setup

name	setup
description	Install sparkrun and configure DGX Spark clusters

Provides complete reference for installing sparkrun, creating and managing cluster configurations, setting up SSH mesh for multi-node inference, configuring CX7 networking, Docker group membership, file permissions, page cache, earlyoom OOM protection, and diagnostics on NVIDIA DGX Spark systems.

<Use_When>

• User wants to install or update sparkrun
• User wants to create, modify, or manage cluster configurations
• User wants to set up SSH for multi-node inference
• User wants to configure CX7 network interfaces
• User wants to fix file permissions on cluster hosts
• User wants to clear page cache on cluster hosts
• User wants to set up Docker group or earlyoom
• User asks about sparkrun configuration or setup
• User is getting started with DGX Spark inference for the first time </Use_When>

<Do_Not_Use_When>

• User wants to run, stop, or monitor workloads -- use the run skill instead
• User wants to manage recipe registries or create recipes -- use the registry skill instead </Do_Not_Use_When>

Installation

# Ensure that uv is installed
uv --version

# uv can be installed with (IF NEEDED)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install sparkrun as a CLI tool
uvx sparkrun setup install

# Update sparkrun + registries (top-level shortcut)
sparkrun update

# Update to latest version (and registries)
sparkrun setup update

# Update sparkrun only (skip registry sync)
sparkrun setup update --no-update-registries

# Update channels (advanced): stable (PyPI, default), beta (develop), alpha (develop-next)
sparkrun update --beta      # switch to and update the beta channel
sparkrun update --alpha     # or --yolo; bleeding-edge develop-next
sparkrun update --stable    # switch back to the stable PyPI release

Setup Wizard (Recommended for First-Time Setup)

The wizard handles all setup steps in a single guided flow:

# Interactive wizard (auto-launches when no default cluster exists)
sparkrun setup wizard

# Pre-populate hosts and cluster name
sparkrun setup wizard --hosts <ip1>,<ip2> --cluster mylab

# Non-interactive (accept all defaults)
sparkrun setup wizard --yes --hosts <ip1>,<ip2>

# Dry-run preview
sparkrun setup wizard --dry-run --hosts <ip1>,<ip2>

The wizard performs these phases:

1. Cluster Setup — detects CX7 peers, creates cluster, sets default
2. SSH Mesh — passwordless SSH across all hosts + control machine
3. CX7 Configuration — high-speed networking (if CX7 detected)
4. Docker Group — ensures user can run Docker without sudo
5. Sudoers Entries — scoped sudoers for fix-permissions and clear-cache
6. earlyoom — OOM protection to prevent system hangs

Running sparkrun setup with no subcommand auto-launches the wizard when no default cluster is configured.

Cluster Management

Clusters are named host groups saved in ~/.config/sparkrun/clusters/.

# Create a cluster (first host = head node)
sparkrun cluster create <name> --hosts <ip1>,<ip2>,... [-d "description"] [--user <ssh_user>]
sparkrun cluster create <name> --hosts <ips> --transfer-mode push --transfer-interface cx7

# Set as default (used when --hosts/--cluster not specified)
sparkrun cluster set-default <name>

# View clusters
sparkrun cluster list
sparkrun cluster show <name>
sparkrun cluster default

# Modify
sparkrun cluster update <name> --hosts <new_hosts> [--user <user>] [-d "desc"]
sparkrun cluster update <name> --add-host 10.0.0.5
sparkrun cluster update <name> --add-host 10.0.0.5,10.0.0.6
sparkrun cluster update <name> --remove-host 10.0.0.2
sparkrun cluster update <name> --transfer-mode push --transfer-interface cx7
sparkrun cluster delete <name>
sparkrun cluster unset-default

Cluster Options

Option	Description
--hosts, -H	Comma-separated host list
--hosts-file	File with hosts (one per line)
--user, -u	SSH username for this cluster
--cache-dir	HuggingFace cache directory for this cluster
--transfer-mode	Resource transfer mode (auto, local, push, delegated)
--transfer-interface	Network interface for transfers (auto, cx7, mgmt)
--add-host	Add host(s) to the cluster (repeatable, comma-ok)
--remove-host	Remove host(s) from the cluster (repeatable, comma-ok)

SSH Setup

Multi-node inference requires passwordless SSH between all hosts. sparkrun bundles a mesh setup script.

# Set up SSH mesh across cluster hosts (interactive -- prompts for passwords)
sparkrun setup ssh --cluster <name>
sparkrun setup ssh --hosts <ip1>,<ip2> [--user <username>]

# Include extra hosts (e.g. control machine) in the mesh
sparkrun setup ssh --cluster <name> --extra-hosts <control_ip>

# Exclude the local machine from the mesh
sparkrun setup ssh --cluster <name> --no-include-self

# Dry-run to see what would happen
sparkrun setup ssh --cluster <name> --dry-run

IMPORTANT: The SSH setup script runs interactively (prompts for passwords on first connection). Do NOT capture its output -- let it pass through to the terminal.

