API Reference

Deprecated: This field is ignored.

Deprecated: This field is ignored.

Deprecated: This field is ignored.

Deprecated: This field is ignored. See HorizontalPodAutoscalerBehavior.

Deprecated: This field is ignored. See MetricSpec.

Name of the ConfigMap containing the desired data. Validation: Required: &#123;&#125; <br />

Key in the ConfigMap to select. If not specified, defaults to “disagg.yaml”.

Name is the desired name for the created DynamoGraphDeployment.<br />If not specified, defaults to the DGDR name. Validation: Optional: &#123;&#125; <br />

Namespace is the desired namespace for the created DynamoGraphDeployment.<br />If not specified, defaults to the DGDR namespace. Validation: Optional: &#123;&#125; <br />

Labels are additional labels to add to the DynamoGraphDeployment metadata.<br />These are merged with auto-generated labels from the profiling process. Validation: Optional: &#123;&#125; <br />

Annotations are additional annotations to add to the DynamoGraphDeployment metadata. Validation: Optional: &#123;&#125; <br />

WorkersImage specifies the container image to use for DynamoGraphDeployment worker components.<br />This image is used for both temporary DGDs created during online profiling and the final DGD.<br />If omitted, the image from the base config file (e.g., disagg.yaml) is used.<br />Example: “nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.2.1” Validation: Optional: &#123;&#125; <br />

Name is the name of the created DynamoGraphDeployment.

Namespace is the namespace of the created DynamoGraphDeployment.

State is the current state of the DynamoGraphDeployment.<br />This value is mirrored from the DGD’s status.state field. See DGDState. Validation: Enum: [initializing pending successful failed] <br />

Created indicates whether the DGD has been successfully created.<br />Used to prevent recreation if the DGD is manually deleted by users.

nvidia.com/v1alpha1

DynamoCheckpoint

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

See DynamoCheckpointSpec.

See DynamoCheckpointStatus.

Model is the model identifier (e.g., “meta-llama/Llama-3-70B”)<br />Deprecated: legacy spec.identity only. Validation: Required: &#123;&#125; <br />

BackendFramework is the runtime framework (vllm, sglang, trtllm)<br />Deprecated: legacy spec.identity only. Validation: Enum: [vllm sglang trtllm] <br />Required: &#123;&#125; <br />

DynamoVersion is the Dynamo platform version (optional).<br />Deprecated: legacy spec.identity only. Validation: Optional: &#123;&#125; <br />

TensorParallelSize is the tensor parallel configuration.<br />Deprecated: checkpoint launch uses the pod template instead. Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

PipelineParallelSize is the pipeline parallel configuration.<br />Deprecated: checkpoint launch uses the pod template instead. Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

Dtype is the data type (fp16, bf16, fp8, etc.).<br />Deprecated: legacy spec.identity only. Validation: Optional: &#123;&#125; <br />

MaxModelLen is the maximum sequence length.<br />Deprecated: legacy spec.identity only. Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

ExtraParameters are additional parameters that affect the checkpoint hash.<br />Use for any framework-specific or custom parameters not covered above.<br />Deprecated: legacy spec.identity only. Validation: Optional: &#123;&#125; <br />

PodTemplateSpec allows customizing the checkpoint Job pod<br />This should include the container that runs the workload to be checkpointed<br />and any workload/runtime env, service account, GMS, or DRA wiring needed<br />by that container. Auto-created checkpoints from DynamoGraphDeployment<br />render Dynamo defaults before creating the DynamoCheckpoint. See PodTemplateSpec. Validation: Required: &#123;&#125; <br />

TargetContainerName is the container in PodTemplateSpec to snapshot. Validation: MaxLength: 63 <br />MinLength: 1 <br />Pattern: ^[a-z0-9]([-a-z0-9]*[a-z0-9])?$ <br />Optional: &#123;&#125; <br />

SharedMemory controls the tmpfs mounted at /dev/shm for the checkpoint Job pod.<br />When omitted, checkpoint Jobs use the same default 8Gi tmpfs as Dynamo components. See SharedMemorySpec. Validation: Optional: &#123;&#125; <br />

ActiveDeadlineSeconds specifies the maximum time the Job can run Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

Deprecated: BackoffLimit is ignored. Checkpoint Jobs never retry. Validation: Minimum: 0 <br />Optional: &#123;&#125; <br />

Deprecated: TTLSecondsAfterFinished is ignored. The operator deletes checkpoint<br />Jobs after recording their terminal outcome. Validation: Minimum: 0 <br />Optional: &#123;&#125; <br />

Deprecated: required by v1alpha1 for standalone checkpoints. Auto<br />checkpoints synthesize it; checkpointRef restores use the referenced CR. See DynamoCheckpointIdentity. Validation: Required: &#123;&#125; <br />

GPUMemoryService records checkpoint-time GPU Memory Service metadata for<br />a prepared checkpoint Job pod. The DynamoCheckpoint controller does not<br />inject GMS/DRA resources; auto-created checkpoints from<br />DynamoGraphDeployment prepare the pod template before creating this object.<br />Manual GMS-enabled checkpoints must provide the prepared pod template; the<br />controller fails the checkpoint if the required GMS/DRA wiring is missing.<br />This field is intentionally outside spec.identity, so it does not affect<br />the checkpoint identity hash or deduplication. See GPUMemoryServiceSpec. Validation: Optional: &#123;&#125; <br />

Job defines the configuration for the checkpoint creation Job See DynamoCheckpointJobConfig. Validation: Required: &#123;&#125; <br />

Phase represents the current phase of the checkpoint lifecycle See DynamoCheckpointPhase. Validation: Enum: [Pending Creating Ready Failed] <br />Optional: &#123;&#125; <br />

CheckpointID is the artifact ID used by the snapshot protocol. Validation: Optional: &#123;&#125; <br />

IdentityHash is the computed hash of the checkpoint identity.<br />Deprecated: use CheckpointID. This field is retained for compatibility<br />with older status consumers. Validation: Optional: &#123;&#125; <br />

Deprecated: Location is ignored and no longer populated. It is retained<br />only so older objects continue to validate. Validation: Optional: &#123;&#125; <br />

Deprecated: StorageType is ignored and no longer populated. It is retained<br />only so older objects continue to validate. See DynamoCheckpointStorageType. Validation: Enum: [pvc s3 oci] <br />Optional: &#123;&#125; <br />

JobName is the name of the checkpoint creation Job Validation: Optional: &#123;&#125; <br />

PodSnapshotName is the name of the PodSnapshot this checkpoint created to drive capture. It is<br />the authoritative pointer to the snapshot (which is otherwise located by label, not by<br />reconstructing its name) and lets the controller distinguish a never-created snapshot (empty)<br />from one that was created and later went missing (set, but no longer found). Validation: Optional: &#123;&#125; <br />

CreatedAt is the timestamp when the checkpoint became ready See Time. Validation: Optional: &#123;&#125; <br />

Message provides additional information about the current state Validation: Optional: &#123;&#125; <br />

DEPRECATED: Conditions are deprecated. Use status.phase instead. See Condition. Validation: Optional: &#123;&#125; <br />

nvidia.com/v1alpha1

DynamoComponentDeployment

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

Spec defines the desired state for this Dynamo component deployment. See DynamoComponentDeploymentSpec.

Annotations to add to generated Kubernetes resources for this component<br />(such as Pod, Service, and Ingress when applicable).

Labels to add to generated Kubernetes resources for this component.

The name of the component

ComponentType indicates the role of this component (for example, “main”).

SubComponentType indicates the sub-role of this component (for example, “prefill”).

GlobalDynamoNamespace indicates that the Component will be placed in the global Dynamo namespace

Resources requested and limits for this component, including CPU, memory,<br />GPUs/devices, and any runtime-specific resources. See Resources.

Deprecated: This field is deprecated and ignored. Use DynamoGraphDeploymentScalingAdapter<br />with HPA, KEDA, or Planner for autoscaling instead. See docs/kubernetes/autoscaling.md<br />for migration guidance. This field will be removed in a future API version. See Autoscaling.

Envs defines additional environment variables to inject into the component containers. See EnvVar.

EnvFromSecret references a Secret whose key/value pairs will be exposed as<br />environment variables in the component containers.

VolumeMounts references PVCs defined at the top level for volumes to be mounted by the component. See VolumeMount.

Ingress config to expose the component outside the cluster (or through a service mesh). See IngressSpec.

ModelRef references a model that this component serves<br />When specified, a headless service will be created for endpoint discovery See ModelReference. Validation: Optional: &#123;&#125; <br />

SharedMemory controls the tmpfs mounted at /dev/shm (enable/disable and size). See SharedMemorySpec.

ExtraPodMetadata adds labels/annotations to the created Pods. See ExtraPodMetadata. Validation: Optional: &#123;&#125; <br />

ExtraPodSpec allows to override the main pod spec configuration.<br />It is a k8s standard PodSpec. It also contains a MainContainer (standard k8s Container) field<br />that allows overriding the main container configuration. See ExtraPodSpec. Validation: Optional: &#123;&#125; <br />

LivenessProbe to detect and restart unhealthy containers. See Probe.

ReadinessProbe to signal when the container is ready to receive traffic. See Probe.

Replicas is the desired number of Pods for this component.<br />When scalingAdapter is enabled, this field is managed by the<br />DynamoGraphDeploymentScalingAdapter and should not be modified directly. Validation: Minimum: 0 <br />

MinAvailable maps to Grove PodClique minAvailable for single-node and<br />Grove PodCliqueScalingGroup minAvailable for multi-node components.<br />This field determines 1) the minimum number of replicas guaranteed to be<br />gang-scheduled, and 2) when violating minAvailable replicas triggers gang<br />termination.<br />For Grove-backed DynamoGraphDeployment components, minAvailable defaults to<br />1 when omitted and is immutable after creation. Positive replica counts must<br />be greater than or equal to minAvailable. Replicas may be scaled to 0 as a<br />special scale-to-zero state; minAvailable remains configured but is not<br />enforced again until replicas is scaled back to a positive value.<br />For non-Grove deployments, setting this field will result in a validation error. Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

Multinode is the configuration for multinode components. See MultinodeSpec.

ScalingAdapter configures whether this service uses the DynamoGraphDeploymentScalingAdapter.<br />When enabled, replicas are managed via DGDSA and external autoscalers can scale<br />the service using the Scale subresource. When disabled, replicas can be modified directly. See ScalingAdapter. Validation: Optional: &#123;&#125; <br />

EPPConfig defines EPP-specific configuration options for Endpoint Picker Plugin components.<br />Only applicable when ComponentType is “epp”. See EPPConfig. Validation: Optional: &#123;&#125; <br />

FrontendSidecar configures an auto-generated frontend sidecar container.<br />When specified, the operator injects a fully configured frontend container<br />with all standard Dynamo environment variables, health probes, and ports.<br />This eliminates the need to manually specify these in extraPodSpec.containers. (GAIE) See FrontendSidecarSpec. Validation: Optional: &#123;&#125; <br />

Checkpoint configures container checkpointing for this service.<br />When enabled, pods can be restored from a checkpoint files for faster cold start. See ServiceCheckpointConfig. Validation: Optional: &#123;&#125; <br />

TopologyConstraint for this service. packDomain is required.<br />When both this and spec.topologyConstraint.packDomain are set, packDomain<br />must be narrower than or equal to the spec-level packDomain. See TopologyConstraint. Validation: Optional: &#123;&#125; <br />

GPUMemoryService configures the GPU Memory Service (GMS) sidecar.<br />When enabled, a GMS sidecar is injected and GPU access is managed via DRA. See GPUMemoryServiceSpec. Validation: Optional: &#123;&#125; <br />

Failover configures GMS (GPU Memory Service) failover for this service.<br />For intraPod mode: the main container is cloned into two engine containers (active + standby).<br />For interPod mode: the operator creates a dedicated GMS weight server pod and<br />multiple engine pods per rank that share GPUs via DRA resource claims. See FailoverSpec. Validation: Optional: &#123;&#125; <br />

BackendFramework specifies the backend framework (e.g., “sglang”, “vllm”, “trtllm”) Validation: Enum: [sglang vllm trtllm] <br />

Annotations to add to generated Kubernetes resources for this component<br />(such as Pod, Service, and Ingress when applicable).

