Helm Deployment

Use this page to deploy the Cosmos3 NIM on Kubernetes with the shared nim-wfm Helm chart. Model selection, hardware floors, and verification are the same as in Deployment.

For the experimental NVIDIA Dynamo stack for the Reasoner, refer to the Dynamo deployment guides. They provide separate plain YAML for regular Dynamo and Dynamo with streaming, and require the Dynamo operator.

Prerequisites

Before you start, make sure you have the following:

• A Kubernetes cluster with GPU nodes that match a row in the Support Matrix, with the NVIDIA GPU Operator installed.
• Helm and kubectl access to the target namespace. The commands below use the default namespace.
• A storage class for the persistent model cache.
• NGC_API_KEY exported as described in Network and NGC Access.

Create Secrets

The chart needs an image pull secret and a generic secret with a key named NGC_API_KEY:

kubectl create secret docker-registry ngc-secret \
  --docker-server=nvcr.io \
  --docker-username='$oauthtoken' \
  --docker-password="$NGC_API_KEY"
kubectl create secret generic ngc-api \
  --from-literal=NGC_API_KEY="$NGC_API_KEY"

Fetch the Chart

This page uses nim-wfm chart version 1.3.0:

export NIM_WFM_VERSION='1.3.0'
helm fetch "https://helm.ngc.nvidia.com/nim/charts/nim-wfm-${NIM_WFM_VERSION}.tgz" \
  --username='$oauthtoken' \
  --password="$NGC_API_KEY"

The chart README lists every value supported by that version:

helm show readme "nim-wfm-${NIM_WFM_VERSION}.tgz"
helm show values "nim-wfm-${NIM_WFM_VERSION}.tgz"

Configure Values

Save the following as cosmos3-values.yaml. It deploys the Nano Generator, optimized for latency, on one GPU:

image:
  repository: nvcr.io/nim/nvidia/cosmos3
  tag: "2.1.0"
imagePullSecrets:
  - name: ngc-secret
model:
  ngcAPISecret: ngc-api
  nimCache: /opt/nim/.cache
persistence:
  enabled: true
  size: 220Gi
sharedMemory:
  sizeLimit: 16Gi
resources:
  limits:
    nvidia.com/gpu: 1
env:
  - name: NIM_MODEL_TYPE
    value: generator
  - name: NIM_MODEL_VARIANT
    value: nano
  - name: NIM_PERF_PROFILE
    value: latency

The chart also accepts these selectors as model.selection.type, model.selection.variant, model.selection.precision, and model.selection.performanceProfile values. Setting a selector both there and in env fails the render. Chart 1.3.0's model.selection does not accept the edge or edge-droid variants, so set those through env.

Adjust it for the chosen configuration:

• Model and GPUs: Change NIM_MODEL_VARIANT, NIM_PERF_PROFILE, and nvidia.com/gpu together using the Generator Profiles. The GPU count does not select a GPU type. Use nodeSelector or affinity to target eligible GPU nodes, and tolerations if those nodes are tainted. On nodes with MIG enabled, request the MIG resource instead of nvidia.com/gpu, for example nvidia.com/mig-4g.24gb: 2.
• Cache size: 220Gi is a starting size for large model files. Set persistence.storageClass if the cluster has no default storage class.
• Memory: A pod memory limit counts as system memory during profile selection. Keep any limit at or above the system-memory floor of the selected row.
• Shared memory: The chart mounts a memory-backed emptyDir at /dev/shm by default. sharedMemory.sizeLimit: 16Gi matches the 16 GiB that the Docker launch allocates.
• Pod user: The chart runs the pod as UID and GID 1000 with fsGroup: 1000 (podSecurityContext). The cache volume must be writable by that user. Storage that ignores fsGroup, such as some NFS exports and hostPath directories, needs matching ownership.

Set these through chart values rather than env:

Chart Value	Replaces
model.nimCache	NIM_CACHE_PATH
model.ngcAPISecret	NGC_API_KEY
model.apiPort	NIM_HTTP_API_PORT
model.jsonLogging	NIM_LOGGING_JSONL
model.logLevel	NIM_LOG_LEVEL

Add other Configuration variables to env.

To check profile compatibility before a cold download, run the Docker pre-download profile preflight on a host with the same GPUs and selectors.

Reasoner

For Reasoner, replace env with the Reasoner selectors and choose the GPU count from the Reasoner Profiles. Reasoner does not use NIM_PERF_PROFILE:

env:
  - name: NIM_MODEL_TYPE
    value: reasoner
  - name: NIM_MODEL_VARIANT
    value: nano

On NVIDIA DGX Spark/GB10 or Jetson AGX Thor nodes, the NIM defaults NIM_GPU_MEMORY_UTILIZATION to 0.70. Set a lower value only from host measurements. Refer to Set the Reasoner Memory Share.

