charts
ollama
Ollama — run large language models locally, with an optional transparent Prometheus exporter/proxy sidecar you toggle with a single switch (exporter.enabled): on, it fronts the API and exports /metrics; off, the API is served directly. This Helm chart packages the Ollama server on the bjw-s common library so behaviour is driven almost entirely from values.yaml. Supports GPU acceleration via RuntimeClass.
TL;DR
helm repo add obeone https://charts.obeone.cloud
helm repo update
helm install ollama obeone/ollama
About
Ollama runs large language models locally (Llama, Mistral, Gemma, and friends)
behind a simple HTTP API. This chart adds two opt-in extras on top of the
stock server: GPU access through a RuntimeClass, and a transparent Prometheus
exporter/proxy sidecar (exporter.enabled) that fronts the API and exposes
/metrics without any change on the client side.
- Application: ollama.com
- Container image:
ollama/ollama, exporter sidecarghcr.io/obeone/ollama-exporter - Chart source:
charts/ollama
Prerequisites
- Helm 3
- A Kubernetes cluster matching the chart’s
kubeVersionconstraint (seeChart.yaml)
Configuration
This chart is built on the bjw-s-labs common library.
Most configuration keys (controllers, service, ingress, persistence,
…) follow its schema; see the common library documentation
for everything it supports beyond what is spelled out in values.yaml.
Defaults are meant to work out of the box on any cluster. The full list of
options lives in values.yaml, is validated by
values.schema.json at install time, and is documented
in the Values section below. Override it with your own values file:
helm install ollama obeone/ollama -f my-values.yaml
GPU acceleration
Point the pod at the RuntimeClass exposing your GPUs (name depends on how the GPU operator or device plugin is installed on your cluster):
defaultPodOptions:
runtimeClassName: nvidia
Model storage
Pulled models land on the persistent volume configured under persistence.
Models are big: size it for what you actually plan to serve.
Prometheus integration
With the exporter sidecar enabled, serviceMonitor.enabled creates a
Prometheus Operator ServiceMonitor scraping its /metrics endpoint.
Upgrading
helm repo update
helm upgrade ollama obeone/ollama
Each release lists its changes in the Artifact Hub changelog; give it a look before jumping across several chart versions.
Uninstalling
helm uninstall ollama
PersistentVolumeClaims created by the chart are kept around: delete them manually if you also want the data gone.
Requirements
Kubernetes: >=1.31.0-0
| Repository | Name | Version |
|---|---|---|
| https://bjw-s-labs.github.io/helm-charts | common | 5.0.1 |
Values
| Key | Type | Default | Description |
|---|---|---|---|
| controllers.main.containers.exporter | object | {"dependsOn":"ollama","enabled":true,"env":{"OLLAMA_HOST":"http://localhost:11434"},"image":{"pullPolicy":"Always","repository":"ghcr.io/obeone/ollama-exporter","tag":"latest"},"ports":[{"containerPort":8000,"name":"proxy"}],"probes":{"liveness":{"custom":true,"enabled":true,"spec":{"failureThreshold":6,"httpGet":{"path":"/metrics","port":8000},"initialDelaySeconds":15,"periodSeconds":30,"timeoutSeconds":5}},"readiness":{"custom":true,"enabled":true,"spec":{"failureThreshold":3,"httpGet":{"path":"/api/version","port":8000},"initialDelaySeconds":10,"periodSeconds":10,"successThreshold":1,"timeoutSeconds":5}},"startup":{"custom":true,"enabled":true,"spec":{"failureThreshold":12,"httpGet":{"path":"/metrics","port":8000},"initialDelaySeconds":5,"periodSeconds":5}}},"resources":{"limits":{"cpu":"500m","memory":"512Mi"},"requests":{"cpu":"100m","memory":"128Mi"}}} |
Prometheus exporter / transparent proxy sidecar. Listens on :8000, exposes /metrics, and forwards API requests to the Ollama server in the same pod, so the http Service port targets :8000 (see service below). Source: https://github.com/frcooper/ollama-exporter (Unlicense). |
