mlx-bun library API
    Preparing search index...

    Interface ModelOpenOptions<Model>

    interface ModelOpenOptions<Model = RuntimeModel> {
        batchSize?: number;
        caches?: Glm52LatentCaches;
        composition?: Composition;
        contextTokens?: number;
        enableMtp?: boolean;
        graph?: string | ((weights: Weights, config: ModelConfig) => Model);
        implementations?: ModelImplementationProvider<Weights, Model, Composition>;
        libraryPath?: string;
        machineBytes?: number;
        maxGenerationTokens?: number;
        memoryBudgetBytes?: number;
        mtpDraftTokens?: number;
        profiles?: ResolveModelProfileOptions;
    }

    Type Parameters

    Hierarchy (View Summary)

    Index
    batchSize? caches? composition? contextTokens? enableMtp? graph? implementations? libraryPath? machineBytes? maxGenerationTokens? memoryBudgetBytes? mtpDraftTokens? profiles?
    batchSize?: number

    Maximum ordinary continuous-batch rows.

    caches?: Glm52LatentCaches

    The composition's KV scheme and paging, which the graph's latent caches compose (default f32, unpaged).

    composition?: Composition

    The composition the graph is built for; absent resolves the default record (resolveComposition(config)). The direct Colibri runtime plans its own resources from the options above and takes none.

    contextTokens?: number

    Context reserved by the exact GLM resource equation.

    enableMtp?: boolean

    Native checkpoint MTP row; on by default.

    graph?: string | ((weights: Weights, config: ModelConfig) => Model)

    A named shipped graph or the caller's own load function. Explicit choices bypass device and configuration recommendations and need no registration.

    implementations?: ModelImplementationProvider<Weights, Model, Composition>
    libraryPath?: string

    Native expert-I/O dylib override. Packaged/dev defaults are automatic.

    machineBytes?: number

    Test/diagnostic override; normal callers use physical RAM.

    maxGenerationTokens?: number

    Generated-token allowance within contextTokens.

    memoryBudgetBytes?: number

    Whole-process ceiling. Defaults to the smaller of the validated 25 GiB preset and physical RAM.

    mtpDraftTokens?: number
    profiles?: ResolveModelProfileOptions

    Settings
    Member Visibility

    On This Page
    Properties