mlx-bun library API
    Preparing search index...

    Full (softmax) attention with output gate + q/k norm + partial RoPE. loadLinear chooses the projection type (the MTP companion loads dense or quantized heads); reads is how the block reads its cache (the cache's own read unless a graph injects its kernels). The graph itself is the same.

    Type Parameters

    Hierarchy (View Summary)

    Index
    headDim: number
    kNorm: RMSNorm
    kProj: L
    nHeads: number
    nKvHeads: number
    oProj: L
    qNorm: RMSNorm
    qProj: L
    reads: AttentionReadStrategy = cacheRead
    ropeBase: number
    ropeDims: number
    scale: number
    vProj: L