DynamoGraphDeploymentRequest (DGDR) Reference

Field reference and lifecycle for the DynamoGraphDeploymentRequest custom resource — Dynamo’s deploy-by-intent generator.

View as Markdown

A DynamoGraphDeploymentRequest (DGDR) is Dynamo’s deploy-by-intent generator for DynamoGraphDeployment (DGD) resources. You describe what you want to run and your performance targets; the profiler determines a configuration and produces the DGD that serves traffic.

For the task-oriented walkthrough — authoring a request step by step — see Auto Deploy with DGDR. For ready-to-copy manifests, see DGDR Examples.

This page documents the nvidia.com/v1beta1 API — the served, storage version.

Spec reference

Only model is required; every other field has a default or is auto-detected.

model
stringRequired

Model to deploy — a HuggingFace ID (for example Qwen/Qwen3-0.6B, meta-llama/Llama-3-70b) or a private model name. Must be non-empty.

image
string

Container image for the profiling job (the planner image), for example nvcr.io/nvidia/ai-dynamo/dynamo-planner:1.2.1. For Dynamo < 1.1.0, use dynamo-frontend.

backend
stringDefaults to auto

Inference backend used for profiling and deployment.

Allowed values: auto vllm sglang trtllm
searchStrategy
stringDefaults to rapid

Profiling depth. rapid performs a fast, AIC-backed sweep (~30s); thorough explores more configurations on real GPUs (2–4h).

Allowed values: rapid thorough
autoApply
booleanDefaults to true

Automatically create a DGD after profiling completes. When false, the generated spec is stored in status for manual review instead of being applied.

workload
WorkloadSpec

Expected workload characteristics for SLA-based profiling.

isl
integerDefaults to 4000

Input Sequence Length — expected average input tokens.

osl
integerDefaults to 1000

Output Sequence Length — expected average output tokens.

concurrency
number

Target concurrency level. Required (or requestRate) when the planner is disabled.

requestRate
number

Target request rate in requests per second. Required (or concurrency) when the planner is disabled.

sla
SLASpec

Service-level agreement targets that drive profiling optimization.

ttft
number

Time To First Token target, in milliseconds.

itl
number

Inter-Token Latency target, in milliseconds.

e2eLatency
number

End-to-end request latency target, in milliseconds. An alternative to specifying ttft + itl; cannot be combined with them.

optimizationType
OptimizationType

Optimization target for SLA profiling.

Allowed values: latency throughput
hardware
HardwareSpec

GPU hardware for profiling and deployment.

gpuSku
GPUSKUType

GPU type to target (see SKU format). When omitted, auto-detected by selecting the GPU with the highest node count, then highest VRAM. In mixed-GPU clusters, set this to choose the GPU type; discovery and totalGpus are then restricted to matching nodes.

vramMb
number

VRAM per GPU in MiB. Auto-detected from cluster GPU nodes when omitted.

totalGpus
integer

GPU budget for profiling and deployment; the profiler uses it to determine parallelism and replica count. When omitted, computed by counting GPUs on discovered nodes (filtered by gpuSku when set), temporarily capped at 32 to limit the profiler search space. Set explicitly to override the cap.

numGpusPerNode
integer

GPUs per node. Auto-detected when omitted.

interconnect
string

Primary GPU-to-GPU interconnect within a node, for example nvlink or pcie. Capability metadata for profiling and planning only — it does not enable any interconnect. Best-effort discovery distinguishes nvlink from pcie when omitted. An over-optimistic value can lead the planner to overestimate intra-node bandwidth and pick overly aggressive parallelism.

rdma
boolean

Whether the cluster has RDMA-capable networking for Dynamo data movement. Capability metadata only — it does not install or configure RDMA. Best-effort discovery via node labels when omitted. A false positive can cause transports to assume RDMA is usable and fail to connect; a false negative falls back to non-RDMA transports.

modelCache
ModelCacheSpec

Optional PVC configuration for pre-downloaded model weights. When set, weights are loaded from the PVC instead of downloading from HuggingFace.

