Disaggregated Serving Guide

AIConfigurator is a performance optimization tool that helps you find the optimal configuration for deploying LLMs with Dynamo. It automatically determines the best number of prefill and decode workers, parallelism settings, and deployment parameters to meet your SLA targets while maximizing throughput.

Why Use AIConfigurator?

When deploying LLMs with Dynamo, you need to make several critical decisions:

• Aggregated vs Disaggregated: Which architecture gives better performance for your workload?
• Worker Configuration: How many prefill and decode workers to deploy?
• Parallelism Settings: What tensor/pipeline parallel configuration to use?
• SLA Compliance: How to meet your TTFT and TPOT targets?

AIConfigurator answers these questions in seconds, providing:

• Recommended configurations that meet your SLA requirements
• Ready-to-deploy Dynamo configuration files (including Kubernetes manifests)
• Performance comparisons between different deployment strategies
• Up to 1.7x better throughput compared to manual configuration

End-to-End Workflow

Aggregated vs Disaggregated Architecture

AIConfigurator evaluates two deployment architectures and recommends the best one for your workload:

When to Use Each Architecture

Quick Start

$
# Install
$
pip3 install aiconfigurator
$
$
# Find optimal configuration for vLLM backend
$
aiconfigurator cli default \
>
  --model_path Qwen/Qwen3-32B-FP8 \
>
  --total_gpus 8 \
>
  --system h200_sxm \
>
  --backend vllm \
>
  --backend_version 0.12.0 \
>
  --isl 4000 \
>
  --osl 500 \
>
  --ttft 600 \
>
  --tpot 16.67 \
>
  --save_dir ./results_vllm
$
$
# Deploy on Kubernetes
$
kubectl apply -f ./results_vllm/agg/top1/agg/k8s_deploy.yaml

Complete Walkthrough: vLLM on H200

This section walks through a validated example deploying Qwen3-32B-FP8 on 8× H200 GPUs using vLLM.

Step 1: Run AIConfigurator

$
aiconfigurator cli default \
>
  --model_path Qwen/Qwen3-32B-FP8 \
>
  --system h200_sxm \
>
  --total_gpus 8 \
>
  --isl 4000 \
>
  --osl 500 \
>
  --ttft 600 \
>
  --tpot 16.67 \
>
  --backend vllm \
>
  --backend_version 0.12.0 \
>
  --save_dir ./results_vllm

Parameters explained:

• --model_path: HuggingFace model ID or local path (e.g., Qwen/Qwen3-32B-FP8)
• --system: GPU system type (h200_sxm, h100_sxm, a100_sxm)
• --total_gpus: Number of GPUs available for deployment
• --isl / --osl: Input/Output sequence lengths in tokens
• --ttft / --tpot: SLA targets - Time To First Token (ms) and Time Per Output Token (ms)
• --backend: Inference backend (vllm, trtllm, or sglang)
• --backend_version: Backend version (e.g., 0.12.0 for vLLM)
• --save_dir: Directory to save generated deployment configs

Step 2: Review the Results

AIConfigurator outputs a comparison of aggregated vs disaggregated deployment strategies:

********************************************************************************
*                     Dynamo aiconfigurator Final Results                      *
********************************************************************************
  ----------------------------------------------------------------------------
  Input Configuration & SLA Target:
    Model: Qwen/Qwen3-32B-FP8 (is_moe: False)
    Total GPUs: 8
    Best Experiment Chosen: disagg at 521.77 tokens/s/gpu
  ----------------------------------------------------------------------------
  Overall Best Configuration:
    - Best Throughput: 4,174.16 tokens/s
    - Per-GPU Throughput: 521.77 tokens/s/gpu
    - Per-User Throughput: 76.96 tokens/s/user
    - TTFT: 388.11ms
    - TPOT: 12.99ms
  ----------------------------------------------------------------------------

