dynamo/docs/components/profiler/profiler-guide.md at main · drivenets/dynamo

title
Profiler Guide

Overview

The Dynamo Profiler analyzes model inference performance and generates optimized deployment configurations (DynamoGraphDeployments). Given a model, hardware, and SLA targets, it determines the best parallelization strategy, selects optimal prefill and decode engine configurations, and produces a ready-to-deploy DGD YAML.

The profiler accepts a DynamoGraphDeploymentRequestSpec (DGDR) as input and uses AI Configurator (AIC) for performance simulation, candidate enumeration, and configuration picking. When the planner is enabled, the profiler additionally generates engine interpolation curves used for runtime autoscaling.

Workflow

What model you want to deploy (model)
How it should perform (SLA targets: sla.ttft, sla.itl)
Where it should run (optional GPU preferences via hardware)
Which backend to use (backend: auto, vllm, sglang, or trtllm)
Which image to use (image)

The profiler follows this pipeline:

flowchart TD
    Input["DGDR Spec"] --> Validate["Validate + Gate Checks"]
    Validate --> Strategy{searchStrategy?}

    Strategy -->|rapid| AICCheck{"AIC supports\nmodel/hw/backend?"}
    Strategy -->|thorough| Enumerate["Enumerate candidates\nvia AIC"]

    AICCheck -->|yes| Simulate["AIC Simulation\n+ Picking"]
    AICCheck -->|no| Naive["Naive Config\nGeneration"]

    Enumerate --> Deploy["Deploy + Benchmark\neach candidate"]
    Deploy --> Pick["AIC Picking"]

    Simulate --> DGDGen["DGD Generation"]
    Pick --> DGDGen
    Naive --> DGDGen

    DGDGen --> Interpolation["Interpolation\nCurves"]

    Interpolation --> MockerCheck{mocker?}
    MockerCheck -->|yes| MockerBase["generate_mocker_config()"]
    MockerCheck -->|no| PlannerCheck
    MockerBase --> PlannerCheck{planner?}
    PlannerCheck -->|yes| AddPlanner["add_planner_to_config()"]
    PlannerCheck -->|no| ProfileCheck
    AddPlanner --> ProfileCheck{"needs profile data?\n(mocker or throughput\nplanner enabled)"}
    ProfileCheck -->|yes| AddProfile["add_profile_data_to_config()"]
    ProfileCheck -->|no| Final
    AddProfile --> Final["final_config.yaml"]

Stage-by-stage walkthrough

Validation: The DGDR spec is validated — required fields checked (image, hardware.gpuSku, hardware.numGpusPerNode), SLA targets verified, and gate checks applied (see Gate Checks).
Search Strategy: The profiler branches based on searchStrategy:
- Rapid: Uses AIC simulation to estimate performance across parallelization configs. No GPUs needed, completes in ~30 seconds.
- Thorough: Enumerates candidate parallelization configs via AIC, deploys each on real GPUs, benchmarks with AIPerf, then picks the best. Takes 2-4 hours, disagg mode only.
Picking: The profiler selects the best configuration using one of three modes, determined automatically from the DGDR spec (see Picking Modes).
DGD Generation: The picked configuration is rendered into a complete DGD YAML via AIC's generator pipeline, including correct parallelization, replica counts, container image, and PVC mounts.
Interpolation (throughput planner/mocker): When the planner is enabled, the profiler generates detailed performance interpolation curves — TTFT vs ISL for prefill, ITL vs KV-cache utilization for decode. These are stored as NPZ files and later packaged into a ConfigMap during final assembly.
Final Assembly (3 composable layers):
1. Mocker base: If mocker is enabled, the base DGD is swapped for the mocker DGD template (generate_mocker_config). Otherwise the AIC-picked DGD is kept.
2. Planner service: If the planner is enabled, the Planner pod and its planner-config ConfigMap are injected into the DGD (add_planner_to_config).
3. Profile data: If mocker is enabled or planner throughput-based scaling is enabled, the interpolation data ConfigMap is created and mounted into all consumers — the Planner service and/or mocker workers (add_profile_data_to_config).
The result is written to final_config.yaml.

