mlx-bun library API
    Preparing search index...
    • adaptiveDepth: each round drafts up to numDraftTokens, chosen from the decayed acceptance rate and the graph's measured verifyRoundCosts (the graph must declare them) by a gate with the settings gate (the graph's adaptiveDraftGates for the drafter's kind); otherwise every round drafts numDraftTokens. The graph's composition names its KV scheme and scheduler switches (a graph without one runs the defaults): every request's, so the binding reads them once, here.

      Parameters

      Returns MlxGatewayBinding