Labels to add to generated Kubernetes resources for this component.

The name of the component

ComponentType indicates the role of this component (for example, “main”).

SubComponentType indicates the sub-role of this component (for example, “prefill”).

GlobalDynamoNamespace indicates that the Component will be placed in the global Dynamo namespace

Resources requested and limits for this component, including CPU, memory,<br />GPUs/devices, and any runtime-specific resources. See Resources.

Deprecated: This field is deprecated and ignored. Use DynamoGraphDeploymentScalingAdapter<br />with HPA, KEDA, or Planner for autoscaling instead. See docs/kubernetes/autoscaling.md<br />for migration guidance. This field will be removed in a future API version. See Autoscaling.

Envs defines additional environment variables to inject into the component containers. See EnvVar.

EnvFromSecret references a Secret whose key/value pairs will be exposed as<br />environment variables in the component containers.

VolumeMounts references PVCs defined at the top level for volumes to be mounted by the component. See VolumeMount.

Ingress config to expose the component outside the cluster (or through a service mesh). See IngressSpec.

ModelRef references a model that this component serves<br />When specified, a headless service will be created for endpoint discovery See ModelReference. Validation: Optional: &#123;&#125; <br />

SharedMemory controls the tmpfs mounted at /dev/shm (enable/disable and size). See SharedMemorySpec.

ExtraPodMetadata adds labels/annotations to the created Pods. See ExtraPodMetadata. Validation: Optional: &#123;&#125; <br />

ExtraPodSpec allows to override the main pod spec configuration.<br />It is a k8s standard PodSpec. It also contains a MainContainer (standard k8s Container) field<br />that allows overriding the main container configuration. See ExtraPodSpec. Validation: Optional: &#123;&#125; <br />

LivenessProbe to detect and restart unhealthy containers. See Probe.

ReadinessProbe to signal when the container is ready to receive traffic. See Probe.

Replicas is the desired number of Pods for this component.<br />When scalingAdapter is enabled, this field is managed by the<br />DynamoGraphDeploymentScalingAdapter and should not be modified directly. Validation: Minimum: 0 <br />

MinAvailable maps to Grove PodClique minAvailable for single-node and<br />Grove PodCliqueScalingGroup minAvailable for multi-node components.<br />This field determines 1) the minimum number of replicas guaranteed to be<br />gang-scheduled, and 2) when violating minAvailable replicas triggers gang<br />termination.<br />For Grove-backed DynamoGraphDeployment components, minAvailable defaults to<br />1 when omitted and is immutable after creation. Positive replica counts must<br />be greater than or equal to minAvailable. Replicas may be scaled to 0 as a<br />special scale-to-zero state; minAvailable remains configured but is not<br />enforced again until replicas is scaled back to a positive value.<br />For non-Grove deployments, setting this field will result in a validation error. Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

Multinode is the configuration for multinode components. See MultinodeSpec.

ScalingAdapter configures whether this service uses the DynamoGraphDeploymentScalingAdapter.<br />When enabled, replicas are managed via DGDSA and external autoscalers can scale<br />the service using the Scale subresource. When disabled, replicas can be modified directly. See ScalingAdapter. Validation: Optional: &#123;&#125; <br />

EPPConfig defines EPP-specific configuration options for Endpoint Picker Plugin components.<br />Only applicable when ComponentType is “epp”. See EPPConfig. Validation: Optional: &#123;&#125; <br />

FrontendSidecar configures an auto-generated frontend sidecar container.<br />When specified, the operator injects a fully configured frontend container<br />with all standard Dynamo environment variables, health probes, and ports.<br />This eliminates the need to manually specify these in extraPodSpec.containers. (GAIE) See FrontendSidecarSpec. Validation: Optional: &#123;&#125; <br />

Checkpoint configures container checkpointing for this service.<br />When enabled, pods can be restored from a checkpoint files for faster cold start. See ServiceCheckpointConfig. Validation: Optional: &#123;&#125; <br />

TopologyConstraint for this service. packDomain is required.<br />When both this and spec.topologyConstraint.packDomain are set, packDomain<br />must be narrower than or equal to the spec-level packDomain. See TopologyConstraint. Validation: Optional: &#123;&#125; <br />

GPUMemoryService configures the GPU Memory Service (GMS) sidecar.<br />When enabled, a GMS sidecar is injected and GPU access is managed via DRA. See GPUMemoryServiceSpec. Validation: Optional: &#123;&#125; <br />

Failover configures GMS (GPU Memory Service) failover for this service.<br />For intraPod mode: the main container is cloned into two engine containers (active + standby).<br />For interPod mode: the operator creates a dedicated GMS weight server pod and<br />multiple engine pods per rank that share GPUs via DRA resource claims. See FailoverSpec. Validation: Optional: &#123;&#125; <br />

nvidia.com/v1alpha1

DynamoGraphDeployment

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

Spec defines the desired state for this graph deployment. See DynamoGraphDeploymentSpec.

Status reflects the current observed state of this graph deployment. See DynamoGraphDeploymentStatus.

KvTransferPolicy configures topology-aware routing for KV-cache<br />transfers between prefill and decode workers. See KvTransferPolicy. Validation: Optional: &#123;&#125; <br />

nvidia.com/v1alpha1

DynamoGraphDeploymentRequest

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

Spec defines the desired state for this deployment request. See DynamoGraphDeploymentRequestSpec.

Status reflects the current observed state of this deployment request. See DynamoGraphDeploymentRequestStatus.

Model specifies the model to deploy (e.g., “Qwen/Qwen3-0.6B”, “meta-llama/Llama-3-70b”).<br />This is a high-level identifier for easy reference in kubectl output and logs.<br />The controller automatically sets this value in profilingConfig.config.deployment.model. Validation: Required: &#123;&#125; <br />

Backend specifies the inference backend for profiling.<br />The controller automatically sets this value in profilingConfig.config.engine.backend.<br />Profiling runs on real GPUs or via AIC simulation to collect performance data. Validation: Enum: [auto vllm sglang trtllm] <br />Required: &#123;&#125; <br />

UseMocker indicates whether to deploy a mocker DynamoGraphDeployment instead of<br />a real backend deployment. When true, the deployment uses simulated engines that<br />don’t require GPUs, using the profiling data to simulate realistic timing behavior.<br />Mocker is available in all backend images and useful for large-scale experiments.<br />Profiling still runs against the real backend (specified above) to collect performance data.

ProfilingConfig provides the complete configuration for the profiling job.<br />Note: GPU discovery is automatically attempted to detect GPU resources from Kubernetes<br />cluster nodes. If the operator has node read permissions (cluster-wide or explicitly granted),<br />discovered GPU configuration is used as defaults when hardware configuration is not manually<br />specified (minNumGpusPerEngine, maxNumGpusPerEngine, numGpusPerNode). User-specified values<br />always take precedence over auto-discovered values. If GPU discovery fails (e.g.,<br />namespace-restricted operator without node permissions), manual hardware config is required.<br />This configuration is passed directly to the profiler.<br />The structure matches the profile_sla config format exactly (see ProfilingConfigSpec for schema).<br />Note: deployment.model and engine.backend are automatically set from the high-level<br />modelName and backend fields and should not be specified in this config. See ProfilingConfigSpec. Validation: Required: &#123;&#125; <br />

EnableGPUDiscovery controls whether the operator attempts to discover GPU hardware from cluster nodes.<br />DEPRECATED: This field is deprecated and will be removed in v1beta1. GPU discovery is now always<br />attempted automatically. Setting this field has no effect - the operator will always try to discover<br />GPU hardware when node read permissions are available. If discovery is unavailable (e.g., namespace-scoped<br />operator without permissions), manual hardware configuration is required regardless of this setting. Validation: Optional: &#123;&#125; <br />

AutoApply indicates whether to automatically create a DynamoGraphDeployment<br />after profiling completes. If false, only the spec is generated and stored in status.<br />Users can then manually create a DGD using the generated spec.

DeploymentOverrides allows customizing metadata for the auto-created DGD.<br />Only applicable when AutoApply is true. See DeploymentOverridesSpec. Validation: Optional: &#123;&#125; <br />

State is a high-level textual status of the deployment request lifecycle. See DGDRState. Validation: Enum: [Initializing Pending Profiling Deploying Ready DeploymentDeleted Failed] <br />

Backend is extracted from profilingConfig.config.engine.backend for display purposes.<br />This field is populated by the controller and shown in kubectl output. Validation: Optional: &#123;&#125; <br />

ObservedGeneration reflects the generation of the most recently observed spec.<br />Used to detect spec changes and enforce immutability after profiling starts.

Conditions contains the latest observed conditions of the deployment request.<br />Standard condition types include: Validation, Profiling, SpecGenerated, DeploymentReady.<br />Conditions are merged by type on patch updates. See Condition.

ProfilingResults contains a reference to the ConfigMap holding profiling data.<br />Format: “configmap/&lt;name&gt;” Validation: Optional: &#123;&#125; <br />

GeneratedDeployment contains the full generated DynamoGraphDeployment specification<br />including metadata, based on profiling results. Users can extract this to create<br />a DGD manually, or it’s used automatically when autoApply is true.<br />Stored as RawExtension to preserve all fields including metadata.<br />For mocker backends, this contains the mocker DGD spec. See RawExtension. Validation: EmbeddedResource: &#123;&#125; <br />Optional: &#123;&#125; <br />

Deployment tracks the auto-created DGD when AutoApply is true.<br />Contains name, namespace, state, and creation status of the managed DGD. See DeploymentStatus. Validation: Optional: &#123;&#125; <br />

nvidia.com/v1alpha1

DynamoGraphDeploymentScalingAdapter

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

See DynamoGraphDeploymentScalingAdapterSpec.

See DynamoGraphDeploymentScalingAdapterStatus.

Replicas is the desired number of replicas for the target service.<br />This field is modified by external autoscalers (HPA/KEDA/Planner) or manually by users. Validation: Minimum: 0 <br />Required: &#123;&#125; <br />

DGDRef references the DynamoGraphDeployment and the specific service to scale. See DynamoGraphDeploymentServiceRef. Validation: Required: &#123;&#125; <br />

Replicas is the current number of replicas for the target service.<br />This is synced from the DGD’s service replicas and is required for the scale subresource. Validation: Optional: &#123;&#125; <br />

Selector is a label selector string for the pods managed by this adapter.<br />Required for HPA compatibility via the scale subresource. Validation: Optional: &#123;&#125; <br />

LastScaleTime is the last time the adapter scaled the target service. See Time. Validation: Optional: &#123;&#125; <br />

Name of the DynamoGraphDeployment Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

ServiceName is the key name of the service within the DGD’s spec.services map to scale Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

Annotations to propagate to all child resources (PCS, DCD, Deployments, and pod templates).<br />Service-level annotations take precedence over these values. Validation: Optional: &#123;&#125; <br />

Labels to propagate to all child resources (PCS, DCD, Deployments, and pod templates).<br />Service-level labels take precedence over these values. Validation: Optional: &#123;&#125; <br />

PriorityClassName is the name of the PriorityClass to use for Grove PodCliqueSets.<br />Requires the Grove pathway. Validation: Optional: &#123;&#125; <br />

PVCs defines a list of persistent volume claims that can be referenced by components.<br />Each PVC must have a unique name that can be referenced in component specifications. See PVC. Validation: MaxItems: 100 <br />Optional: &#123;&#125; <br />

Services are the services to deploy as part of this deployment. See DynamoComponentDeploymentSharedSpec. Validation: MaxProperties: 25 <br />Optional: &#123;&#125; <br />

Envs are environment variables applied to all services in the deployment unless<br />overridden by service-specific configuration. See EnvVar. Validation: Optional: &#123;&#125; <br />

BackendFramework specifies the backend framework (e.g., “sglang”, “vllm”, “trtllm”). Validation: Enum: [sglang vllm trtllm] <br />

Restart specifies the restart policy for the graph deployment. See Restart. Validation: Optional: &#123;&#125; <br />

TopologyConstraint is the deployment-level topology constraint.<br />When set, topologyProfile is required and names the ClusterTopology CR to use.<br />packDomain is optional here — it can be omitted when only services carry constraints.<br />Services without their own topologyConstraint inherit from this value. See SpecTopologyConstraint. Validation: Optional: &#123;&#125; <br />

Experimental groups graph-level preview features whose API shape and<br />behavior may change in breaking ways between releases. See DynamoGraphDeploymentExperimentalSpec. Validation: Optional: &#123;&#125; <br />

ObservedGeneration is the most recent generation observed by the controller. Validation: Optional: &#123;&#125; <br />

State is a high-level textual status of the graph deployment lifecycle. See DGDState. Validation: Enum: [initializing pending successful failed] <br />

Conditions contains the latest observed conditions of the graph deployment.<br />The slice is merged by type on patch updates. See Condition.

