mlx-bun library API
    Preparing search index...

    Qwen3 attention (per-head q/k RMSNorm over head_dim before full-head RoPE, GQA). qwen3 and qwen3_moe share it; they differ only in the rope_theta default used when the config omits one.

    Index
    headDim: number
    kNorm: RMSNorm
    nHeads: number
    nKvHeads: number
    qNorm: RMSNorm
    reads: AttentionReadStrategy = cacheRead
    ropeBase: number
    scale: number