Images and templates
The image catalog is what the session wizard offers. An image that is not here (or not compatible with the chosen card) cannot be run.
| Column | Meaning |
|---|---|
| Name | What users see |
| Kind | Catalog image, or a user's own |
| Source | The registry reference |
| CUDA | The runtime inside — matched against the offering's minimum |
| Supported GPUs | Models verified for it |
| Public | Visible to everyone, or private to its owner — toggled from the row |
Two tabs: the catalog itself, and builds — every console-driven build with its state and log, whether it succeeded or not.
Importing
Import registers an image that already exists in a registry: the reference, a display name, its CUDA version, and which GPU models it is for.
Two things decide whether it will actually run:
- The nodes must be able to pull it. A private registry needs per-node trust (see Cluster setup → Local registry); on a multi-cluster fleet every cluster needs to reach it.
- The CUDA version must be honest. It is what the wizard filters on. Declaring 12.8 on an image built for 12.4 does not make it work on a Blackwell card — it makes the failure happen later and less clearly.
Builds
The Builds tab follows image builds run through the platform:
queued → building → pushing → scanning → succeeded, with the log of each.
A build is started through the API (POST /api/v1/image-builds, from an inline Dockerfile or a
public git repository); the console has no form for it. It also needs the chart's
operator.kanikoImage to be set — without it the operator fails builds immediately. The result
is pushed to api.buildRegistry and registered automatically.
Images a member imported or built themselves are private to that member (up to 20 each) and tagged My image in the wizard.
Seeded images
With api.seedSessionImages (the default) the catalog seeds boanlab/gshare-session for the
CUDA 12.4/12.5 line and the Blackwell 12.8/12.9 line at startup. Those tags exist on Docker Hub
only after the publish workflow has run for a release.
A fleet of Blackwell cards needs the 12.8 line; older images are refused with
incompatible_image before the session starts, which is the behavior you want — the alternative
is a container that starts and then cannot see the GPU.
Keeping the catalog small
Every extra image is a decision for the user and 5–15 GiB on every node that pulls it. One maintained image per framework generation beats a dozen variants; let people build their own for the rest.