mlx-bun library API
    Preparing search index...

    The defaults of each setting a graph may declare for its machine (AdaptiveDraftGateSettings):

    • priorAccepted, priorRejected: every position's counts before any round is observed, 4.5 and 0.5 (an acceptance of 0.9, so a new gate drafts deeply and backs off as rejections arrive);
    • decay 0.85, and idleDecay 1: a round without drafts leaves the counts as they are;
    • margin 0.03;
    • probeInterval: rounds without drafts before next drafts one token to observe acceptance again (8); each probe after which the gate still drafts nothing doubles the interval, up to maxProbeInterval (32);
    • probeDepth: the depth a drafter that proposes without committing drafts a probe at (the round's budget).
    interface AdaptiveDraftGateOptions {
        decay?: number;
        idleDecay?: number;
        margin?: number;
        maxDraftTokens: number;
        maxProbeInterval?: number;
        priorAccepted?: number;
        priorRejected?: number;
        probeDepth?: number;
        probeInterval?: number;
    }

    Hierarchy (View Summary)

    Index
    decay?: number

    The factor a position's counts keep per round that observes it, and every count per round without drafts.

    idleDecay?: number
    margin?: number

    How much a round's expected rate must exceed a step without drafts.

    maxDraftTokens: number
    maxProbeInterval?: number
    priorAccepted?: number

    Accepted and rejected counts every draft position starts from: their ratio is the acceptance the gate assumes before it observes the text, their sum how many rounds of evidence that assumption weighs.

    priorRejected?: number
    probeDepth?: number

    The depth of a probe a drafter proposes without a verify round.

    probeInterval?: number

    Rounds without drafts before a probe, doubling after each probe that leaves the gate drafting nothing, up to maxProbeInterval.