Overview
ArcHub is the Grid's self-hosted model and artifact registry. It replaced any reliance on external registries (Hugging Face Hub, Docker Hub for model images). The decision was Option C: Postgres metadata + MinIO blob storage + TypeScript API + a Hugging Face Hub API shim so existing tooling (vLLM, transformers) works without modification.
ARCHUB.0 (buy vs build decision) and ARCHUB.1 (mirror + pull-through cache, port 8080, brain PR #2) are complete. ARCHUB.2 (full registry with namespace typing + admission rules) and ARCHUB.3 (ArcRouter + SLRP + Vera wired) are planned.
Architecture
Storage layer
MinIO on ArcHelm stores all blob artifacts (model weights, LoRA adapters, eval artifacts). Postgres stores metadata: model cards, versions, namespace records, admission decisions, license status.
API layer
TypeScript service exposes a Hugging Face Hub-compatible REST API (GET /api/models/{repo}, GET /{repo}/resolve/{ref}/{file}). Existing HF clients work by pointing HF_ENDPOINT at ArcHub.
Pull-through cache
On a cache miss, ArcHub fetches from HF Hub, stores in MinIO, and serves. Subsequent pulls are local. The cache is the only permitted path to pull a new model into the fleet — no direct HF downloads from ArcHelm nodes.
Milestones
- ARCHUB.0 (complete) - buy vs build decision: Option C selected
- ARCHUB.1 (complete) - mirror + pull-through cache live at port 8080 on ArcHelm; brain PR #2 merged
- ARCHUB.2 (planned) - full registry with namespace typing, version locking, admission rules, license gate wired
- ARCHUB.3 (planned) - ArcRouter, SLRP, and Vera wired; scoring-based auto-promotion; Noema catalog as first-class entries
API endpoints
# HF-compatible model info
GET http://archelm:8080/api/models/{namespace}/{name}
# Resolve a specific file
GET http://archelm:8080/{namespace}/{name}/resolve/{ref}/{filename}
# List all models in namespace
GET http://archelm:8080/api/models?search={query}
# Push a new model version (ARCHUB.2+)
POST http://archelm:8080/api/models/{namespace}/{name}/versionsPull-through cache usage
Set the HF endpoint env var to point all HF tooling at ArcHub. The cache transparently fetches on miss.
export HF_ENDPOINT=http://archelm:8080
# vLLM, transformers, and sentence-transformers will all use ArcHubAdmission rules (ARCHUB.2)
Planned for ARCHUB.2: every model pushed or pulled-through must pass admission checks before it enters the served registry. Checks include: license clearance (via license_gate.py), namespace typing (base model / adapter / eval artifact), and a quorum approval for external pulls in production namespace.