Module concurrency 

2026-09-26: The concurrency sweep: for every (ISL × concurrency) cell, send conc streaming requests at once and report client TTFT, TPOT and E2E percentiles plus the batch’s aggregate output tok/s. Each next() runs one cell. Every request’s delivered evidence is kept, and a cell that is vacuous (concurrency_vacuity.rs), errored or cache-uncontrolled is left out of the gate metrics.

Owner: bench (concurrency). Invariants: the per-rung and peak throughput keys in ConcurrencySweep’s metrics come only from cells for which CellRow::comparable holds.

Structs

ConcurrencySweep

Constants

DESCRIPTOR

DFLASH2_DESCRIPTOR

2026-09-26: The concurrency gate served with the DFlash2 drafter armed. A separate gate id rather than a BENCH.toml variant: a required gate counts only records of its declared default checkpoint (gate::check::record_is_required_subject). It shares this driver and its floor params with concurrency-sweep; its BENCH.toml entry names a DFlash2 recipe and its own instrument and floors.

METADATA

MOE_DESCRIPTOR

2026-09-26: The concurrency gate on the Qwen3.6-35B-A3B MoE. A separate gate id, for the reason the DFlash2 one is: a required gate scores only its declared default checkpoint per box class (gate::check::record_is_required_subject), and gate::bench::baseline_for refuses two defaults on one box class. Its BENCH.toml entry pins the published ladder’s instrument (ISL 128, OSL 1024, the essay fixture, C=1..16), the one the repository’s vLLM baseline for this checkpoint (bench/baselines/qwen36-35b-a3b/published.json) was measured on.