/**
 * multimodal_vocab.ts — one flat token space shared by text and video.
 *
 * Evermind's generator ({@link ../lm/evermind_lm.EvermindLM}) is a pure
 * next-token predictor: it neither knows nor cares whether a token id is a BPE
 * text piece or a quantised video-patch code. So "make Evermind generate video"
 * is a *vocabulary* problem, not a model problem — lay text tokens, a handful of
 * modality control tokens, and the video codebook codes out on a single integer
 * axis, and the same model handles `(caption … <vid> … frames … </vid>)`.
 *
 * Layout (low → high):
 *
 *   [ 0 .. textVocabSize )                     text (BPE) — passthrough
 *   textVocabSize + 0                          <vid>   (BOS_VIDEO)
 *   textVocabSize + 1                          </vid>  (EOS_VIDEO)
 *   textVocabSize + 2                          <key>   (start of an intra frame)
 *   textVocabSize + 3                          <delta> (start of an inter frame)
 *   codeBase + bank·levels·K + level·K + code  residual-VQ code tokens
 *
 * where `bank ∈ {INTRA, INTER}`, `level ∈ [0, levels)`, `code ∈ [0, K)` and
 * `K = codebookSize`. A frame is therefore a marker followed by
 * `patchesPerFrame · levels` code tokens — a deterministic, self-delimiting
 * structure the codec can parse straight out of the generator's output.
 */
export declare const VIDEO_BANK_INTRA = 0;
export declare const VIDEO_BANK_INTER = 1;
export interface MultimodalVocabConfig {
    /** Size of the text (BPE) region; 0 for a pure-video model. */
    textVocabSize: number;
    /** Residual-VQ depth (codes emitted per patch). */
    levels: number;
    /** Entries per codebook per level per bank. */
    codebookSize: number;
}
/** What a token id decodes to when parsing a mixed stream. */
export type TokenKind = {
    kind: "text";
    id: number;
} | {
    kind: "control";
    control: "bosVideo" | "eosVideo" | "frameKey" | "frameDelta";
} | {
    kind: "code";
    bank: number;
    level: number;
    code: number;
};
/**
 * Bijective map between (text ids, control tokens, video codes) and a single
 * flat token id space. Owns the offset arithmetic so nothing else has to.
 */
export declare class MultimodalVocab {
    readonly textVocabSize: number;
    readonly levels: number;
    readonly codebookSize: number;
    readonly bosVideo: number;
    readonly eosVideo: number;
    readonly frameKey: number;
    readonly frameDelta: number;
    readonly codeBase: number;
    /** Total vocabulary size — pass this as EvermindLM's `vocabSize`. */
    readonly size: number;
    private readonly perBank;
    constructor(cfg: MultimodalVocabConfig);
    /** Global token id for a residual-VQ code. */
    codeToken(bank: number, level: number, code: number): number;
    isText(tok: number): boolean;
    isCode(tok: number): boolean;
    isVideoMarker(tok: number): boolean;
    /** Decompose a code token into (bank, level, code). Level/code are meaningful even if the bank differs. */
    decodeCode(tok: number): {
        bank: number;
        level: number;
        code: number;
    };
    /** Classify any token id in the flat space. */
    classify(tok: number): TokenKind;
}
//# sourceMappingURL=multimodal_vocab.d.ts.map