OpenTelemetry

To export traces and metrics over OTLP, add these entries to env. The endpoint assumes an OpenTelemetry Collector DaemonSet that listens on host port 4318. Change OTEL_EXPORTER_OTLP_ENDPOINT for other collector layouts:

  - name: NIM_ENABLE_OTEL
    value: "1"
  - name: OTEL_SERVICE_NAME
    value: cosmos3
  - name: OTEL_TRACES_EXPORTER
    value: otlp
  - name: OTEL_METRICS_EXPORTER
    value: otlp
  - name: HOST_IP
    valueFrom:
      fieldRef:
        fieldPath: status.hostIP
  - name: OTEL_EXPORTER_OTLP_ENDPOINT
    value: "http://$(HOST_IP):4318"

HOST_IP must come before OTEL_EXPORTER_OTLP_ENDPOINT so that Kubernetes can expand it.

Other Chart Features

The chart README documents the following optional features, which are all off by default:

• Prometheus metrics: Set metrics.serviceMonitor.enabled: true to create a Prometheus Operator ServiceMonitor that scrapes /v1/metrics on the http-api port. Refer to Metrics.
• Ingress: ingress.enabled, ingress.className, ingress.hosts, and ingress.tls. The NIM does not authenticate requests, so protect any ingress outside the cluster.
• Autoscaling: autoscaling.enabled with minReplicas, maxReplicas, and metrics. CPU and memory metrics are of limited use for scaling a NIM. Use custom metrics, for example through prometheus-adapter. Each new replica must download or mount the model before it becomes ready. Refer to Storage and Scaling.

Install

Review the rendered manifests, then install the release:

helm template cosmos3 "nim-wfm-${NIM_WFM_VERSION}.tgz" -f cosmos3-values.yaml
helm install cosmos3 "nim-wfm-${NIM_WFM_VERSION}.tgz" -f cosmos3-values.yaml
kubectl get pods -w

Cold download, materialization, and model load can take much longer than pod startup. Send requests only after the pod reports Ready. Kubernetes restarts the container if it is not ready within startupProbe.initialDelaySeconds plus startupProbe.failureThreshold × startupProbe.periodSeconds. Raise failureThreshold for slow first downloads.

Verify

The chart creates a ClusterIP Service with no ingress, and the NIM does not handle authentication. Port-forward the Service for testing:

kubectl port-forward service/cosmos3-nim-wfm 8000:http-api

In another terminal, check readiness and metadata:

export NIM_URL=http://localhost:8000
curl -fsS "$NIM_URL/v1/health/ready"
curl -fsS "$NIM_URL/v1/metadata" | python3 -m json.tool

Confirm that metadata reports generator and /v1/infer, or reasoner and /v1/chat/completions. Then run a first request from Use an Existing NIM Endpoint.

Storage and Scaling

The chart supports the following cache options. Use only one of them:

Option	Values	Notes
StatefulSet volume claims	persistence.enabled: true with the default statefulSet.enabled: true	Default. Each replica gets its own claim and downloads the model before it becomes ready
Shared volume claim	persistence.enabled: true, statefulSet.enabled: false, persistence.accessMode: ReadWriteMany	One cache shared by all replicas; needs a ReadWriteMany storage class
Existing claim	persistence.enabled: true, persistence.existingClaim: <claim-name>	Reuse a pre-filled cache. Run one replica unless the claim is ReadWriteMany
Direct NFS	nfs.enabled: true, nfs.server, nfs.path	Mount options cannot be set per pod; an NFS-backed claim is usually a better choice
hostPath	hostPath.enabled: true, hostPath.path	Ties pods to one node and has security implications

Set replicaCount for a fixed number of replicas. With the default StatefulSet, you can pre-populate caches by scaling up to the maximum replica count once and then scaling down. The retained claims let later scale-ups start without a new download.

With statefulSet.enabled: false and a ReadWriteOnce volume, scaling or rolling upgrades beyond one pod are likely to fail. Use a ReadWriteMany storage class or writable direct NFS. Without persistence, the cache is an emptyDir and every pod start downloads the model again.

