Source installs and testing¶

For normal installation, use the setup guides. This page covers terminal installs, WSL2, and testing without a comma.

Mac (Apple silicon)¶

Use the Mac app from Releases. The server is built in.

From a terminal¶

The app's server as a command. Needs Xcode 26:

git clone https://github.com/zoompilot/jetlink.git && cd jetlink
swift build -c release --package-path JetlinkKit --product jetlink-server
caffeinate -i JetlinkKit/.build/release/jetlink-server --usb

• Quit the app first: one server holds the comma at a time.
• caffeinate -i keeps the Mac awake while it serves.
• Models and prepared engines go in ~/Library/Caches/jetlink, about 3 GB per model. --cache DIR moves them; the app's folder is ~/Library/Application Support/Jetlink/cache.
• CoreML prepares a model in about 20 seconds the first time, then loads it in under a second to about 10 seconds.

Options (jetlink-server --help lists the rest):

# TCP for a bench tool instead of the comma (see Test without a comma)
JetlinkKit/.build/release/jetlink-server --listen

# the GPU only, if another app keeps the Neural Engine busy
JetlinkKit/.build/release/jetlink-server --usb --device coreml

# prepare a model ahead of time, then exit
JetlinkKit/.build/release/jetlink-server build /path/to/big_driving_supercombo.onnx

Every command and option: the server command. Performance: backends.

Linux (NVIDIA GPU)¶

Use the Linux PC installer guide. To build the server yourself, see From a checkout.

Windows (NVIDIA GPU)¶

Untested. Run the Linux installer in Ubuntu 22.04 or 24.04 on WSL2:

1. Update the NVIDIA driver in Windows to 580 or newer. Never install one inside WSL.
2. Turn on systemd in WSL: add these lines to /etc/wsl.conf, then run wsl --shutdown in Windows.

[boot]
systemd=true

1. Run the Linux installer in Ubuntu.
2. Attach the comma to WSL with usbipd-win.

Start with a TCP test.

From a checkout¶

For development on Linux. Build the server tarball (in a Docker or podman container, or natively on a Jetson or PC), then install it:

scripts/build-linux.sh linux-x86_64       # linux-aarch64 for a Jetson
sudo ./install.sh --binary dist/jetlink-server-<version>-linux-x86_64.tar.gz

Or run it in a terminal without installing. The host needs:

• libcurl4, which the server links;
• TensorRT: JetPack's on a Jetson; on a PC, TensorRT 11.3.0.99's libraries named with --tensorrt-libs (an installed PC has them in /opt/jetlink/tensorrt/11.3.0.99; otherwise see installing by hand);
• the udev rule, which grants USB access without root. Replug the comma after installing it.

sudo install -m 644 scripts/99-jetlink-host.rules /etc/udev/rules.d/
sudo udevadm control --reload-rules
tar -xzf dist/jetlink-server-<version>-linux-x86_64.tar.gz
jetlink-server-<version>-linux-x86_64/bin/jetlink-server --usb --backend trt \
  --tensorrt-libs /opt/jetlink/tensorrt/11.3.0.99

On a Jetson, leave out --tensorrt-libs. Without root the server cannot turn off USB 3 link power management on the comma's port, which costs about 2 ms a frame; the log says so.

CPU only¶

Checks the protocol and model loading without a GPU. Too slow for driving. The tarball has no onnxruntime; the build script fetches it and prints its folder:

ORT=$(scripts/build-linux.sh linux-x86_64 ort)
LD_LIBRARY_PATH=$ORT/lib jetlink-server-<version>-linux-x86_64/bin/jetlink-server \
  --backend ort --device cpu --listen

On a Mac, --device cpu does the same.

Test without a comma¶

You need a large driving-model ONNX file and a server listening on TCP: on an installed Jetson or PC, jetlink run --listen (it stops the service while it runs). In a second terminal, from a checkout, replace /path/to/big_model.onnx with your model:

python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -e . onnx
python3 scripts/bench_link.py --host 127.0.0.1 --onnx /path/to/big_model.onnx --rate 20

• It uploads the model, waits for the build, and reports round-trip latency and frames over the 50 ms budget at 20 Hz.
• For a server on another machine, use its wired-network IP.
• TCP has no authentication: use a trusted network.
• Wi-Fi does not meet the frame budget.

Troubleshooting¶

Cache folders: model management. Comma-side alerts: troubleshooting.

From the Jetlink documentation.