| controllers.main.containers.exporter.enabled | bool | true |
Single switch for the whole metrics/proxy path. Set to false to drop the sidecar: the chart then routes the http Service port straight to Ollama (11434) and disables the metrics port. No other value needs changing (handled in templates/common.yaml). |
| controllers.main.containers.exporter.env.OLLAMA_HOST | string | "http://localhost:11434" |
Upstream Ollama URL the exporter proxies to (same pod, localhost). |
| controllers.main.containers.exporter.image.repository | string | "ghcr.io/obeone/ollama-exporter" |
Exporter image. Multi-arch build published to GHCR and signed with cosign (keyless). Repoint to ghcr.io/frcooper/ollama-exporter if the upstream build workflow is merged there. |
| controllers.main.containers.exporter.image.tag | string | "latest" |
Image tag. Pin to a released version once tags are published. |
| controllers.main.containers.ollama.env.NVIDIA_DRIVER_CAPABILITIES | string | "compute,utility" |
|
| controllers.main.containers.ollama.env.NVIDIA_VISIBLE_DEVICES | string | "all" |
|
| controllers.main.containers.ollama.env.OLLAMA_HOST | string | "0.0.0.0" |
Bind on all interfaces so the Service and the exporter sidecar can reach Ollama. Never set this to 127.0.0.1 in-cluster. |
| controllers.main.containers.ollama.env.OLLAMA_NOHISTORY | string | "true" |
Disable prompt history persistence. |
| controllers.main.containers.ollama.env.OLLAMA_ORIGINS | string | "*" |
Allowed CORS origins. |
| controllers.main.containers.ollama.env.OLLAMA_TMPDIR | string | "/tmp/ollama" |
Scratch directory (mounted as an emptyDir below). |
| controllers.main.containers.ollama.image.pullPolicy | string | "Always" |
|
| controllers.main.containers.ollama.image.repository | string | "ollama/ollama" |
Ollama server image. |
| controllers.main.containers.ollama.image.tag | string | "latest" |
Image tag. Pin to a release (e.g. “0.6.0”) for reproducibility. |
| controllers.main.containers.ollama.probes.liveness.custom | bool | true |
|
| controllers.main.containers.ollama.probes.liveness.enabled | bool | true |
|
| controllers.main.containers.ollama.probes.liveness.spec.failureThreshold | int | 6 |
|
| controllers.main.containers.ollama.probes.liveness.spec.httpGet.path | string | "/" |
|
| controllers.main.containers.ollama.probes.liveness.spec.httpGet.port | int | 11434 |
|
| controllers.main.containers.ollama.probes.liveness.spec.initialDelaySeconds | int | 30 |
|
| controllers.main.containers.ollama.probes.liveness.spec.periodSeconds | int | 15 |
|
| controllers.main.containers.ollama.probes.liveness.spec.timeoutSeconds | int | 5 |
|
| controllers.main.containers.ollama.probes.readiness.custom | bool | true |
|
| controllers.main.containers.ollama.probes.readiness.enabled | bool | true |
|
| controllers.main.containers.ollama.probes.readiness.spec.failureThreshold | int | 6 |
|
| controllers.main.containers.ollama.probes.readiness.spec.httpGet.path | string | "/" |
|
| controllers.main.containers.ollama.probes.readiness.spec.httpGet.port | int | 11434 |
|
| controllers.main.containers.ollama.probes.readiness.spec.initialDelaySeconds | int | 10 |
|
| controllers.main.containers.ollama.probes.readiness.spec.periodSeconds | int | 10 |
|
| controllers.main.containers.ollama.probes.readiness.spec.timeoutSeconds | int | 5 |
|
| controllers.main.containers.ollama.probes.startup.enabled | bool | false |
|
| controllers.main.containers.ollama.resources.requests.cpu | string | "100m" |
|
| controllers.main.containers.ollama.resources.requests.memory | string | "2Gi" |
Bump to match the models you intend to serve (e.g. 10Gi). |
| controllers.main.pod.labels | object | {} |
Extra pod labels. Example for Sablier scale-to-zero on a GPU group: |
| controllers.main.strategy | string | "Recreate" |
|
| controllers.main.type | string | "deployment" |
|
| defaultPodOptions | object | {"automountServiceAccountToken":false,"nodeSelector":{},"runtimeClassName":""} |
