mlx-bun library API
    Preparing search index...
    • Return bf16 K and rotated V (f32 for eager inverse rotation, bf16 for deferred rotation). Null means the ordinary codec must handle this input. Input arrays and the shared centroid table remain borrowed.

      Parameters

      • inputs: PackedKvArrays
      • centroids: MlxArray
      • kBits: number
      • vBits: number
      • headDim: number
      • deferV: boolean
      • stream: bigint = gpuStream

      Returns [MlxArray, MlxArray] | null