Install and Use NVIDIA PAIR

Start the terminal interface

Use the PAIR terminal interface on a system without a desktop environment or over SSH. If the desktop application is running, quit it first. Do not run the desktop application and terminal interface on the same system. They start competing services and can conflict over ports, engines, and settings.

Install PAIR with the appropriate platform package, then run:

nvpair

nvpair starts the bundled terminal interface and its PAIR service process tree. No other PAIR process needs to be running first.

The header changes to broker ready v<version> when the terminal interface connects to the service. If it remains on connecting to broker..., open the Logs tab and inspect the service output.

Pair this system

PAIR uses the same six-digit PIN exchange as the desktop application. Pair only systems on a trusted network. The PIN is a short-lived setup code, not a long-term credential.

To invite a discovered system:

1. Open Nodes (tab 3).
2. Select the system with j or k.
3. Press i.
4. Give the displayed PIN to the person operating the other system.

To invite a system by address:

1. Open Cluster (tab 7).
2. Press i.
3. Enter the other system's hostname or IP address. Use host:port when it does not use the default pairing port.
4. Press enter, then give the displayed PIN to the other operator.

Use address-based pairing when network discovery is unavailable, such as when a network filters multicast.

To accept an invitation, open Cluster, wait for invite received from <name>, press a, enter the PIN, and press enter. Press d to decline an invitation.

The peer appears under Members when pairing is complete.

Start an engine and add a model

Open Engines (tab 6). Select an engine with j or k, then use these keys:

After you press p, enter a model name such as llama3.2, then press enter. Download progress appears on the status line.

A system can serve a request when it is online, the engine is running, and the requested model is available there. Add the same model to several systems when you want any of them to serve it.

Check the service and endpoint

Open Overview (tab 1). Confirm that the header shows broker ready v<version>. An ok worker state means no crash was reported; it does not prove that the worker is responding. DOWN means the broker reported a crash.

Open Proxies (tab 4), press g until the engine that you prepared is selected, and read its listening port. Use http://127.0.0.1:<port> as <PAIR_BASE_URL>. The default port is 11434 for the Ollama-compatible proxy and 1234 for the LM Studio / OpenAI-compatible proxy.

Ask the endpoint what the cluster can serve:

curl <PAIR_BASE_URL>/v1/models

The response lists the cluster's model inventory, not only the local system's models. The terminal interface does not send inference requests. Configure a compatible client with the local PAIR endpoint to send a request.

Check routing activity

Open Workloads (tab 5) to see live inference activity, including the workload ID, model, engine, state, and age. The terminal interface does not show which system served a workload.

Open Proxies (tab 4) to check each proxy's listening port and selected system. selected=auto means that routing is automatic. Press g to switch between engines, enter to pin the highlighted upstream, and a to restore automatic routing.

Leave automatic routing enabled unless you are testing one system.

Next steps

Keep PAIR running after an SSH disconnect

If an SSH session closes, the terminal interface exits and the system stops serving requests. Use a terminal multiplexer when PAIR must stay running:

tmux new -s pair
nvpair

Detach from tmux with Ctrl-b d. Reattach with:

tmux attach -t pair

GNU Screen also works. Start it with screen -S pair, detach with Ctrl-a d, and reattach with screen -r pair. tmux and Screen are not included with PAIR.

Move around the terminal interface

The second line shows the numbered tabs. The footer shows the keys for the current tab. If the screen shows starting..., enlarge the terminal window.

When you enter a PIN, address, port, or model name, all keys go to that field. Press enter to submit or esc to cancel.

Use the terminal interface tabs

Inspect errors and logs

In Errors (tab 2), select an error and press c to clear it.

In Logs (tab 10), scroll with j, k, and the page keys. Change the service log level with d for debug, i for info, w for warn, or e for error. Check this tab first when a service does not start.

Change settings

In Settings (tab 9), select a row with j or k and press enter. Boolean settings change immediately. Text fields open for editing; press enter to save or esc to cancel.

In Proxies, press p to change proxy ports. Engine ports are read-only in the terminal interface. Use the desktop application to change them.

Command locations and flags

The installer adds nvpair to PATH. It places the command in these locations:

Open a new terminal after installation if the command is not found.

The terminal interface writes its logs to stderr. Service logs appear on the Logs tab. Press q to quit. Quitting also shuts down the PAIR services cleanly.

Terminal-interface limits

The terminal interface is an operations tool. It cannot:

• List or delete models. It can download a model but does not show a model inventory.
• Change an engine's port.
• Update an engine.
• Control engines on other cluster systems.
• Show which system served a workload.
• Send an inference request. Use a compatible client with the local endpoint.