Search Strategies

Rapid

Uses AIC's performance simulation to estimate optimal configurations without deploying real engines. Completes in ~30 seconds.

searchStrategy: rapid

Supports all backends: vLLM, SGLang, TensorRT-LLM
If the model/hardware/backend combination is not supported by AIC, falls back to a naive config (memory-fit TP calculation)
No GPU resources consumed during profiling

Thorough

Enumerates candidate parallelization configs, deploys each as a real K8s workload, and benchmarks with AIPerf.

searchStrategy: thorough

Only disaggregated mode is supported
Does not support auto backend — specify vllm, sglang, or trtllm
Takes 2-4 hours depending on the number of candidates
Provides highest accuracy since measurements come from real hardware

Picking Modes

The profiler automatically selects a picking mode based on the DGDR spec:

Autoscale

Triggered when the planner is enabled (scaling enabled in features.planner). Picks prefill and decode engines independently, each with 1 replica. The planner handles scaling at runtime.

Load Match

Triggered when a target load is specified (workload.requestRate or workload.concurrency). Finds the configuration that serves the target load with the minimum number of GPUs under SLA.

workload:
  requestRate: 5.0   # target 5 req/s

Default

Triggered when there is no planner and no target load. Maximizes throughput for the available GPU budget under SLA.

Planner Integration

When the planner is enabled, the profiler generates engine interpolation data needed for throughput-based autoscaling. The pre_deployment_sweeping_mode field controls how this data is produced:

features:
  planner:
    pre_deployment_sweeping_mode: rapid   # rapid | thorough | none
    enable_throughput_scaling: true

rapid: Uses AIC simulation to generate interpolation curves (~30s, no GPUs)
thorough: Deploys the selected engine config on real GPUs and sweeps across ISL/concurrency ranges (2-4h)
none: Skips interpolation. Only valid when using load-based scaling without throughput-based scaling.

The profiler saves two ConfigMaps into the generated DGD:

planner-config-XXXX: Serialized PlannerConfig JSON (with profile_results_dir pointing to the profiling data mount)
planner-profile-data-XXXX: Prefill and decode interpolation data (JSON)

See the Planner Guide for the full PlannerConfig reference.

Mocker

When features.mocker.enabled: true, the profiler outputs a mocker DGD that simulates engine behavior without real GPUs. This is useful for testing planner behavior and validating configurations at scale.

Mocker requires pre-deployment sweeping to generate simulated performance profiles — pre_deployment_sweeping_mode cannot be none when mocker is enabled.

Gate Checks and Constraints

The profiler enforces these rules at startup:

Condition	Behavior
`searchStrategy: thorough` + `backend: auto`	Rejected. Specify a concrete backend.
AIC unsupported + `enable_throughput_scaling: true`	Rejected. Throughput planner requires AIC support.
AIC unsupported + `pre_deployment_sweeping_mode: rapid`	Falls back to `none` with a warning.
`e2eLatency` provided without `ttft: null, itl: null`	Rejected by SLA validator. When using `e2eLatency`, explicitly null out `ttft` and `itl`.
SLA unachievable	Warning logged, SLA updated to best achievable value.
Load-match needs more GPUs than available	Warning logged.

Support Matrix

Backend	Dense Models	MoE Models
vLLM	✅	🚧
SGLang	✅	✅
TensorRT-LLM	✅	🚧

The profiler sweeps over the following parallelization mappings for prefill and decode:

Model Architecture	Prefill Parallelization Mapping	Decode Parallelization Mapping
MLA+MoE (DeepseekV3ForCausalLM, DeepseekV32ForCausalLM)	TEP, DEP	TEP, DEP
GQA+MoE (Qwen3MoeForCausalLM)	TP, TEP, DEP	TP, TEP, DEP
Other Models	TP	TP

Note

Exact model x parallelization mapping support is dependent on the backend. The profiler does not guarantee that the recommended P/D engine configuration is supported and bug-free by the backend.

