---

name: multinode-setup description: Automates setup of remote GPU rental machines (single or multi-node). Triggered when user provides a mapping of public IPs to private IPs for nodes to set up, or asks to configure remote machines for distributed training.

Multi-Node Remote Machine Setup

This skill automates the setup of rented GPU machines using a single setup script.

Quick Start

When user provides a node mapping, run:

~/.claude/skills/multinode-setup/scripts/setup_cluster.sh PUBLIC1:PRIVATE1 PUBLIC2:PRIVATE2 ...

Example:

~/.claude/skills/multinode-setup/scripts/setup_cluster.sh 147.185.41.18:10.15.22.9 147.185.41.19:10.15.22.17

The first node is always the head node (rental0).

What the Script Does

1. Updates ~/.ssh/config - Creates rental0, rental1, ... aliases (overwrites existing)
2. Copies setup script - SCPs multinode_claude.sh to all nodes in parallel
3. Runs setup on all nodes - Executes with correct NID, IHN, and RAY_HEAD_IP vars
4. Verifies setup - Checks Docker and Ray cluster status
5. Prints final report - Node table and quick access commands

Input Format

User provides mapping in various formats. Parse into PUBLIC:PRIVATE pairs:

Public IP       Private IP
147.185.40.110  10.15.17.105
147.185.40.111  10.15.17.106

Becomes:

./setup_cluster.sh 147.185.40.110:10.15.17.105 147.185.40.111:10.15.17.106

Script Variables

The underlying multinode_claude.sh uses:

• RAY_HEAD_IP: Private IP of the head node (auto-set to first node's private IP)
• NID: Node ID (0 for head, 1, 2, ... for workers)
• IHN: "Is Head Node" - true only for node 0

After Setup

Quick access commands:

ssh rental0  # Head node
ssh rental1  # Worker node
ssh rental0 "ray status"  # Check Ray cluster

Troubleshooting

If setup fails, check logs:

cat /tmp/setup_rental0.log  # Local log for head node
cat /tmp/setup_rental1.log  # Local log for worker node
ssh rental0 "cat /workspace/onstart.log"  # Remote log

The setup script is idempotent - safe to re-run.