Weights mirror
Keep one copy in object storage of the pretrained checkpoints that training uses when it starts
Required permission: Global admin
Training pods download pretrained checkpoints (for example RF-DETR Large, 1.5GB) from the internet on every run. The weights mirror keeps one copy of these files in object storage so that training pods download them over the internal network. If a file is not in the mirror, the pod downloads it directly from the internet, so training is only slower; it does not fail.
Look at the badges at the top first.
| Badge | Meaning |
|---|---|
| Object storage connected (오브젝트 스토리지 연결됨) / No object storage (오브젝트 스토리지 없음) | Whether there is a place to keep the mirror |
| Prefetch on (프리페치 켜짐) / Prefetch off (프리페치 꺼짐) | Whether the feature that places files in training pods in advance is on |
| First training fills the mirror (첫 학습이 미러를 채웁니다) | The first training run that uses a file stores it in the mirror as it downloads — nothing else to do |
| "Auto-fetch off — download them here (자동 확보 꺼짐 — 여기서 받아 둬야 합니다)" | The mirror is not filled automatically, so download files yourself from the table below |
Download in advance
- In the table, find models whose mirror column says "Not yet (아직 없음)" (a missing file is not a fault, so it is not shown in a warning color).
- Press Download to mirror (받아 두기). The download starts in the background, and when it finishes the mirror column becomes Present (있음). If it fails, the reason appears with Download again (다시 받기).
- Press Refresh (다시 확인) at the top to reload the state.
Clean up orphaned objects
Old files that the catalog no longer points to (when upstream changed a file or a framework was upgraded) appear with their count and size on the Orphaned objects (고아 객체) card below. Press Clean up (정리) to delete them after confirmation. Weights uploaded by tenants are not affected.