Docs
Runs
How does a training run work on Onrup?
A run is created in draft, then submitted. Submission estimates the cost, reserves it against your spend limit, and queues the job. It then moves through matching, provisioning, training, checkpointing and uploading, streaming events throughout.
Draft, then submit
Creating a run costs nothing and commits nothing. Submission is the moment of consequence: the cost is estimated, reserved against your available headroom, and the run queued — or refused, if the reservation would breach your limit.
Reservations expire. A run that is abandoned releases its hold after a fixed window rather than sitting on your budget indefinitely.
The states, and what they mean
| State | What is happening |
|---|---|
| Draft | Created, not submitted. Nothing is reserved and nothing is queued. |
| Validating | Template, model and dataset compatibility are being checked. |
| Queued | Cost authorised and reserved. Waiting for capacity in the chosen class. |
| Matching | Selecting capacity that meets the class requirements. |
| Provisioning | Capacity acquired; the environment is being prepared. |
| Training | Steps are running. Metrics and logs stream from here. |
| Checkpointing | Durable state is being written so the run can resume. |
| Uploading | Weights and artefacts are being written to your storage. |
| Succeeded | Finished. Weights downloadable, eligible for gating and deployment. |
| Failed | Stopped with a typed reason. The reservation is released. |
| Cancelled | Stopped on request. Billed for time used, reservation released. |
Interruption is survivable
Checkpoints are written periodically, so an interruption costs the time since the last one rather than the whole run. Resumption is idempotent — a resume request that arrives twice does not start two continuations.
This is what makes running on capacity that can be reclaimed a reasonable proposition rather than a gamble.
Cost, live
A run reports what it has cost so far, not just what it was estimated to cost. The gap between those two numbers is the useful signal: consistently over-estimating means your forecasts are pessimistic, and consistently under means something is running longer than expected.
Frequently asked questions
What happens if I cancel mid-run?
Compute stops, you are billed for the time used, and the unused part of the reservation is released. The last checkpoint remains, so the run can be resumed later rather than restarted.
Can I run several at once?
Up to your concurrency quota. Each one reserves its own cost, and reservations are counted together against your spend limit — so three concurrent runs cannot each pass a check that they collectively breach.
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.