spark-setup

DGX Spark Setup

Reproducible provisioning for the NVIDIA DGX Spark. Each phase is idempotent — safe to re-run.

Prerequisites

• DGX Spark powered on and accessible via SSH
• SSH key configured for passwordless access
• Mac-side .env configured with SPARK_HOST and SPARK_USER

Phases

Execute in order. Skip phases that are already complete.

Phase 1: Connectivity

# Verify SSH access
ssh -o ConnectTimeout=5 ${SPARK_USER}@${SPARK_HOST} "echo 'SSH OK' && uname -a"

If SSH fails, guide the user through NVIDIA Sync setup or manual SSH key configuration.

Phase 2: System Updates

ssh ${SPARK_USER}@${SPARK_HOST} "sudo apt update && sudo apt upgrade -y"
ssh ${SPARK_USER}@${SPARK_HOST} "nvidia-smi && nvcc --version"

Record CUDA version, driver version, and DGX OS version.

Phase 3: Ollama

Ollama comes pre-installed on DGX Spark via snap.

# Verify Ollama is running
ssh ${SPARK_USER}@${SPARK_HOST} "ollama --version && systemctl status snap.ollama.daemon"

# Configure for remote access (bind to all interfaces)
ssh ${SPARK_USER}@${SPARK_HOST} "sudo snap set ollama bind=0.0.0.0:11434"

# Pull a starter model
ssh ${SPARK_USER}@${SPARK_HOST} "ollama pull llama3.1:8b"

Phase 4: vLLM

Pull NVIDIA's custom vLLM container optimized for DGX Spark (Blackwell architecture, sm_121a).

# Log in to NVIDIA container registry
ssh ${SPARK_USER}@${SPARK_HOST} "docker login nvcr.io"

# Pull the vLLM image
ssh ${SPARK_USER}@${SPARK_HOST} "docker pull nvcr.io/nvidia/vllm:latest"

Do NOT use the standard vllm/vllm-openai image — it produces erroneous OOM errors on DGX Spark.

Phase 5: Tailscale

ssh ${SPARK_USER}@${SPARK_HOST} "curl -fsSL https://tailscale.com/install.sh | sh"
ssh ${SPARK_USER}@${SPARK_HOST} "sudo tailscale up"
# User completes auth in browser
ssh ${SPARK_USER}@${SPARK_HOST} "sudo tailscale set --ssh"

After auth, update Mac-side .env with SPARK_MCP_URL_TAILSCALE.

Phase 6: Docker

Docker and NVIDIA Container Runtime are pre-installed on DGX Spark.

# Verify
ssh ${SPARK_USER}@${SPARK_HOST} "docker info | grep -i runtime"

# Ensure user is in docker group (no sudo required)
ssh ${SPARK_USER}@${SPARK_HOST} "groups | grep -q docker || sudo usermod -aG docker \$USER"

Phase 7: SSH Hardening

# Disable password auth
ssh ${SPARK_USER}@${SPARK_HOST} "sudo sed -i 's/^#*PasswordAuthentication.*/PasswordAuthentication no/' /etc/ssh/sshd_config && sudo systemctl restart ssh"

# Enable UFW with SSH + MCP + Ollama + vLLM
ssh ${SPARK_USER}@${SPARK_HOST} "sudo ufw allow ssh && sudo ufw allow 3100/tcp && sudo ufw allow 11434/tcp && sudo ufw allow 8000/tcp && sudo ufw --force enable"

Phase 8: MCP Server Deployment

Run the deploy script from the Mac:

./deploy/install.sh

This rsyncs the project to the Spark, builds the Docker container, and starts the MCP server.

Phase 9: Validation

# Health check
curl http://${SPARK_HOST}:3100/health

# GPU status via MCP
# Use /spark-status command

# Test Ollama
curl http://${SPARK_HOST}:11434/api/tags

# Test vLLM (if a model is running)
curl http://${SPARK_HOST}:8000/v1/models

Output

After all phases, generate spark-setup-report.md with:

• DGX OS version, CUDA version, driver version
• Ollama version and installed models
• vLLM container image tag
• Tailscale IP and hostname
• MCP server status
• Network ports and firewall rules