한국어

First sign-in and initial setup

Sign in as the global administrator, then create a tenant, register its first admin, wire GPUs and publish the shared images

Required permission: Global admin

Once installation is done, sign in as the bootstrap administrator (global administrator) and create the first tenant. The order is sign in → create a tenant → register the first admin → wire GPUs → check shared images. The Administration chapter has the details.

1. Sign in

  1. Go to the web UI address and sign in with the GEO_MLOPS_RBAC_BOOTSTRAP_ADMIN_EMAIL account from app.env and its password.

    Sign-in screen — sign in with the bootstrap administrator email

  2. Open System console (시스템 콘솔) from the avatar menu at the top right. First check that System diagnostics (시스템 진단) shows no CRITICAL — Required security settings.

    System console — opened from the avatar menu

2. Create a tenant

A tenant is a workspace for one team or customer. Datasets, models, experiments and images are separated per tenant.

  1. In the system console, press Tenants (테넌트) → Create tenant (테넌트 생성).

  2. Enter the Code (코드): 2–16 letters and digits, with hyphens allowed only in the middle. It is stored in upper case and cannot be changed after creation. The code determines the names of the Kubernetes namespace (tenant-acme) and the MLflow workspace (acme). The reserved word MLOPS is not allowed.

  3. Enter the Name (이름) shown in the UI and turn off any Features (기능) you will not use (turning one off blocks its menu and API in the tenant). Leave Default shared images (기본 공유 이미지) as is. Press Create (생성).

    Create tenant — code, name, features and default shared images

When creation finishes, the notification shows an Assign first admin (첫 관리자 지정) link. Press it to go to the next step.

3. Register the first admin

  1. In the system console, press Users (사용자) → Register user (사용자 등록).

  2. Under Create new account (새 계정 만들기), enter an email and a password (8 characters or more), choose the tenant you just created and the role ADMIN, then press Register (등록). If the person already has an account, use Grant role to existing account (기존 계정에 역할 부여).

    Register user — email, password, tenant and role

As the tenant ADMIN, this person now registers the other members and assigns their roles. The Permission matrix (권한 매트릭스) button shows each role's permissions.

4. Wire GPUs

A new tenant starts with its GPUs unwired. An unwired tenant cannot choose a GPU on the training and serving screens, and GPU requests are rejected (CPU still works). To use GPUs, set the wiring.

ModeMeaning
UnwiredNo GPU use
SharedTenants bound to a shared pool share the pool's slots. When total pool occupancy exceeds the slot count, new submissions are rejected
DedicatedA set number of slots is always reserved. Other tenants cannot use them

  1. Under "Wiring (배선)" at the bottom of the system console's GPU slots (GPU 슬롯) screen, if you will use a shared pool, first set the pool name and slot count with ① Register pool (풀 등록).

    GPU wiring — ① Register pool ② per-tenant settings

  2. Press ② Settings (설정) on the tenant's row, choose a mode and save. For dedicated, choose how many slots to reserve permanently; for shared, choose the pool to bind. The server rejects wiring that exceeds the total advertised slots.

    Tenant GPU wiring — choose unwired, shared or dedicated

5. Check shared images

Tool images such as the training runtime, dataset stager and serving builder are shared images that the global administrator builds once and provides to many tenants. The first start creates only their entries; they are not yet published (built), so training submissions are rejected until you publish them.

  1. Open Shared images (공유 이미지) in the system console. If ① "Per-tenant defaults (테넌트별 기본)" shows a tenant with an empty default, fill it with the "Fill … (… 채우기)" button.

  2. On each tab (Training runtime (학습 런타임), Dataset stager (데이터셋 스테이저), Serving builder (서빙 빌더)), press ② Publish (게시) on every image marked "Unpublished (미게시)". Publishing runs as a build Job on k3s; when it finishes, the image appears in the "Reserved repository (예약 저장소)" (mlops). The training runtime has one image per framework, so you only need to publish the ones for the frameworks you will use.

    Shared images — ① per-tenant defaults ② Publish

Next steps

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

© MLOps Central Platform