Services contains per-service replica status information.<br />The map key is the service name from spec.services. See ServiceReplicaStatus. Validation: Optional: &#123;&#125; <br />

Restart contains the status of the restart of the graph deployment. See RestartStatus. Validation: Optional: &#123;&#125; <br />

Checkpoints contains per-service checkpoint status information.<br />The map key is the service name from spec.services. See ServiceCheckpointStatus. Validation: Optional: &#123;&#125; <br />

RollingUpdate tracks the progress of operator manged rolling updates.<br />Currently only supported for singl-node, non-Grove deployments (DCD/Deployment). See RollingUpdateStatus. Validation: Optional: &#123;&#125; <br />

Placement groups DGD-level scheduler placement signals (score, reporting<br />state, and any future placement fields). See PlacementStatus. Validation: Optional: &#123;&#125; <br />

nvidia.com/v1alpha1

DynamoModel

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

See DynamoModelSpec.

See DynamoModelStatus.

ModelName is the full model identifier (e.g., “meta-llama/Llama-3.3-70B-Instruct-lora”) Validation: Required: &#123;&#125; <br />

BaseModelName is the base model identifier that matches the service label<br />This is used to discover endpoints via headless services Validation: Required: &#123;&#125; <br />

ModelType specifies the type of model (e.g., “base”, “lora”, “adapter”) Validation: Enum: [base lora adapter] <br />Optional: &#123;&#125; <br />

Source specifies the model source location (only applicable for lora model type) See ModelSource. Validation: Optional: &#123;&#125; <br />

Endpoints is the current list of all endpoints for this model See EndpointInfo. Validation: Optional: &#123;&#125; <br />

ReadyEndpoints is the count of endpoints that are ready

LoRAFallbackCoveredEndpoints is the count of legacy prefill endpoints<br />covered by a capable prefill during a rolling upgrade. These endpoints are<br />excluded from ReadyEndpoints because they cannot serve the adapter directly. Validation: Optional: &#123;&#125; <br />

TotalEndpoints is the total count of endpoints

Conditions represents the latest available observations of the model’s state See Condition. Validation: Optional: &#123;&#125; <br />

ConfigMapRef references a user-provided ConfigMap containing EPP configuration.<br />The ConfigMap should contain EndpointPickerConfig YAML.<br />Mutually exclusive with Config. See ConfigMapKeySelector. Validation: Optional: &#123;&#125; <br />

Config allows specifying EPP EndpointPickerConfig directly as a structured object.<br />The operator will marshal this to YAML and create a ConfigMap automatically.<br />Mutually exclusive with ConfigMapRef.<br />One of ConfigMapRef or Config must be specified (no default configuration).<br />Uses the upstream type from github.com/kubernetes-sigs/gateway-api-inference-extension Validation: Type: object <br />Optional: &#123;&#125; <br />

Address is the full address of the endpoint (e.g., “http://10.0.1.5:9090”)

PodName is the name of the pod serving this endpoint Validation: Optional: &#123;&#125; <br />

Ready indicates whether this endpoint is ready to serve traffic.<br />For LoRA models: true only if this endpoint’s lifecycle request succeeded.<br />For base models: always false (no probing performed).

LoRAFallbackCovered indicates a legacy prefill endpoint that cannot manage<br />LoRAs itself is covered by a capable prefill in the same topology during a<br />rolling upgrade. It does not make this endpoint ready to serve the adapter. Validation: Optional: &#123;&#125; <br />

See Container.

Enabled activates failover mode.

Mode selects the failover deployment topology.<br />intraPod: engine containers run within the same pod (requires gpuMemoryService.enabled).<br />interPod: a dedicated GMS weight server pod + engine pods per rank (requires Grove). See GPUMemoryServiceMode. Validation: Enum: [intraPod interPod] <br />Optional: &#123;&#125; <br />

NumShadows is the number of shadow (standby) engine pods per rank.<br />Total engine pods per rank = NumShadows + 1 (1 primary + NumShadows shadows).<br />NumShadows is only meaningful for mode=interPod; intraPod uses a fixed<br />1 primary + 1 shadow sidecar layout and any value other than 1 is<br />rejected at admission time. Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

Image is the container image for the frontend sidecar. Validation: Required: &#123;&#125; <br />

Args overrides the default frontend arguments. When specified, these replace<br />the default [“-m”, “dynamo.frontend”] entirely.<br />For example, [“-m”, “dynamo.frontend”, “—router-mode”, “direct”] for GAIE deployments. Validation: Optional: &#123;&#125; <br />

EnvFromSecret references a Secret whose key/value pairs will be exposed as<br />environment variables in the frontend sidecar container. Validation: Optional: &#123;&#125; <br />

Envs defines additional environment variables for the frontend sidecar.<br />These are merged with (and can override) the auto-generated Dynamo env vars. See EnvVar. Validation: Optional: &#123;&#125; <br />

Name identifies this client pod. Validation: MaxLength: 63 <br />MinLength: 1 <br />Pattern: ^[a-z0-9]([-a-z0-9]*[a-z0-9])?$ <br />

PodTemplate configures the pod to run as a GMS client. See PodTemplateSpec. Validation: Schemaless: &#123;&#125; <br />Type: object <br />

Enabled activates GMS wiring. GPU resources on client containers are<br />replaced with a DRA ResourceClaim for shared GPU access.

Mode selects the GMS deployment topology. See GPUMemoryServiceMode. Validation: Enum: [intraPod interPod] <br />Optional: &#123;&#125; <br />

DeviceClassName is the DRA DeviceClass to request GPUs from. Validation: Optional: &#123;&#125; <br />

ExtraClientContainers lists additional user-declared containers that should<br />be wired as GMS clients in pods rendered from the enclosing spec.<br />DGD/DCD services apply this to service pods. Auto-created checkpoints<br />apply checkpoint job clients before creating the DynamoCheckpoint; manual<br />DynamoCheckpoint users must provide an already-prepared pod template.<br />In each rendered pod, only matching container names are wired; absent<br />names are ignored. Validation: items:MaxLength: 63 <br />items:MinLength: 1 <br />items:Pattern: ^[a-z0-9]([-a-z0-9]*[a-z0-9])?$ <br />Optional: &#123;&#125; <br />

ExtraClientPods declares additional GMS client pods for inter-pod GMS. This field is<br />reserved for future use and is rejected until inter-pod client orchestration is wired. See GMSClientPodSpec. Validation: Optional: &#123;&#125; <br />

Enabled exposes the component through an ingress or virtual service when true.

Host is the base host name to route external traffic to this component.

UseVirtualService indicates whether to configure a service-mesh VirtualService instead of a standard Ingress.

VirtualServiceGateway optionally specifies the gateway name to attach the VirtualService to.

HostPrefix is an optional prefix added before the host.

Annotations to set on the generated Ingress/VirtualService resources.

Labels to set on the generated Ingress/VirtualService resources.

TLS holds the TLS configuration used by the Ingress/VirtualService. See IngressTLSSpec.

HostSuffix is an optional suffix appended after the host.

IngressControllerClassName selects the ingress controller class (e.g., “nginx”).

SecretName is the name of a Kubernetes Secret containing the TLS certificate and key.

ClusterTopologyName references a Grove ClusterTopology CR. The operator<br />reads the CR’s topology levels and projects them through Dynamo-owned pod<br />labels for worker topology metadata. Validation: MinLength: 1 <br />Optional: &#123;&#125; <br />

LabelKey is a Kubernetes node label key (e.g.<br />“topology.kubernetes.io/zone”) whose value identifies the topology<br />domain for each worker. The operator copies the node label onto worker<br />pods so the runtime can publish it as worker metadata. The label<br />should correspond to the topology level named in domain. Validation: MaxLength: 317 <br />MinLength: 1 <br />Pattern: ^(([a-z0-9]([-a-z0-9]\{0,61\}[a-z0-9])?)(\.[a-z0-9]([-a-z0-9]\{0,61\}[a-z0-9])?)*/)?([A-Za-z0-9]([-A-Za-z0-9_.]\{0,61\}[A-Za-z0-9])?)$ <br />Optional: &#123;&#125; <br />

Domain is the logical name for the topology level to enforce<br />(e.g. “zone”, “rack”). The router uses this to match workers that<br />share the same value for the label identified by labelKey. See TopologyDomain. Validation: Pattern: ^[a-z0-9]([a-z0-9-]*[a-z0-9])?$ <br />

Enforcement controls how the selected prefill worker’s topology is<br />applied to decode routing. “required” only allows decode workers in the<br />same topology domain as the selected prefill worker. “preferred” keeps<br />all decode workers eligible, but biases selection toward workers in the<br />same topology domain. Defaults to “required”. See KvTransferEnforcement. Validation: Enum: [required preferred] <br />Optional: &#123;&#125; <br />

PreferredWeight is required and used only when enforcement is<br />“preferred”. Higher values create a stronger same-domain routing<br />preference, but do not guarantee same-domain selection. The value is not<br />a probability; worker selection still depends on load and other routing<br />inputs. A value of 0 disables the topology preference; 1 is the strongest<br />supported preference. Validation: Maximum: 1 <br />Minimum: 0 <br />Optional: &#123;&#125; <br />

Name is the base model identifier (e.g., “llama-3-70b-instruct-v1”) Validation: Required: &#123;&#125; <br />

Revision is the model revision/version (optional) Validation: Optional: &#123;&#125; <br />

URI is the model source URI<br />Supported formats:<br />- S3: s3://bucket/path/to/model<br />- HuggingFace: hf://org/model@revision_sha<br />- Local filesystem: file:///path/to/model Validation: Required: &#123;&#125; <br />

Indicates the number of nodes to deploy for multinode components.<br />Total number of GPUs is NumberOfNodes * GPU limit.<br />Must be greater than 1. Validation: Minimum: 2 <br />

Create indicates to create a new PVC

Name is the name of the PVC Validation: Required: &#123;&#125; <br />

StorageClass to be used for PVC creation. Required when create is true.

Size of the volume in Gi, used during PVC creation. Required when create is true. See Quantity.

VolumeAccessMode is the volume access mode of the PVC. Required when create is true. See PersistentVolumeAccessMode.

Score is the DGD-level scheduler placement score. Normalized to [0.0, 1.0]<br />where higher is better and 1.0 is the best possible placement. Validation: Maximum: 1 <br />Minimum: 0 <br />Optional: &#123;&#125; <br />

State indicates placement score reporting state. See PlacementScoreState. Validation: Enum: [Reported Partial Unsupported Unknown] <br />Optional: &#123;&#125; <br />

Name of the source pod. Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

UID of the source pod, recorded so the node agent dumps that specific<br />pod and not a same-named recreation. See UID. Validation: Optional: &#123;&#125; <br />

nvidia.com/v1alpha1

PodSnapshot

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

See PodSnapshotSpec.

See PodSnapshotStatus.

nvidia.com/v1alpha1

PodSnapshotContent

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

See PodSnapshotContentSpec.

See PodSnapshotContentStatus.

