mlx-bun library API
    Preparing search index...

    Interface RecurrentMtpTarget

    What a recurrent multi-token-prediction head reads from its target: the final hidden layer's output (tapped before the final norm), the token embedding, and the output head.

    interface RecurrentMtpTarget {
        hiddenSize: number;
        layerCount: number;
        logits: MlxRowTable<(hidden: MlxArray) => MlxArray>;
        vocabularyHead?: {
            biases: MlxArray | null;
            scales: MlxArray;
            spec: { bits: number; groupSize: number; mode: string };
            vocabSize: number;
            w: MlxArray;
        };
        embed(ids: MlxArray): MlxArray;
    }
    Index
    hiddenSize layerCount logits vocabularyHead?
    embed
    hiddenSize: number
    layerCount: number
    logits: MlxRowTable<(hidden: MlxArray) => MlxArray>

    The output head per row count or range, indexed by the rows the drafter projects.

    vocabularyHead?: {
        biases: MlxArray | null;
        scales: MlxArray;
        spec: { bits: number; groupSize: number; mode: string };
        vocabSize: number;
        w: MlxArray;
    }

    The untied quantized vocabulary projection, when the drafter may project onto a row subset of it (frequency-ranked draft vocabulary). Borrowed.

    Settings
    Member Visibility

    On This Page
    Properties
    Methods