Training runtime · Dataset stager · Serving builder
Add, publish and set default tool image variants for training and builds, plus external-image variants
Required permission: SettingsADMIN
Training and model serving use three kinds of tool images. The three admin console menus share one screen; only the kind differs.
| Menu | Image | When the default variant is used |
|---|---|---|
| Training runtime | The training container. A separate one per framework (YOLO, RF-DETR, Pointcept and so on) | Training submission |
| Dataset stager | An init container that places the dataset at /geo/dataset before the training pod starts. Storage credentials are injected only into this image | Training submission — without a published default stager, container training submissions are rejected |
| Serving builder | A builder that builds a registered model into a serving image | Model image build |
Within one kind you keep several variants (images with different tags and build-args) and set one of them as the default. There is one default per (kind, and framework for training runtimes).
The serving builder also builds the serving image of Pointcept models, but uses a training runtime image as its base. The base is the training image recorded on the model; if there is no record or that image is gone, it is this tenant's default Pointcept training runtime variant (own variant first, then a shared one). If neither exists the build is rejected, so publish that variant and set it as default as well.
Shared variants and your own variants
The variant list mixes two kinds.
- Shared variants — built once by a global administrator and shared with this tenant ("Shared (공유)" badge). Here you can only set them as default; editing, deleting and publishing are done by a global administrator in Shared images.
- Own variants — created by this tenant itself. You edit, delete and publish them.
Using a shared variant as the default is usually enough. If the list is empty, ask a global administrator for shared images, or create one yourself as below.
Create and publish a variant
- Press Add variant (변형 추가).
- Enter the Training framework (학습 프레임워크) (for a training runtime), Name (이름), Tag (태그), description and build-args. The tag becomes the image tag as is (
training-runtime:cpu-slim). If you leave the build-args empty, the Dockerfile defaults are used. To use it as the default right away, turn on Set as default (기본으로 지정). The ① Use external image (외부 이미지 사용) switch is covered under External-image variants below.Create a training runtime variant — ① the "Use external image" switch - Press Create (만들기), then press Publish (게시) in the list. The cluster builds the image, and the publish column changes from Publishing (게시 중) to Published (게시됨). Press the row to see the publish progress and log. The Publish (게시) button appears only with the
deploypermission (OPERATOR, ADMIN). - Press the star icon (Set as default (기본으로 지정)) of the variant to use as default. The confirmation dialog shows the default before and after.
External-image variants (training runtime only)
Images with no Dockerfile in the platform, such as a training image handed over by a partner, are not built; the variant points to them as they are.
- First import the image as a tar on the workspace's Container Registry (컨테이너 레지스트리) screen.
- Under Add variant (변형 추가) for the training runtime, turn on Use external image (외부 이미지 사용).
- Enter the framework (leave it empty to make it selectable under any framework, or use Enter own key (자체 키 입력) for the image's own key), name and tag, and choose the imported image from the Image (이미지) list. If it is not in the list, use Enter address manually… (주소 직접 입력…) to enter the address the nodes will pull.
External-image variant — points at an image imported into the registry - Press Create (만들기). An external variant has nothing to publish, so the publish column shows External image (외부 이미지) and it is usable right away.
Catalog and image label columns
The training runtime list shows a Catalog (카탈로그) column and an Image label (이미지 라벨) column. Catalog (카탈로그) tells where the wizard gets the catalog for the variant's framework key, and Image label (이미지 라벨) tells how the catalog label in the image relates to the DB. Selecting a row shows the training catalog detail below. See Training catalog for input and rules.