# Updating (/docs/getting-started/updating)



New llama-dash builds are published from the `main` branch. The [System page](/docs/operations/updates) shows when your running commit is behind `main`. Database migrations run automatically when llama-dash starts, so updating means fetching the new version and restarting.

Your data is kept across updates:

* **`.env`** and **`config/config.yaml`** are ignored by Git, so `git pull` does not overwrite them.
* **The database** lives in the `llama-dash-data` Docker volume, or at `DATABASE_PATH` (default `data/dash.db`) in a manual setup.

<Callout type="info">
  Migrations are not reversed when you roll back. Before a large update, stop
  llama-dash and back up its data directory so you can restore it.
</Callout>

## Docker Compose [#docker-compose]

Run these commands from your `llama-dash` checkout. Use the Compose file you started with.

<Tabs items="[&#x22;AMD / Vulkan&#x22;, &#x22;NVIDIA / CUDA&#x22;]">
  <Tab value="AMD / Vulkan">
    ```bash
    git pull
    docker compose -f docker-compose.amd.yaml pull
    docker compose -f docker-compose.amd.yaml up -d
    ```
  </Tab>

  <Tab value="NVIDIA / CUDA">
    ```bash
    git pull
    docker compose -f docker-compose.nvidia.yaml pull
    docker compose -f docker-compose.nvidia.yaml up -d
    ```
  </Tab>
</Tabs>

* `git pull` updates the Compose files and example config. It can fail if you edited a tracked file such as the Compose file; move local changes into `.env` where possible, or commit them on a local branch.
* `pull` fetches the latest llama-dash image and the latest llama-swap image. llama-swap rebuilds its `unified-*` images nightly from the latest upstream sources, so this also updates `llama-server`.
* `up -d` recreates only the containers whose image or configuration changed.

Remove old images afterwards with `docker image prune`.

### Pinning versions [#pinning-versions]

To stay on a known llama-dash build, set `LLAMA_DASH_IMAGE_TAG` in `.env` to a [published image tag](https://github.com/ndom91/llama-dash/pkgs/container/llama-dash), such as a `sha-` commit tag. Compose reads this value when it resolves the image. `pull` and `up -d` then keep that build until you change the tag.

The llama-swap image tag is set in the Compose file. To pin llama-swap and llama.cpp, replace the floating tag with a fixed one from [llama-swap's image list](https://github.com/mostlygeek/llama-swap/pkgs/container/llama-swap): a release tag such as `unified-cuda13-268`, or a nightly build such as `unified-vulkan-2026-10-31`.

### Backing up the database [#backing-up-the-database]

The database uses SQLite WAL mode, so copy the whole `/data` directory while llama-dash is stopped:

```bash
docker compose -f docker-compose.amd.yaml stop llama-dash
docker compose -f docker-compose.amd.yaml cp llama-dash:/data ./llama-dash-backup
docker compose -f docker-compose.amd.yaml start llama-dash
```

## Manual setup [#manual-setup]

From your checkout, for example `/opt/llama-dash`:

```bash
git pull
pnpm install --frozen-lockfile
pnpm build
sudo systemctl restart llama-dash
```

Run these as the account that owns the checkout. Check that the service started with `sudo systemctl status llama-dash`. To stay on a specific build, check out that commit instead of pulling `main`.

To back up the database, stop the service and copy the directory that holds `data/dash.db` (or your `DATABASE_PATH`), including any `-wal` and `-shm` files, before you restart it.

### llama-swap and llama.cpp [#llama-swap-and-llamacpp]

In a manual setup, llama-swap and llama.cpp update separately from llama-dash:

* **llama-swap**: run `brew upgrade llama-swap`, or replace the binary with a newer [release](https://github.com/mostlygeek/llama-swap/releases), then restart llama-swap.
* **llama.cpp**: update `llama-server` the same way you installed it, through your package manager, a newer [release binary](https://github.com/ggml-org/llama.cpp/releases), or a rebuild from source. llama-swap uses the new binary the next time it starts a model; unload running models from the **Models** page to pick it up immediately.

llama-dash talks to llama-swap over its HTTP API, so you can update each component independently.