Your first node.
A clear path forward.

Start with the right foundation.

Use a Linux x86-64 host with Docker Engine, Docker Compose 2.30 or newer, and enough memory and storage for your selected model.

• NVIDIA / CUDA: a working NVIDIA driver and NVIDIA Container Toolkit.
• AMD / Vulkan: a compatible AMD GPU with host DRM drivers.
• Intel / SYCL: a compatible Intel GPU with host DRM drivers.

Install Docker and your host drivers first. The installer creates persistent configuration and model directories; it does not install drivers or download model weights. Intel SYCL has been verified on Arc Pro B65 hardware; CUDA and Vulkan recipes still need validation on your target GPU.

Install for your GPU.

Run this command for NVIDIA. Replace cuda with vulkan for AMD or sycl for Intel.

curl -fsS https://accolade-grid.online/install/helm-node/install.sh -o install-helm-node.sh && sh install-helm-node.sh --backend cuda

Read or download the installer before running it. Use --prepare-only to review the generated configuration before starting containers.

Uses /srv/apps/helm-node for the application, /srv/models for shared model storage, and port 13010. Run with write access to those paths, or use --directory /absolute/path, --models-directory /absolute/path and --port 14000. Existing installation directories are never overwritten. The service listens on localhost only.

From a prepared installation directory:

docker compose config --quiet
docker compose pull
docker compose up -d --wait
docker compose logs --tail=50 local-service

These commands use compose.yaml and the private .env in that directory. Keep both with your configuration backup.

Make this node yours.

Open the service at its configured local address or through an SSH tunnel. The public installer stores the first-use setup token in the private .env file in your installation directory. Manual deployments also show it in the container startup logs. Use it to create your administrator account.

Select the GPUs the service may use, verify the configured model directory and keep a backup of your configuration. Configuration exports do not contain model weights. A healthy management service does not establish GPU or model readiness; test your first inference request.

From model to endpoint.

1. Register a model. Add a supported model from mounted storage or use the model transfer tools.
2. Create a profile. Select the engine, GPU placement, context length and concurrency.
3. Start and verify. Check readiness and logs before connecting an application.
4. Connect securely. Create a scoped inference token and use the URL in Connections.

llama.cpp is the default engine. vLLM requires a separate controller and supported recipes. Model registration alone does not establish runtime compatibility.

Your local node comes first.

Local setup and inference do not require workspace enrollment. Licensing and enrollment controls are currently paused in Helm Node and Helm Fleet.

You can sign in to Helm Control to manage your account. Fleet retains a read-only node registry; remote orchestration and cloud-provider integrations are planned.