Make it selectable in the wizard
Registering an imported image as an external-image variant of a training runtime and choosing it in step 10 of the new training wizard
Required permission: SettingsADMIN
Having an image in the registry does not make it usable for training right away. The new training wizard can choose the image only after it is registered as a training runtime variant. An image with no Dockerfile in the platform repository, such as a partner's image, is registered as an external-image variant — it is not built; it points at the imported image as is.
The training runtime screen is in the admin console of the tenant administrator ( SettingsADMIN).
Create an external-image variant
- Enter the admin console from the avatar menu, open Training runtime (학습 런타임), and press Add variant (변형 추가).
- ① Turn on Use external image (외부 이미지 사용). ② Choose the Training framework (학습 프레임워크). If left empty (Unrestricted (제한 없음)), the variant can be chosen for training in any framework. ③ Tag (태그) is the name the wizard uses for this variant (unrelated to the image tag; it cannot be changed after creation). ④ In Image (이미지), choose the imported image. For an external registry address that is not in the list, use Enter address manually… (주소 직접 입력…). Press Create (만들기).
Create an external-image variant — ① Use external image ② framework (empty = Unrestricted) ③ the tag chosen in the wizard ④ the imported image - The variant appears in the list. An external-image variant has no publish (build) step and is marked External image (외부 이미지). If the image is not in the registry, it shows "No image (이미지 없음)" and cannot be chosen.
Training runtimes — ① an external variant is not published; it points at the imported image as is
Choose it in wizard step 10
In 10. Runtime (실행 환경) of the new training wizard, open Training runtime variant (학습 런타임 변형). This list shows only variants for the framework chosen in steps 1–3 and external variants with no framework restriction. ① If you choose the external-image variant, that image's ENTRYPOINT runs in the training step.
Choosing "Auto (default: …) (자동 (기본: …))" uses the tenant's default variant at submission time. At submission the server resolves the variant to an image address and records it in the experiment, so later changes to the variant do not change the image that an experiment already run points at.
What else must be in place for training to run
An external variant alone is not enough. The items below are usually prepared once by the platform operator; if one is missing, submission or the first step stops with a reason.
| Required | If missing | Who |
|---|---|---|
| Dataset stager variant (published) | Submission is rejected with 데이터셋 스테이저 변형이 없습니다 ("no dataset stager variant") | Tenant administrator or system administrator (shared image) |
Tenant namespace tenant-<code> | Fails in the Prepare training environment (학습 환경 준비) step with namespaces "tenant-…" not found | Automatic when the tenant is created. If the cluster was down at that time, the operator creates it |
| GPU slot wiring | Submission is rejected if one or more GPUs are requested (CPU training works) | System administrator |
Operator-side preparation is in the Administration guide.
Results of an actual run
We took the MNIST image of this section through the steps above: import → register an external variant → submit in the wizard.
- Submitting the example unchanged, steps 1–3 (Queued · Prepare training environment · Prepare dataset) complete, and the training step fails with
no labeled images in manifest.json. This is because the example reads labels frommeta.labelin the local verification data (see the warning in the MNIST example). - Resubmitting with an image in which only the label-reading part was changed, all six steps completed, and the model was registered in the registry as a Staging version.