pvcName
string

Name of a ReadWriteMany PVC containing cached model weights. The PVC must exist in the same namespace as the DGDR.

pvcModelPath
string

Path to the model checkpoint directory within the PVC, for example deepseek-r1 or models/Llama-3.1-405B-FP8.

pvcMountPath
stringDefaults to /opt/model-cache

Mount path for the PVC inside the container.

features
FeaturesSpec

Optional Dynamo platform features in the generated deployment.

planner
runtime.RawExtension

A PlannerConfig object passed to the Planner service. The presence of this field (non-null) enables the planner in the generated DGD. DGDR passes it through without field-level validation; the Planner service validates it at startup. See Planner for the full field list.

runtime.RawExtension
mocker
MockerSpec

Configures the simulated (mocker) backend for testing without GPUs.

enabled
boolean

Deploy mocker workers instead of real inference workers.

DGDR does not currently expose a features.kvRouter field. To configure router mode or KV-aware routing, use a direct DGD, a tuned recipe, or overrides.dgd (see Generated DGD overrides).

Overrides

overrides
OverridesSpec

Customizes the profiling job and the generated DGD.

profilingJob
batch/v1.JobSpec

Overrides merged into the controller-generated profiling Job spec (for example, tolerations or node selectors). Only customizes the profiling Job — not the deployed worker pods.

batch/v1.JobSpec
dgd
runtime.RawExtension

A partial, versioned DGD merged into the generated deployment. Use nvidia.com/v1beta1 for new overrides; nvidia.com/v1alpha1 remains supported for compatibility. See Generated DGD overrides.

runtime.RawExtension

Planner

spec.features.planner is a PlannerConfig object, passed through to the Planner service without field-level validation; the Planner validates it at startup. Setting it (non-null) enables the planner in the generated DGD. The same object is documented standalone at Planner Configuration; its fields are reproduced here. Nested list-and-object types are expanded under Additional Types. All fields are optional.

Core settings

mode
stringDefaults to disagg

Planner operating mode, matching the deployment topology.

Allowed values: disagg prefill decode agg
backend
stringDefaults to vllm

Inference backend the Planner scales.

Allowed values: vllm sglang trtllm mocker
environment
stringDefaults to kubernetes

Runtime environment that determines how the Planner applies scaling actions.

Allowed values: kubernetes virtual global-planner
namespace
stringDefaults to env DYN_NAMESPACE, else dynamo

Dynamo namespace of the deployment the Planner manages. Defaults from DYN_NAMESPACE, which the operator injects as {k8s_namespace}-{dgd_name}. Excluded from serialized output.

model_name
stringDefaults to null

Optional model name override. Auto-detected from the deployment when unset.

global_planner_namespace
stringDefaults to null

Namespace where the GlobalPlanner runs. Required when environment is global-planner.

log_dir
stringDefaults to null

Optional directory for Planner log output.

Optimization target

optimization_target
stringDefaults to throughput

Scaling strategy. throughput and latency use static thresholds on queue depth and KV cache utilization — no SLA targets or profiling required. load uses user-defined prefill queue-token and decode KV-utilization thresholds. sla uses the Rust engine performance model to target specific ttft_ms / itl_ms values. Under any target other than sla, the Planner forces load scaling on, throughput scaling off, and ignores ttft_ms / itl_ms.

Allowed values: throughput latency load sla
enable_throughput_scaling
booleanDefaults to true

Enable predictive, traffic-based scaling. Only honored when optimization_target is sla.

enable_load_scaling
booleanDefaults to false

Enable reactive, load-based scaling. Only honored when optimization_target is sla. At least one scaling mode must be enabled.

ttft_ms
numberDefaults to 500.0

Time To First Token SLA target, in milliseconds. Also accepts the alias ttft. Must be greater than 0. Used only under optimization_target: sla.

itl_ms
numberDefaults to 50.0

Inter-Token Latency SLA target, in milliseconds. Also accepts the alias itl. Used only under optimization_target: sla.