PodRef identifies the pod to dump. Its UID guards against dumping a<br />same-named recreation of the pod. See PodReference. Validation: Required: &#123;&#125; <br />

NodeName is the node the source pod runs on, denormalized from the live<br />pod so it travels with PodRef as one immutable unit and selects the node<br />agent that performs the dump. Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

PodSnapshotRef is the back-pointer to the bound PodSnapshot. It may span<br />namespaces because PodSnapshotContent is cluster-scoped. See PodSnapshotReference. Validation: Required: &#123;&#125; <br />

Source describes what to capture: the source pod and the node it runs on. See PodSnapshotContentSource. Validation: Required: &#123;&#125; <br />

Conditions reflect the latest observations of the PodSnapshotContent’s state.<br />Standard types are Ready and Failed. See Condition. Validation: Optional: &#123;&#125; <br />

Namespace of the referenced PodSnapshot. Validation: Required: &#123;&#125; <br />

Name of the referenced PodSnapshot. Validation: Required: &#123;&#125; <br />

UID of the referenced PodSnapshot, recorded at binding time to detect a<br />stale reference after a delete and recreate. See UID. Validation: Optional: &#123;&#125; <br />

PodRef references the pod, in the PodSnapshot’s namespace, that is captured.<br />The operator prepares the pod (control volume, target-container annotation,<br />checkpoint storage mount) before creating the PodSnapshot. See PodReference. Validation: Required: &#123;&#125; <br />

Source identifies the captured workload. It is a struct (rather than an<br />inlined reference) so future source variants can be added additively. See PodSnapshotSource. Validation: Required: &#123;&#125; <br />

BoundPodSnapshotContentName is the name of the cluster-scoped PodSnapshotContent<br />this PodSnapshot is bound to. It is nil until the agent has created the<br />content and recorded the binding. Validation: Optional: &#123;&#125; <br />

Conditions reflect the latest observations of the PodSnapshot’s state.<br />Standard types are Ready and Failed. See Condition. Validation: Optional: &#123;&#125; <br />

Config is the profiling configuration as arbitrary JSON/YAML. This will be passed directly to the profiler.<br />The profiler will validate the configuration and report any errors. See JSON. Validation: Optional: &#123;&#125; <br />Type: object <br />

ConfigMapRef is an optional reference to a ConfigMap containing the DynamoGraphDeployment<br />base config file (disagg.yaml). This is separate from the profiling config above.<br />The path to this config will be set as engine.config in the profiling config. See ConfigMapKeySelector. Validation: Optional: &#123;&#125; <br />

ProfilerImage specifies the container image to use for profiling jobs.<br />This image contains the profiler code and dependencies needed for SLA-based profiling.<br />Example: “nvcr.io/nvidia/ai-dynamo/dynamo-planner:1.2.1” Validation: Required: &#123;&#125; <br />

OutputPVC is an optional PersistentVolumeClaim name for storing profiling output.<br />If specified, all profiling artifacts (logs, plots, configs, raw data) will be written<br />to this PVC instead of an ephemeral emptyDir volume. This allows users to access<br />complete profiling results after the job completes by mounting the PVC.<br />The PVC must exist in the same namespace as the DGDR.<br />If not specified, profiling uses emptyDir and only essential data is saved to ConfigMaps.<br />Note: ConfigMaps are still created regardless of this setting for planner integration. Validation: Optional: &#123;&#125; <br />

Resources specifies the compute resource requirements for the profiling job container.<br />If not specified, no resource requests or limits are set. See ResourceRequirements. Validation: Optional: &#123;&#125; <br />

Tolerations allows the profiling job to be scheduled on nodes with matching taints.<br />For example, to schedule on GPU nodes, add a toleration for the nvidia.com/gpu taint. See Toleration. Validation: Optional: &#123;&#125; <br />

NodeSelector is a selector which must match a node’s labels for the profiling pod to be scheduled on that node.<br />For example, to schedule on ARM64 nodes, use &#123;“kubernetes.io/arch”: “arm64”&#125;. Validation: Optional: &#123;&#125; <br />

CPU specifies the CPU resource request/limit (e.g., “1000m”, “2”)

Memory specifies the memory resource request/limit (e.g., “4Gi”, “8Gi”)

GPU indicates the number of GPUs to request.<br />Total number of GPUs is NumberOfNodes * GPU in case of multinode deployment.

GPUType can specify a custom GPU type, e.g. “gpu.intel.com/xe”<br />By default if not specified, the GPU type is “nvidia.com/gpu”

Custom specifies additional custom resource requests/limits

Requests specifies the minimum resources required by the component See ResourceItem.

Limits specifies the maximum resources allowed for the component See ResourceItem.

Claims specifies resource claims for dynamic resource allocation See ResourceClaim.

ID is an arbitrary string that triggers a restart when changed.<br />Any modification to this value will initiate a restart of the graph deployment according to the strategy. Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

Strategy specifies the restart strategy for the graph deployment. See RestartStrategy. Validation: Optional: &#123;&#125; <br />

ObservedID is the restart ID that has been observed and is being processed.<br />Matches the Restart.ID field in the spec.

Phase is the phase of the restart. See RestartPhase.

InProgress contains the names of the services that are currently being restarted. Validation: Optional: &#123;&#125; <br />

Type specifies the restart strategy type. See RestartStrategyType. Validation: Enum: [Sequential Parallel] <br />

Order specifies the order in which the services should be restarted. Validation: Optional: &#123;&#125; <br />

Phase indicates the current phase of the rolling update. See RollingUpdatePhase. Validation: Enum: [Pending InProgress Completed Failed ] <br />Optional: &#123;&#125; <br />

StartTime is when the rolling update began. See Time. Validation: Optional: &#123;&#125; <br />

EndTime is when the rolling update completed (successfully or failed). See Time. Validation: Optional: &#123;&#125; <br />

UpdatedServices is the list of services that have completed the rolling update.<br />A service is considered updated when its new replicas are all ready and old replicas are fully scaled down.<br />Only services of componentType Worker (or Prefill/Decode) are considered. Validation: Optional: &#123;&#125; <br />

Enabled indicates whether the ScalingAdapter should be enabled for this service.<br />When true, a DGDSA is created and owns the replicas field.<br />When false (default), no DGDSA is created and replicas can be modified directly in the DGD. Validation: Optional: &#123;&#125; <br />

Enabled indicates whether checkpointing is enabled for this service Validation: Optional: &#123;&#125; <br />

Deprecated: omit mode. Use enabled=true without checkpointRef for a<br />DGD-managed automatic checkpoint, or use checkpointRef to restore the<br />named checkpoint. See CheckpointMode. Validation: Enum: [Auto Manual] <br />Optional: &#123;&#125; <br />

StartupPolicy defines when normal worker replicas are started relative to<br />automatic checkpoint readiness.<br />- Immediate: start workers cold immediately; later Pods restore from the<br /> checkpoint once it is Ready.<br />- WaitForCheckpoint: keep worker replicas at zero until the checkpoint is<br /> Ready, then start them from the checkpoint. See CheckpointStartupPolicy. Validation: Enum: [Immediate WaitForCheckpoint] <br />Optional: &#123;&#125; <br />

DeletionPolicy defines whether a DGD-managed automatic checkpoint CR and<br />artifact are deleted or retained when the owning DGD is deleted.<br />Explicit checkpointRef checkpoints are never owned or deleted by the DGD. See CheckpointDeletionPolicy. Validation: Enum: [Delete Retain] <br />Optional: &#123;&#125; <br />

CheckpointRef references an existing DynamoCheckpoint CR by metadata.name.<br />If specified, this service’s Identity is ignored and the referenced checkpoint is used directly. Validation: Optional: &#123;&#125; <br />

Deprecated: omit for DGD-managed checkpoints; no action is needed.<br />Use CheckpointRef to restore an existing checkpoint. See DynamoCheckpointIdentity. Validation: Optional: &#123;&#125; <br />

TargetContainerName is the workload container to snapshot and restore. Validation: MaxLength: 63 <br />MinLength: 1 <br />Pattern: ^[a-z0-9]([-a-z0-9]*[a-z0-9])?$ <br />Optional: &#123;&#125; <br />

Job customizes the DGD-managed checkpoint Job. See ServiceCheckpointJobConfig. Validation: Optional: &#123;&#125; <br />

GMSClientContainers lists checkpoint Job containers that should receive<br />GMS client wiring. Requires gpuMemoryService on the service. Validation: items:MaxLength: 63 <br />items:MinLength: 1 <br />items:Pattern: ^[a-z0-9]([-a-z0-9]*[a-z0-9])?$ <br />Optional: &#123;&#125; <br />

PodTemplate customizes the checkpoint Job pod. The operator starts from the<br />selected workload container and merges this template so users can add helper<br />containers such as gms-saver. See PodTemplateSpec. Validation: Schemaless: &#123;&#125; <br />Type: object <br />Optional: &#123;&#125; <br />

CheckpointName is the name of the associated Checkpoint CR Validation: Optional: &#123;&#125; <br />

CheckpointID is the artifact ID used by the snapshot protocol Validation: Optional: &#123;&#125; <br />

IdentityHash is the computed hash of the checkpoint identity<br />Deprecated: automatic checkpoints use CheckpointID. This field is retained<br />for older status consumers. Validation: Optional: &#123;&#125; <br />

Ready indicates the checkpoint artifact is ready for future pods to restore. Validation: Optional: &#123;&#125; <br />

ComponentKind is the underlying resource kind (e.g., “PodClique”, “PodCliqueScalingGroup”, “Deployment”, “LeaderWorkerSet”). See ComponentKind. Validation: Enum: [PodClique PodCliqueScalingGroup Deployment LeaderWorkerSet] <br />

ComponentName is the name of the primary underlying resource.<br />DEPRECATED: Use ComponentNames instead. This field will be removed in a future release.<br />During rolling updates, this reflects the new (target) component name.

ComponentNames is the list of underlying resource names for this service.<br />During normal operation, this contains a single name.<br />During rolling updates, this contains both old and new component names. Validation: Optional: &#123;&#125; <br />

RuntimeNamespace is the effective Dynamo runtime namespace for this<br />component. Worker components may include a generation suffix; non-workers and<br />Grove-backed workers use the base namespace. During rolling updates, worker<br />status keeps the old active revision namespace until cutover completes. Validation: Optional: &#123;&#125; <br />

Replicas is the total number of non-terminated replicas.<br />Required for all component kinds. Validation: Minimum: 0 <br />

UpdatedReplicas is the number of replicas at the current/desired revision.<br />Required for all component kinds. Validation: Minimum: 0 <br />

ReadyReplicas is the number of ready replicas.<br />Populated for PodClique, Deployment, and LeaderWorkerSet.<br />Not available for PodCliqueScalingGroup.<br />When nil, the field is omitted from the API response. Validation: Minimum: 0 <br />Optional: &#123;&#125; <br />

AvailableReplicas is the number of available replicas.<br />For Deployment: replicas ready for >= minReadySeconds.<br />For PodCliqueScalingGroup: replicas where all constituent PodCliques have >= MinAvailable ready pods.<br />Not available for PodClique or LeaderWorkerSet.<br />When nil, the field is omitted from the API response. Validation: Minimum: 0 <br />Optional: &#123;&#125; <br />

ScheduledReplicas is the number of replicas the backend scheduler has<br />scheduled, in Dynamo component-replica units. Optional; omitted (nil)<br />when the backend cannot derive it reliably. A nil value means “not<br />reported”, never “zero scheduled”. Validation: Minimum: 0 <br />Optional: &#123;&#125; <br />

Disabled, when true, opts out of mounting a shared-memory medium for the<br />component. When false (or unset), shared memory is enabled and Size is<br />required (enforced by the validating webhook). Size is ignored when<br />Disabled is true.

See Quantity.