AIC evaluates both aggregated and disaggregated architectures and outputs ranked configurations for each:

agg Top Configurations: (Sorted by tokens/s/gpu)
+------+--------------+---------------+--------+-----------------+-------------+-------------------+----------+----------+----+
| Rank | tokens/s/gpu | tokens/s/user |  TTFT  | request_latency | concurrency | total_gpus (used) | replicas | parallel | bs |
+------+--------------+---------------+--------+-----------------+-------------+-------------------+----------+----------+----+
|  1   |    397.31    |     60.66     | 509.14 |     8734.68     | 56 (=14x4)  |    8 (8=4x2)      |    4     |  tp2pp1  | 14 |
|  2   |    349.90    |     60.98     | 412.58 |     8596.28     | 48 (=24x2)  |    8 (8=2x4)      |    2     |  tp4pp1  | 24 |
|  3   |    235.62    |     62.71     | 482.57 |     8439.41     | 32 (=32x1)  |    8 (8=1x8)      |    1     |  tp8pp1  | 32 |
+------+--------------+---------------+--------+-----------------+-------------+-------------------+----------+----------+----+
disagg Top Configurations: (Sorted by tokens/s/gpu)
+------+--------------+---------------+--------+-----------------+-------------+-------------------+----------+-------------+----------+
| Rank | tokens/s/gpu | tokens/s/user |  TTFT  | request_latency | concurrency | total_gpus (used) | replicas | (p)parallel | (d)parallel |
+------+--------------+---------------+--------+-----------------+-------------+-------------------+----------+-------------+----------+
|  1   |    521.77    |     76.96     | 388.11 |     6871.61     | 60 (=60x1)  |    8 (8=1x8)      |    1     |   tp2pp1    |  tp4pp1  |
|  2   |    521.77    |     63.29     | 388.11 |     8272.31     | 80 (=40x2)  |    8 (8=2x4)      |    2     |   tp2pp1    |  tp2pp1  |
|  3   |    260.89    |     62.81     | 388.11 |     8332.18     | 42 (=42x1)  |    8 (8=1x8)      |    1     |   tp2pp1    |  tp1pp1  |
+------+--------------+---------------+--------+-----------------+-------------+-------------------+----------+-------------+----------+

Reading the output:

• tokens/s/gpu: Overall throughput efficiency — higher is better
• tokens/s/user: Per-request generation speed (inverse of TPOT)
• TTFT: Predicted time to first token
• concurrency: Total concurrent requests across all replicas (e.g., 56 (=14x4) means batch size 14 × 4 replicas)
• agg Rank 1 recommends TP2 with 4 replicas — simpler to deploy
• disagg Rank 1 recommends 2 prefill workers (TP2) + 1 decode worker (TP4) — higher throughput but requires RDMA

Step 3: Deploy on Kubernetes

The --save_dir generates ready-to-use Kubernetes manifests:

results_vllm/
├── agg/
│   └── top1/
│       └── agg/
│           ├── k8s_deploy.yaml    # Kubernetes DynamoGraphDeployment
│           └── agg_config.yaml    # Engine configuration
├── disagg/
│   └── top1/
│       └── disagg/
│           ├── k8s_deploy.yaml
│           ├── prefill_config.yaml
│           └── decode_config.yaml
└── pareto_frontier.png

Prerequisites

Before deploying, ensure you have:

1. HuggingFace Token Secret (for gated models):

   $
   kubectl create secret generic hf-token-secret \
   >
     -n your-namespace \
   >
     --from-literal=HF_TOKEN="your-huggingface-token"

2. Model Cache PVC (recommended for faster restarts):

   1
   apiVersion: v1
   2
   kind: PersistentVolumeClaim
   3
   metadata:
   4
     name: model-cache
   5
     namespace: your-namespace
   6
   spec:
   7
     accessModes:
   8
       - ReadWriteMany
   9
     resources:
   10
       requests:
   11
         storage: 100Gi

Deploy the Configuration

The generated k8s_deploy.yaml provides a starting point. You’ll typically need to customize it for your environment:

$
kubectl apply -f ./results_vllm/agg/top1/agg/k8s_deploy.yaml

Complete deployment example with model cache and production settings:

1
apiVersion: nvidia.com/v1alpha1
2
kind: DynamoGraphDeployment
3
metadata:
4
  name: dynamo-agg
5
  namespace: your-namespace
6
spec:
7
  backendFramework: vllm
8
  pvcs:
9
    - name: model-cache
10
      create: false           # Use existing PVC
11
  services:
12
    Frontend:
13
      componentType: frontend
14
      replicas: 1
15
      volumeMounts:
16
        - name: model-cache
17
          mountPoint: /opt/models
18
      envs:
19
        - name: HF_HOME
20
          value: /opt/models
21
      extraPodSpec:
22
        mainContainer:
23
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.8.0
24
          imagePullPolicy: IfNotPresent
25
26
    VLLMWorker:
27
      envFromSecret: hf-token-secret
28
      componentType: worker
29
      replicas: 4
30
      resources:
31
        limits:
32
          gpu: "2"
33
      sharedMemory:
34
        size: 16Gi            # Required for vLLM
35
      volumeMounts:
36
        - name: model-cache
37
          mountPoint: /opt/models
38
      envs:
39
        - name: HF_HOME
40
          value: /opt/models
41
      extraPodSpec:
42
        mainContainer:
43
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.8.0
44
          workingDir: /workspace
45
          imagePullPolicy: IfNotPresent
46
          command:
47
            - python3
48
            - -m
49
            - dynamo.vllm
50
          args:
51
            - --model
52
            - "Qwen/Qwen3-32B-FP8"
53
            - "--no-enable-prefix-caching"
54
            - "--tensor-parallel-size"
55
            - "2"
56
            - "--pipeline-parallel-size"
57
            - "1"
58
            - "--data-parallel-size"
59
            - "1"
60
            - "--kv-cache-dtype"
61
            - "fp8"
62
            - "--max-model-len"
63
            - "6000"
64
            - "--max-num-seqs"
65
            - "1024"

Key deployment settings:

Step 4: Validate with AIPerf

After deployment, validate the predictions against actual performance using AIPerf.

Deriving AIPerf Parameters from AIC Output

To use AIPerf to benchmark an AIC-recommended configuration, you’ll need to translate AIC parameters into AIPerf profiling arguments (we are working to automate this):

Note on concurrency: AIC reports concurrency as total (=bs × replicas). When benchmarking through the frontend (which routes to all replicas), use the total value. If benchmarking a single replica directly, use the per-replica bs value instead.

1
apiVersion: batch/v1
2
kind: Job
3
metadata:
4
  name: aiperf-benchmark
5
  namespace: your-namespace
6
spec:
7
  template:
8
    spec:
9
      restartPolicy: Never
10
      containers:
11
      - name: aiperf
12
        image: python:3.10
13
        command:
14
        - /bin/bash
15
        - -c
16
        - |
17
          pip install aiperf
18
          aiperf profile \
19
            -m Qwen/Qwen3-32B-FP8 \
20
            --endpoint-type chat \
21
            -u http://dynamo-agg-frontend:8000 \
22
            --isl 4000 --isl-stddev 0 \
23
            --osl 500 --osl-stddev 0 \
24
            --num-requests 800 \
25
            --concurrency 56 \
26
            --streaming \
27
            --extra-inputs "ignore_eos:true" \
28
            --num-warmup-requests 40 \
29
            --ui-type simple

$
kubectl apply -f aiperf-job.yaml
$
kubectl logs -f -l job-name=aiperf-benchmark

Validated results (Qwen3-32B-FP8, 8× H200, TP2×4 replicas, aggregated):

Fine-Tuning Your Deployment

AIConfigurator provides a strong starting point. Here’s how to iterate for production:

Adjusting for Actual Workload

If your real workload differs from the benchmark parameters:

$
# For longer outputs (chat/code generation):
$
# increase OSL, relax TTFT target
$
aiconfigurator cli default \
>
  --model_path Qwen/Qwen3-32B-FP8 \
>
  --total_gpus 8 \
>
  --system h200_sxm \
>
  --backend vllm \
>
  --backend_version 0.12.0 \
>
  --isl 2000 \
>
  --osl 2000 \
>
  --ttft 1000 \
>
  --tpot 10 \
>
  --save_dir ./results_long_output

Exploring Alternative Configurations

Use exp mode to compare custom configurations:

1
# custom_exp.yaml
2
exps:
3
  - exp_tp2
4
  - exp_tp4
5
6
exp_tp2:
7
  mode: "patch"
8
  serving_mode: "agg"
9
  model_path: "Qwen/Qwen3-32B-FP8"
10
  total_gpus: 8
11
  system_name: "h200_sxm"
12
  backend_name: "vllm"
13
  backend_version: "0.12.0"
14
  isl: 4000
15
  osl: 500
16
  ttft: 600
17
  tpot: 16.67
18
  config:
19
    agg_worker_config:
20
      tp_list: [2]
21
22
exp_tp4:
23
  mode: "patch"
24
  serving_mode: "agg"
25
  model_path: "Qwen/Qwen3-32B-FP8"
26
  total_gpus: 8
27
  system_name: "h200_sxm"
28
  backend_name: "vllm"
29
  backend_version: "0.12.0"
30
  isl: 4000
31
  osl: 500
32
  ttft: 600
33
  tpot: 16.67
34
  config:
35
    agg_worker_config:
36
      tp_list: [4]

$
aiconfigurator cli exp --yaml_path custom_exp.yaml --save_dir ./results_custom

Critical: Disaggregated deployments require RDMA for KV cache transfer. Without RDMA, performance degrades by 40x (TTFT increases from 355ms to 10+ seconds). See the Disaggregated Deployment section below.

Deploying Disaggregated (RDMA Required)

Disaggregated deployments transfer KV cache between prefill and decode workers. Without RDMA, this transfer becomes a severe bottleneck, causing 40x performance degradation.

Prerequisites for Disaggregated

1. RDMA-capable network (InfiniBand or RoCE)
2. RDMA device plugin installed on the cluster (provides rdma/ib resources)
3. ETCD and NATS deployed (for coordination)