Deployment

Kubernetes Deployment (DGDR)

The recommended deployment method is through DGDRs. Sample configurations are provided in components/src/dynamo/profiler/deploy/:

Sample	Description
`profile_sla_dgdr.yaml`	Standard online profiling with AIPerf
`profile_sla_aic_dgdr.yaml`	Fast offline profiling with AI Configurator
`profile_sla_moe_dgdr.yaml`	MoE model profiling (SGLang)

Container Images

Each DGDR requires a container image for profiling and deployment:

image (Optional): Container image for the profiling job. Must contain the profiler code and dependencies.

spec:
  image: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.9.0"

Quick Start: Deploy with DGDR

Step 1: Create Your DGDR

Use a sample configuration or create your own:

apiVersion: nvidia.com/v1beta1
kind: DynamoGraphDeploymentRequest
metadata:
  name: my-model-profiling
spec:
  model: "Qwen/Qwen3-0.6B"
  backend: vllm
  image: "nvcr.io/nvidia/ai-dynamo/dynamo-frontend:1.0.0"

Step 2: Apply the DGDR

export NAMESPACE=your-namespace
kubectl apply -f my-profiling-dgdr.yaml -n $NAMESPACE

Step 3: Monitor Progress

# View status
kubectl get dgdr -n $NAMESPACE

# Detailed status
kubectl describe dgdr my-model-profiling -n $NAMESPACE

# Watch profiling job logs
kubectl logs -f job/profile-my-model-profiling -n $NAMESPACE

DGDR Status Phases:

Pending: Initial state, preparing to profile
Profiling: Running profiling job (20-30 seconds for AIC, 2-4 hours for online)
Ready: Profiling complete, generated DGD spec available in status
Deploying: Generating and applying DGD configuration
Deployed: DGD successfully deployed and running
Failed: Error occurred (check events for details)

Step 4: Access Your Deployment

# Find the frontend service
kubectl get svc -n $NAMESPACE | grep frontend

# Port-forward to access locally
kubectl port-forward svc/<deployment>-frontend 8000:8000 -n $NAMESPACE

# Test the endpoint
curl http://localhost:8000/v1/models

Note

DGDRs are immutable. To update SLAs or configuration, delete the existing DGDR and create a new one.

Profiling Method

The profiler follows a 5-step process:

Hardware Setup: Uses defaults or user-specified hardware configuration. Optionally, cluster-scoped operators can enable automatic GPU discovery to detect specifications from cluster nodes.
Identify Sweep Ranges: Automatically determine minimum and maximum number of GPUs per engine. Minimum is determined by the model size and GPU VRAM. Maximum is set to one node for dense models and 4 nodes for MoE models.
Parallelization Mapping Sweep: Test performance of engines with different parallelization mappings using the input ISL and OSL.
- For dense models, test different TP sizes for both prefill and decode.
- For MoE models (SGLang), evaluate both TEP and DEP as candidates for prefill and decode.
- Prefill:
  - TP/TEP: Measure TTFT with batch size = 1 (assuming ISL is long enough to saturate compute) without KV reuse.
  - DEP: Attention uses data parallelism. Send a single burst with total concurrency attention_dp_size × attn_dp_num_req_ratio (defaults to 4) and compute the reported TTFT as time_to_first_token.max / attn_dp_num_req_ratio from the AIPerf summary of that burst.
- Decode: Measure the ITL under different numbers of in-flight requests, from 1 to the maximum the KV cache can hold. To measure ITL without being affected by piggy-backed prefill requests, the script enables KV-reuse and warms up the engine by issuing the same prompts before measuring.
Recommendation: Select optimal parallelization mapping for prefill and decode that achieves the highest per-GPU throughput while adhering to the SLA on TTFT and ITL.
In-Depth Profiling on the Recommended P/D Engine: Interpolate TTFT with ISL and ITL with active KV cache and decode context length for more accurate performance estimation.
- Prefill: Measures TTFT and throughput per GPU across different input lengths with batch size=1.
- Decode: Measures ITL and throughput per GPU under various KV cache loads and decode context lengths.