TopologyProfile is the name of the ClusterTopology CR that defines the<br />topology hierarchy for this deployment. Validation: MinLength: 1 <br />

PackDomain is the default topology domain to pack pods within.<br />Optional — omit when only services carry constraints. See TopologyDomain. Validation: Pattern: ^[a-z0-9]([a-z0-9-]*[a-z0-9])?$ <br />Optional: &#123;&#125; <br />

PackDomain is the topology domain to pack pods within. Must match a<br />domain defined in the referenced ClusterTopology CR. See TopologyDomain. Validation: Pattern: ^[a-z0-9]([a-z0-9-]*[a-z0-9])?$ <br />

Name references a PVC name defined in the top-level PVCs map Validation: Required: &#123;&#125; <br />

MountPoint specifies where to mount the volume.<br />If useAsCompilationCache is true and mountPoint is not specified,<br />a backend-specific default will be used.

UseAsCompilationCache indicates this volume should be used as a compilation cache.<br />When true, backend-specific environment variables will be set and default mount points may be used.

pvcName references a user-created PVC by name. The PVC must exist in<br />the same namespace as the DynamoGraphDeployment. Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

mountPath overrides the backend-specific default mount path. When<br />empty, the operator selects a default appropriate for the backend<br />framework. Validation: Optional: &#123;&#125; <br />

enabled indicates whether checkpointing is enabled for this component.<br />When true, omit checkpointRef for a DGD-managed automatic checkpoint or<br />set checkpointRef to restore an existing checkpoint. Omit the checkpoint<br />block, or set enabled=false, to disable checkpointing. Validation: Required: &#123;&#125; <br />

Deprecated: omit mode. Use enabled=true without checkpointRef for a<br />DGD-managed automatic checkpoint, or use checkpointRef to restore the<br />named checkpoint. See CheckpointMode. Validation: Enum: [Auto Manual] <br />Optional: &#123;&#125; <br />

startupPolicy defines when normal worker replicas are started relative to<br />automatic checkpoint readiness.<br />Immediate (default): start workers cold immediately; later Pods restore<br />from the checkpoint once it is Ready.<br />WaitForCheckpoint: keep worker replicas at zero until the checkpoint is<br />Ready, then start them from the checkpoint. See CheckpointStartupPolicy. Validation: Enum: [Immediate WaitForCheckpoint] <br />Optional: &#123;&#125; <br />

DeletionPolicy defines whether a DGD-managed automatic checkpoint CR and<br />artifact are deleted or retained when the owning DGD is deleted.<br />Explicit checkpointRef checkpoints are never owned or deleted by the DGD. See CheckpointDeletionPolicy. Validation: Enum: [Delete Retain] <br />Optional: &#123;&#125; <br />

checkpointRef references an existing DynamoCheckpoint CR by metadata.name.<br />When set, this component’s identity is ignored and the referenced<br />checkpoint is used directly. Validation: Optional: &#123;&#125; <br />

Deprecated: omit for DGD-managed checkpoints; no action is needed.<br />Use checkpointRef to restore an existing checkpoint. See DynamoCheckpointIdentity. Validation: Optional: &#123;&#125; <br />

targetContainerName is the workload container to snapshot and restore. Validation: MaxLength: 63 <br />MinLength: 1 <br />Pattern: ^[a-z0-9]([-a-z0-9]*[a-z0-9])?$ <br />Optional: &#123;&#125; <br />

job customizes the DGD-managed checkpoint Job. See ComponentCheckpointJobConfig. Validation: Optional: &#123;&#125; <br />

gmsClientContainers lists checkpoint Job containers that should receive<br />GMS client wiring. Requires gpuMemoryService on the component. Validation: items:MaxLength: 63 <br />items:MinLength: 1 <br />items:Pattern: ^[a-z0-9]([-a-z0-9]*[a-z0-9])?$ <br />Optional: &#123;&#125; <br />

podTemplate customizes the checkpoint Job pod. The operator starts from the<br />selected workload container and merges this template so users can add helper<br />containers such as gms-saver. See PodTemplateSpec. Validation: Schemaless: &#123;&#125; <br />Type: object <br />Optional: &#123;&#125; <br />

checkpointName is the name of the associated DynamoCheckpoint CR. Validation: Optional: &#123;&#125; <br />

checkpointID is the artifact ID used by the snapshot protocol. Validation: Optional: &#123;&#125; <br />

identityHash is the computed hash of the checkpoint identity.<br />Deprecated: automatic checkpoints use checkpointID. This field is retained<br />for older status consumers. Validation: Optional: &#123;&#125; <br />

ready indicates the checkpoint artifact is ready for future pods to restore. Validation: Optional: &#123;&#125; <br />

componentKind is the underlying resource kind (e.g. PodClique,<br />Deployment, LeaderWorkerSet). See ComponentKind. Validation: Enum: [PodClique PodCliqueScalingGroup Deployment LeaderWorkerSet] <br />

componentNames is the list of underlying Kubernetes resource names for<br />this Dynamo component. During normal operation this contains a single<br />name; during rolling updates it contains both old and new resource names. Validation: Optional: &#123;&#125; <br />

runtimeNamespace is the effective Dynamo runtime namespace for this<br />component. Worker components may include a generation suffix; non-workers and<br />Grove-backed workers use the base namespace. During rolling updates, worker<br />status keeps the old active revision namespace until cutover completes. Validation: Optional: &#123;&#125; <br />

replicas is the total number of non-terminated replicas. Validation: Minimum: 0 <br />

updatedReplicas is the number of replicas at the current/desired revision. Validation: Minimum: 0 <br />

readyReplicas is the number of ready replicas. Populated for<br />PodClique, Deployment, and LeaderWorkerSet; not available for<br />PodCliqueScalingGroup. Validation: Minimum: 0 <br />Optional: &#123;&#125; <br />

availableReplicas is the number of available replicas. Populated for<br />Deployment and PodCliqueScalingGroup; not available for<br />PodClique or LeaderWorkerSet. Validation: Minimum: 0 <br />Optional: &#123;&#125; <br />

scheduledReplicas is the number of replicas the backend scheduler has<br />scheduled, expressed strictly in Dynamo component-replica units (not<br />raw backend pod counts). It is a diagnostic aid for distinguishing<br />capacity/scheduling shortfalls from runtime readiness.<br />It is optional and omitted (nil) when the active backend cannot derive<br />it reliably in component-replica units — for example before the backing<br />resource’s status has been observed, or for backends that do not report<br />a scheduling count. A nil value therefore means “not reported”, never<br />“zero scheduled”; consumers must not treat absence as a scheduling<br />failure. Validation: Minimum: 0 <br />Optional: &#123;&#125; <br />

Replicas is the desired number of replicas. Validation: Optional: &#123;&#125; <br />

AvailableReplicas is the number of replicas that are available and ready. Validation: Optional: &#123;&#125; <br />

model is the model identifier (e.g. “meta-llama/Llama-3-70B”).<br />Deprecated: legacy identity only. Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

backendFramework is the runtime framework (vllm, sglang, trtllm).<br />Deprecated: legacy identity only. Validation: Enum: [vllm sglang trtllm] <br />Required: &#123;&#125; <br />

dynamoVersion is the Dynamo platform version.<br />Deprecated: legacy identity only. Validation: Optional: &#123;&#125; <br />

tensorParallelSize is the tensor parallel configuration.<br />Deprecated: checkpoint launch uses the pod template instead. Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

pipelineParallelSize is the pipeline parallel configuration.<br />Deprecated: checkpoint launch uses the pod template instead. Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

dtype is the data type (fp16, bf16, fp8, etc.).<br />Deprecated: legacy identity only. Validation: Optional: &#123;&#125; <br />

maxModelLen is the maximum sequence length.<br />Deprecated: legacy identity only. Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

extraParameters are additional parameters that affect the checkpoint hash.<br />Deprecated: legacy identity only. Validation: Optional: &#123;&#125; <br />

nvidia.com/v1beta1

DynamoComponentDeployment

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

spec defines the desired state for this Dynamo component deployment. See DynamoComponentDeploymentSpec.

name is the stable logical identifier for this component within its<br />DynamoGraphDeployment. It must be unique within the parent’s<br />spec.components list.<br />For standalone DynamoComponentDeployment objects, the defaulting webhook<br />populates name from metadata.name on admission, so users<br />typically do not need to set it explicitly.<br />name is decoupled from the underlying Kubernetes resource name so that<br />the operator can rename child workloads (e.g. suffixing worker DCDs with<br />a hash during rolling updates) without losing the stable identity that<br />downstream consumers (labels, status maps, DGDSA references, planner<br />RBAC, EPP filters) depend on. Validation: MaxLength: 63 <br />MinLength: 1 <br />Pattern: ^[A-Za-z0-9]([-A-Za-z0-9]*[A-Za-z0-9])?$ <br />Required: &#123;&#125; <br />

type indicates the role of this component within a Dynamo graph. Drives<br />port mapping, frontend detection, planner RBAC, and the pod label<br />nvidia.com/dynamo-component-type. Because prefill and decode are<br />first-class values, users can set them directly. See ComponentType. Validation: Enum: [frontend worker prefill decode planner epp] <br />Optional: &#123;&#125; <br />

globalDynamoNamespace places the component in the global Dynamo<br />namespace rather than the per-deployment namespace derived from the<br />DGD name. Validation: Optional: &#123;&#125; <br />

podTemplate is the pod template used to create the component’s pods.<br />The operator injects its defaults (image, command, env, ports, probes,<br />resources, volume mounts) into the container named "main" inside<br />podTemplate.spec.containers, merging user overrides by name. If no<br />container named "main" is present, the operator auto-generates it<br />with standard defaults. All other containers in podTemplate.spec.containers<br />are treated as user-managed sidecars: the operator does not inject<br />defaults into them, so sidecars must specify required fields (e.g. image)<br />themselves. The validation webhook rejects pod templates where a<br />non-"main" container is missing a required field such as image. See PodTemplateSpec. Validation: Optional: &#123;&#125; <br />

replicas is the desired number of Pods for this component. When<br />scalingAdapter is set on this component, this field is managed by<br />the DynamoGraphDeploymentScalingAdapter and should not be modified<br />directly. Validation: Minimum: 0 <br />Optional: &#123;&#125; <br />

minAvailable maps to Grove PodClique minAvailable for single-node and<br />Grove PodCliqueScalingGroup minAvailable for multi-node components.<br />This field determines 1) the minimum number of replicas guaranteed to be<br />gang-scheduled, and 2) when violating minAvailable replicas triggers gang<br />termination.<br />For Grove-backed DynamoGraphDeployment components, minAvailable defaults to<br />1 when omitted and is immutable after creation. Positive replica counts must<br />be greater than or equal to minAvailable. Replicas may be scaled to 0 as a<br />special scale-to-zero state; minAvailable remains configured but is not<br />enforced again until replicas is scaled back to a positive value.<br />For non-Grove deployments, setting this field will result in a validation error. Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

multinode configures multinode components. See MultinodeSpec. Validation: Optional: &#123;&#125; <br />

sharedMemorySize controls the size of the tmpfs mounted at /dev/shm.<br />nil selects the operator default (8Gi), a positive quantity sets a<br />custom size, and "0" disables the shared-memory volume entirely.<br />Simpler replacement for v1alpha1’s SharedMemorySpec struct with its<br />disabled bool + size Quantity pattern. See Quantity. Validation: Optional: &#123;&#125; <br />

modelRef references a model served by this component. When specified,<br />a headless service is created for endpoint discovery. See ModelReference. Validation: Optional: &#123;&#125; <br />

scalingAdapter opts this component into using the<br />DynamoGraphDeploymentScalingAdapter. When set (even as an empty object,<br />scalingAdapter: \{\}), a DGDSA is created and owns the replicas field<br />so that external autoscalers (HPA/KEDA/Planner) can drive scaling via<br />the Scale subresource. Omit the field to opt out. See ScalingAdapter. Validation: Optional: &#123;&#125; <br />

eppConfig holds EPP-specific configuration for Endpoint Picker Plugin<br />components. Only meaningful when type is epp. See EPPConfig. Validation: Optional: &#123;&#125; <br />