Performance model and pre-deployment sweeping

pre_deployment_sweeping_mode
stringDefaults to rapid

How to generate optional bootstrap performance data. none skips bootstrap data; rapid uses AIConfigurator to simulate engine performance (~30s); thorough measures on real GPUs (several hours).

Allowed values: none rapid thorough
profile_results_dir
stringDefaults to profiling_results

Directory holding profiler-generated performance data (npz or JSON), used to bootstrap or tune the performance model when the get_perf_metrics endpoint is unavailable.

aic_perf_model
AICPerfModelSpecDefaults to null

Native AIConfigurator forward-pass model identity for the Rust engine performance shim, enabling real-time AIC estimates with online correction. Unsupported native configs fall back to FPM regression. Does not trigger an AIC interpolation sweep.

hf_id
stringRequired

HuggingFace model id, for example Qwen/Qwen3-32B.

system
stringRequired

AIC system identifier, for example h200_sxm.

backend
stringRequired
Allowed values: trtllm vllm sglang
backend_version
stringDefaults to null

Optional backend version string.

prefill_pick
PickedParallelConfigDefaults to null

Parallelism pick for the prefill engine. Required for mode disagg or prefill. See PickedParallelConfig.

decode_pick
PickedParallelConfigDefaults to null

Parallelism pick for the decode engine. Required for mode disagg, decode, or agg. See PickedParallelConfig.

model_arch
stringDefaults to null

Optional model architecture override.

weight_dtype
stringDefaults to null

Optional weight dtype.

moe_dtype
stringDefaults to null

Optional Mixture-of-Experts dtype.

activation_dtype
stringDefaults to null

Optional activation dtype.

kv_cache_dtype
stringDefaults to null

Optional KV cache dtype.

aic_interpolation
AICInterpolationSpecDefaults to null

AIConfigurator interpolation spec. Populated by the profiler in rapid mode and written onto the Planner ConfigMap; you do not normally set this by hand. When present, the Planner runs the AIC sweep in-process at bootstrap to seed the performance model. See AICInterpolationSpec.

GPU budget

max_gpu_budget
integerDefaults to 8

Maximum total GPUs the Planner may allocate across worker types.

min_gpu_budget
integerDefaults to -1

Per-DGD GPU floor enforced by the local Planner. -1 disables it. When set with max_gpu_budget such that min == max, the Planner pins the per-DGD total and only redistributes replicas between prefill and decode.

min_endpoint
integerDefaults to 1

Minimum number of engine endpoints (replicas) to maintain per worker type.

decode_engine_num_gpu
integerDefaults to null

GPUs per decode engine replica. Auto-detected from the deployment when unset.

prefill_engine_num_gpu
integerDefaults to null

GPUs per prefill engine replica. Auto-detected from the deployment when unset.

Throughput-based scaling

throughput_adjustment_interval_seconds
integerDefaults to 180

Seconds between throughput-based scaling decisions. Also accepts the alias throughput_adjustment_interval.

throughput_metrics_source
stringDefaults to frontend

Prometheus traffic source. frontend reads dynamo_frontend_* metrics from the public Frontend; router reads dynamo_component_router_* from a LocalRouter (use for a pool-local Planner in GlobalPlanner deployments).

Allowed values: frontend router

Load-based scaling

load_adjustment_interval_seconds
integerDefaults to 5

Seconds between FPM tuning updates and load-based scaling decisions. Even when only throughput scaling is enabled, live FPM observations feed the performance model at this interval. Also accepts the alias load_adjustment_interval. Must be greater than 0 and, when both scaling modes are on, shorter than throughput_adjustment_interval_seconds.

max_num_fpm_samples
integerDefaults to 64

Maximum retained ForwardPassMetrics observations for online tuning and regression fallback.

fpm_sample_bucket_size
integerDefaults to 16

Number of buckets for observation retirement. Must be a perfect square.

load_scaling_down_sensitivity
integerDefaults to 80

Scale-down sensitivity from 0 to 100 (0 = never scale down, 100 = aggressive).

load_min_observations
integerDefaults to 5

Minimum observations before load-based scaling decisions begin (cold-start threshold).

