This screen manages training results as versions of named models. Each version carries a stage (being validated, in operation and so on). Promotion to operation (Production) always goes through approval. You also build the container image for serving and field deployment on the version screen.

Model list

Model registry — ① filter by Production ② search ③ versions per stage ④ a model (opens detail)
#AreaDescription
①Production filterAll models (전체 모델) · Has Production (Production 있음) · No Production (Production 없음).
②Search models (모델 검색)Find by model name.
③Stage columnsThe version in each stage. "—" if none.
④ModelClick to open the model detail. Models whose registered-model tag geo.safety_critical is true get a Safety-critical (안전 필수) badge.

Where MLflow is not connected, the notice "MLflow is not configured (MLflow가 설정되지 않았습니다)" appears instead of the list.

Model detail

Model detail — ① drift, monitoring, delete model ② promote to Production ③ version history ④ archive ⑤ compare versions ⑥ comments
#AreaDescription
①Header actionsShortcuts to Drift (드리프트) and Monitoring (모니터링), Delete model (모델 삭제) (model_register). The delete button is disabled while there is a Production version. The server also refuses while an endpoint is serving the model. Only the registry entries and comments are deleted; the artifacts of the runs that produced each version remain.
②Promote to production (프로덕션 승격)Promotion request dialog (production_promote, APPROVER, ADMIN).
③Version history (버전 히스토리)Version, stage, status, source run, created. Click a row to open the version detail.
④ArchiveMoves the version down to Archived (staging_promote; production_promote for a Production version). Not shown on versions already Archived. The trash icon on the right deletes the version (model_register). On a Production version the trash icon is greyed out; archive it first to delete it.
⑤Compare versions (버전 비교)Compare the run metrics of two versions side by side.
⑥Comments (코멘트)Notes per version. Writing needs develop.

A safety-critical model gets a Safety-critical · 2 approvals (안전 필수 · 승인 2건) badge next to the title. If the approvals feature is on and there is a pending promotion request, a link such as "Awaiting approval: v3 (0/1) (승인 대기 중: v3 (0/1))" appears under the title.

Model version detail

Model version — ① version info (download, archive, delete) ② container image (serving image build) ③ comments
#AreaDescription
①Version info (버전 정보)Version, stage, status, run, created. Download model (모델 다운로드) (a single file as is, a model folder as zip), Archive (아카이브), Delete version (버전 삭제) (model_register; refused for a Production version).
②Container image (컨테이너 이미지)The serving image of this version. Create it with Build image (이미지 빌드) (deploy); if one already exists, the button becomes Rebuild (다시 빌드). The button is disabled while a build runs or where building is not possible. Under Serving builder (서빙 빌더) you can choose a builder variant; leave it empty (automatic) to use the tenant's default builder. When ready, the registry image link and the builder variant are shown. The built image is reused by API services and field deployment.
③Comments (코멘트)Same as the comments on the model detail.

If there is a pending promotion request for this version, an Approval (승인) card below ① shows the approval count (for example 0/1), the requester and a View request (요청 보기) link. If the last request was rejected, it shows the rejection reason.

The model is checked before building, and problems show as a warning in ② (for example, a PyTorch model logged without a signature builds fine but fails on the first inference). If building is not possible in this environment, a yellow notice appears in ②. For "There is no serving builder variant (서빙 빌더 변형이 없습니다)", ask the system administrator for a shared image, or have the tenant administrator create and publish a dedicated variant on the serving builder screen. For "Builder image … is not in the registry (빌더 이미지 … 가 레지스트리에 없습니다)", publish it on the serving builder screen; if a shared builder is unpublished, the system administrator publishes it on the shared images screen. Where the serving pipeline is turned off, a dashed notice appears and you cannot build. The serving image of a Pointcept model is built on top of the training runtime image. If there is no record of the image used for training and the tenant has no default Pointcept training runtime variant either, the build is refused. In that case, publish a Pointcept variant on the training runtime screen, set it as the default, and build again. During a build, the elapsed time is shown and View logs (로그 보기) shows the log live. If you are on this screen when the build ends, a notice pops up. If it fails, the first line of the error is shown; check the cause with Build logs (빌드 로그).

Main tasks

Request Production promotion

  1. On the model detail, press Promote to production (프로덕션 승격).
  2. Choose the version to promote (승격 버전) (any version not in Production) and enter a reason (사유) (required). Rollback conditions (롤백 조건) and scope (범위) give the approver grounds for the decision. Press Request promotion (승격 요청).
    Request Production promotion — it goes to Approvals and becomes Production once approved
  3. The request goes to the Approvals screen. Once the required approvals are collected, that version becomes Production and the previous Production version moves down to Archived.

Even after promotion, deployment does not start automatically. Serving and field deployment are done separately in API services and Deployments.

Archive

Click the archive icon on a version row and confirm to move that version down to Archived. Only the registry label changes. Nothing is deleted.

Archive a version — only the registry label changes; nothing is deleted

Statuses and badges

StageMeaningHow a version gets there
NoneNot assignedA version registered directly in MLflow without going through the platform.
StagingBeing validatedVersions go here when you register a model from a run or training registers it automatically.
ProductionIn operationPromotion request → approval. One per model.
ArchivedKept for referenceArchived by hand, or pushed out when a new version becomes Production.

Image build status: Building (빌드 중) (BUILDING) · Ready (준비됨) (READY) · Failed (실패) (FAILED).

Written for the platform as of 2026-10-02.

© MLOps Central Platform