AIPerf on Real Engines

Profiles your model by creating real test deployments in Kubernetes and measuring their performance.

Duration: 2-4 hours
Accuracy: Highest (real measurements)
GPU Requirements: Full access to test different parallelization mappings
Backends: vLLM, SGLang, TensorRT-LLM

AIPerf-based profiling is the default behavior. Use searchStrategy: thorough for comprehensive real-engine profiling:

spec:
  searchStrategy: thorough  # Deep exploration with real engine profiling

AI Configurator Simulation

Uses performance simulation to rapidly estimate optimal configurations without running real deployments.

Duration: 20-30 seconds
Accuracy: Estimated (may have errors for unusual configurations)
GPU Requirements: None
Backends: TensorRT-LLM only (vLLM/SGLang coming soon)

AI Configurator is used by default with searchStrategy: rapid:

spec:
  searchStrategy: rapid  # Fast profiling with AI Configurator simulation (default)

Note

aicBackendVersion specifies the TensorRT-LLM version that AI Configurator simulates. See the AI Configurator supported features for available versions.

Currently supports:

Backends: TensorRT-LLM (versions 0.20.0, 1.0.0rc3, 1.0.0rc6)
Systems: H100 SXM, H200 SXM, B200 SXM, GB200 SXM, A100 SXM
Models: Wide range including GPT, Llama, Mixtral, DeepSeek, Qwen, and more

See AI Configurator documentation for the full list.

Automatic GPU Discovery

The operator automatically discovers GPU resources from cluster nodes, providing hardware info (GPU model, VRAM, GPUs per node) and automatic profiling search space calculation.

Requirements:

Cluster-scoped operators: Have node read permissions by default
Namespace-scoped operators: GPU discovery is enabled by default when installing via Helm — the chart provisions the required ClusterRole/ClusterRoleBinding automatically

For namespace-scoped operators, GPU discovery is controlled by a Helm value:

# GPU discovery enabled (default) — Helm provisions read-only node access automatically
helm install dynamo-platform ... --set dynamo-operator.gpuDiscovery.enabled=true

# GPU discovery disabled — you must provide hardware config manually in each DGDR
helm install dynamo-platform ... --set dynamo-operator.gpuDiscovery.enabled=false

If GPU discovery is disabled, provide hardware config manually in the DGDR:

spec:
  hardware:
    numGpusPerNode: 8
    gpuSku: "H100-SXM5-80GB"
    vramMb: 81920

If GPU discovery is disabled and no manual hardware config is provided, the DGDR will be rejected at admission time.

Configuration

DGDR Configuration Structure

All profiler configuration is provided through the v1beta1 DGDR spec fields:

apiVersion: nvidia.com/v1beta1
kind: DynamoGraphDeploymentRequest
metadata:
  name: my-deployment
spec:
  model: "Qwen/Qwen3-0.6B"
  backend: vllm
  image: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.9.0"

  searchStrategy: rapid  # or thorough
  autoApply: true

  workload: { ... }
  sla: { ... }
  hardware: { ... }
  features: { ... }
  overrides: { ... }

SLA Configuration (Optional)

workload:
  isl: 3000      # Average input sequence length (tokens)
  osl: 150       # Average output sequence length (tokens)

sla:
  ttft: 200.0    # Target Time To First Token (milliseconds)
  itl: 20.0      # Target Inter-Token Latency (milliseconds)

ISL/OSL: Based on your expected traffic patterns
TTFT: First token latency target (lower = more GPUs needed, affects prefill engine)
ITL: Token generation latency target (lower = more GPUs needed, affects decode engine)
Trade-offs: Tighter SLAs require more GPU resources

Hardware Configuration (Optional)

hardware:
  gpuSku: h200_sxm            # GPU SKU identifier (auto-detected)
  vramMb: 81920               # VRAM per GPU in MiB
  totalGpus: 16               # Total GPUs available in the cluster
  numGpusPerNode: 8           # GPUs per node (for multi-node MoE)

numGpusPerNode: Determine the upper bound of GPUs per node for dense models and configure Grove for multi-node MoE engines
gpuSku: GPU SKU identifier, auto-detected by the controller

Tip

If you don't specify hardware constraints, the controller auto-detects based on your model size and available cluster resources.

