Expand description
2026-09-26: Per-benchmark baselines: the metrics of an earlier run, stored as
JSON in the artifact store’s runs/<benchmark_id>/ directory, for the
benchmarks that compare a run against their own history (ttft,
serve_matrix). Each file records the endpoint and model it was measured
against.
Owner: bench.
Invariants: load_for with a model named never returns a baseline whose
model is a different one.
Structs§
Functions§
- load
- 2026-09-26: The stored baseline, if any. An unreadable or corrupt file reads as absent: a run without a baseline still works, a refusal to start would not.
- load_
for - 2026-09-26: The baseline for one model: its model-keyed file, else
baseline.jsonif that records the same model. With no model named,baseline.jsonalone. - save
- 2026-09-26: Writes the model-keyed baseline, replacing any earlier one. Call only after a run whose numbers should become the bar: a failed or partial leg stored here is compared against by every later run.