Disaggregated DGD with RDMA

1
apiVersion: nvidia.com/v1alpha1
2
kind: DynamoGraphDeployment
3
metadata:
4
  name: dynamo-disagg
5
  namespace: your-namespace
6
spec:
7
  backendFramework: vllm
8
  pvcs:
9
    - name: model-cache
10
      create: false
11
  services:
12
    Frontend:
13
      componentType: frontend
14
      replicas: 1
15
      volumeMounts:
16
        - name: model-cache
17
          mountPoint: /opt/models
18
      envs:
19
        - name: HF_HOME
20
          value: /opt/models
21
      extraPodSpec:
22
        mainContainer:
23
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.8.0
24
          imagePullPolicy: IfNotPresent
25
26
    VLLMPrefillWorker:
27
      envFromSecret: hf-token-secret
28
      componentType: worker
29
      subComponentType: prefill
30
      replicas: 2
31
      resources:
32
        limits:
33
          gpu: "2"
34
      sharedMemory:
35
        size: 16Gi
36
      volumeMounts:
37
        - name: model-cache
38
          mountPoint: /opt/models
39
      envs:
40
        - name: HF_HOME
41
          value: /opt/models
42
        - name: UCX_TLS
43
          value: "rc_x,rc,dc_x,dc,cuda_copy,cuda_ipc"  # Enable RDMA transports
44
        - name: UCX_RNDV_SCHEME
45
          value: "get_zcopy"
46
        - name: UCX_RNDV_THRESH
47
          value: "0"
48
      extraPodSpec:
49
        mainContainer:
50
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.8.0
51
          workingDir: /workspace
52
          imagePullPolicy: IfNotPresent
53
          securityContext:
54
            capabilities:
55
              add: ["IPC_LOCK"]  # Required for RDMA memory registration
56
          resources:
57
            limits:
58
              rdma/ib: "2"      # Request RDMA resources
59
            requests:
60
              rdma/ib: "2"
61
          command: ["python3", "-m", "dynamo.vllm"]
62
          args:
63
            - --model
64
            - "Qwen/Qwen3-32B-FP8"
65
            - "--tensor-parallel-size"
66
            - "2"
67
            - "--kv-cache-dtype"
68
            - "fp8"
69
            - "--max-num-seqs"
70
            - "1"               # Prefill workers use batch size 1
71
            - --is-prefill-worker
72
73
    VLLMDecodeWorker:
74
      envFromSecret: hf-token-secret
75
      componentType: worker
76
      subComponentType: decode
77
      replicas: 1
78
      resources:
79
        limits:
80
          gpu: "4"
81
      sharedMemory:
82
        size: 16Gi
83
      volumeMounts:
84
        - name: model-cache
85
          mountPoint: /opt/models
86
      envs:
87
        - name: HF_HOME
88
          value: /opt/models
89
        - name: UCX_TLS
90
          value: "rc_x,rc,dc_x,dc,cuda_copy,cuda_ipc"
91
        - name: UCX_RNDV_SCHEME
92
          value: "get_zcopy"
93
        - name: UCX_RNDV_THRESH
94
          value: "0"
95
      extraPodSpec:
96
        mainContainer:
97
          image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.8.0
98
          workingDir: /workspace
99
          imagePullPolicy: IfNotPresent
100
          securityContext:
101
            capabilities:
102
              add: ["IPC_LOCK"]
103
          resources:
104
            limits:
105
              rdma/ib: "4"
106
            requests:
107
              rdma/ib: "4"
108
          command: ["python3", "-m", "dynamo.vllm"]
109
          args:
110
            - --model
111
            - "Qwen/Qwen3-32B-FP8"
112
            - "--tensor-parallel-size"
113
            - "4"
114
            - "--kv-cache-dtype"
115
            - "fp8"
116
            - "--max-num-seqs"
117
            - "1024"            # Decode workers handle high concurrency
118
            - --is-decode-worker

Critical RDMA settings:

Verifying RDMA is Active

After deployment, check the worker logs for UCX initialization:

$
kubectl logs <prefill-worker-pod> | grep -i "UCX\|NIXL"

You should see:

NIXL INFO Backend UCX was instantiated

If you see only TCP transports, RDMA is not active - check your RDMA device plugin and resource requests.

Tuning vLLM-Specific Parameters

Override vLLM engine parameters with --generator-set:

$
aiconfigurator cli default \
>
  --model_path Qwen/Qwen3-32B-FP8 \
>
  --total_gpus 8 \
>
  --system h200_sxm \
>
  --backend vllm \
>
  --backend_version 0.12.0 \
>
  --isl 4000 --osl 500 \
>
  --ttft 600 --tpot 16.67 \
>
  --save_dir ./results_tuned \
>
  --generator-set Workers.agg.kv_cache_free_gpu_memory_fraction=0.85 \
>
  --generator-set Workers.agg.max_num_seqs=2048

Run aiconfigurator cli default --generator-help to see all available parameters.

Prefix Caching Considerations

For workloads with repeated prefixes (e.g., system prompts):

• Enable prefix caching when you have high prefix hit rates
• Disable prefix caching (--no-enable-prefix-caching) for diverse prompts

AIConfigurator’s default predictions assume no prefix caching. Enable it post-deployment if your workload benefits.

