Deploying Inference Using NVIDIA Dynamo and vLLM

NVIDIA Dynamo is an open-source, high-throughput, low-latency inference framework for deploying large-scale generative AI and reasoning models across multi-node, multi-GPU environments. It boosts LLM inference efficiency and reduces time-to-first-token (TTFT) through intelligent resource management, dynamic GPU allocation, and distributed orchestration, and it supports disaggregated serving — separating prompt processing (prefill) and response generation (decode) across different GPUs so each phase can be optimized independently. Dynamo integrates with backends including vLLM, SGLang, and NVIDIA TensorRT-LLM. This guide deploys NVIDIA Dynamo with the vLLM backend, covering infrastructure setup, container deployment, and two serving patterns: aggregated serving for single-GPU configurations and disaggregated serving for multi-GPU setups with independent prefill and decode workers. By the end, you'll have a running Dynamo + vLLM deployment serving chat completion requests in both aggregated and disaggregated configurations. Prerequisites Access to a GPU-enabled server with NVIDIA GPUs installed and the NVIDIA Container Toolkit configured, using a non-root user with sudo privileges. 1 GPU minimum for aggregated serving. 4 GPUs for disaggregated serving. Docker Engine and Docker Compose installed. A Hugging Face account with an access token generated for gated models like Llama. Key Components NVIDIA Dynamo is the orchestration layer, managing GPU resources, routing requests, and coordinating between workers. It includes a frontend service that receives inference requests, a smart router that directs traffic based on KV cache awareness, and a GPU planner that dynamically adjusts resource allocation. vLLM is the inference backend that executes model computations on GPUs. Prefill workers process incoming prompts and generate initial tokens, while decode workers handle sequential token generation. The vLLM backend integrates with Dynamo through metrics reporting and KV cache event publishing. etcd provides distributed service discovery so Dynamo components can locate and communicate with each other, maintaining a registry of active workers and their capabilities. NATS handles message passing between components, particularly KV cache events — prefill workers publish KV cache information through NATS so the router can make intelligent request-placement decisions. NIXL (NVIDIA Inter-GPU Exchange Library) manages efficient data transfer between GPUs during disaggregated serving, letting prefill workers transfer KV cache data to decode workers with minimal latency. 1. Clone the Dynamo Repository NVIDIA Dynamo provides deployment scripts, container utilities, and orchestration modules required to run inference workloads. 1. Clone the repository: $ git clone https://github.com/ai-dynamo/dynamo.git 2. Navigate to the repository directory: $ cd dynamo 3. Switch to the latest stable release: $ git checkout release/0.9.0 Visit the Dynamo releases page to find the latest stable release version. 2. Start Infrastructure Services Dynamo's distributed architecture relies on etcd for worker registry and service discovery, and NATS for KV cache event propagation between prefill and decode workers. The Docker Compose configuration launches both with exposed ports for client connections (etcd: 2379-2380, NATS: 4222, 6222, 8222). These services must run continuously for Dynamo to coordinate worker resources and route inference requests. 1. Start the infrastructure services: $ docker compose -f deploy/docker-compose.yml up -d 2. Verify the services are running: $ docker compose -f deploy/docker-compose.yml ps The output displays the running etcd and NATS containers: NAME IMAGE COMMAND SERVICE CREATED STATUS PORTS deploy-etcd-server-1 bitnamilegacy/etcd:3.6.1 "/opt/bitnami/script…" etcd-server 6 seconds ago Up 6 seconds 0.0.0.0:2379-2380->2379-2380/tcp, [::]:2379-2380->2379-2380/tcp deploy-nats-server-1 nats:2.11.4 "/nats-server -c /et…" nats-server 6 seconds ago Up 6 seconds 0.0.0.0:4222->4222/tcp, [::]:4222->4222/tcp, 0.0.0.0:6222->6222/tcp, [::]:6222->6222/tcp, 0.0.0.0:8222->8222/tcp, [::]:8222->8222/tcp 3. Verify CUDA Version and Pull the Container Image The vLLM container requires a CUDA version match between the host driver and container runtime to prevent GPU kernel incompatibilities. 