mlx-bun library API
    Preparing search index...

    One re-forwarded sequence: ids, the response-region start, and the m-layer hiddens over the FULL sequence (row-major [L, m*H] bf16 bytes).

    interface DflashRecord {
        hiddenMlBf16: Uint8Array;
        ids: number[];
        respStart: number;
    }
    Index
    hiddenMlBf16: Uint8Array
    ids: number[]
    respStart: number