Status: 🔮 Possible future — no commitment, nothing being built ·
Date: 2026-07-22
The Model Registry pillar is
future state: for the MVP and the current phase, vLLM pulls weights directly
from Hugging Face with a PVC cache (chart configuration — explicitly not a
registry component). We are not building a registry for the MVP.
When a real registry becomes worth building, a candidate emerged in the
2026-07-22 runtime-team discussion: MSR (Mirantis Secure Registry).
Hugging Face ──pull──▶ MSR ──pull──▶ vLLM (runtime)
- MSR acts as the on-prem/sovereign system of record for model artifacts:
vLLM pulls from MSR; MSR mirrors/pulls from Hugging Face (or takes customer
uploads).
- Attractive because it's an existing Mirantis product with the enterprise
properties the registry pillar wants anyway: private storage, access
control, scanning, air-gap friendliness — and OCI artifacts are one of the
registry approaches already under evaluation for weights.
- Only LLMs and tokens matter right now; HF-pull + PVC cache covers the MVP
and the first NeoCloud phase.
- Open questions before this becomes an ADR with an "Accepted" status:
multi-hundred-GB artifact behavior, chunked/parallel pulls, LoRA/fine-tune
artifact layout, and whether MSR's roadmap aligns.
- No repo, no issues, no staffing for the registry — unchanged.
- When the registry pillar gets staffed, evaluating MSR is the first task,
against the alternatives already listed on the pillar page (Harbor/ORAS,
purpose-built store).