Stellarator is a control plane for fine-tuning and reinforcement learning workloads on Tinker. Operators launch sandbox runs, validate them through a structured pre-flight gate, then promote to scale — with cost projections, budgets, and live alert streams threaded through every step.
Screenshots
How it works
Each run is owned by the identity that created it. Other identities can read the metrics, alerts, and notes but cannot mutate the run — pause, cancel, edit hyperparameters all require ownership. This makes the platform safe for multi-operator workspaces where several pipelines may share a Tinker organization.
Runs cannot reach scale GPUs without first passing through a sandbox. The sandbox is a small, cheap smoke test (CPU or single A10G, capped step count). Its result feeds a structured pre-flight payload — model, method, dataset mixture, hyperparameters, projected cost, citations — which the server validates before accepting any production submission. Stale lineage (over 24 hours) or missing fields are rejected with HTTP 412 and an actionable hint.
Training scripts emit trackio-style alerts (error, warn, info) that surface in real time on the run detail page and trigger toasts via a server-sent event stream. A high-concurrency Rust supervisor handles per-job polling, metric ingestion, cost computation, and websocket fan-out — keeping the orchestration layer free from per-tick work.
A research subsystem fetches Hugging Face papers, arXiv records, and GitHub code examples on demand so methodology choices can be cited per run, with a global research log auditing every lookup.
Stack
- FastAPI + SQLAlchemy + SQLite (WAL) with Alembic migrations
- Rust supervisor: axum, tokio, sqlx, subtle (constant-time auth)
- Next.js 15 + Tailwind v4 + shadcn/ui (App Router, dark mode default)
- MultiFernet encryption for OAuth tokens and per-identity provider keys
- OpenTelemetry instrumentation, Prometheus + Grafana dashboards
- Locust load-regression gating in CI
- Tinker REST surface for training job orchestration
Status
Active