Skip to content

6. Register your runs

Produces: one run ID per run folder, each with its folder uploaded to MLCommons.

Before you begin

A submission isn't built from folders on your disk. It's built from runs, which are records on the MLCommons server that each hold one uploaded run folder. You register every folder from step 4 as a run, then step 8 creates the submission from the run IDs.

What you'll do

  • Give the CLI your API key and check it authenticates
  • Preview and register each folder, and note the run IDs
  • Pin the runs you intend to submit
  • Confirm the list

How each command works, with its flags and output, is in the CLI's runs command docs. This page covers what a submission needs from them.

Steps

1. Set your token

Every command authenticates with your PRISM API key. Set it as described in Authentication, then check it with runs list. Treat the key as a credential: don't commit it or paste it into an issue.

2. Preview and register each folder

Preview each folder with runs create --dry-run, then register it with runs create. Register every folder: each concurrency point, the dedicated Offline run if you have one, and any accuracy run you ran as its own folder. A --mode both run keeps performance and accuracy in one folder, so it's one run. Keep the returned IDs, because step 8 takes them as --run-ids.

The folder layout the CLI expects is in Submission package layout.

Decide on --test before you register

--test excludes a run from published reporting, and it can't be changed after the run is created. Use it for rehearsals, never for runs you mean to publish.

3. Pin the runs you'll submit

Runs expire under a server policy that isn't published (C12 in Open questions), so pin the ones you intend to submit, at creation or afterwards with runs pin.

4. If you change a folder after registering it

The uploaded archive is a copy, and a run can't be updated. Register the folder again, use the new run ID, and delete the old run with runs delete.

Verify

runs list shows one row per folder, with the concurrencies from your plan in step 3. runs get on each run shows Test Run: No; runs list doesn't show the test flag.

Next

→ 7. Validate locally

Problems? See Troubleshooting.