int version = TFE_CUSTOM_DEVICE_VERSION;

  // Computes the rank of the tensor handle.
  //
  // Shapes are specified via callbacks because retrieving the shape of a tensor
  // is a blocking operation for async eager; custom devices should avoid
  // retrieving shapes of tensors they wrap until the custom device tensor's
  // shape is explicitly requested where possible.

      TFE_Context* context, const std::vector<ParallelTensor*>& inputs,
      const char* operation_name, const TFE_OpAttrs* attributes,
      int expected_max_outputs, TF_Status* status) const;

  // A non-blocking version of `Execute`. After each call, `Join` must be called
  // before `StartExecute` is called again. Using `StartExecute` with `Join`
  // allows the caller to schedule computation on multiple ParallelDevices