| # Connecting machines with Tailscale (recommended) |
|
|
| DaisyChain needs every node reachable by a **stable IP on one interface**. |
| Tailscale gives you exactly that: a private mesh where each machine gets a |
| `100.x.y.z` address on the `tailscale0` interface, reachable P2P across NATs and |
| different networks — no port-forwarding. |
|
|
| ## 1. Install on every machine |
| - Linux: `curl -fsSL https://tailscale.com/install.sh | sh` |
| - Windows / macOS: <https://tailscale.com/download> |
|
|
| Bring each up (same tailnet / account) and note its IP: |
|
|
| ```bash |
| sudo tailscale up |
| tailscale ip -4 # e.g. 100.101.102.10 |
| tailscale status # see every machine + IP |
| ``` |
|
|
| ## 2. Configure the cluster |
| Pick one machine as **rank 0** and use its Tailscale IP as `MASTER_ADDR`. On |
| every machine: |
|
|
| ```bash |
| export MASTER_ADDR=100.101.102.10 |
| export MASTER_PORT=29560 |
| export WORLD_SIZE=3 |
| export GLOO_SOCKET_IFNAME=tailscale0 # pin gloo to the mesh NIC |
| export RANK=0 # 1, 2, ... on the others |
| daisychain-train |
| ``` |
|
|
| `GLOO_SOCKET_IFNAME=tailscale0` is the important line — it stops gloo from |
| wandering onto a LAN/VPN/Docker interface. |
|
|
| ## 3. Verify |
| Run `daisychain-dashboard` (point `config/nodes.example.json` at the Tailscale |
| IPs). Green banner = every node reachable and ports open → launch training. |
|
|
| > On Windows the Tailscale interface name differs and gloo tensor collectives are |
| > unstable anyway — prefer Linux nodes or Docker for the actual training. |
|
|
