mlx-bun library API
    Preparing search index...
    uncomposedTurboQuantKernels: TurboQuantKernels = ...

    The set of plain storage whose graph composes no TurboQuant read: every graph but Qwen 3.5's, Gemma 4's and MiniCPM5's reads converted TurboQuant storage through the deprecated fetches, which take no set. A TurboQuant cache converted from such storage serves those fetches; its named reads throw.