mlx-bun library API
    Preparing search index...

    Interface Glm52StreamedOpenOptions

    interface Glm52StreamedOpenOptions {
        autoPin?: boolean;
        budgetBytes: number;
        caches?: Glm52LatentCaches;
        decodeKernel?: "stock" | "metal";
        dsaBenchmarkPrefillTokens?: number;
        dsaPrefillMode?: "exact" | "dense-benchmark";
        enableDsa?: boolean;
        enableMtp?: boolean;
        libraryPath?: string;
        liveRepin?: boolean;
        maxSlotsPerLayer?: number;
        mtpDraftTokens?: number;
        noCache?: boolean;
        pilotHintK?: number;
        pilotMeasure?: boolean;
        pilotTwoStep?: boolean;
        pinned?: readonly { expertId: number; layer: number }[];
        reserveBytes: number;
        routeObserver?: (
            layer: number,
            routes: readonly ExpertUsageRoute[],
        ) => void;
        usagePath?: string | false;
        workers?: number;
        workingSlots?: number;
    }

    Hierarchy

    • Omit<Glm52ExpertRuntimeOptions, "fixedBytes">
      • Glm52StreamedOpenOptions
    Index
    autoPin? budgetBytes caches? decodeKernel? dsaBenchmarkPrefillTokens? dsaPrefillMode? enableDsa? enableMtp? libraryPath? liveRepin? maxSlotsPerLayer? mtpDraftTokens? noCache? pilotHintK? pilotMeasure? pilotTwoStep? pinned? reserveBytes routeObserver? usagePath? workers? workingSlots?
    autoPin?: boolean

    Opt-in G6 candidate. Explicit pinned entries take precedence.

    budgetBytes: number
    caches?: Glm52LatentCaches

    The latent caches' scheme and paging (default f32, unpaged).

    decodeKernel?: "stock" | "metal"
    dsaBenchmarkPrefillTokens?: number

    Exact prompt length covered by dense-benchmark; required in that mode.

    dsaPrefillMode?: "exact" | "dense-benchmark"

    Stage-2 benchmark seam: build long contexts with dense batched prefill, then exercise exact DSA on one-token decode. Normal serving stays exact.

    enableDsa?: boolean

    Internal counterfactual control; indexer weights remain resident.

    enableMtp?: boolean
    libraryPath?: string
    liveRepin?: boolean

    Opt-in G6 candidate: adapt at generation safe points, max four swaps.

    maxSlotsPerLayer?: number
    mtpDraftTokens?: number

    Native MTP verify width. The int8 working tier reserves topK*gamma scratch slots plus one persistent LRU slot.

    noCache?: boolean
    pilotHintK?: number

    Opt-in G6 advisory prefetch width. Zero disables WILLNEED hints.

    pilotMeasure?: boolean

    Opt-in G6 observation: score next-layer router predictions, no I/O.

    pilotTwoStep?: boolean

    Opt-in G6 measurement: score Colibri's shared-expert-corrected route.

    pinned?: readonly { expertId: number; layer: number }[]
    reserveBytes: number

    KV/transient/allocator/Bun/OS reserve, excluding resident weights.

    routeObserver?: (layer: number, routes: readonly ExpertUsageRoute[]) => void

    Opt-in G6 route trace observer for the main target tier only.

    usagePath?: string | false

    Persistent Colibri-compatible route profile. Defaults to <modelDir>/.coli_usage; false disables persistence.

    workers?: number
    workingSlots?: number

    Settings
    Member Visibility

    On This Page
    Properties