Interface MLACacheGeometry

interface MLACacheGeometry {
    dsa?: DSAIndexGeometry;
    kvLoraRank: number;
    maxTokens?: number;
    role?: "mtp" | "target";
    ropeHeadDim: number;
}

Properties

kvLoraRank: number

maxTokens?: number

Optional hard context bound from the model config.

role?: "mtp" | "target"

Persistence discriminator; native MTP owns an independent MLA row.

ropeHeadDim: number