Skip to navigation

This page covers Helm values for the nvidia-studio-voice-h4m-operator chart (controller pod) and fields for the NvidiaStudioVoiceMediaFunction custom resource (NIM workload).

Pipeline tuning semantics on the custom resource match the Pipeline Configuration Helm keys.


Operator Helm Chart (nvidia-studio-voice-h4m-operator)

Values for the operator controller pod (not the NIM deployment):

ConfigurationHelm KeyComment
Operator imageimage.repository, image.tag, image.pullPolicyDefault: repo nvcr.io/nim/nvidia/studio-voice-h4m-operator, tag 1.2.0, pull Always.
Inline registry credentials (operator pod)imageCredentials.*Defaults: registry nvcr.io, username $oauthtoken, password "", email ""; empty password skips chart-created pull secret.
Inline registry credentials (NIM operands)mediaFunction.imageCredentials.*Same shape as imageCredentials; empty password skips the media-function registry secret.
Pull secretsimagePullSecretsDefault: [{name: ngc-api-key}] (operator pod image pull).
Default NIM image for operandsmediaFunction.image, mediaFunction.imagePullSecretsDefaults: image nvcr.io/nim/nvidia/studio-voice-h4m:1.2.0, pull secrets [{name: ngc-api-key}].
Default NIM schedulermediaFunction.schedulerNameDefault: "" (cluster default); applied to operator-managed NIM pods when the CR omits spec.schedulerName. The standalone service and sample Helm charts default to schedulerName: topo-aware-scheduler for NUMA-aligned placement and expose it as a Helm value (per-component, or via global.schedulerName on the umbrella chart).
Operator replicasreplicasDefault: 1.
Leader electionleaderElectionDefault: true.
Operator pod resourcesresources.*Defaults: limits 500m CPU / 128Mi memory; requests 10m CPU / 64Mi memory.
Metricsmetrics.enabled, metrics.portDefault: true, port 8443.
Service accountserviceAccount.create, serviceAccount.nameDefault: create true, name "".

The media-function image is configured when the operator is installed, so NvidiaStudioVoiceMediaFunction instances managed by an operator use the same validated operator and media-function image combination. This approach ensures that the operator and workload are always supported together and avoids compatibility issues.

As a result, CRs managed by an operator share the chart-level mediaFunction.image and mediaFunction.imagePullSecrets configured at install time. To run multiple custom resources with different NIM image tags, deploy multiple operators, each configured with its own mediaFunction.image. For more context, refer to Known Limitations.


(operator-media-function-cr)=

Custom Resource (NvidiaStudioVoiceMediaFunction)

Full YAML example: Example Manifest.

Scheduling and Cluster

ConfigurationCR PathComment
Node selectorspec.nodeSelectorKubernetes node labels (for example, kubernetes.io/hostname).
Schedulerspec.schedulerNameOptional. Kubernetes schedulerName for the Studio Voice NIM operand pod (for example, topo-aware-scheduler for NUMA-aligned placement of the GPU, SR-IOV virtual function, and CPU pinning, or default-scheduler for clusters that lack a topology-aware scheduler). The operator does not currently validate this field. On a NUMA-spacious cluster, default-scheduler typically schedules successfully but is not guaranteed to be NUMA-aligned, so per-packet latency can vary slightly from the Support Matrix figures while remaining within the SMPTE ST 2110-21 envelope. On a NUMA-booked cluster, the kubelet might reject the pod with TopologyAffinityError. Refer to Scheduler and NUMA Topology.

Parameters: Network, Resources, Secrets, Security

ConfigurationCR PathComment
High-speed networksspec.parameters.highSpeedNetworkDefault name: media-a-tx-net; list of {name} for Multus.
NIM pod resourcesspec.parameters.resourcesCPU, memory, hugepages, GPU.
NGC model downloadspec.parameters.ngcSecretsecretName, secretKey; default key: NGC_API_KEY.
Model cache / logsspec.parameters.nimModelCache, nimLogsEach has a path (default /opt/nim/.cache and /workspace/nim-logs) and a nested pvc.{enabled,claimName,size,storageClassName} block (default pvc.enabled: false). When pvc.enabled: true, the operator binds an existing claim with that name or creates one (annotated helm.sh/resource-policy: keep). Refer to Advanced Usage.
Security contextspec.parameters.securityContextExample: runAsUser: 1000, capabilities for Rivermax and RTP; refer to Getting Started.
Pod security contextspec.parameters.podSecurityContextfsGroup, seccompProfileType.
NMOS metadataspec.parameters.label, spec.parameters.descriptionNMOS IS-04 registration strings.

Parameters: CR

ConfigurationCR Path (spec.parameters.*)Comment
NMOS enablednmosEnabledDefault: true.
NIM enablednimEnabledDefault: true; false disables inference at runtime but does not relax the pod’s resource requests (GPU and SR-IOV VF are still reserved).
Log levelloggingLevel0 FATAL, 1 ERROR, 2 WARN, 3 INFO (default), 4 DEBUG, 5 VERBOSE.

Inputs and Outputs

AreaCR PathComment
Audio inspec.inputs.audio_inputcontrol, transport, format.audio; single mono audio input. See subfields in the next table.
Audio outspec.outputs.audio_outputEnhanced audio output; same format as input.

Subfields for format.audio

These subfields apply to audio input and audio output.

FieldExample ValueComment
sample_rate48000Audio sample rate in Hz. Fixed at 48000; the CRD rejects other values (enum: [48000]).
channel_count1Mono per stream. (Studio Voice expects a single mono speech stream.)
sample_depth24Bits; corresponds to PCM L24.
media_typeaudio/L24Fixed for ST 2110-30 mono L24.

See Also