Options applied to every pod created by this chart. |
| defaultPodOptions.nodeSelector | object | {} |
Pin the workload to specific node(s), typically a GPU host. |
| defaultPodOptions.runtimeClassName | string | "" |
RuntimeClass exposing the GPU to the pod (e.g. “nvidia”). Leave empty for CPU-only clusters. When set, the NVIDIA_* env vars on the ollama container enable GPU acceleration. |
| ingress.main.annotations | object | {} |
Ingress annotations. Example Traefik whitelist middleware: |
| ingress.main.className | string | "" |
|
| ingress.main.enabled | bool | false |
|
| ingress.main.hosts[0].host | string | "chart-example.local" |
|
| ingress.main.hosts[0].paths[0].path | string | "/" |
|
| ingress.main.hosts[0].paths[0].pathType | string | "Prefix" |
|
| ingress.main.hosts[0].paths[0].service.identifier | string | "main" |
|
| ingress.main.hosts[0].paths[0].service.port | string | "http" |
|
| ingress.main.tls | list | [] |
|
| persistence.data.accessMode | string | "ReadWriteOnce" |
|
| persistence.data.advancedMounts | object | {"main":{"ollama":[{"path":"/root/.ollama"}]}} |
Reuse a pre-created PVC instead of provisioning a new one. existingClaim: ollama |
| persistence.data.enabled | bool | true |
|
| persistence.data.size | string | "100Gi" |
|
| persistence.data.type | string | "persistentVolumeClaim" |
|
| persistence.tmp.advancedMounts.main.ollama[0].path | string | "/tmp/ollama" |
|
| persistence.tmp.enabled | bool | true |
|
| persistence.tmp.type | string | "emptyDir" |
|
| route | object | {"main":{"enabled":false,"hostnames":["chart-example.local"],"kind":"HTTPRoute","parentRefs":[{"name":"gateway","namespace":"gateway-system"}],"rules":[{"backendRefs":[{"identifier":"main","port":"http"}],"matches":[{"path":{"type":"PathPrefix","value":"/"}}]}]}} |
Gateway API HTTPRoute, mirroring the Ingress above. Disabled by default: pick either Ingress or HTTPRoute, not both. Requires the Gateway API CRDs and an existing Gateway in the cluster. |
| route.main.enabled | bool | false |
Enable the HTTPRoute. Mutually exclusive with ingress.main.enabled. |
| route.main.hostnames | list | ["chart-example.local"] |
Hostnames served by this route. |
| route.main.kind | string | "HTTPRoute" |
Route kind. HTTPRoute, GRPCRoute, TCPRoute, TLSRoute or UDPRoute. |
| route.main.parentRefs | list | [{"name":"gateway","namespace":"gateway-system"}] |
Gateways this route attaches to. |
| route.main.rules | list | [{"backendRefs":[{"identifier":"main","port":"http"}],"matches":[{"path":{"type":"PathPrefix","value":"/"}}]}] |
Routing rules. identifier refers to a Service defined above. |
| service.main.controller | string | "main" |
|
| service.main.ports.http.port | int | 11434 |
|
| service.main.ports.http.primary | bool | true |
|
| service.main.ports.http.protocol | string | "HTTP" |
|
| service.main.ports.http.targetPort | string | "proxy" |
|
| service.main.type | string | "ClusterIP" |
|
| serviceMonitor | object | {"metrics":{"enabled":false,"endpoints":[{"interval":"30s","path":"/metrics","port":"http","scrapeTimeout":"10s"}],"serviceName":""}} |
Prometheus Operator ServiceMonitor scraping the exporter’s /metrics. Disabled by default; enable it (or set it in your override) when a Prometheus Operator stack is present. |
Verifying the chart signature
Charts in this repository are signed with GPG and every release ships a
provenance file. The public key is available at
charts.obeone.cloud/public_key.gpg,
fingerprint B9FE852F28888D27F8C9A11CD33E04CD22E335CE.
# Import the signing key into a legacy keyring (helm verifies with GnuPG v1 keyrings)
curl -fsSL https://charts.obeone.cloud/public_key.gpg | gpg --import
gpg --export > ~/.gnupg/pubring.gpg
# Pull the chart and check it against its provenance file
helm pull --verify obeone/ollama
Support
This is a personal chart repository, maintained on a best-effort basis. Bug reports and contributions are welcome on GitHub.
Autogenerated from chart metadata using helm-docs v1.14.2