ComponentsPlanner

Planner Examples

View as Markdown

Practical examples for deploying the SLA Planner with different configurations. For deployment concepts, see the Planner Guide. For a quick overview, see the Planner README.

Basic Examples

Minimal DGDR with AIC (Fastest)

The simplest way to deploy with the SLA planner. Uses AI Configurator for offline profiling (20-30 seconds instead of hours):

1apiVersion: nvidia.com/v1alpha1
2kind: DynamoGraphDeploymentRequest
3metadata:
4  name: sla-aic
5spec:
6  model: Qwen/Qwen3-32B
7  backend: vllm
8
9  profilingConfig:
10    profilerImage: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.6.1"
11    config:
12      sla:
13        isl: 3000
14        osl: 150
15        ttft: 200
16        itl: 20
17      sweep:
18        useAiConfigurator: true
19        aicSystem: h200_sxm
20        aicHfId: Qwen/Qwen3-32B
21        aicBackendVersion: "0.20.0"
22
23  deploymentOverrides:
24    workersImage: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.6.1"
25
26  autoApply: true

Deploy:

$export NAMESPACE=your-namespace
$kubectl apply -f benchmarks/profiler/deploy/profile_sla_aic_dgdr.yaml -n $NAMESPACE

Online Profiling (Real Measurements)

Standard online profiling runs real GPU measurements for more accurate results. Takes 2-4 hours:

1apiVersion: nvidia.com/v1alpha1
2kind: DynamoGraphDeploymentRequest
3metadata:
4  name: sla-online
5spec:
6  model: meta-llama/Llama-3.3-70B-Instruct
7  backend: vllm
8
9  profilingConfig:
10    profilerImage: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.6.1"
11    config:
12      sla:
13        isl: 3000
14        osl: 150
15        ttft: 200
16        itl: 20
17      sweep:
18        useAiConfigurator: false
19        prefillInterpolationGranularity: 16
20        decodeInterpolationGranularity: 6
21
22  deploymentOverrides:
23    workersImage: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.6.1"
24
25  autoApply: true

Deploy:

$kubectl apply -f benchmarks/profiler/deploy/profile_sla_dgdr.yaml -n $NAMESPACE

Available sample DGDRs in benchmarks/profiler/deploy/:

  • profile_sla_dgdr.yaml: Standard online profiling for dense models
  • profile_sla_aic_dgdr.yaml: Fast offline profiling using AI Configurator
  • profile_sla_moe_dgdr.yaml: Online profiling for MoE models (SGLang)

Profiling Config Cases: Prior to 0.8.1, fields under profilingConfig.config use snake_case. Starting 0.8.1, fields use camelCase. There is backwards compatibility to snake_case, but example DGDRs use camelCase.

Kubernetes Examples

MoE Models (SGLang)

For Mixture-of-Experts models like DeepSeek-R1, use SGLang backend:

1apiVersion: nvidia.com/v1alpha1
2kind: DynamoGraphDeploymentRequest
3metadata:
4  name: sla-moe
5spec:
6  model: deepseek-ai/DeepSeek-R1
7  backend: sglang
8
9  profilingConfig:
10    profilerImage: "nvcr.io/nvidia/ai-dynamo/sglang-runtime:0.6.1"
11    config:
12      sla:
13        isl: 4000
14        osl: 500
15        ttft: 300
16        itl: 10
17      sweep:
18        useAiConfigurator: false
19
20  deploymentOverrides:
21    workersImage: "nvcr.io/nvidia/ai-dynamo/sglang-runtime:0.6.1"
22
23  autoApply: true

Deploy:

$kubectl apply -f benchmarks/profiler/deploy/profile_sla_moe_dgdr.yaml -n $NAMESPACE

Using Existing DGD Configs (Custom Setups)

Reference an existing DynamoGraphDeployment config via ConfigMap:

Step 1: Create ConfigMap from your DGD config:

$kubectl create configmap deepseek-r1-config \
>  --from-file=disagg.yaml=/path/to/your/disagg.yaml \
>  --namespace $NAMESPACE \
>  --dry-run=client -o yaml | kubectl apply -f -

Step 2: Reference it in your DGDR:

1apiVersion: nvidia.com/v1alpha1
2kind: DynamoGraphDeploymentRequest
3metadata:
4  name: deepseek-r1
5spec:
6  model: deepseek-ai/DeepSeek-R1
7  backend: sglang
8
9  profilingConfig:
10    profilerImage: "nvcr.io/nvidia/ai-dynamo/sglang-runtime:0.6.1"
11    configMapRef:
12      name: deepseek-r1-config
13      key: disagg.yaml  # Must match the key used in --from-file
14    config:
15      sla:
16        isl: 4000
17        osl: 500
18        ttft: 300
19        itl: 10
20      sweep:
21        useAiConfigurator: true
22        aicSystem: h200_sxm
23        aicHfId: deepseek-ai/DeepSeek-V3
24        aicBackendVersion: "0.20.0"
25
26  deploymentOverrides:
27    workersImage: "nvcr.io/nvidia/ai-dynamo/sglang-runtime:0.6.1"
28
29  autoApply: true

The profiler uses the DGD config from the ConfigMap as a base template, then optimizes it based on your SLA targets. The controller automatically injects spec.model and spec.backend into the final configuration.