prefill_scale_up_queue_tokens
integerDefaults to null

Prefill queue-token count that triggers scale-up. Required (with the scale-down field) when optimization_target is load and mode includes prefill. Must be greater than 0 and greater than prefill_scale_down_queue_tokens.

prefill_scale_down_queue_tokens
integerDefaults to null

Prefill queue-token count that allows scale-down. Required alongside prefill_scale_up_queue_tokens.

decode_scale_up_kv_rate
numberDefaults to null

Decode KV-utilization percentage (0–100) that triggers scale-up. Required (with the scale-down field) when optimization_target is load and mode includes decode. Must be greater than decode_scale_down_kv_rate.

decode_scale_down_kv_rate
numberDefaults to null

Decode KV-utilization percentage (0–100) that allows scale-down. Required alongside decode_scale_up_kv_rate.

speculative_nextn
integerDefaults to 0

Manual fallback speculative-decoding depth. A worker’s published runtime_config.runtime_data.spec_decode.nextn takes precedence when present. Must be 0 or greater.

Load prediction

load_predictor
stringDefaults to arima

Prediction method for request count, ISL, and OSL.

Allowed values: constant arima kalman prophet
load_predictor_log1p
booleanDefaults to false

Apply a log1p transform to predicted request count, ISL, and OSL.

prophet_window_size
integerDefaults to 50

Window size (seconds) for the Prophet predictor.

load_predictor_warmup_trace
stringDefaults to null

Path to a warmup trace file for bootstrapping predictions.

kalman_q_level
numberDefaults to 1.0

Kalman process noise for the level component.

kalman_q_trend
numberDefaults to 0.1

Kalman process noise for the trend component.

kalman_r
numberDefaults to 10.0

Kalman measurement noise.

kalman_min_points
integerDefaults to 5

Minimum data points before Kalman predictions activate.

Prometheus metrics

These fields default from environment variables and are excluded from serialized output.

metric_pulling_prometheus_endpoint
stringDefaults to env PROMETHEUS_ENDPOINT

Prometheus endpoint the Planner queries for traffic metrics. Defaults to PROMETHEUS_ENDPOINT, else the in-cluster kube-prometheus service URL.

metric_pulling_prometheus_token
stringDefaults to env PROMETHEUS_TOKEN

Optional bearer token sent as Authorization: Bearer <token> on every PromQL request. Read once at startup.

metric_pulling_prometheus_token_file
stringDefaults to env PROMETHEUS_TOKEN_FILE

Path to a file containing a bearer token, re-read before every request so rotated tokens are picked up without a restart.

metric_pulling_prometheus_ssl_verify
booleanDefaults to false

Verify the upstream Prometheus TLS certificate. Defaults from PROMETHEUS_SSL_VERIFY (1/true/yes enable it). Pair with a CA bundle for a private CA.

metric_pulling_prometheus_extra_query_params
objectDefaults to env PROMETHEUS_EXTRA_QUERY_PARAMS

Fixed key/value pairs appended as URL query parameters on every PromQL request. Set PROMETHEUS_EXTRA_QUERY_PARAMS as a URL query string, for example namespace=my-ns&tenant=foo.

metric_pulling_prometheus_ca_bundle
stringDefaults to env PROMETHEUS_CA_BUNDLE

Path to a CA bundle for verifying the upstream Prometheus TLS certificate. No-op unless SSL verification is enabled. Must point to an existing file when set.

metric_reporting_prometheus_port
integerDefaults to 0

Port on which the Planner exposes its own dynamo_planner_* metrics. Defaults from PLANNER_PROMETHEUS_PORT. 0 disables.