Troubleshooting

Symptom	Check
Pod stays Pending	Run kubectl describe pod <pod> and check Events for insufficient GPUs, untolerated taints, or an unbound volume claim
Pod restarts while preparing the model workspace	The startup probe ran out before the download finished; increase startupProbe.failureThreshold
Scaling or upgrade fails with statefulSet.enabled: false	The cache volume is ReadWriteOnce; refer to Storage and Scaling
Profile selection fails	Compare GPU type, count, free memory, and pod memory limit with the Support Matrix; refer to Troubleshooting

Remove the Deployment

helm uninstall cosmos3
kubectl get pvc

StatefulSet volume claims are retained by default and keep the model cache. Delete them and the ngc-secret and ngc-api secrets only when they are no longer needed.

Chart Parameters

These are the nim-wfm chart values used with the chart version pinned on this page (1.3.0). For the authoritative list, run helm show values "nim-wfm-${NIM_WFM_VERSION}.tgz".

Deployment Parameters

Name	Description	Value
affinity	Affinity settings for deployment.	{}
containerSecurityContext	Sets privilege and access control settings for container (Only affects the main container, not pod-level).	{}
customCommand	Overrides command line options sent to the NIM with the array listed here.	[]
customArgs	Overrides command line arguments of the NIM container with the array listed here.	[]
env	Adds arbitrary environment variables to the main container.	[]
extraEnvFrom	Adds arbitrary ConfigMap or Secret sources to the main container environment.	[]
extraVolumes	Adds arbitrary additional volumes to the deployment set definition.	{}
extraVolumeMounts	Adds volume mounts to the main container from extraVolumes.	{}
sharedMemory.enabled	Mounts a RAM-backed emptyDir at /dev/shm.	true
sharedMemory.sizeLimit	Optional Kubernetes quantity limiting the shared-memory volume.	""
image.repository	Specifies the NIM-WFM image to deploy. The chart default is another NIM; set nvcr.io/nim/nvidia/cosmos3 as in Configure Values.	nvcr.io/nvidia/nemo-microservices/nemoretriever-parse
image.tag	Specifies the image tag or version.	""
image.pullPolicy	Sets the image pull policy.	IfNotPresent
imagePullSecrets	Specifies a list of secret names that are needed for the main container and any init containers.	[{name: ngc-secret}]
initContainers	Specifies model init containers, if needed.	Refer to subkeys
initContainers.extraInit	Fully specify any additional init containers your use case requires.	[]
nodeSelector	Sets node selectors for the NIM—for example nvidia.com/gpu.present: "true".	{}
runtimeClassName	Selects the container runtime class for workload pods.	""
priorityClassName	Selects the Kubernetes priority class for workload pods.	""
topologySpreadConstraints	Configures pod topology spreading.	[]
podAnnotations	Sets additional annotations on the main deployment pods.	{}
podSecurityContext	Specifies security context settings for pod.	Refer to subkeys
podSecurityContext.runAsUser	Specify user UID for pod.	1000
podSecurityContext.runAsGroup	Specify group ID for pod.	1000
podSecurityContext.fsGroup	Specify file system owner group id.	1000
proxyCA	Specify a certificate for a custom proxy. When proxyCA is set native TLS is used for downloading from NGC.	Refer to subkeys
proxyCA.enabled	Specify true if NIM is run behind a proxy.	false
proxyCA.secretName	Specify a name of the Kubernetes secret containing the certificate. The secret is created before the deployment. Must be used together with proxyCA.keyName	""
proxyCA.keyName	Specify a name of the key inside the secret that contains the certificate. Must be used together with proxyCA.secretName	""
terminationGracePeriodSeconds	Seconds Kubernetes allows the NIM to terminate gracefully.	60
replicaCount	Specify static replica count for deployment.	1
resources	Specify resources limits and requests for the running service.	{limits: {nvidia.com/gpu: 1}}
resources.limits.nvidia.com/gpu	Specify number of GPUs to present to the running service.	1
serviceAccount	Options to specify service account for the deployment.	Refer to subkeys
serviceAccount.create	Specifies whether a service account should be created.	false
serviceAccount.annotations	Sets annotations to be added to the service account.	{}
serviceAccount.name	Specifies the name of the service account to use. If it is not set and create is true, a name is generated using a fullname template.	""
statefulSet.enabled	Enables statefulset deployment. Enabling statefulSet allows PVC templates for scaling. If using central PVC with RWX accessMode, this is not needed.	true
tolerations	Specify tolerations for pod assignment. Allows the scheduler to schedule pods with matching taints.	[{key: nvidia.com/gpu, operator: Exists, effect: NoSchedule}]

Autoscaling Parameters

Values used for creating a Horizontal Pod Autoscaler. If autoscaling is not enabled, the rest are ignored. NVIDIA recommends usage of the custom metrics API, commonly implemented with the Prometheus Adapter. Standard metrics of CPU and memory are of limited use in scaling NIM.