Inline Configuration (Simple Use Cases)

For simple use cases without a custom DGD config, provide profiler configuration directly. The profiler auto-generates a basic DGD configuration:

1profilingConfig:
2  config:
3    sla:
4      isl: 8000
5      osl: 200
6      ttft: 200.0
7      itl: 10.0
8
9    hardware:
10      minNumGpusPerEngine: 2
11      maxNumGpusPerEngine: 8
12      gpuType: h200_sxm
13
14    sweep:
15      prefillInterpolationGranularity: 16
16      decodeInterpolationGranularity: 6

Mocker Deployment (Testing)

Deploy a mocker backend that simulates GPU timing behavior without real GPUs. Useful for:

  • Large-scale experiments without GPU resources
  • Testing planner behavior and infrastructure
  • Validating deployment configurations
1spec:
2  model: <model-name>
3  backend: trtllm  # Real backend for profiling
4  useMocker: true  # Deploy mocker instead of real backend
5
6  profilingConfig:
7    profilerImage: "nvcr.io/nvidia/dynamo/trtllm-runtime:<image-tag>"
8    config:
9      sla:
10        isl: 3000
11        osl: 150
12        ttft: 200
13        itl: 20
14      sweep:
15        useAiConfigurator: true
16        aicSystem: h100_sxm
17  autoApply: true

Profiling runs against the real backend (via GPUs or AIC). The mocker deployment then uses profiling data to simulate realistic timing.

Model Cache PVC (0.8.1+)

For large models, use a pre-populated PVC instead of downloading from HuggingFace:

See SLA-Driven Profiling for configuration details.

Advanced Examples

Custom Load Predictors

Warm-starting with Trace Data

Pre-load predictors with historical request patterns before live traffic:

1# In planner arguments
2args:
3  - --load-predictor arima
4  - --load-predictor-warmup-trace /data/trace.jsonl
5  - --load-predictor-log1p

The trace file should be in mooncake-style JSONL format with request-count, ISL, and OSL samples.

Kalman Filter Tuning

For workloads with rapid changes, tune the Kalman filter:

1args:
2  - --load-predictor kalman
3  - --kalman-q-level 2.0      # Higher = more responsive to level changes
4  - --kalman-q-trend 0.5      # Higher = trend changes faster
5  - --kalman-r 5.0            # Lower = trusts new measurements more
6  - --kalman-min-points 3     # Fewer points before forecasting starts
7  - --load-predictor-log1p    # Often helps with request-rate series

Prophet for Seasonal Workloads

For workloads with daily/weekly patterns:

1args:
2  - --load-predictor prophet
3  - --prophet-window-size 100   # Larger window for seasonal detection
4  - --load-predictor-log1p

Virtual Connector

For non-Kubernetes environments, use the VirtualConnector to communicate scaling decisions:

1from dynamo._core import DistributedRuntime, VirtualConnectorClient
2
3# Initialize client
4client = VirtualConnectorClient(distributed_runtime, namespace)
5
6# Main loop: watch for planner decisions and execute them
7while True:
8    # Block until the planner makes a new scaling decision
9    await client.wait()
10
11    # Read the decision
12    decision = await client.get()
13    print(f"Scale to: prefill={decision.num_prefill_workers}, "
14          f"decode={decision.num_decode_workers}, "
15          f"id={decision.decision_id}")
16
17    # Execute scaling in your environment
18    scale_prefill_workers(decision.num_prefill_workers)
19    scale_decode_workers(decision.num_decode_workers)
20
21    # Report completion
22    await client.complete(decision)

See components/planner/test/test_virtual_connector.py for a full working example.

Planner Configuration Passthrough

Pass planner-specific settings through the DGDR:

1profilingConfig:
2  config:
3    planner:
4      plannerMinEndpoint: 2

Review Before Deploy (autoApply: false)

Disable auto-deployment to inspect the generated DGD:

1spec:
2  autoApply: false

After profiling completes:

$# Extract and review generated DGD
$kubectl get dgdr sla-aic -n $NAMESPACE \
>  -o jsonpath='{.status.generatedDeployment}' > my-dgd.yaml
$
$# Review and modify as needed
$vi my-dgd.yaml
$
$# Deploy manually
$kubectl apply -f my-dgd.yaml -n $NAMESPACE

Profiling Artifacts with PVC

Save detailed profiling artifacts (plots, logs, raw data) to a PVC:

1spec:
2  profilingConfig:
3    outputPVC: "dynamo-pvc"
4    config:
5      sla:
6        isl: 3000
7        osl: 150
8        ttft: 200
9        itl: 20

Setup:

$export NAMESPACE=your-namespace
$deploy/utils/setup_benchmarking_resources.sh

Access results:

$kubectl apply -f deploy/utils/manifests/pvc-access-pod.yaml -n $NAMESPACE
$kubectl wait --for=condition=Ready pod/pvc-access-pod -n $NAMESPACE --timeout=60s
$kubectl cp $NAMESPACE/pvc-access-pod:/data ./profiling-results
$kubectl delete pod pvc-access-pod -n $NAMESPACE