1. Check the installed CUDA version: $ nvidia-smi The output displays the CUDA version in the top-right corner of the table: +-----------------------------------------------------------------------------------------+ | NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 | +-----------------------------------------+------------------------+----------------------+ .... 2. Pull the vLLM container image from NGC, matching the image tag to your system's CUDA version: For CUDA 13.x: $ docker pull nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.9.0-cuda13 For CUDA 12.x: $ docker pull nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.9.0 Visit the NVIDIA NGC Catalog to view all available image tags and CUDA versions. 3. (Optional) Build the container from source instead of pulling the pre-built image: $ ./container/build.sh --framework VLLM This creates an image named dynamo:latest-vllm. If you use this locally built image, replace nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.9.0-cuda13 with dynamo:latest-vllm in all subsequent commands. 4. Configure Hugging Face Cache Permissions The container runs as UID 1000 and requires write access to the Hugging Face cache directory for model downloads. Incorrect permissions prevent the container from accessing cached model weights, causing worker initialization failures. 1. Create the cache directory if it does not exist: $ mkdir -p container/.cache/huggingface 2. Set ownership to the container user (UID 1000): $ sudo chown -R 1000:1000 container/.cache/huggingface 3. Set appropriate permissions: $ sudo chmod -R 775 container/.cache/huggingface 5. Deploy Aggregated Serving Aggregated serving combines prefill and decode phases on a single worker, eliminating inter-GPU KV cache transfers and reducing request latency. This suits single-GPU environments or workloads prioritizing response time over throughput. 1. Export your Hugging Face token to avoid rate limitations when downloading large models (replace YOUR_HF_TOKEN with your actual token): $ export HF_TOKEN=YOUR_HF_TOKEN 2. Run the vLLM container with GPU access and workspace mounting, using the image tag that matches your CUDA version: $ ./container/run.sh -it --framework VLLM --mount-workspace --image nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.9.0-cuda13 -e HF_TOKEN=$HF_TOKEN 3. Inside the container, create a custom launch script for aggregated serving with the NVIDIA Nemotron model: $ cat << 'EOF' > ~/nemotron_agg.sh #!/bin/bash set -e trap 'echo Cleaning up...; kill 0' EXIT # Set deterministic hash for KV event IDs export PYTHONHASHSEED=0 # Model configuration - use command line argument or default MODEL="${1:-nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1}" echo "Starting Dynamo Frontend..." python -m dynamo.frontend & echo "Starting vLLM Worker with model: $MODEL" DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT:-8081} \ python -m dynamo.vllm \ --model "$MODEL" \ --trust-remote-code \ --enforce-eager \ --connector none EOF 4. Make the script executable: $ chmod +x ~/nemotron_agg.sh 5. Run the aggregated serving script: $ ~/nemotron_agg.sh This starts a frontend service on port 8000 and a vLLM worker that loads the specified model (defaults to NVIDIA Nemotron Nano 4B). To deploy the larger NVIDIA Nemotron Super 49B model instead, pass the model name as an argument: $ ~/nemotron_agg.sh "nvidia/Llama-3_3-Nemotron-Super-49B-v1_5" The 49B model requires high-memory GPUs such as B200 or GB200 class devices. Ensure sufficient VRAM and consider tensor parallelism for production deployments. 6. Open a new terminal session on your server (outside the container) and test with a chat completion request: $ curl -X POST http://localhost:8000/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "model": "nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1", "messages": [{"role": "user", "content": "Hello! Tell me about AI."}], "max_tokens": 100 }' The output displays the model's chat response in JSON format. 6. Deploy Disaggregated Serving Disaggregated serving assigns prefill and decode phases to separate GPU workers, enabling independent scaling and optimization of each phase. Prefill workers process incoming prompts and transfer KV cache data to decode workers via NIXL. This maximizes throughput by letting multiple decode workers share prefill resources, improving GPU utilization across the cluster. 