Advisory and diagnostics

advisory
booleanDefaults to false

Suggestion-only mode. The Planner computes, logs, exports, and reports recommended replica counts without executing scaling actions or changing the deployment. Use it to evaluate a new configuration or validate SLA targets against production traffic.

report_interval_hours
numberDefaults to 24.0

Generate an HTML diagnostics report every N hours (simulated time). Set to null to disable. Must be a positive finite number or null.

report_output_dir
stringDefaults to ./planner_reports

Directory for HTML diagnostics reports.

report_filename
stringDefaults to null

Fixed report filename written under report_output_dir. When unset, a timestamped name is used.

report_write_gzip_log
booleanDefaults to true

Write a compressed JSONL diagnostics log (.log.jsonl.gz) next to each HTML report.

live_dashboard_port
integerDefaults to 8080

Port for the live diagnostics dashboard HTTP server. 0 disables. When enabled, visit http://host:port/ for a real-time Plotly report.

Scheduling and plugin pipeline

scheduling
SchedulingConfig

Plugin-pipeline scheduling. Rejects unknown fields.

scale_interval_seconds
numberDefaults to gcd of enabled builtin intervals

Base pipeline cadence. The pipeline wakes once per interval; each plugin’s execution_interval_seconds decides whether it fires. When unset, computed as the gcd of load_adjustment_interval_seconds and, when throughput scaling is enabled, throughput_adjustment_interval_seconds. Must be greater than 0 and evenly divide both adjustment intervals.

tick_max_duration_seconds
numberDefaults to 30.0

Outer deadline wrapping the full 4-stage pipeline. Exceeding it aborts the tick; the next tick runs from a clean state. Must be greater than 0.

external_plugins
[]ExternalPluginEntryDefaults to []

Static external-plugin registrations applied at startup. Per-entry failures are logged but do not crash the Planner. See ExternalPluginEntry.

gateway
GatewayConfig

gRPC registration gateway for self-registering plugins. Disabled by default.

enabled
booleanDefaults to false

Open the gRPC gateway at listen. Not needed for static external_plugins.

listen
stringDefaults to unix:///var/run/dynamo/planner/registry.sock

Bind address. A unix: socket path for in-Pod registration, or host:port for TCP.

allow_insecure
booleanDefaults to false

Permit binding a plaintext (no-TLS) gateway on a TCP listen. Fails closed by default because the gateway receives plugins’ shared-secret tokens. unix: listens are always allowed.

plugin_registration
PluginRegistrationConfig

Plugin registry configuration — auth, transport, heartbeat, and in-process plugins. Rejects unknown fields.

auth
AuthConfig

Registry authentication. An empty trusted_sources fails closed (every token rejected).

trusted_sources
[]stringDefaults to []

Accepted auth sources. Use ['allow_unauthenticated'] for development, or ['static_secret'] with static_secrets for production.

Allowed values: static_secret allow_unauthenticated
static_secrets
objectDefaults to {}

Map of secret_value to subject_label. Populate from a mounted Kubernetes Secret rather than hard-coding in the ConfigMap.

transport
TransportConfig

Outbound gRPC transport settings for calling plugins.

allow_insecure_grpc
booleanDefaults to false

Allow plaintext grpc:// plugin endpoints. Required to use any grpc:// endpoint; mTLS support is not yet shipped.

request_timeout_seconds
numberDefaults to 5.0

Per-RPC timeout applied to every plugin call. Must be greater than 0.

keepalive_time_ms
integerDefaults to 30000

gRPC keepalive time, in milliseconds.

max_message_size_bytes
integerDefaults to 10485760

Maximum gRPC message size (10 MB).

protocol_version_min
stringDefaults to 1.0

Minimum accepted plugin protocol version.

protocol_version_max
stringDefaults to 1.0

Maximum accepted plugin protocol version.

heartbeat_timeout_seconds
numberDefaults to 15.0

Seconds without a heartbeat before a plugin is considered missed.

heartbeat_missed_threshold
integerDefaults to 2

Consecutive missed heartbeats before a plugin is dropped.

in_process_plugins
[]InProcessPluginSpecDefaults to []

In-process plugins loaded from Python modules at startup. See InProcessPluginSpec.

admin
AdminAuthConfig

Admin (ListPlugins) RBAC config. Parsed but currently inert — the gateway default-denies ListPlugins regardless of mode.

mode
stringDefaults to allow_all
Allowed values: allow_all k8s_rbac

spec.overrides.dgd is not required to enable the Planner. Use spec.features.planner for Planner enablement and configuration; use spec.overrides.dgd only to customize the generated DGD after DGDR assembles it. For the deployment workflow, scaling-mode concepts, and cross-field validation rules, see the Planner Configuration reference and Planner Guide.

