both --values and --features="values=true" during deployment. The Helm Chart values merge order is: Chart defaults -> in-package values/*.yaml -> --values -> --set-values. The prometheus and gpu-operator components in the air-gap package already have in-package values configurations, so no additional configuration is needed unless you have specific requirements.

Prepare Custom Values

The in-package hami/README.md provides complete values documentation for hami-enterprise. Common configuration items at a glance:

ParameterDescriptionDefault
dra.enabledWhether to enable and deploy DRAfalse
scheduler.leaderElectWhether to enable leader election for hami-scheduler across multiple nodes. Strongly recommended to disable for single-node clusters.true
scheduler.replicasAdjust the number of hami-scheduler instances1
scheduler.kubeScheduler.image.registryImage registry for the kube-scheduler image used by hami-schedulerregistry.cn-hangzhou.aliyuncs.com
scheduler.kubeScheduler.image.repositoryImage repository name for the kube-scheduler image used by hami-schedulergoogle_containers/kube-scheduler
scheduler.kubeScheduler.image.tagVersion of the kube-scheduler image used by hami-scheduler; should match the target cluster""

Minimal configuration example: my-overrides.yaml

dra:
enabled: false

scheduler:
leaderElect: true

The HAMi scheduler depends on a kube-scheduler image that matches the target Kubernetes cluster version. Additional configuration is only required when the target cluster's kube-scheduler is not running inside the cluster (and its image cannot be reused directly from kube-system); this is common in cloud-managed Kubernetes control plane clusters. If a kube-scheduler Pod exists in the cluster, you can skip the following configuration:

scheduler:
kubeScheduler:
image:
registry: your-registry.example.com
repository: google_containers/kube-scheduler
tag: v1.29.8

The built-in kube-scheduler image version in this air-gap package is v1.36.0. If the target cluster's Kubernetes version differs and you cannot reuse the existing in-cluster kube-scheduler, you must prepare and import a matching kube-scheduler image in the offline environment yourself.

Before deployment, it is recommended to check the merged values result offline:

zarf package inspect values-files hami-enterprise-v0.0.2-airgap-amd64.tar.zst \
--components=hami-deploy-scripts,tools,prometheus-crds,prometheus,gpu-operator,hami \
--values=my-overrides.yaml \
--features="values=true"

Execute Deployment

Minimal installation, deploying only the HAMi Enterprise core components:

zarf package deploy hami-enterprise-v0.0.2-airgap-amd64.tar.zst \
--components=hami-deploy-scripts,hami \
--values=my-overrides.yaml \
--features="values=true" \
--confirm

Full installation, including tools, HAMi, PrometheusCRD, Prometheus, and GPU Operator:

zarf package deploy hami-enterprise-v0.0.2-airgap-amd64.tar.zst \
--components=hami-deploy-scripts,tools,prometheus-crds,prometheus,gpu-operator,hami \
--values=my-overrides.yaml \
--features="values=true" \
--confirm

If a component remains stuck for a long time, the installation has encountered an issue. You can use zarf tools helm to diagnose the component; if the failure is caused by incorrect values during Helm rendering or installation, fix my-overrides.yaml and rerun the same zarf package deploy ... command.

After an interrupted deployment, resolve the issue and resume with the same zarf package deploy ... --components=... --values=... command. Zarf will skip re-importing images when the digest has not changed; it will perform a Helm upgrade when Helm charts or values have changed.

If zarf package deploy fails due to different Helm resource ownership leading to Kubernetes resource fields conflict, try adding the --force-conflicts flag to force an overwrite.

Enable GPU Nodes

HAMi device-plugin only starts on nodes with the gpu=on label (can be applied manually):

kubectl label nodes <node-name> gpu=on

Verify: kubectl -n hami-system get pods should show hami-device-plugin-* and hami-scheduler-* in Running state.

Monitoring Integration

Ensure the monitoring metric system in the cluster (kube-prometheus-stack Prometheus, VictoriaMetrics vmagent, etc.) can collect HAMi and DCGM-Exporter metrics.

If using Prometheus, the metadata.labels of the ServiceMonitor resource must match the spec.serviceMonitorSelector field of the Prometheus resource; otherwise Prometheus will not collect these metrics.

If using VictoriaMetrics, the metadata.labels of the ServiceMonitor resource must match the spec.serviceScrapeSelector field of the VMServiceScrape resource; otherwise vmagent will not collect these metrics.

Verify Metric Collection

ExporterQuery MetricExpected
dcgm-exporterDCGM_FI_DEV_GPU_UTILReturns a non-empty value
hami-exporterHostCoreUtilizationReturns a non-empty value
hami-device-plugin-exporterGPUDeviceCoreAllocatedReturns a non-empty value

Licence Acquisition

Please complete the installation tasks above and ensure that all component Pods are running normally before starting the activation process.

Execute the following script to collect licence information (requires kubectl, jq):

# Online script acquisition
curl -fsSL https://public.hami.run/collect-hami-license-info.sh | bash

# Offline installation (included in the package)
bash collect-hami-license-info.sh

After execution, you will see the following JSON content:

{
"esn": "96565d61-986a-4918-aafb-448ff6e3746b",
"deviceInstances": [
{
"uuid": "GPU-ceee905d-48ac-93de-a81b-17c00e1e5e02",
"deviceType": "NVIDIA A10"
}
]
}

Send the above JSON to Dynamia.ai's pre-sales / technical support to obtain the licence.

