Skip to content

models.md

Latest commit

 

History

History
106 lines (81 loc) · 3.86 KB

models.md

File metadata and controls

106 lines (81 loc) · 3.86 KB
Raw

Choose and prepare models

While offroad and online, open Settings > Models > Big Model on the comma. Start with the default. The comma downloads it on any connection, and picks up where it left off if the download is interrupted. The Jetson then prepares it, about 3 minutes, the next time it is on while the car is parked. Until a newly picked model is ready, the model the device already has loaded drives (on a phone or Mac, the one picked in the app), else the last one it prepared; with none, the small model does. See when it switches.

If the Jetson only has power while the car is on, prepare a new model by starting the car, forcing offroad, and waiting for the Big Model panel to say it is ready.

If the list is empty, use Refresh Model List. On Jetson, the 766 MB models leave more time in the 50 ms frame budget than the 1.7 GB Lebowski.

Preview Models are comma models that are not merged into openpilot yet. comma may still change or withdraw one. When a preview leaves the list, a comma that had picked it goes back to the default.

Prepare ahead of time (optional)

Download on the server's internet connection before connecting the comma. This saves time when the comma is on LTE.

On Mac, open Models, click Use Model, and wait for In Use. On iPhone, iPad, or Android, open Models and tap Get. Keep the iPhone app open until it finishes. Pick the same model on the comma.

On a Jetson or an installed PC

List the models, then fetch one by its full 40-character ref from the JSON (plain list shortens refs):

jetlink models list --json
jetlink models fetch <ref>

To prepare it too, stop the server first, with any connected comma offroad:

jetlink stop
jetlink models prepare <ref>
jetlink start

Do not prepare a model in a separate process while the server uses the same cache. Concurrent builds are unsupported and can exhaust memory; jetlink models prepare refuses while the service runs. The apps prepare through their running server, so they need no stop and start.

From a checkout

The same commands on a built jetlink-server: jetlink-server models list --json, jetlink-server models fetch <ref>, jetlink-server models prepare <ref>. Stop any server using that cache before prepare.

Downloads, prepared engines, and disk space

  • A download is the ONNX model, most about 766 MB.
  • A prepared engine is built for your backend and device. Jetlink keeps both.
  • A runtime update may prepare again; the download is kept.
  • On a Mac, allow about 3 GB per model with the default backend.
Installation Default cache folder
Jetson installer /mnt/data/jetlink
PC installer /var/lib/jetlink
Mac app ~/Library/Application Support/Jetlink/cache
jetlink-server on a Mac ~/Library/Caches/jetlink
jetlink-server on Linux /mnt/data/jetlink on a Jetson, else /var/lib/jetlink as root, else ~/.cache/jetlink
  • Mac app: choose a cache folder in Settings.
  • jetlink-server: JETLINK_CACHE or --cache DIR (the rule).
  • The installer's jetlink models uses the server's cache.

Disk use: jetlink models inventory (installer), jetlink-server models inventory, or Models in the apps.

Commands

Listing, fetching, importing, preparing and deleting models: model command reference.