Jetson AI Lab
LLM Launcher → Jetson AI Lab mirrors NVIDIA's Jetson AI Lab catalogue inside Cordatus, so you can run the latest generative AI models on your Jetson devices without translating a published recipe into a container launch by hand.
The catalogue
The page loads the model list from Jetson AI Lab. Two views:
- Cards — grouped by family, with Featured and New Releases sections.
- Table — the same models, sortable.
Filters:
| Filter | Effect |
|---|---|
| All Devices | Narrow to one of your Jetson modules. |
| All Models | Narrow to one model family. |
| More filters | Engine, Quantization and Use case. |
| Search | By name, family or size. |
The result header tells you what you are looking at — for example "{n} model(s) matching "qwen" on Orin Nano".
If Jetson AI Lab cannot be reached, Cordatus shows the last cached catalogue and says so, with a Retry button, rather than showing an empty page.
Running a model
Run on a model opens the launch dialog in express mode: a single panel instead of the full wizard.
Step 1 — pick a device
Only Jetson devices are listed, each with an Online / Offline marker.
If you have none:
This model runs on Jetson hardware only. None of your connected devices match — add a Jetson device or bring an offline one online to continue.
Add a device and Rescan are on the same screen.
Step 2 — Cordatus resolves the recipe
The panel reports what it is doing: "Resolving the container image, model and engine arguments from the Jetson AI Lab recipe…", then shows the result:
- Device
- Image — the container image chosen for this device, and how many were available for it.
- Model — the model the recipe names.
- Configuration — how many engine arguments and environment variables came from the recipe.
When the model is gated on Hugging Face, the panel asks for a token — Use a saved token or Paste a token instead.
Step 3 — Start
Start launches it and takes you to your containers.
Seeing the original recipe
Show the Jetson AI Lab recipe displays what Jetson AI Lab publishes, unchanged, with a Copy button.
This is what Jetson AI Lab publishes, unchanged. Cordatus builds the container from it — docker-level flags come from your application instead, and the model reference is passed through the model selection.
When express cannot finish
Cordatus is deliberately strict about image compatibility: an image built for a newer JetPack is never offered to an older device. When that means it cannot prepare the launch, the panel explains exactly which half is missing rather than failing silently.
| Message | What it means |
|---|---|
| "No suitable recipe is available for this device." | Jetson AI Lab publishes no configuration for this device with this engine. |
| "The Jetson AI Lab catalog publishes this model on this engine for X only, and that recipe cannot be reused here" | A different GPU architecture means different container images and quantization formats. |
| "No container image is available for this device" | The recipe itself is fine — what is missing is an image Cordatus can offer to this device. Upgrading the device's JetPack usually resolves it. |
| "The recipe was applied, but something is still missing" | The recipe resolved, but a required value is unset. |
In every case you can continue by hand: Open it in the full editor keeps your device selected and the model chosen, and lets you pick the image yourself. See the Launch Guide.
After it starts
The containers appear on the Containers page like any other deployment, and Go To Application takes you to the application they were launched from.