.Modules,
    whose construction can be complex and prohibitively cumbersome to coerce
    into a `setup` string. Finally, most builtin types are sufficiently well
    behaved and sufficiently common to warrant allowing as well. (e.g.
    `globals={"n": 1}` is very convenient.)

    Fortunately, all have well defined serialization semantics. This class
    is responsible for enabling the Valgrind subprocess to use elements in
    `globals` so long as they are an allowed type.

    Caveats:
        The user is required to acknowledge this serialization by wrapping
        elements in `globals` with `CopyIfCallgrind`.

        While ScriptFunction and ScriptModule are expected to save and load
        quite robustly, it is up to the user to ensure that an nn.Module can
        un-pickle successfully.

        `torch.Tensor` and `np.ndarray` are deliberately excluded. The
        serialization/deserialization process perturbs the representation of a
        tensor in ways that could result in incorrect measurements. For example,
        if a tensor lives in pinned CPU memory, this fact would not be preserved
        by a dump, and that will in turn change the performance of certain CUDA
        operations.
    rŪ