Generated DGD overrides

spec.overrides.dgd is a partial, versioned DGD merged into the profiler-generated deployment after Dynamo selects a configuration. Use apiVersion: nvidia.com/v1beta1 and kind: DynamoGraphDeployment for new overrides. Components, pod-template containers, and container environment variables merge by name; atomic lists such as graph-level spec.env and container args replace their generated values.

For example, set an environment variable on one generated component:

1apiVersion: nvidia.com/v1beta1
2kind: DynamoGraphDeploymentRequest
3metadata:
4 name: qwen3-sglang
5spec:
6 model: Qwen/Qwen3-30B-A3B
7 backend: sglang
8 overrides:
9 dgd:
10 apiVersion: nvidia.com/v1beta1
11 kind: DynamoGraphDeployment
12 spec:
13 components:
14 - name: SGLangDecodeWorker # replace with the generated component name
15 podTemplate:
16 spec:
17 containers:
18 - name: main
19 env:
20 - name: TRITON_PTXAS_PATH
21 value: /usr/local/cuda/bin/ptxas

Inspect .status.profilingResults.selectedConfig with autoApply: false to find component names. Overrides can modify only components present in the generated DGD; unknown names are ignored with a warning, and an override cannot add deployment topology. Labels and annotations merge, metadata.name selects the final DGD name, and other identity metadata is ignored. status, null values for typed fields, and field deletion are not supported.

Existing nvidia.com/v1alpha1 overrides remain supported. They use the older spec.services shape, merge services by name, and append extraPodSpec.mainContainer.args to generated worker arguments. In v1beta1, container args is atomic and replaces the generated list. Use v1beta1 for new overrides and provide the complete desired argument list when overriding args.

overrides.profilingJob customizes only the profiling Job. Use overrides.dgd for settings that must appear on deployed components.

For worked examples, see DGDR Examples. For router configuration and EPP topologies, see the Router Guide and Gateway API Inference Extension.

SKU format

hardware.gpuSku uses lowercase underscore format (h100_sxm, not H100-SXM5-80GB).

Allowed values: 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

Not all SKUs are supported by the AIC profiler for rapid mode.

PCIe variants not yet supported by the profiler. The CRD admits PCIe SKUs (h100_pcie, a100_pcie, v100_pcie), but the profiler does not currently ship training data for them. You can submit a DGDR with a PCIe value; the operator accepts it but profiler-assisted sizing falls back to defaults. Profiler support for PCIe SKUs is tracked as an engineering follow-up.

Status

The operator maintains observed state under status. See Lifecycle for how phase and profilingPhase progress, and for the full condition-type list.

status.phase
DGDRPhase

High-level lifecycle phase of the deployment request.

Allowed values: Pending Profiling Ready Deploying Deployed Failed
status.profilingPhase
ProfilingPhase

Current sub-phase of the profiling pipeline. Only meaningful while phase is Profiling; cleared when profiling completes or fails.

Allowed values: Initializing SweepingPrefill SweepingDecode SelectingConfig BuildingCurves GeneratingDGD Done
status.dgdName
string

Name of the generated or created DynamoGraphDeployment.

status.profilingJobName
string

Name of the Kubernetes Job running the profiler.

status.conditions
[]metav1.Condition

Latest observed conditions. Standard types include Succeeded, Validation, Profiling, SpecGenerated, and DeploymentReady — see Conditions.

metav1.Condition
status.profilingResults
ProfilingResultsStatus

Output of the profiling process — Pareto-optimal configurations and the selected deployment configuration.

pareto
[]ParetoConfig

Pareto-optimal deployment configurations discovered during profiling. Each entry represents a different cost/performance trade-off.

