# Run Models on llama.cpp

This recipe runs a model with `llama-server` inside a persistent `tmux` session, downloading it directly from Hugging Face.

## Install tmux

```bash
sudo apt install -y tmux
```

## Start a tmux Session

```bash
tmux new -s llama
```

## Run the Model Server

Replace:

- `<REPO>` with the Hugging Face model repo (e.g. `deepreinforce-ai/Ornith-1.0-9B-GGUF`)
- `<HOST>` by default `0.0.0.0` hear all network interfaces
- `<PORT>` with the port to serve on (e.g. `8000`)
- `<CONTEXT_SIZE>` with the context window size (e.g. `262144`)

```bash
llama-server -hf <REPO> --host <HOST> --port <PORT> -c <CONTEXT_SIZE>
```

## Detach from the Session

Press:

```text
Ctrl+B, D
```

## List tmux Sessions

```bash
tmux ls
```

Expected output:

```text
llama: 1 windows (created ...)
```

## Reattach to the Session

```bash
tmux attach -t llama
```

## Verify Server is Running

Replace `<PORT>` with the configured port.

```bash
curl http://localhost:<PORT>/health
```

Expected output:

```json
{"status":"ok"}
```

## Troubleshooting

### Port already in use

Verify that no other process is bound to the port.

```bash
sudo ss -tulpn | grep <PORT>
```

### Session not found

List active sessions and verify the name.

```bash
tmux ls
```

### Server not responding

Reattach to the session and check the server logs.

```bash
tmux attach -t llama
```