frontendSidecar optionally designates a container in<br />podTemplate.spec.containers as the frontend sidecar. The value must<br />match the name of a container in that list; the operator merges its<br />frontend-sidecar defaults (auto-generated Dynamo env vars, ports,<br />health probes) into that container the same way it merges into "main".<br />The full container definition (image, args, envFrom, env) lives in<br />podTemplate — this eliminates the redundant image, args,<br />envFromSecret, and envs fields from v1alpha1’s FrontendSidecarSpec.<br />The validation webhook rejects values that do not match any container<br />name in podTemplate.spec.containers. Validation: Optional: &#123;&#125; <br />

compilationCache configures a PVC-backed compilation cache. The operator<br />handles backend-specific mount paths and environment variables, so<br />users do not need to hand-wire them into podTemplate. Extracted from<br />v1alpha1’s volumeMount.useAsCompilationCache flag. See CompilationCacheConfig. Validation: Optional: &#123;&#125; <br />

topologyConstraint applies to this component.<br />topologyConstraint.packDomain is required. When both this and<br />spec.topologyConstraint.packDomain are set, this field’s packDomain<br />must be narrower than or equal to the spec-level value. See TopologyConstraint. Validation: Optional: &#123;&#125; <br />

experimental groups opt-in preview features whose API shape and<br />behavior may change in breaking ways between v1beta1 releases,<br />including disappearing without a name-preserving graduation path.<br />In v1beta1 this block holds gpuMemoryService and failover (which<br />remain tightly coupled — failover requires GMS — and are expected to<br />evolve together as the DRA-based GPU sharing story matures), and<br />checkpoint (whose API shape is still settling). Fields here are<br />explicitly NOT covered by the normal v1beta1 deprecation policy; do not<br />depend on them for production workloads. See ExperimentalSpec. Validation: Optional: &#123;&#125; <br />

backendFramework specifies the backend framework. Validation: Enum: [sglang vllm trtllm] <br />

name is the stable logical identifier for this component within its<br />DynamoGraphDeployment. It must be unique within the parent’s<br />spec.components list.<br />For standalone DynamoComponentDeployment objects, the defaulting webhook<br />populates name from metadata.name on admission, so users<br />typically do not need to set it explicitly.<br />name is decoupled from the underlying Kubernetes resource name so that<br />the operator can rename child workloads (e.g. suffixing worker DCDs with<br />a hash during rolling updates) without losing the stable identity that<br />downstream consumers (labels, status maps, DGDSA references, planner<br />RBAC, EPP filters) depend on. Validation: MaxLength: 63 <br />MinLength: 1 <br />Pattern: ^[A-Za-z0-9]([-A-Za-z0-9]*[A-Za-z0-9])?$ <br />Required: &#123;&#125; <br />

type indicates the role of this component within a Dynamo graph. Drives<br />port mapping, frontend detection, planner RBAC, and the pod label<br />nvidia.com/dynamo-component-type. Because prefill and decode are<br />first-class values, users can set them directly. See ComponentType. Validation: Enum: [frontend worker prefill decode planner epp] <br />Optional: &#123;&#125; <br />

globalDynamoNamespace places the component in the global Dynamo<br />namespace rather than the per-deployment namespace derived from the<br />DGD name. Validation: Optional: &#123;&#125; <br />

podTemplate is the pod template used to create the component’s pods.<br />The operator injects its defaults (image, command, env, ports, probes,<br />resources, volume mounts) into the container named "main" inside<br />podTemplate.spec.containers, merging user overrides by name. If no<br />container named "main" is present, the operator auto-generates it<br />with standard defaults. All other containers in podTemplate.spec.containers<br />are treated as user-managed sidecars: the operator does not inject<br />defaults into them, so sidecars must specify required fields (e.g. image)<br />themselves. The validation webhook rejects pod templates where a<br />non-"main" container is missing a required field such as image. See PodTemplateSpec. Validation: Optional: &#123;&#125; <br />

replicas is the desired number of Pods for this component. When<br />scalingAdapter is set on this component, this field is managed by<br />the DynamoGraphDeploymentScalingAdapter and should not be modified<br />directly. Validation: Minimum: 0 <br />Optional: &#123;&#125; <br />

minAvailable maps to Grove PodClique minAvailable for single-node and<br />Grove PodCliqueScalingGroup minAvailable for multi-node components.<br />This field determines 1) the minimum number of replicas guaranteed to be<br />gang-scheduled, and 2) when violating minAvailable replicas triggers gang<br />termination.<br />For Grove-backed DynamoGraphDeployment components, minAvailable defaults to<br />1 when omitted and is immutable after creation. Positive replica counts must<br />be greater than or equal to minAvailable. Replicas may be scaled to 0 as a<br />special scale-to-zero state; minAvailable remains configured but is not<br />enforced again until replicas is scaled back to a positive value.<br />For non-Grove deployments, setting this field will result in a validation error. Validation: Minimum: 1 <br />Optional: &#123;&#125; <br />

multinode configures multinode components. See MultinodeSpec. Validation: Optional: &#123;&#125; <br />

sharedMemorySize controls the size of the tmpfs mounted at /dev/shm.<br />nil selects the operator default (8Gi), a positive quantity sets a<br />custom size, and "0" disables the shared-memory volume entirely.<br />Simpler replacement for v1alpha1’s SharedMemorySpec struct with its<br />disabled bool + size Quantity pattern. See Quantity. Validation: Optional: &#123;&#125; <br />

modelRef references a model served by this component. When specified,<br />a headless service is created for endpoint discovery. See ModelReference. Validation: Optional: &#123;&#125; <br />

scalingAdapter opts this component into using the<br />DynamoGraphDeploymentScalingAdapter. When set (even as an empty object,<br />scalingAdapter: \{\}), a DGDSA is created and owns the replicas field<br />so that external autoscalers (HPA/KEDA/Planner) can drive scaling via<br />the Scale subresource. Omit the field to opt out. See ScalingAdapter. Validation: Optional: &#123;&#125; <br />

eppConfig holds EPP-specific configuration for Endpoint Picker Plugin<br />components. Only meaningful when type is epp. See EPPConfig. Validation: Optional: &#123;&#125; <br />

frontendSidecar optionally designates a container in<br />podTemplate.spec.containers as the frontend sidecar. The value must<br />match the name of a container in that list; the operator merges its<br />frontend-sidecar defaults (auto-generated Dynamo env vars, ports,<br />health probes) into that container the same way it merges into "main".<br />The full container definition (image, args, envFrom, env) lives in<br />podTemplate — this eliminates the redundant image, args,<br />envFromSecret, and envs fields from v1alpha1’s FrontendSidecarSpec.<br />The validation webhook rejects values that do not match any container<br />name in podTemplate.spec.containers. Validation: Optional: &#123;&#125; <br />

compilationCache configures a PVC-backed compilation cache. The operator<br />handles backend-specific mount paths and environment variables, so<br />users do not need to hand-wire them into podTemplate. Extracted from<br />v1alpha1’s volumeMount.useAsCompilationCache flag. See CompilationCacheConfig. Validation: Optional: &#123;&#125; <br />

topologyConstraint applies to this component.<br />topologyConstraint.packDomain is required. When both this and<br />spec.topologyConstraint.packDomain are set, this field’s packDomain<br />must be narrower than or equal to the spec-level value. See TopologyConstraint. Validation: Optional: &#123;&#125; <br />

experimental groups opt-in preview features whose API shape and<br />behavior may change in breaking ways between v1beta1 releases,<br />including disappearing without a name-preserving graduation path.<br />In v1beta1 this block holds gpuMemoryService and failover (which<br />remain tightly coupled — failover requires GMS — and are expected to<br />evolve together as the DRA-based GPU sharing story matures), and<br />checkpoint (whose API shape is still settling). Fields here are<br />explicitly NOT covered by the normal v1beta1 deprecation policy; do not<br />depend on them for production workloads. See ExperimentalSpec. Validation: Optional: &#123;&#125; <br />

nvidia.com/v1beta1

DynamoGraphDeployment

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

spec defines the desired state for this graph deployment. See DynamoGraphDeploymentSpec.

status reflects the current observed state of this graph deployment. See DynamoGraphDeploymentStatus.

name is the metadata.name of the target DynamoGraphDeployment. Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

componentName is the componentName of the entry within the target<br />DGD’s spec.components list to scale. Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

kvTransferPolicy configures topology-aware routing for KV-cache<br />transfers between prefill and decode workers. See KvTransferPolicy. Validation: Optional: &#123;&#125; <br />

nvidia.com/v1beta1

DynamoGraphDeploymentRequest

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

Spec defines the desired state for this deployment request. See DynamoGraphDeploymentRequestSpec.

Status reflects the current observed state of this deployment request. See DynamoGraphDeploymentRequestStatus.

Model specifies the model to deploy (e.g., “Qwen/Qwen3-0.6B”, “meta-llama/Llama-3-70b”).<br />Can be a HuggingFace ID or a private model name. Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

Backend specifies the inference backend to use for profiling and deployment. See BackendType. Validation: Enum: [auto sglang trtllm vllm] <br />Optional: &#123;&#125; <br />

Image is the container image reference for the profiling job (planner image).<br />Example: “nvcr.io/nvidia/ai-dynamo/dynamo-planner:1.2.1”.<br />For Dynamo < 1.1.0, use dynamo-frontend. Validation: Optional: &#123;&#125; <br />

ModelCache provides optional PVC configuration for pre-downloaded model weights.<br />When provided, weights are loaded from the PVC instead of downloading from HuggingFace. See ModelCacheSpec. Validation: Optional: &#123;&#125; <br />

Hardware describes the hardware resources available for profiling and deployment.<br />Typically auto-filled by the operator from cluster discovery. See HardwareSpec. Validation: Optional: &#123;&#125; <br />

Workload defines the expected workload characteristics for SLA-based profiling. See WorkloadSpec. Validation: Optional: &#123;&#125; <br />

SLA defines service-level agreement targets that drive profiling optimization. See SLASpec. Validation: Optional: &#123;&#125; <br />

Overrides allows customizing the profiling job and the generated DynamoGraphDeployment. See OverridesSpec. Validation: Optional: &#123;&#125; <br />

Features controls optional Dynamo platform features in the generated deployment. See FeaturesSpec. Validation: Optional: &#123;&#125; <br />

SearchStrategy controls the profiling search depth.<br />“rapid” performs a fast sweep; “thorough” explores more configurations. See SearchStrategy. Validation: Enum: [rapid thorough] <br />Optional: &#123;&#125; <br />

AutoApply indicates whether to automatically create a DynamoGraphDeployment<br />after profiling completes. If false, the generated spec is stored in status<br />for manual review and application. Validation: Optional: &#123;&#125; <br />

Phase is the high-level lifecycle phase of the deployment request. See DGDRPhase. Validation: Enum: [Pending Profiling Ready Deploying Deployed Failed] <br />Optional: &#123;&#125; <br />

ProfilingPhase indicates the current sub-phase of the profiling pipeline.<br />Only meaningful when Phase is “Profiling”. Cleared when profiling completes or fails. See ProfilingPhase. Validation: Enum: [Initializing SweepingPrefill SweepingDecode SelectingConfig BuildingCurves GeneratingDGD Done] <br />Optional: &#123;&#125; <br />

DGDName is the name of the generated or created DynamoGraphDeployment. Validation: Optional: &#123;&#125; <br />

ProfilingJobName is the name of the Kubernetes Job running the profiler. Validation: Optional: &#123;&#125; <br />

Conditions contains the latest observed conditions of the deployment request.<br />Standard condition types include: Succeeded, Validation, Profiling, SpecGenerated, DeploymentReady. See Condition. Validation: Optional: &#123;&#125; <br />

ProfilingResults contains the selected deployment configuration produced by profiling.<br />Deprecated compatibility fields may remain on objects created by older releases. See ProfilingResultsStatus. Validation: Optional: &#123;&#125; <br />

DeploymentInfo tracks the state of the deployed DynamoGraphDeployment.<br />Populated when a DGD has been created (either via autoApply or manually). See DeploymentInfoStatus. Validation: Optional: &#123;&#125; <br />

ObservedGeneration is the most recent generation observed by the controller. Validation: Optional: &#123;&#125; <br />

nvidia.com/v1beta1

DynamoGraphDeploymentScalingAdapter

Refer to Kubernetes API documentation for fields of metadata. See ObjectMeta.