See: ParetoConfig

selectedConfig
runtime.RawExtension

Recommended configuration chosen by the profiler based on the SLA targets. This is the configuration used for deployment when autoApply is true.

runtime.RawExtension
status.deploymentInfo
DeploymentInfoStatus

State of the deployed DynamoGraphDeployment. Populated once a DGD has been created (via autoApply or manually).

replicas
integer

Desired number of replicas.

availableReplicas
integer

Number of replicas that are available and ready.

status.observedGeneration
integer

Most recent metadata.generation observed by the controller.

Additional Types

Nested struct types referenced by the fields above, broken out here to keep the field lists shallow. Types prefixed batch/v1., metav1., or runtime. are standard Kubernetes types and link to their Go package documentation instead of being expanded.

ParetoConfig

A single Pareto-optimal deployment configuration discovered during profiling. Referenced by status.profilingResults.pareto.

config
runtime.RawExtension

Full deployment configuration for this Pareto point.

runtime.RawExtension

PickedParallelConfig

A parallelism pick emitted by AIConfigurator. Referenced by features.planner.aic_perf_model.prefill_pick / decode_pick and by AICInterpolationSpec.

tp
integerDefaults to 1

Tensor-parallel size.

pp
integerDefaults to 1

Pipeline-parallel size.

dp
integerDefaults to 1

Data-parallel (attention-DP) size.

moe_tp
integerDefaults to 1

Mixture-of-Experts tensor-parallel size.

moe_ep
integerDefaults to 1

Mixture-of-Experts expert-parallel size.

AICInterpolationSpec

Everything the Planner needs to reproduce a rapid-mode AIC sweep. Written by the profiler onto the Planner ConfigMap; not normally hand-authored. Referenced by features.planner.aic_interpolation.

hf_id
stringRequired

HuggingFace model id.

system
stringRequired

AIC system identifier.

backend
stringRequired
Allowed values: trtllm vllm sglang
isl
integerRequired

Input sequence length for the sweep. Must be greater than 0.

osl
integerRequired

Output sequence length for the sweep. Must be greater than 0.

sweep_max_context_length
integerRequired

Maximum context length swept. Must be greater than 0.

prefill_interpolation_granularity
integerRequired

Number of prefill interpolation points. Must be greater than 0.

decode_interpolation_granularity
integerRequired

Number of decode interpolation points. Must be greater than 0.

prefill_pick
PickedParallelConfigRequired

Prefill parallelism pick. See PickedParallelConfig.

decode_pick
PickedParallelConfigRequired

Decode parallelism pick. See PickedParallelConfig.

ExternalPluginEntry

One static external-plugin registration under features.planner.scheduling.external_plugins. Rejects unknown fields.

plugin_id
stringRequired

Unique identifier. Must not collide with a builtin plugin id.

plugin_type
stringRequired

Pipeline stage the plugin participates in.

Allowed values: predict propose reconcile constrain
priority
integerRequired

Stage priority. Smaller number is more authoritative.

endpoint
stringRequired

Wire endpoint. Must start with grpc://host:port; inproc:// is rejected here.

auth_token
string

Bearer token validated by the registry. Populate from a mounted Secret.

protocol_version
stringDefaults to 1.0

Plugin protocol version. Must fall within the registry’s supported range.

version
stringDefaults to v1

Plugin’s own version string, surfaced in ListPlugins.

execution_interval_seconds
numberDefaults to 0.0

0.0 runs every tick; a positive value throttles to every N seconds. Must be 0 or greater.

hold_policy
stringDefaults to HOLD_LAST

Behavior when throttled. HOLD_LAST reuses the cached result; ACCEPT_WHEN_IDLE treats the plugin as no-opinion when not due. Accepts the name (case-insensitive) or its integer value.

Allowed values: HOLD_LAST ACCEPT_WHEN_IDLE
needs
[]stringDefaults to []

Capability list consumed by the type-aware merge.

requires_produced_fields
[]stringDefaults to []

Dot-paths into the pipeline context (for example predictions, observations.traffic) that must be set for the plugin to fire on a tick. Empty means no gating.

observation_window_seconds
numberDefaults to 0.0

Aggregation window for windowed observation types in needs. 0.0 uses scale_interval freshness; a positive value aggregates over the last N seconds. Enforced to be greater than or equal to scale_interval_seconds.