Search Strategy (Optional)

Controls the profiling search depth:

spec:
  searchStrategy: rapid   # "rapid" (default) for fast sweep; "thorough" for deeper exploration

rapid: Performs a fast sweep over parallelization mappings (default)
thorough: Explores more configurations for potentially better results

Planner Configuration (Optional)

Pass arguments to the SLA planner via the features section:

features:
  planner:
    planner_min_endpoint: 2                    # Minimum endpoints to maintain
    planner_adjustment_interval: 60            # Adjustment interval (seconds)
    planner_load_predictor: linear             # Load prediction method

Note

Planner arguments use planner_ prefix. See SLA Planner documentation for full list.

Model Cache PVC (Advanced)

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

modelCache:
  pvcName: "model-cache"
  pvcModelPath: "hub/models--deepseek-ai--DeepSeek-R1"
  pvcMountPath: "/opt/model-cache"

Requirements:

The PVC must exist in the same namespace as the DGDR
The model weights must be accessible at {mountPath}/{pvcPath}

Engine Configuration (Auto-configured)

The controller automatically handles model and backend configuration from high-level fields:

# You specify:
spec:
  model: "Qwen/Qwen3-0.6B"
  backend: vllm

# Controller auto-injects into the profiling job

You should not manually set model or backend in profiling config overrides.

Using Existing DGD Configs

Provide a base DGD config via the overrides section:

overrides:
  dgd:
    apiVersion: nvidia.com/v1alpha1
    kind: DynamoGraphDeployment
    metadata:
      name: my-dgd
    spec:
      # ... your base DGD spec

The profiler uses the DGD config as a base template, then optimizes it based on your SLA targets.

Integration

With SLA Planner

The Profiler generates interpolation data that the SLA Planner uses for autoscaling decisions.

Prefill Interpolation (selected_prefill_interpolation/raw_data.npz):

prefill_isl: 1D array of input sequence lengths tested
prefill_ttft: 1D array of TTFTs (ms) at each ISL
prefill_thpt_per_gpu: 1D array of throughput (tokens/s/GPU) at each ISL

Decode Interpolation (selected_decode_interpolation/raw_data.npz):

max_kv_tokens: Total KV tokens capacity in decode engine
x_kv_usage: 1D array of active KV usage percentages [0, 1]
y_context_length: 1D array of average context lengths tested
z_itl: 1D array of ITLs (ms) at each (KV usage, context length) point
z_thpt_per_gpu: 1D array of throughput (tokens/s/GPU) at each point

With Dynamo Operator

When using DGDR, the Dynamo Operator:

Creates profiling jobs automatically
Stores profiling data in ConfigMaps (planner-profile-data)
Generates optimized DGD configurations
Deploys the DGD with SLA Planner integration

The generated DGD is tracked via labels:

metadata:
  labels:
    dgdr.nvidia.com/name: my-deployment
    dgdr.nvidia.com/namespace: your-namespace

With Observability

Monitor profiling jobs:

kubectl logs -f job/profile-<dgdr-name> -n $NAMESPACE
kubectl describe dgdr <name> -n $NAMESPACE

Advanced Topics

Manual Deployment Control

Disable auto-deployment to review the generated DGD before applying:

spec:
  autoApply: false

Then manually extract and apply:

# Extract generated DGD from DGDR status
kubectl get dgdr my-deployment -n $NAMESPACE -o jsonpath='{.status.profilingResults.selectedConfig}' | kubectl apply -f -

# Or save to file for review
kubectl get dgdr my-deployment -n $NAMESPACE -o jsonpath='{.status.profilingResults.selectedConfig}' > my-dgd.yaml

Mocker Deployment

Deploy a mocker deployment that simulates engines without GPUs:

spec:
  model: <model-name>
  backend: trtllm
  features:
    mocker:
      enabled: true    # Deploy mocker instead of real backend
  autoApply: true

Profiling still runs against the real backend to collect performance data. The mocker uses this data to simulate realistic timing behavior. Useful for large-scale experiments, testing Planner behavior, and validating configurations.