See DynamoGraphDeploymentScalingAdapterSpec.

See DynamoGraphDeploymentScalingAdapterStatus.

replicas is the desired number of replicas for the target component.<br />This field is modified by external autoscalers (HPA/KEDA/Planner) or<br />manually by users. Validation: Minimum: 0 <br />Required: &#123;&#125; <br />

dgdRef references the DynamoGraphDeployment and the specific component to scale. See DynamoGraphDeploymentComponentRef. Validation: Required: &#123;&#125; <br />

replicas is the current number of replicas for the target component.<br />This is synced from the DGD’s component replicas and is required for<br />the scale subresource.

selector is a label selector string for the pods managed by this<br />adapter. Required for HPA compatibility via the scale subresource. Validation: Optional: &#123;&#125; <br />

lastScaleTime is the last time the adapter scaled the target component. See Time. Validation: Optional: &#123;&#125; <br />

annotations to propagate to all child resources (PCS, DCD, Deployments,<br />and pod templates). Component-level (podTemplate) values take precedence<br />on conflict. Validation: Optional: &#123;&#125; <br />

labels to propagate to all child resources. Same precedence rules as annotations. Validation: Optional: &#123;&#125; <br />

priorityClassName is the name of the PriorityClass to use for Grove PodCliqueSets.<br />Requires the Grove pathway. Validation: Optional: &#123;&#125; <br />

components are the components deployed as part of this graph. Each entry<br />carries its own stable logical name, and names must be unique within<br />the list. Component types are generally repeatable, except type: epp<br />which may appear at most once. See DynamoComponentDeploymentSharedSpec. Validation: MaxItems: 25 <br />Optional: &#123;&#125; <br />

env is prepended to every component’s environment. Component-specific<br />env entries with the same name take precedence and may reference values<br />from this list. See EnvVar. Validation: Optional: &#123;&#125; <br />

backendFramework specifies the backend framework (e.g. “sglang”, “vllm”, “trtllm”). Validation: Enum: [sglang vllm trtllm] <br />

restart specifies the restart policy for the graph deployment. See Restart. Validation: Optional: &#123;&#125; <br />

topologyConstraint is the deployment-level topology constraint. When<br />set, spec.topologyConstraint.clusterTopologyName names the ClusterTopology<br />CR to use. spec.topologyConstraint.packDomain is optional at this<br />level and can be omitted when only components carry constraints.<br />Components without their own topologyConstraint inherit from this value. See SpecTopologyConstraint. Validation: Optional: &#123;&#125; <br />

experimental groups graph-level preview features whose API shape and<br />behavior may change in breaking ways between v1beta1 releases. See DynamoGraphDeploymentExperimentalSpec. Validation: Optional: &#123;&#125; <br />

observedGeneration is the most recent generation observed by the controller. Validation: Optional: &#123;&#125; <br />

state is a high-level textual status of the graph deployment lifecycle. See DGDState. Validation: Enum: [initializing pending successful failed] <br />

conditions contains the latest observed conditions of the graph deployment.<br />Merged by type on patch updates. See Condition. Validation: Optional: &#123;&#125; <br />

components contains per-component replica status information, keyed by component name. See ComponentReplicaStatus. Validation: Optional: &#123;&#125; <br />

restart contains the status of a graph-level restart. See RestartStatus. Validation: Optional: &#123;&#125; <br />

checkpoints contains per-component checkpoint status, keyed by component name. See ComponentCheckpointStatus. Validation: Optional: &#123;&#125; <br />

rollingUpdate tracks the progress of operator-managed rolling updates.<br />Currently only supported for single-node, non-Grove deployments (DCD/Deployment). See RollingUpdateStatus. Validation: Optional: &#123;&#125; <br />

placement groups DGD-level scheduler placement signals (score, reporting<br />state, and any future placement fields). See PlacementStatus. Validation: Optional: &#123;&#125; <br />

configMapRef references a user-provided ConfigMap containing EPP<br />configuration. Mutually exclusive with config. See ConfigMapKeySelector. Validation: Optional: &#123;&#125; <br />

config allows specifying EPP EndpointPickerConfig directly as a<br />structured object. The operator marshals this to YAML and creates a<br />ConfigMap automatically. Mutually exclusive with configMapRef. One of<br />configMapRef or config must be specified. Validation: Type: object <br />Optional: &#123;&#125; <br />

gpuMemoryService configures the GPU Memory Service (GMS). When set, GPU<br />access for GMS clients is managed via DRA. See GPUMemoryServiceSpec. Validation: Optional: &#123;&#125; <br />

failover configures active-passive GPU failover for this component.<br />Requires gpuMemoryService to also be set, and failover.mode must<br />match gpuMemoryService.mode (enforced by the validation webhook). See FailoverSpec. Validation: Optional: &#123;&#125; <br />

checkpoint configures container-image snapshotting and restore for<br />this component. Set checkpoint.enabled: true to opt in. Without<br />checkpointRef, the DGD controller creates a DGD-scoped DynamoCheckpoint<br />CR and later restores pods in the same DGD generation from that<br />checkpoint. With checkpointRef, the DGD restores from that existing<br />checkpoint instead. The user-facing shape of this field is still settling,<br />which is why it lives under experimental in v1beta1 instead of at the<br />top level. See ComponentCheckpointConfig. Validation: Optional: &#123;&#125; <br />

mode selects the failover deployment topology. Must match<br />spec.experimental.gpuMemoryService.mode (or<br />spec.components[*].experimental.gpuMemoryService.mode inside a<br />DynamoGraphDeployment). See GPUMemoryServiceMode. Validation: Enum: [IntraPod InterPod] <br />Optional: &#123;&#125; <br />

numShadows is the number of shadow (standby) engine containers per<br />rank. Reserved for future use; the operator currently creates exactly<br />one shadow. Validation: Maximum: 1 <br />Minimum: 1 <br />Optional: &#123;&#125; <br />

Planner contains the raw Planner configuration passed to the Planner service.<br />Its schema is defined by dynamo.planner.config.planner_config.PlannerConfig.<br />See Planner Guide.<br />DGDR passes this object through without field-level validation; the Planner<br />service validates it at startup.<br />The presence of this field (non-null) enables the planner in the generated DGD. See RawExtension. Validation: Type: object <br />Optional: &#123;&#125; <br />

Mocker configures the simulated (mocker) backend for testing without GPUs. See MockerSpec. Validation: Optional: &#123;&#125; <br />

name identifies this client pod. Validation: MaxLength: 63 <br />MinLength: 1 <br />Pattern: ^[a-z0-9]([-a-z0-9]*[a-z0-9])?$ <br />

podTemplate configures the pod to run as a GMS client. See PodTemplateSpec. Validation: Schemaless: &#123;&#125; <br />Type: object <br />

mode selects the GMS deployment topology. See GPUMemoryServiceMode. Validation: Enum: [IntraPod InterPod] <br />Optional: &#123;&#125; <br />

deviceClassName is the DRA DeviceClass to request GPUs from. Validation: Optional: &#123;&#125; <br />

extraClientContainers lists additional user-declared containers that should<br />be wired as GMS clients in service pods. Checkpoint Job clients are declared<br />under checkpoint.job.gmsClientContainers. In each rendered pod, only<br />matching container names are wired; absent names are ignored. Validation: items:MaxLength: 63 <br />items:MinLength: 1 <br />items:Pattern: ^[a-z0-9]([-a-z0-9]*[a-z0-9])?$ <br />Optional: &#123;&#125; <br />

extraClientPods declares additional GMS client pods for inter-pod GMS. This field is<br />reserved for future use and is rejected until inter-pod client orchestration is wired. See GMSClientPodSpec. Validation: Optional: &#123;&#125; <br />

GPUSKU selects the GPU type to target.<br />When omitted, auto-detected by selecting the GPU with the highest<br />node count, then highest VRAM. In mixed-GPU clusters, set this to<br />choose which GPU type to use. Discovery and totalGpus are then<br />restricted to nodes matching this SKU. See GPUSKUType. Validation: Enum: [gb200_sxm gb10 b200_sxm h200_sxm h100_sxm h100_pcie a100_sxm a100_pcie a30 l40s l40 l4 v100_sxm v100_pcie t4 mi200 mi300] <br />Optional: &#123;&#125; <br />

VRAMMB is the VRAM per GPU in MiB.<br />When omitted, auto-detected from cluster GPU nodes. Validation: Optional: &#123;&#125; <br />

TotalGPUs is the GPU budget for profiling and deployment.<br />The profiler uses this to determine parallelism and replica count.<br />When omitted, computed by counting GPUs on discovered nodes<br />(filtered by gpuSku when set), temporarily capped at 32 to<br />limit profiler search space. This cap may be removed in a future<br />release. Set this field explicitly to override. Validation: Optional: &#123;&#125; <br />

NumGPUsPerNode is the number of GPUs per node.<br />When omitted, auto-detected from cluster GPU nodes. Validation: Optional: &#123;&#125; <br />

Interconnect describes the primary GPU-to-GPU interconnect within a node.<br />Semantics / usage:<br /> - This is capability metadata used for profiling, planning, and deployment decisions.<br /> - It does NOT configure or enable any GPU interconnect; it only describes what is available/assumed.<br /> - When omitted, the operator may attempt best-effort discovery (currently distinguishes “nvlink”<br /> vs “pcie” based on DCGM NVLink link count). If discovery is unavailable, it may remain empty.<br />Impact of wrong / missing values:<br /> - If set more optimistically than reality (e.g., “nvlink” when only PCIe is present), performance<br /> models may overestimate intra-node bandwidth and choose overly aggressive parallelism or layouts,<br /> resulting in degraded performance compared to expectations.<br /> - If set more pessimistically than reality (e.g., “pcie” when NVLink is present), the system may<br /> choose conservative plans and leave performance on the table.<br /> - If unset and undiscovered, consumers should treat the interconnect as unknown and fall back to<br /> conservative assumptions.<br />Example values: “pcie”, “nvlink”. Other values may be accepted but may not be auto-detected. Validation: Optional: &#123;&#125; <br />

RDMA indicates whether the cluster has RDMA-capable networking available for Dynamo data movement.<br />Semantics / usage:<br /> - This is capability metadata used for profiling, planning, and deployment decisions.<br /> - It does NOT install, enable, or configure RDMA (e.g., drivers, SR-IOV, NVIDIA network operator,<br /> GPUDirect settings). It only expresses availability/intent.<br /> - When omitted, the operator may attempt best-effort discovery (e.g., via node labels indicating<br /> RDMA/SR-IOV capability and/or presence of NVIDIA network-operator RDMA components). If discovery<br /> is unavailable, it may remain unset.<br />Impact of wrong / missing values:<br /> - False positive (set true when RDMA is not actually usable end-to-end) may cause plans or<br /> deployments to assume RDMA is available; depending on the runtime transport selection and<br /> fallback behavior, this can lead to connection/setup failures or performance regressions.<br /> - False negative (set false when RDMA is available) will typically avoid RDMA-optimized paths and<br /> fall back to non-RDMA transports, usually remaining functional but potentially slower.<br /> - If unset and undiscovered, consumers should treat RDMA availability as unknown and use<br /> conservative defaults / fallback transports. Validation: Optional: &#123;&#125; <br />

clusterTopologyName references a Grove ClusterTopology CR. The operator<br />reads the CR’s topology levels and projects them through Dynamo-owned pod<br />labels for worker topology metadata. Validation: MinLength: 1 <br />Optional: &#123;&#125; <br />

labelKey is a Kubernetes node label key (e.g.<br />“topology.kubernetes.io/zone”) whose value identifies the topology<br />domain for each worker. The operator copies the node label onto worker<br />pods so the runtime can publish it as worker metadata. The label<br />should correspond to the topology level named in domain. Validation: MaxLength: 317 <br />MinLength: 1 <br />Pattern: ^(([a-z0-9]([-a-z0-9]\{0,61\}[a-z0-9])?)(\.[a-z0-9]([-a-z0-9]\{0,61\}[a-z0-9])?)*/)?([A-Za-z0-9]([-A-Za-z0-9_.]\{0,61\}[A-Za-z0-9])?)$ <br />Optional: &#123;&#125; <br />

domain is the logical name for the topology level to enforce<br />(e.g. “zone”, “rack”). The router uses this to match workers that<br />share the same value for the label identified by labelKey. See TopologyDomain. Validation: Pattern: ^[a-z0-9]([a-z0-9-]*[a-z0-9])?$ <br />

enforcement controls how the selected prefill worker’s topology is<br />applied to decode routing. “required” only allows decode workers in the<br />same topology domain as the selected prefill worker. “preferred” keeps<br />all decode workers eligible, but biases selection toward workers in the<br />same topology domain. Defaults to “required”. See KvTransferEnforcement. Validation: Enum: [required preferred] <br />Optional: &#123;&#125; <br />

preferredWeight is required and used only when enforcement is<br />“preferred”. Higher values create a stronger same-domain routing<br />preference, but do not guarantee same-domain selection. The value is not<br />a probability; worker selection still depends on load and other routing<br />inputs. A value of 0 disables the topology preference; 1 is the strongest<br />supported preference. Validation: Maximum: 1 <br />Minimum: 0 <br />Optional: &#123;&#125; <br />