CX7 Networking

Configure ConnectX-7 network interfaces on cluster hosts for high-speed transfers.

# Auto-detect CX7 interfaces and configure with defaults
sparkrun setup cx7 --cluster <name>
sparkrun setup cx7 --hosts <ip1>,<ip2>

# Override subnets
sparkrun setup cx7 --cluster <name> --subnet1 192.168.11.0/24 --subnet2 192.168.12.0/24

# Force reconfiguration and set MTU
sparkrun setup cx7 --cluster <name> --force --mtu 9000

# Dry-run
sparkrun setup cx7 --cluster <name> --dry-run

Requires passwordless sudo on target hosts. Will prompt for sudo password if needed.

Docker Group Membership

Ensure the SSH user can run Docker commands without sudo.

sparkrun setup docker-group --cluster <name>
sparkrun setup docker-group --hosts <ip1>,<ip2> [--user <username>]
sparkrun setup docker-group --cluster <name> --dry-run

Fix File Permissions

Fix file ownership in HuggingFace cache directories on cluster hosts.

# Fix permissions on default cache directory
sparkrun setup fix-permissions --cluster <name>

# Custom cache directory
sparkrun setup fix-permissions --cluster <name> --cache-dir /data/hf-cache

# Install sudoers entry for passwordless future runs
sparkrun setup fix-permissions --cluster <name> --save-sudo

# Dry-run
sparkrun setup fix-permissions --cluster <name> --dry-run

Clear Page Cache

Drop the Linux page cache on cluster hosts to free memory for inference.

sparkrun setup clear-cache --cluster <name>
sparkrun setup clear-cache --cluster <name> --save-sudo
sparkrun setup clear-cache --cluster <name> --dry-run

earlyoom OOM Protection

Install earlyoom on cluster hosts to prevent system hangs from memory pressure.

sparkrun setup earlyoom --cluster <name>
sparkrun setup earlyoom --hosts <ip1>,<ip2> [--user <username>]
sparkrun setup earlyoom --cluster <name> --dry-run

Diagnostics

Collect diagnostic information from cluster hosts (hidden command, useful for debugging).

sparkrun setup diagnose --cluster <name>
sparkrun setup diagnose --cluster <name> --output diag.json
sparkrun setup diagnose --cluster <name> --json   # JSON to stdout
sparkrun setup diagnose --cluster <name> --sudo   # include sudo-level checks

Configuration

Config file: ~/.config/sparkrun/config.yaml

Key settings:

• cluster.hosts: Default host list (used when no --hosts/--cluster given)
• ssh.user: Default SSH username
• ssh.key: Path to SSH private key
• ssh.options: Additional SSH options list
• cache_dir: sparkrun cache directory (default: ~/.cache/sparkrun)
• hf_cache_dir: HuggingFace cache directory (default: ~/.cache/huggingface)

<Tool_Usage> All sparkrun commands are executed via the Bash tool. No MCP tools are required.

When running SSH setup, the command is interactive and must be run with inherited stdio -- do NOT use capture_output or pipe through other commands. </Tool_Usage>

<Important_Notes>

• The setup wizard is recommended for first-time setup — it handles all steps automatically
• Always create a cluster and set it as default for the user's lab setup
• The first host in a cluster is the head node for multi-node jobs
• SSH mesh must be set up before multi-node inference will work
• sparkrun setup ssh is interactive -- let it pass through to the terminal
• DGX Spark systems have 1 GPU per host, so tensor_parallel maps to node count
• uv is the recommended Python package manager; install with curl -LsSf https://astral.sh/uv/install.sh | sh
• sparkrun setup cx7 requires passwordless sudo; use --force to reconfigure already-valid hosts
• sparkrun setup fix-permissions and clear-cache try non-interactive sudo first, then prompt if needed
• Use --save-sudo to install scoped sudoers entries for passwordless future runs
• sparkrun update is a top-level shortcut that upgrades sparkrun (if uv-installed) and updates registries
• Update channels: --stable (PyPI, default), --beta (develop), --alpha/--yolo (develop-next) on update, setup install, and setup update; no flag keeps the current channel, and switching to stable from a preview build may downgrade
• Cluster --transfer-mode options: auto (default), local (no transfer), push (head pushes to workers), delegated (workers pull)
• Cluster --transfer-interface options: auto (default), cx7 (use CX7 IPs), mgmt (use management IPs)
• Use --add-host / --remove-host for incremental cluster changes instead of replacing the full host list </Important_Notes>

Task: {{ARGUMENTS}}