1. Exit the container if you are still inside from the previous section — press Ctrl+C to terminate the running process, then Ctrl+D to exit. 2. Export your Hugging Face token (replace YOUR_HF_TOKEN with your actual token): $ export HF_TOKEN=YOUR_HF_TOKEN 3. Run the container with the image tag that matches your CUDA version: $ ./container/run.sh -it --framework VLLM --mount-workspace --image nvcr.io/nvidia/ai-dynamo/vllm-runtime:0.9.0-cuda13 -e HF_TOKEN=$HF_TOKEN 4. Inside the container, create a custom launch script for disaggregated serving with the NVIDIA Nemotron model: $ cat << 'EOF' > ~/nemotron_disagg.sh #!/bin/bash # Kill any existing processes pkill -f "dynamo.frontend" pkill -f "dynamo.vllm" sleep 2 # Set deterministic hash for KV event IDs export PYTHONHASHSEED=0 # Model configuration - use command line argument or default MODEL="${1:-nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1}" echo "Starting Dynamo Frontend..." python -m dynamo.frontend & echo "Starting Decode Workers with model: $MODEL" # Decode Worker 1 - Uses GPU 0 CUDA_VISIBLE_DEVICES=0 python3 -m dynamo.vllm \ --model "$MODEL" \ --trust-remote-code \ --is-decode-worker \ --max-model-len 2048 & # Decode Worker 2 - Uses GPU 1 VLLM_NIXL_SIDE_CHANNEL_PORT=20097 \ CUDA_VISIBLE_DEVICES=1 python3 -m dynamo.vllm \ --model "$MODEL" \ --trust-remote-code \ --is-decode-worker \ --max-model-len 2048 & echo "Starting Prefill Workers with model: $MODEL" # Prefill Worker 1 - Uses GPU 2 CUDA_VISIBLE_DEVICES=2 python3 -m dynamo.vllm \ --model "$MODEL" \ --trust-remote-code \ --is-prefill-worker \ --max-model-len 2048 \ --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20082","enable_kv_cache_events":true}' & # Prefill Worker 2 - Uses GPU 3 VLLM_NIXL_SIDE_CHANNEL_PORT=20099 \ CUDA_VISIBLE_DEVICES=3 python3 -m dynamo.vllm \ --model "$MODEL" \ --trust-remote-code \ --is-prefill-worker \ --max-model-len 2048 \ --kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:20083","enable_kv_cache_events":true}' & echo "All services starting... waiting for initialization..." sleep 30 echo "Deployment ready!" # Keep script running wait EOF 5. Make the script executable: $ chmod +x ~/nemotron_disagg.sh 6. Run the disaggregated serving script: $ ~/nemotron_disagg.sh This starts the frontend service on port 8000, two decode workers on GPUs 0 and 1, and two prefill workers on GPUs 2 and 3 with the specified model (defaults to NVIDIA Nemotron Nano 4B). To deploy the larger NVIDIA Nemotron Super 49B model instead, pass the model name as an argument: $ ~/nemotron_disagg.sh "nvidia/Llama-3_3-Nemotron-Super-49B-v1_5" The 49B model requires high-memory GPUs such as B200 or GB200 class devices. Ensure sufficient VRAM and consider tensor parallelism for production deployments. 7. Open a new terminal session on your server (outside the container) and test with multiple sequential requests to observe worker distribution: $ for i in {1..5}; do echo "Request $i:" curl -s http://localhost:8000/v1/chat/completions \ -H "Content-Type: application/json" \ -d "{ \"model\": \"nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1\", \"messages\": [{\"role\": \"user\", \"content\": \"Test request $i\"}], \"max_tokens\": 10 }" | jq '.id' sleep 1 done Each request returns a unique ID, and the logs inside the container show which workers process each request. 8. Test with concurrent requests to verify load distribution: $ for i in {1..10}; do curl -s http://localhost:8000/v1/chat/completions \ -H "Content-Type: application/json" \ -d "{ \"model\": \"nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1\", \"messages\": [{\"role\": \"user\", \"content\": \"Concurrent test $i\"}], \"max_tokens\": 20 }" & done wait echo "All requests completed" Dynamo's router distributes the requests across available prefill and decode workers. Next Steps Explore KV-aware routing and speculative decoding in the official NVIDIA Dynamo documentation. Scale the disaggregated deployment to more GPUs or nodes for higher throughput. Deploy Dynamo on a Kubernetes cluster for production-grade orchestration and autoscaling. Benchmark aggregated vs. disaggregated serving for your specific model and traffic pattern to choose the right architecture. For the full guide with additional tips, visit the original article on Vultr Docs.

Read full story →