Enabled indicates whether to deploy mocker workers instead of real inference workers.<br />Useful for large-scale testing without GPUs. Validation: Optional: &#123;&#125; <br />

PVCName is the name of the PersistentVolumeClaim containing model weights.<br />The PVC must exist in the same namespace as the DGDR. Validation: Optional: &#123;&#125; <br />

PVCModelPath is the path to the model checkpoint directory within the PVC<br />(e.g. “deepseek-r1” or “models/Llama-3.1-405B-FP8”). Validation: Optional: &#123;&#125; <br />

PVCMountPath is the mount path for the PVC inside the container. Validation: Optional: &#123;&#125; <br />

name is the base model identifier (e.g. “llama-3-70b-instruct-v1”). Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

revision is the model revision/version. Validation: Optional: &#123;&#125; <br />

nodeCount is the number of nodes to deploy for the multinode component.<br />Total GPUs used is nodeCount * container GPU request. Validation: Minimum: 2 <br />Optional: &#123;&#125; <br />

ProfilingJob allows overriding the profiling Job specification.<br />Fields set here are merged into the controller-generated Job spec. See JobSpec. Validation: Optional: &#123;&#125; <br />

DGD provides a partial, versioned DynamoGraphDeployment override for the<br />profiler-generated deployment. Set apiVersion to nvidia.com/v1alpha1 or<br />nvidia.com/v1beta1 and kind to DynamoGraphDeployment.<br />The profiler merges the override using the schema for its declared version.<br />If the generated DGD uses another supported version, the complete DGD is<br />converted before the merge and converted back afterward. The final DGD<br />selected or created by a DGDR is nvidia.com/v1beta1.<br />The override can update DGD fields, but topology entries are limited to<br />services or components already present in the generated DGD. Metadata labels<br />and annotations are merged, metadata.name selects the final DGD name, and<br />other identity or runtime metadata is ignored.<br />V1alpha1 worker argument lists retain legacy append behavior. V1beta1 follows<br />structural schema merge behavior, including map-list merging and atomic-list<br />replacement.<br />The raw embedded resource preserves either supported schema. The API server<br />validates that it has apiVersion and kind; override processing validates the<br />DGD kind, supported version, and field schema. See RawExtension. Validation: EmbeddedResource: &#123;&#125; <br />Optional: &#123;&#125; <br />

Config is the full deployment configuration for this Pareto point. See RawExtension. Validation: Type: object <br />

score is the DGD-level scheduler placement score aggregated from<br />relevant scheduler placement units. Normalized to [0.0, 1.0] where higher<br />is better and 1.0 represents the best possible placement. Aggregation<br />uses the minimum across placement units so the value is a worst-placement<br />signal for the graph. Scores are only comparable across DGDs that share<br />the same scheduler scoring contract and version. Validation: Maximum: 1 <br />Minimum: 0 <br />Optional: &#123;&#125; <br />

state indicates placement score reporting state. See PlacementScoreState<br />for the semantics of each value. See PlacementScoreState. Validation: Enum: [Reported Partial Unsupported Unknown] <br />Optional: &#123;&#125; <br />

Pareto is retained for compatibility with existing status objects.<br />Deprecated: The controller no longer populates this field. See ParetoConfig. Validation: Optional: &#123;&#125; <br />

SelectedConfig is the recommended configuration chosen by the profiler<br />based on the SLA targets. This is the configuration used for deployment<br />when autoApply is true. See RawExtension. Validation: Type: object <br />Optional: &#123;&#125; <br />

id is an arbitrary string that triggers a restart when changed. Any<br />modification to this value initiates a restart of the graph deployment<br />according to the configured strategy. Validation: MinLength: 1 <br />Required: &#123;&#125; <br />

strategy specifies the restart strategy for the graph deployment. See RestartStrategy. Validation: Optional: &#123;&#125; <br />

observedID is the restart ID currently being processed. Matches Restart.id in the spec.

phase is the phase of the restart. See RestartPhase.

inProgress contains the names of the components currently being restarted. Validation: Optional: &#123;&#125; <br />

type specifies the restart strategy type. See RestartStrategyType. Validation: Enum: [Sequential Parallel] <br />Optional: &#123;&#125; <br />

order is the complete ordered set of component names for sequential<br />restarts. Omit or leave empty to use the controller’s default order.<br />This field must not be set for parallel restarts. Validation: Optional: &#123;&#125; <br />

phase indicates the current phase of the rolling update. See RollingUpdatePhase. Validation: Enum: [Pending InProgress Completed Failed ] <br />Optional: &#123;&#125; <br />

startTime is when the rolling update began. See Time. Validation: Optional: &#123;&#125; <br />

endTime is when the rolling update completed (successfully or failed). See Time. Validation: Optional: &#123;&#125; <br />

updatedComponents is the list of components that have completed the<br />rolling update. Validation: Optional: &#123;&#125; <br />

TTFT is the Time To First Token target in milliseconds. Validation: Optional: &#123;&#125; <br />

ITL is the Inter-Token Latency target in milliseconds. Validation: Optional: &#123;&#125; <br />

E2ELatency is the target end-to-end request latency in milliseconds.<br />Alternative to specifying TTFT + ITL. Validation: Optional: &#123;&#125; <br />

OptimizationType is the optimization target for SLA profiling.<br />Valid values: latency, throughput. See OptimizationType. Validation: Enum: [latency throughput] <br />Optional: &#123;&#125; <br />

clusterTopologyName is the name of the ClusterTopology resource that<br />defines the topology hierarchy for this deployment. Validation: MinLength: 1 <br />

packDomain is the default topology domain to pack pods within.<br />Optional; omit when only components carry constraints. See TopologyDomain. Validation: Pattern: ^[a-z0-9]([a-z0-9-]*[a-z0-9])?$ <br />Optional: &#123;&#125; <br />

packDomain is the topology domain to pack pods within. Must match a<br />domain defined in the referenced ClusterTopology CR. See TopologyDomain. Validation: Pattern: ^[a-z0-9]([a-z0-9-]*[a-z0-9])?$ <br />

ISL is the Input Sequence Length (number of tokens). Validation: Optional: &#123;&#125; <br />

OSL is the Output Sequence Length (number of tokens). Validation: Optional: &#123;&#125; <br />

Concurrency is the target concurrency level.<br />Required (or RequestRate) when the planner is disabled. Validation: Optional: &#123;&#125; <br />

RequestRate is the target request rate (req/s).<br />Required (or Concurrency) when the planner is disabled. Validation: Optional: &#123;&#125; <br />

Enabled indicates if checkpoint functionality is enabled

Seccomp controls the localhost seccomp profile applied to checkpoint and<br />restore pods. A nil value means “use the default profile”; set<br />Seccomp.Disabled=true to disable seccomp injection entirely. See CheckpointSeccompConfiguration.

Storage optionally configures the namespace-local checkpoint PVC that<br />workload pods mount. When omitted, the operator preserves the legacy<br />behavior of discovering storage from a snapshot-agent DaemonSet in the<br />workload namespace. See CheckpointStorageConfiguration.

CleanupImage is the image used by best-effort artifact cleanup Jobs for<br />automatically-created checkpoints. It must provide a POSIX shell and rm.

URI is the legacy OCI URI (oci://registry/repository).

CredentialsSecretRef is the legacy docker config secret name.

PVCName is the PVC name in each workload namespace.

BasePath is the mount path inside checkpoint and restore workload pods.

Create tells the operator to create the PVC in workload namespaces when<br />it is missing. When false, the PVC must already exist.

Size is the storage request used when Create is true.

StorageClassName is the optional StorageClass name used when Create is true.

AccessMode is the PVC access mode used when Create is true.

URI is the legacy S3 URI (s3://[endpoint/]bucket/prefix).

CredentialsSecretRef is the legacy credentials secret name.

Disabled, when true, suppresses seccomp profile injection entirely.<br />Use this for clusters where custom localhost profiles are not allowed<br />(e.g. OpenShift’s restricted-v2 SCC) or for CRIU builds that handle<br />io_uring natively.

Profile is the localhost seccomp profile path. Empty falls back to<br />DefaultSeccompProfile. Ignored when Disabled is true.

Type is the storage backend type. Only pvc is implemented today.

PVC configuration for pvc-based settings. See CheckpointPVCConfig.

Deprecated: S3 is retained for compatibility and ignored. See CheckpointS3Config.

Deprecated: OCI is retained for compatibility and ignored. See CheckpointOCIConfig.

Enabled overrides auto-detection of the resource.k8s.io/v1 API.<br />nil = auto-detect. Setting true requires detection to also succeed (the<br />operator will exit at startup otherwise).

Backend is the discovery backend: “kubernetes” or “etcd” See DiscoveryBackend.

DiscoveryEnabled indicates whether GPU discovery is enabled

Enabled overrides auto-detection. nil = auto-detect.

TerminationDelay configures the termination delay for Grove PodCliqueSets See Duration.

NATSAddress is the address of the NATS server

ETCDAddress is the address of the etcd server

ModelExpressURL is the URL of the Model Express server to inject into all pods

PrometheusEndpoint is the URL of the Prometheus endpoint to use for metrics

VirtualServiceGateway is the name of the Istio virtual service gateway

ControllerClassName is the ingress controller class name

ControllerTLSSecretName is the TLS secret for the ingress controller

HostSuffix is the suffix for ingress hostnames

TLSMode is the Istio TLS mode for DestinationRules.<br />Supported values: “DISABLE”, “SIMPLE”, “ISTIO_MUTUAL”, “MUTUAL”.<br />Defaults to “SIMPLE”.

InsecureSkipVerify skips TLS certificate verification in DestinationRules.<br />Defaults to true (matching upstream GAIE behavior with self-signed certs).

ClientCertificate is the path (in the istio-proxy sidecar’s filesystem)<br />to the file holding the client-side TLS certificate used for mTLS.<br />REQUIRED when TLSMode is “MUTUAL”; ignored for other modes.

PrivateKey is the path (in the istio-proxy sidecar’s filesystem) to the<br />file holding the client-side TLS private key used for mTLS.<br />REQUIRED when TLSMode is “MUTUAL”; ignored for other modes.

CaCertificates is the optional path (in the istio-proxy sidecar’s<br />filesystem) to the file holding CA certificates used to verify the<br />server certificate. Used only when TLSMode is “MUTUAL”; for other modes<br />the field is ignored.

Enabled overrides auto-detection. nil = auto-detect.

Enabled overrides auto-detection. nil = auto-detect.

Enabled enables leader election for controller manager

ID is the leader election resource identity

Namespace is the namespace for the leader election resource

Level is the log level (e.g., “info”, “debug”)

Format is the log format (e.g., “json”, “text”)

SSHSecretName is the name of the secret containing the SSH key for MPI

SSHSecretNamespace is the namespace where the MPI SSH secret is located