Skip to content

point.yaml

Authored in step 5.

Fields

This page lists the values the checker accepts for the §8.3 fields it validates by value, and where the tooling adds a condition the rules don't state.

Field What the checker accepts
region low_latency, low_concurrency, med_concurrency, high_concurrency or submitters_choice (region-declared)
dataset_type Any of the §8.3 values, but the bundle builder only uses it when it is exactly Accuracy or Performance.
offline dedicated, elected or none, or absent. elected only on the point whose concurrency equals max_supported_concurrency (offline-point-present)
target_cohort YYYY-MM-C0 or YYYY-MM-C1 (target-cohort)
shared_src, shared_docs Must resolve to directories under the submission root.

region on a dedicated Offline run

Its concurrency is the size of the performance dataset and it counts toward no region, but the rules don't say what region should hold. submitters_choice passes without a placement warning. Tracked as C7 in Open questions.

Per-point power fields

Two optional fields come from §4.5.3, with their shape in §8.3. This page hasn't confirmed what the checker accepts for them.

  • nodes_used: the nodes engaged at the point, per node set, for per-point power normalization. It can't exceed what system_power.json provisions for that node set. A point without it is normalized by the full provisioned power.
  • dp_shortfall: only for a point that ran fewer data-parallel replicas than fit, recording why. Reviewers can raise a Methodology objection to it.

The steady_state block

What the block records is defined in §8.3, and how the status and verdict are decided in §4.4. Background: What your numbers are measured over.

Two things the checker reads that §8.3 doesn't name: the window sub-keys are super_pass_start, super_pass_end, super_pass_size, n_samples and duration_s, and the block also carries a verdict, which takes one of the §4.4 verdict values.

You copy this block in yourself

Run the benchmark with --steady-state and the client finds the window during the run and reports it in result_summary.json and report.txt. Nothing copies it into point.yaml, so you fill in this block from that output. To re-check a finished run, point python -m inference_endpoint.metrics.steady_state_diagnostics at its run directory. The client doesn't compute steady state for agentic or Offline runs, and a dedicated Offline run sits outside the rules' steady-state scope too. Tracked as B9 in Open questions.

The warmup block

What it must declare is set by §6.3.3, and the sub-field names by §8.3.

See also

Last verified against: mlcommons/endpoints_policies@main (1818269), mlcommons/endpoints-submission-cli@main (a42a056, v1.1.0.0) and mlcommons/endpoints@main (f1100cf), 2026-10-02.