InProcessPluginSpec

One in-process plugin under features.planner.plugin_registration.in_process_plugins. Rejects unknown fields.

module
stringRequired

Python module path containing the plugin class.

class
stringRequired

Plugin class name within module.

plugin_id
stringRequired

Unique plugin identifier.

plugin_type
stringRequired
Allowed values: predict propose reconcile constrain
priority
integerRequired

Stage priority. Smaller number is more authoritative.

execution_interval_seconds
numberDefaults to 0.0

0.0 runs every tick; a positive value throttles to every N seconds.

hold_policy
stringDefaults to ACCEPT_WHEN_IDLE

Behavior when throttled.

Allowed values: ACCEPT_WHEN_IDLE HOLD_LAST
needs
[]stringDefaults to []

Capability list consumed by the type-aware merge.

requires_produced_fields
[]stringDefaults to []

Dot-paths into the pipeline context that must be set for the plugin to fire on a tick.

observation_window_seconds
numberDefaults to 0.0

Aggregation window for windowed observation types in needs. Must be 0 or greater.

kwargs
objectDefaults to {}

Keyword arguments passed to the plugin class constructor.

Lifecycle

When you create a DGDR, it progresses through these phases (status.phase):

PhaseWhat is happening
PendingSpec validated; operator is discovering GPU hardware and preparing the profiling job
ProfilingProfiling job running — sub-phases in status.profilingPhase: Initializing, SweepingPrefill, SweepingDecode, SelectingConfig, BuildingCurves, GeneratingDGD, Done
ReadyProfiling complete; optimal config stored in status.profilingResults.selectedConfig. Terminal state when autoApply: false.
DeployingCreating the DGD (only when autoApply: true)
DeployedDGD is running and healthy
FailedUnrecoverable error — profiling failures are not retried (backoffLimit: 0); check events and conditions for details

Conditions

The operator maintains these conditions on status.conditions:

ConditionMeaning
ValidationSpec validation passed or failed
ProfilingProfiling job is running, succeeded, or failed
SpecGeneratedGenerated DGD spec is available
DeploymentReadyDGD is deployed and healthy
SucceededAggregate condition — true when the DGDR has reached its target state

Monitoring

$# Watch phase transitions
$kubectl get dgdr my-model -n $NAMESPACE -w
$
$# Detailed status, conditions, and events
$kubectl describe dgdr my-model -n $NAMESPACE
$
$# Profiling sub-phase
$kubectl get dgdr my-model -n $NAMESPACE -o jsonpath='{.status.profilingPhase}'
$
$# Profiling job logs
$kubectl get pods -n $NAMESPACE -l nvidia.com/dgdr-name=my-model
$kubectl logs -f <profiling-pod-name> -n $NAMESPACE
$
$# View generated DGD spec (when autoApply: false)
$kubectl get dgdr my-model -n $NAMESPACE \
> -o jsonpath='{.status.profilingResults.selectedConfig}' | python3 -m json.tool
$
$# View Pareto-optimal configs from profiling
$kubectl get dgdr my-model -n $NAMESPACE \
> -o jsonpath='{.status.profilingResults.pareto}'

Resource ownership

  • The DGDR does not set an owner reference on the DGD it creates. Deleting a DGDR does not delete the DGD — it persists independently so it can continue serving traffic.
  • The relationship is tracked via labels: dgdr.nvidia.com/name and dgdr.nvidia.com/namespace.
  • Additional resources (planner ConfigMaps) are created in the same namespace and labeled with dgdr.nvidia.com/name.

Known issues

  • pareto_analysis.py produces NaN for some configurations. Tracked as an engineering follow-up. Workaround: re-run with a narrower sweep; narrow sweeps bypass the NaN path in practice.
  • PCIe profiler data not yet available. See the PCIe callout under SKU format.