MlxCacheLayer:
| {
kind: "full";
kv4DecodeBlock?: number;
kv4Head256?: boolean;
turboQuant: "dense"
| "deferred";
windowRead?: "folded";
}
| { kind: "sliding"; window: number }
| { kind: "recurrent" }
| {
geometry: {
indexHeadDim?: number;
kvLoraRank: number;
maxTokens?: number;
ropeHeadDim: number;
};
kind: "latent";
}
Cache layout written by the graph, independent of request storage policy.