mlx-bun library API
    Preparing search index...

    How a TurboQuant cache attends, fixed when the cache is built, like the affine kernel sets. The cache leases the mask for the phase and offers its fetches; the set decides which fetch runs and what attends over it. Both sets below are a graph's served numbers today, so the graph that composes the cache names its set: denseTurboQuantKernels for Qwen 3.5's, deferredTurboQuantKernels for Gemma 4's and MiniCPM5's.

    interface TurboQuantKernels {
        fetch: "rotated" | "decoded";
        read(fetches: TurboQuantFetches, mask: MaskLease): AttentionRead;
    }
    Index
    fetch: "rotated" | "decoded"

    The fetch read calls: the window decoded to the values' own domain, or with values left rotated. A compiled decode step decodes the stored prefix in this domain outside its recorded graph (TurboQuantKVCache).