Name	Description	Value
autoscaling.enabled	Enables horizontal pod autoscaler.	false
autoscaling.minReplicas	Specify minimum replicas for autoscaling.	1
autoscaling.maxReplicas	Specify maximum replicas for autoscaling.	10
autoscaling.metrics	Array of metrics for autoscaling.	[]

Ingress Parameters

Name	Description	Value
ingress.enabled	Enables ingress.	false
ingress.className	Specify class name for Ingress.	""
ingress.annotations	Specify additional annotations for ingress.	{}
ingress.hosts	Specify list of hosts each containing lists of paths.	Refer to subkeys
ingress.hosts[0].host	Specify name of host.	chart-example.local
ingress.hosts[0].paths[0].path	Specify ingress path.	/
ingress.hosts[0].paths[0].pathType	Specify path type.	ImplementationSpecific
ingress.tls	Specify list of pairs of TLS secretName and hosts.	[]

Probe Parameters

Name	Description	Value
livenessProbe.enabled	Enables livenessProbe.	true
livenessProbe.method	LivenessProbe http or script, but no script is currently provided.	http
livenessProbe.command	LivenessProbe script command to use (unsupported at this time).	["myscript.sh"]
livenessProbe.path	LivenessProbe endpoint path.	/v1/health/live
livenessProbe.initialDelaySeconds	Initial delay seconds for livenessProbe.	15
livenessProbe.timeoutSeconds	Timeout seconds for livenessProbe.	1
livenessProbe.periodSeconds	Period seconds for livenessProbe.	10
livenessProbe.successThreshold	Success threshold for livenessProbe.	1
livenessProbe.failureThreshold	Failure threshold for livenessProbe.	3
readinessProbe.enabled	Enables readinessProbe.	true
readinessProbe.path	Readiness Endpoint Path.	/v1/health/ready
readinessProbe.initialDelaySeconds	Initial delay seconds for readinessProbe.	15
readinessProbe.timeoutSeconds	Timeout seconds for readinessProbe.	1
readinessProbe.periodSeconds	Period seconds for readinessProbe.	10
readinessProbe.successThreshold	Success threshold for readinessProbe.	1
readinessProbe.failureThreshold	Failure threshold for readinessProbe.	3
startupProbe.enabled	Enables startupProbe.	true
startupProbe.path	StartupProbe Endpoint Path.	/v1/health/ready
startupProbe.initialDelaySeconds	Initial delay seconds for startupProbe.	40
startupProbe.timeoutSeconds	Timeout seconds for startupProbe.	1
startupProbe.periodSeconds	Period seconds for startupProbe.	10
startupProbe.successThreshold	Success threshold for startupProbe.	1
startupProbe.failureThreshold	Failure threshold for startupProbe.	180

Helm Test Parameters

The readiness test always uses the HTTP management endpoint, including when gRPC inference is selected. The inference test is opt-in because request payloads are model-specific, and it is only rendered for model.inferenceProtocol: http.

Name	Description	Value
tests.image.repository	Container image repository used by Helm test pods.	curlimages/curl
tests.image.tag	Container image tag used by Helm test pods.	8.6.0
tests.image.pullPolicy	Container image pull policy used by Helm test pods.	IfNotPresent
tests.readiness.enabled	Enables the HTTP readiness Helm test.	true
tests.readiness.path	HTTP path used by the readiness Helm test.	/v1/health/ready
tests.readiness.timeoutSeconds	Timeout in seconds for the readiness request.	30
tests.metadata.enabled	Enables the runtime metadata Helm test.	false
tests.metadata.path	HTTP path used by the metadata Helm test.	/v1/metadata
tests.metadata.timeoutSeconds	Timeout in seconds for the metadata request.	30
tests.metadata.expectedModelType	Optional expected model_type value.	""
tests.metadata.expectedInferenceEndpoint	Optional expected inference_endpoint value.	""
tests.inference.enabled	Enables the HTTP inference Helm test. This test is only rendered when model.inferenceProtocol is http.	false
tests.inference.path	HTTP path used by the inference Helm test.	/v1/infer
tests.inference.timeoutSeconds	Timeout in seconds for the inference request.	300
tests.inference.expectedStatus	Expected HTTP status code from the inference request.	200
tests.inference.payload	Request body sent by the inference Helm test. Required when the inference test is enabled.	""
tests.inference.expectedResponseField	Optional response field required in the inference response.	""