Post-Activation Verification

# 1. Pod status
kubectl -n hami-system get pods

# 2. GPU resources registered by the Device Plugin
kubectl describe node <gpu-node> | grep -A 5 'Capacity:'
# Expected to see: nvidia.com/gpu: <N> and nvidia.com/gpumem: <MB>

# 3. Submit a test Pod to verify scheduling
cat <<EOF | kubectl apply -f -
apiVersion: v1
kind: Pod
metadata:
name: hami-smoke
spec:
restartPolicy: Never
containers:
- name: cuda
image: nvidia/cuda:12.4.0-base-ubuntu22.04
command: ["nvidia-smi"]
resources:
limits:
nvidia.com/gpu: 1
nvidia.com/gpumem: 2000
EOF

kubectl logs hami-smoke

Expected: the nvidia-smi output should show GPU information, and the video memory should be limited to 2000 MiB.

Sample Workload Verification

The outer air-gap package already contains two independent Zarf sample packages, so there is no need to manually kubectl apply local YAML.

GPU Burn Verification

zarf package deploy zarf-package-hami-example-gpu-burn-amd64-v0.0.1.tar.zst --confirm

After deployment, check the Deployment / Pod status:

kubectl -n default get deploy turbo-gpu-burn
kubectl -n default get pods -l app=turbo-gpu-burn
kubectl -n default logs -l app=turbo-gpu-burn --tail=50

This sample creates default/turbo-gpu-burnDeployment; clean it up as needed after verification:

kubectl -n default delete deploy turbo-gpu-burn

vLLM + Qwen Verification

zarf package deploy zarf-package-hami-example-vllm-qwen-amd64-v0.0.2.tar.zst --confirm

After deployment, check the inference service status:

kubectl -n default get deploy vllm-qwen3
kubectl -n default get pods -l app=vllm-qwen3
kubectl -n default get svc vllm-qwen3-webui

Once the Pod is ready, access Open WebUI via any node IP + NodePort (30081):

# Get cluster node IP (any reachable node is fine)
kubectl get nodes -o wide

# Access via browser
# http://<node-ip>:30081

Open WebUI is automatically connected to the vLLM sidecar in the same Pod; open the page to start interacting.

If the Pod remains Pending, first check whether the license is activated, whether GPU nodes are labeled with gpu=on, and whether the node GPU drivers are normal.

Troubleshooting

Common inspection commands:

kubectl get pods -A | grep -E 'hami|gpu-operator|prometheus|vllm|gpu-burn'
kubectl get events -A --sort-by=.lastTimestamp | tail -50
zarf package list

Collect diagnostic information:

bash collect-cluster-info.sh

Manually inspect the main package contents:

# Extract the main package to a temporary directory; ensure sufficient space
zarf tools archiver decompress hami-enterprise-v0.0.2-airgap-amd64.tar.zst /tmp/hami-enterprise-pkg

# View package definition
cat /tmp/hami-enterprise-pkg/zarf.yaml

# View rendered manifests offline
zarf package inspect manifests hami-enterprise-v0.0.2-airgap-amd64.tar.zst --components=hami

# View rendered values offline
zarf package inspect values-files hami-enterprise-v0.0.2-airgap-amd64.tar.zst \
--components=hami \
--values=my-overrides.yaml \
--features="values=true"

FAQ

SymptomPossible CauseResolution
Image cannot be pulledNode has no external network or poor connectivity to ghcr.io.Contact Dynamia.ai's pre-sales / technical support for a domestic mirror registry address or the All-in-One offline package.
hami-device-plugin Pod Pending or does not existNode is not labeled with gpu=onkubectl label nodes <node> gpu=on
hami-device-plugin Pod CrashLoopBackOffConflicts with the NVIDIA default device-pluginDisable GPU Operator's devicePlugin (--set devicePlugin.enabled=false).
HAMi metrics not foundThe serviceMonitorSelector of the Prometheus resource does not match the labels in the ServiceMonitor resourceAlign the spec.serviceMonitorSelector of prometheus/prometheus-kube-prometheus-prometheus with the serviceMonitor labels of hami-enterprise.
nvidia-smi errorGPU driver not readyCheck the driver Pod status under the gpu-operator namespace.
Example workload PendingLicence not activated, insufficient GPU, or missing node labelsCheck the licence, GPU node labels, and kubectl describe pod events

Limitations of Installing Components with Zarf

Cannot mix internal and external image registries in the same namespace (using multiple registries in the same cluster has limitations)

Zarf uses an admission webhook to silently replace the images of all containers in a managed namespace with versions from the private registry. In a cluster, some namespaces can bypass Zarf and access other registries normally, but you cannot use images from different registries within the same namespace.You can ignore this behavior by labeling the namespace with zarf.dev/agent=ignore, allowing you to use other registries in that namespace instead of the registry specified by zarf init --registry-url ....

Native Helm cannot be used to manage the same set of resources

Once Zarf has taken over a set of components, attempting to manage the same resources with another Helm workflow can easily lead to ownership conflicts, upgrade collisions, or ambiguous state. This is not a Zarf-specific issue; it is an almost inevitable result when delivery boundaries are not clearly defined. Zarf makes it convenient to spin up a set of components, but you cannot install them and then kick Zarf aside and return to Helm for daily iteration.Therefore, unless necessary, please rely on the Migua AI team for regular version upgrades and do not iterate HAMi Enterprise on your own.

Difficult to iterate flexibly

Zarf's design means that every transfer carries the full software stack. This inherently precludes "flexibility."

Get Support

  • Email: info@dynamia.ai
  • Pre-sales / Technical Support: 400-026-7800
  • Customers with signed commercial contracts please submit issues through the dedicated support channel