mlx-bun library API
    Preparing search index...

    DeepSeek's released DSpark drafter (DeepSpec): a block forward over cached, already projected context K/V.

    interface DeepspecDrafterModel {
        cfg: { num_hidden_layers: number; num_target_layers: number };
        gamma: number;
        hidden: number;
        projectContext: MlxRowTable<DrafterContextProjection>;
        projectContextKVRows: MlxRowTable<DrafterContextKV>;
        tapLayers: number[];
        dispose(): void;
        draftRows(
            context: DeepspecContextAttention,
            anchors: readonly number[],
            position: number | MlxArray,
            collectLogits?: boolean,
        ): DeepspecDraftRows;
    }

    Implemented by

    Index
    cfg: { num_hidden_layers: number; num_target_layers: number }
    gamma: number

    The trained block width.

    hidden: number

    The context projection per row count or range of the rows projected (B × L), indexed by the caller (B × L): MLX's quantized matmul (matmul for a BF16 checkpoint) at every class.

    projectContextKVRows: MlxRowTable<DrafterContextKV>

    Each layer's context K/V per row count or range of the context rows (B × L), indexed like projectContext: the same kernel at every class.

    tapLayers: number[]

    The target layers whose hiddens the drafter conditions on.