Metrics Parameters

Name	Description	Value
metrics	Configures Prometheus integration for the NIM HTTP metrics endpoint.	Refer to subkeys
metrics.serviceMonitor	Options for serviceMonitor to use the Prometheus Operator and the primary service object.	Refer to subkeys
metrics.serviceMonitor.enabled	Enables serviceMonitor creation.	false
metrics.serviceMonitor.additionalLabels	Specify additional labels for serviceMonitor.	{}
metrics.serviceMonitor.port	Named Service port scraped by ServiceMonitor.	http-api
metrics.serviceMonitor.path	HTTP metrics path scraped by ServiceMonitor.	/v1/metrics
metrics.serviceMonitor.interval	Optional Prometheus scrape interval.	""
metrics.serviceMonitor.scrapeTimeout	Optional Prometheus scrape timeout.	""

Models Parameters

Name	Description	Value
model.nimCache	Path to mount writable storage or a pre-filled model cache for the NIM.	/model-store
model.name	Optional model name retained for compatibility with NIM-specific values files.	nvidia/Cosmos3-Nano
model.ngcAPISecret	Name of a secret containing NGC_API_KEY for model downloads. When model.ngcAPIKey is set, the chart generates this secret; leave the name empty to generate only the image pull secret.	ngc-api
model.ngcAPIKey	NGC API key literal used to generate the image pull secret and, when model.ngcAPISecret is nonempty, the model-download secret.	""
model.hfTokenSecret	Name of a pre-existing secret with a key named HF_TOKEN for Hugging Face model downloads.	""
model.apiPort	Specifies the HTTP API and management port.	8000
model.grpcPort	Specifies the gRPC inference port.	50051
model.inferenceProtocol	Selects http or grpc inference. The HTTP port remains available for health and management endpoints in gRPC mode.	http
model.selection.type	Optional runtime selector mapped to NIM_MODEL_TYPE.	""
model.selection.variant	Optional model variant mapped to NIM_MODEL_VARIANT.	""
model.selection.precision	Optional precision mapped to NIM_PRECISION; empty preserves runtime auto-selection.	""
model.selection.performanceProfile	Optional Generator objective mapped to NIM_PERF_PROFILE.	""
model.labels	Specifies extra labels to be added on deployed pods.	{}
model.jsonLogging	Turns JSON lines logging on or off.	true
model.logLevel	Log level of the NIM service. Possible values are TRACE, DEBUG, INFO, WARNING, ERROR, and CRITICAL.	INFO

Storage Parameters

Name	Description	Value
persistence	Specify settings to modify the behavior and use of persistent volumes for model weights.	Refer to subkeys
persistence.enabled	Enables the use of persistent volumes.	false
persistence.existingClaim	Specifies an existing persistent volume claim. If using existingClaim, run only one replica or use a ReadWriteMany storage setup.	""
persistence.storageClass	Specifies the persistent volume storage class. If set to "-", this disables dynamic provisioning. If left undefined or set to null, the cluster default storage provisioner is used.	""
persistence.accessMode	Specify accessMode. If using an NFS or similar setup, you can use ReadWriteMany.	ReadWriteOnce
persistence.stsPersistentVolumeClaimRetentionPolicy.whenDeleted	Specifies persistent volume claim retention policy when deleted. Only used with Stateful Set volume templates.	Retain
persistence.stsPersistentVolumeClaimRetentionPolicy.whenScaled	Specifies persistent volume claim retention policy when scaled. Only used with Stateful Set volume templates.	Retain
persistence.size	Specifies the size of the persistent volume claim (for example 40Gi).	50Gi
persistence.annotations	Adds annotations to the persistent volume claim.	{}
hostPath	Configures model cache on local disk on the nodes using hostPath—for special cases. You should understand the security implications before using this option.	Refer to subkeys
hostPath.enabled	Enable hostPath.	false
hostPath.path	Specifies path on the node used as a hostPath volume.	/model-store
nfs	Configures the model cache to sit on shared direct-mounted NFS. NOTE: you cannot set mount options using direct NFS mount to pods without a node-intalled nfsmount.conf. An NFS-based PersistentVolumeClaim is likely better in most cases.	Refer to subkeys
nfs.enabled	Enables direct pod NFS mount.	false
nfs.path	Specify path on NFS server to mount.	/exports
nfs.server	Specify NFS server address.	nfs-server.example.com
nfs.readOnly	Set to true to mount as read-only.	false

Service Parameters

Name	Description	Value
service.type	Specifies the service type for the deployment.	ClusterIP
service.name	Overrides the default service name.	""
service.apiPort	Specifies the HTTP API and management port for the service.	8000
service.grpcPort	Specifies the gRPC inference port for the service. It is exposed when model.inferenceProtocol=grpc.	50051
service.annotations	Specifies additional annotations to be added to the service.	{}
service.labels	Specifies additional labels to be added to the service.	{}