Supported Configurations

Backends and Versions

Systems

Models

• Dense: GPT, LLAMA2/3, QWEN2.5/3
• MoE: Mixtral, DEEPSEEK_V3

Common Use Cases

$
# Strict latency SLAs (real-time chat)
$
aiconfigurator cli default \
>
  --model_path meta-llama/Llama-3.1-70B \
>
  --total_gpus 16 \
>
  --system h200_sxm \
>
  --backend vllm \
>
  --backend_version 0.12.0 \
>
  --ttft 200 --tpot 8
$
$
# High throughput (batch processing)
$
aiconfigurator cli default \
>
  --model_path Qwen/Qwen3-32B-FP8 \
>
  --total_gpus 32 \
>
  --system h200_sxm \
>
  --backend trtllm \
>
  --ttft 2000 --tpot 50
$
$
# Request latency constraint (end-to-end SLA)
$
aiconfigurator cli default \
>
  --model_path Qwen/Qwen3-32B-FP8 \
>
  --total_gpus 16 \
>
  --system h200_sxm \
>
  --backend vllm \
>
  --backend_version 0.12.0 \
>
  --request_latency 12000 \
>
  --isl 4000 --osl 500

Additional Options

$
# Web interface for interactive exploration
$
pip3 install aiconfigurator[webapp]
$
aiconfigurator webapp  # Visit http://127.0.0.1:7860
$
$
# Quick config generation (no parameter sweep)
$
aiconfigurator cli generate \
>
  --model_path Qwen/Qwen3-32B-FP8 \
>
  --total_gpus 8 \
>
  --system h200_sxm \
>
  --backend vllm
$
$
# Check model/system support
$
aiconfigurator cli support \
>
  --model_path Qwen/Qwen3-32B-FP8 \
>
  --system h200_sxm \
>
  --backend vllm

Troubleshooting

AIConfigurator Issues

Model not found: Use the full HuggingFace path (e.g., Qwen/Qwen3-32B-FP8 not QWEN3_32B)

Backend version mismatch: Check supported versions with aiconfigurator cli support --model_path <model> --system <system> --backend <backend>

Deployment Issues

Pods crash with “Permission denied” on cache directory:

• Mount the PVC at /opt/models instead of /root/.cache/huggingface
• Set HF_HOME=/opt/models environment variable
• Ensure the PVC has ReadWriteMany access mode

Workers stuck in CrashLoopBackOff:

• Check logs: kubectl logs <pod-name> --previous
• Verify sharedMemory.size is set (16Gi for vLLM, 80Gi for TRT-LLM)
• Ensure HuggingFace token secret exists and is named correctly

Model download slow on every restart:

• Add PVC for model caching (see deployment example above)
• Verify volumeMounts and HF_HOME are configured on workers

“Context stopped or killed” errors (disaggregated only):

• Deploy ETCD and NATS infrastructure (required for KV cache transfer)
• See Dynamo Kubernetes Guide for platform setup

Performance Issues

OOM errors: Reduce --max-num-seqs or increase tensor parallelism

Performance below predictions:

• Verify warmup requests are sufficient (40+ recommended)
• Check for competing workloads on the cluster
• Ensure KV cache memory fraction is optimized
• Run benchmarks from inside the cluster to eliminate network latency

Disaggregated TTFT extremely high (10+ seconds): This is almost always caused by missing RDMA configuration. Without RDMA, KV cache transfer falls back to TCP and becomes a severe bottleneck.

To diagnose:

$
# Check if RDMA resources are allocated
$
kubectl get pod <worker-pod> -o yaml | grep -A5 "resources:"
$
$
# Check UCX transport in logs
$
kubectl logs <worker-pod> | grep -i "UCX\|transport"

To fix:

1. Ensure your cluster has RDMA device plugin installed
2. Add rdma/ib resource requests to worker pods
3. Add IPC_LOCK capability to security context
4. Add UCX environment variables (see Disaggregated Deployment section)

Disaggregated working but throughput lower than aggregated: For balanced workloads (ISL/OSL ratio between 2:1 and 10:1), aggregated is often better. Disaggregated shines for:

• Very long inputs (ISL > 8000) with short outputs
• Workloads needing independent prefill/decode scaling

Learn More

• AIConfigurator CLI Guide
• Dynamo Deployment Guide
• Dynamo Installation Guide
• Benchmarking Guide