Accessing Profiling Artifacts

By default, profiling data is stored in ConfigMaps. For detailed artifacts (plots, logs, raw data), attach a PVC via overrides:

overrides:
  profilingJob:
    template:
      spec:
        volumes:
        - name: profiling-output
          persistentVolumeClaim:
            claimName: "dynamo-pvc"

ConfigMaps (always created):

dgdr-output-<name>: Generated DGD configuration
planner-profile-data: Profiling data for Planner (JSON)

PVC artifacts (optional):

Performance plots (PNGs)
DGD configurations for each profiled deployment
AIPerf profiling artifacts
Raw profiling data (.npz files)
Profiler logs

Access PVC 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

Output Performance Plots

The profiler generates plots to visualize performance data:

Parallelization Mapping Sweep Plots:

prefill_performance.png: TTFT vs Parallelization Mapping size
decode_performance.png: ITL vs Parallelization Mapping size and in-flight requests

In-Depth Profiling Plots:

selected_prefill_interpolation/prefill_ttft_interpolation.png: TTFT vs ISL
selected_prefill_interpolation/prefill_throughput_interpolation.png: Throughput vs ISL
selected_decode_interpolation/decode_itl_interplation.png: ITL vs KV usage and context length
selected_decode_interpolation/decode_throughput_interpolation.png: Throughput vs KV usage and context length

Runtime Profiling (SGLang)

SGLang workers expose profiling endpoints for runtime performance analysis:

# Start profiling
curl -X POST http://localhost:9090/engine/start_profile \
  -H "Content-Type: application/json" \
  -d '{"output_dir": "/tmp/profiler_output"}'

# Run inference requests...

# Stop profiling
curl -X POST http://localhost:9090/engine/stop_profile

View traces using Chrome's chrome://tracing, Perfetto UI, or TensorBoard.

Troubleshooting

SLA Cannot Be Met

The profiler logs a warning and updates the SLA to the best achievable value. To improve results:

Relax SLA targets (increase TTFT/ITL)
Add more GPU resources
Try a different backend
Use a smaller or quantized model

Profiling Takes Too Long

Use searchStrategy: rapid for ~30s profiling
Reduce interpolation granularity
Reduce the GPU search space via hardware constraints

Out of Memory During Profiling

Reduce max_batch_size in engine config
Skip larger TP configurations by constraining hardware
Use a quantized model variant

Image Pull Errors

Ensure image pull secrets are configured in your namespace for the container registry.

FilesExpand file tree

profiler-guide.md

Latest commit

History

profiler-guide.md

File metadata and controls

Overview

Workflow

Stage-by-stage walkthrough

Search Strategies

Rapid

Thorough

Picking Modes

Autoscale

Load Match

Default

Planner Integration

Mocker

Gate Checks and Constraints

Support Matrix

Deployment

Kubernetes Deployment (DGDR)

Container Images

Quick Start: Deploy with DGDR

Profiling Method

AIPerf on Real Engines

AI Configurator Simulation

Automatic GPU Discovery

Configuration

DGDR Configuration Structure

SLA Configuration (Optional)

Hardware Configuration (Optional)

Search Strategy (Optional)

Planner Configuration (Optional)

Model Cache PVC (Advanced)

Engine Configuration (Auto-configured)

Using Existing DGD Configs

Integration

With SLA Planner

With Dynamo Operator

With Observability

Advanced Topics

Manual Deployment Control

Mocker Deployment

Accessing Profiling Artifacts

Output Performance Plots

Runtime Profiling (SGLang)

Troubleshooting

SLA Cannot Be Met

Profiling Takes Too Long

Out of Memory During Profiling

Image Pull Errors

See Also