Interface DsparkContextAttention

Storage owns context validity; the DSpark graph supplies only Q and the block's K/V.

interface DsparkContextAttention {
    attend(
        layer: number,
        query: MlxArray,
        keys: MlxArray,
        values: MlxArray,
        scale: number,
    ): MlxArray;
}

Index

Methods

Methods

attend

attend(
    layer: number,
    query: MlxArray,
    keys: MlxArray,
    values: MlxArray,
    scale: number,
): MlxArray
Parameters
- layer: number
- query: MlxArray
- keys: MlxArray
- values: MlxArray
- scale: number
Returns MlxArray
- Defined in packages/inference/src/contracts/mlx/drafter.ts:56