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

  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.

//
// Questions left to address:
// * Might at some point need a way for callers to provide their own Env.
// * Maybe add TF_TensorShape that encapsulates dimension info.
//
// Design decisions made:
// * Backing store for tensor memory has an associated deallocation
//   function.  This deallocation function will point to client code
//   for tensors populated by the client.  So the client can do things
//   like shadowing a numpy array.

  // ExtendSessionGraphHelper manually.
  std::atomic<bool> extend_before_run;
};

struct TF_ImportGraphDefOptions {
  tensorflow::ImportGraphDefOptions opts;

  // Backing memory for TensorId fields in opts.
  // TODO(skyewm): it'd be better if ImportGraphDefOptions owned this.
  std::vector<tensorflow::string> tensor_id_data;
};

struct TF_ImportGraphDefResults {