mlx-bun library API
    Preparing search index...

    Interface ServedComposition

    The configurations for which a graph is recommended by default. Each field present participates in default selection; explicit graph choices bypass it.

    interface ServedComposition {
        device?: string;
        draftDepth?: DraftDepth;
        fingerprint?: string;
        kv?: readonly ServedKv[];
        maxRows?: number;
        maxVerifyRows?: number;
        prefillChunkTokens?: number;
    }
    Index
    device? draftDepth? fingerprint? kv? maxRows? maxVerifyRows? prefillChunkTokens?
    device?: string

    The GPU architecture (Composition.device).

    draftDepth?: DraftDepth
    fingerprint?: string

    The artifact's fingerprint, as the table's fingerprint computes it from the config.

    kv?: readonly ServedKv[]

    The KV schemes served. A composition matches one with the same kind and widths (every layer's, for a per-layer scheme).

    maxRows?: number

    The most rows per forward the graph was built for: a composition with at most this many matches.

    maxVerifyRows?: number

    The most verification rows the file lists. The scheduler's row cap times its maximum drafted-window length must fit for default selection.

    prefillChunkTokens?: number

    Settings
    Member Visibility

    On This Page
    Properties