# Set Up on Linux

Actual provides Linux x86_64 builds for NVIDIA CUDA and AMD ROCm, and Linux ARM64 builds for NVIDIA CUDA, including DGX Spark. Install the appropriate GPU driver and runtime for your hardware before loading models.

## Requirements

You need shell access, `curl`, `bash`, and an Actual account for online platform use. The installer checks the host platform and runtime compatibility, including the system C library. If it rejects your system, follow its reported requirement rather than forcing an incompatible binary.

The installer can detect an installed ROCm runtime on supported x86_64 hosts. Advanced installations can select a published backend with `ACTUAL_BACKEND`; setting an override does not install the required GPU runtime.

## Install

1. Sign in and open [Computers](/console/computers).
2. Choose **Add lead computer**, or select a cluster and choose **Add child computer**.
3. Copy the macOS / Linux command and run it within 10 minutes on the target machine.
4. Wait for installation and authorization, then open a fresh shell.

Use the command shown by the website. Do not assume beta or employee-only channels are available to your account.

## Verify and start

```bash
command -v actual
actual --version
actual whoami
actual
actual status
```

The generic installer stores the CLI under `~/.actual/bin`; a launcher in `~/.local/bin` is also valid. Check the version resolved by your shell.

Bare `actual` starts the background daemon and returns to the shell. This works on headless servers too. The model picker is interactive, so automation should use the stable selector from `actual models list`.

Confirm the computer is online in Computers. [Download and load a model](/docs/model-discovery) before expecting inference to work.

## Troubleshooting

If `curl` or `bash` is missing, install it with your distribution's package manager. If `actual` is missing from PATH after opening a new shell, try:

```bash
"$HOME/.actual/bin/actual" --version
```

If authorization is incomplete, run `actual login` and open the printed URL on a device with a browser. The code expires after 5 minutes; generate a fresh one if needed.

If the daemon starts but the model fails, check the GPU driver, runtime, available memory, and `actual logs --tail 100 --once`.

For multiple machines, use [Clustering](/docs/clustering). A Spark and a Mac appearing in the same account does not itself configure distributed inference.