Skip to content

birch-setup.md

Latest commit

 

History

History
89 lines (69 loc) · 4.03 KB

birch-setup.md

File metadata and controls

89 lines (69 loc) · 4.03 KB
Raw

PANIC! Birch setup

The Birch Level 3 installation is four TCL 4K TVs, each driven by a Raspberry Pi attached to it, plus four Vestaboards. Whatever network is showing there, it runs through an installation with eight watchers:

display watcher shows
TV 1--4 tv1--tv4 https://panic.fly.dev/i/<installation>/tvN
Vestaboards panic1--panic4 driven by the server, no browser

The watchers share a stride of 8 and sit at different offsets (TVs on 0, 2, 4, 6; Vestaboards on 1, 3, 5, 7, with panic1 also showing the prompt), so the TVs show the images and the Vestaboards the text, each a different step of the run. That layout suits a network of eight models alternating text-to-image and image-to-text. Captioners for the boards should keep their text to 6 lines of 22 characters.

Choosing what the TVs show

The Birch setup is one installation, "Birch Level 3", and each Pi's kiosk boots into its watcher's URL (in the Pi's kiosk-url.txt; see kiosk/README.md). To show another network, pick it on the installation's edit page: the TVs switch to it straight away, and the Vestaboards show its runs, with nothing to change on the Pis. Don't make a second installation for the new network: every installation with Vestaboard watchers drives the same four boards, so two of them running at once would fight over the boards.

Waking the installation up

The TVs need turning on at the wall. The Pis boot straight into their kiosk URL, so there is nothing to type on the TVs. If the screens are blank, the kiosks are stopped; bring them back with

./kiosk/kiosk.sh start

from a machine on the tailnet. ./kiosk/kiosk.sh status shows what each Pi is doing and which URL it's on. The Pis stay on the tailnet whether or not the kiosks are running, and rejoin by themselves after a power cycle.

Starting a run

On any laptop or tablet, log in at the panic landing page, open the network's terminal page (/networks/<id>/terminal) and type a prompt. Admins can also start one over the JSON API with the network's start_run route. The TVs and Vestaboards pick the run up on their own.

Letting visitors prompt

Set the network's "Anonymous terminal (minutes)" (anon_terminal_minutes in the API) and the installation's first TV watcher shows a small QR code in its bottom right. Scanning it opens a terminal that can re-prompt the network, without logging in, for that many minutes; the code changes every minute, so a photo of it soon stops working. The network's lockout still applies. Clear the setting to take the QR code down.

Stopping

Hit "Stop" on the network's page (/networks/<id>). A run slows down as it ages (see Panic.Engine.Schedule) but never stops by itself, so stop it when you're done to save AI platform credits.

If the installation is finished for a while, also stop the kiosks:

./kiosk/kiosk.sh stop

Each running kiosk holds a websocket open to the server, which keeps the Fly machine awake and billing even when no run is going. Stopping them lets it sleep until someone visits the site again.

Other notes

  • a model that fails (a refusal, a hang, an outage) has a stand-in model take its turn, or failing that its last output reused, so a run can pause for a few minutes but carries on (see Panic.Engine.Changes.InvokeModel)
  • TVs 2 and 3 won't take 3840×2160 at 60 Hz (probably an HDMI input setting on those TVs), and left alone their Pis pick 4096×2160 at 24 Hz, which is wider than 16:9. So those two Pis pin 3840×2160 at 30 Hz in ~/.config/kanshi/config; TVs 1 and 4 run at 60 Hz with no config. kanshi only applies the profile at boot, so to change the mode on a running Pi, also run wlr-randr --output HDMI-A-1 --mode <mode> in its Wayland session
  • the Vestaboards sometimes get "behind" the TVs. You don't need to do anything; they'll catch back up