| .. | ||
| README.md | ||
NIM on Spark
Deploy a NIM on Spark
Table of Contents
Overview
Basic idea
NVIDIA NIM is containerized software for fast, reliable AI model serving and inference on NVIDIA GPUs. This playbook demonstrates how to run NIM microservices for LLMs on DGX Spark devices, enabling local GPU inference through a simple Docker workflow. You'll authenticate with NVIDIA's registry, launch the NIM inference microservice, and perform basic inference testing to verify functionality.
What you'll accomplish
You'll launch a NIM container on your DGX Spark device to expose a GPU-accelerated HTTP endpoint for text completions. While these instructions feature working with the Llama 3.1 8B NIM, additional NIM including the Qwen3-32 NIM are available for DGX Spark (see them here).
What to know before starting
- Working in a terminal environment
- Using Docker commands and GPU-enabled containers
- Basic familiarity with REST APIs and curl commands
- Understanding of NVIDIA GPU environments and CUDA
Prerequisites
- DGX Spark device with NVIDIA drivers installed
nvidia-smi - Docker with NVIDIA Container Toolkit configured, instructions here
docker run -it --gpus=all nvcr.io/nvidia/cuda:13.0.1-devel-ubuntu24.04 nvidia-smi - NGC account with API key from here
echo $NGC_API_KEY | grep -E '^[a-zA-Z0-9]{86}==' - Sufficient disk space for model caching (varies by model, typically 10-50GB)
df -h ~
Time & risk
- Estimated time: 15-30 minutes for setup and validation
- Risks:
- Large model downloads may take significant time depending on network speed
- GPU memory requirements vary by model size
- Container startup time depends on model loading
- Rollback: Stop and remove containers with
docker stop <CONTAINER_NAME> && docker rm <CONTAINER_NAME>. Remove cached models from~/.cache/nimif disk space recovery is needed. - Last Updated: 12/22/2025
- Update docker container version to cuda:13.0.1-devel-ubuntu24.04
- Add docker container permission setup instructioins
Instructions
Step 1. Verify environment prerequisites
Check that your system meets the basic requirements for running GPU-enabled containers.
nvidia-smi
docker --version
docker run --rm --gpus all nvcr.io/nvidia/cuda:13.0.1-devel-ubuntu24.04 nvidia-smi
If you see a permission denied error (something like permission denied while trying to connect to the Docker daemon socket), add your user to the docker group so that you don't need to run the command with sudo .
sudo usermod -aG docker $USER
newgrp docker
Step 2. Configure NGC authentication
Set up access to NVIDIA's container registry using your NGC API key.
export NGC_API_KEY="<YOUR_NGC_API_KEY>"
echo "$NGC_API_KEY" | docker login nvcr.io --username '$oauthtoken' --password-stdin
Step 3. Select and configure NIM container
Choose a specific LLM NIM from NGC and set up local caching for model assets.
export CONTAINER_NAME="nim-llm-demo"
export IMG_NAME="nvcr.io/nim/meta/llama-3.1-8b-instruct-dgx-spark:latest"
export LOCAL_NIM_CACHE=~/.cache/nim
export LOCAL_NIM_WORKSPACE=~/.local/share/nim/workspace
mkdir -p "$LOCAL_NIM_WORKSPACE"
chmod -R a+w "$LOCAL_NIM_WORKSPACE"
mkdir -p "$LOCAL_NIM_CACHE"
chmod -R a+w "$LOCAL_NIM_CACHE"
Step 4. Launch NIM container
Start the containerized LLM service with GPU acceleration and proper resource allocation.
docker run -it --rm --name=$CONTAINER_NAME \
--gpus all \
--shm-size=16GB \
-e NGC_API_KEY=$NGC_API_KEY \
-v "$LOCAL_NIM_CACHE:/opt/nim/.cache" \
-v "$LOCAL_NIM_WORKSPACE:/opt/nim/workspace" \
-p 8000:8000 \
$IMG_NAME
The container will download the model on first run and may take several minutes to start. Look for startup messages indicating the service is ready.
Step 5. Validate inference endpoint
Test the deployed service with a basic completion request to verify functionality. Run the following curl command in a new terminal.
curl -X 'POST' \
'http://0.0.0.0:8000/v1/chat/completions' \
-H 'accept: application/json' \
-H 'Content-Type: application/json' \
-d '{
"model": "meta/llama-3.1-8b-instruct",
"messages": [
{
"role":"system",
"content":"detailed thinking on"
},
{
"role":"user",
"content":"Can you write me a song?"
}
],
"top_p": 1,
"n": 1,
"max_tokens": 15,
"frequency_penalty": 1.0,
"stop": ["hello"]
}'
Expected output should be a JSON response containing a completion field with generated text.
Step 6. Cleanup and rollback
Remove the running container and optionally clean up cached model files.
Warning
Removing cached models will require re-downloading on next run.
docker stop $CONTAINER_NAME
docker rm $CONTAINER_NAME
To remove cached models and free disk space:
rm -rf "$LOCAL_NIM_CACHE"
Step 7. Next steps
With a working NIM deployment, you can:
- Integrate the API endpoint into your applications using the OpenAI-compatible interface
- Experiment with different models available in the NGC catalog
- Scale the deployment using container orchestration tools
- Monitor resource usage and optimize container resource allocation
Test the integration with your preferred HTTP client or SDK to begin building applications.
Troubleshooting
| Symptom | Cause | Fix |
|---|---|---|
| Container fails to start with GPU error | NVIDIA Container Toolkit not configured | Install nvidia-container-toolkit and restart Docker |
| "Invalid credentials" during docker login | Incorrect NGC API key format | Verify API key from NGC portal, ensure no extra whitespace |
| Model download hangs or fails | Network connectivity or insufficient disk space | Check internet connection and available disk space in cache directory |
| API returns 404 or connection refused | Container not fully started or wrong port | Wait for container startup completion, verify port 8000 is accessible |
| runtime not found | NVIDIA Container Toolkit not properly configured | Run sudo nvidia-ctk runtime configure --runtime=docker and restart Docker |
Note
DGX Spark uses a Unified Memory Architecture (UMA), which enables dynamic memory sharing between the GPU and CPU. With many applications still updating to take advantage of UMA, you may encounter memory issues even when within the memory capacity of DGX Spark. If that happens, manually flush the buffer cache with:
sudo sh -c 'sync; echo 3 > /proc/sys/vm/drop_caches'