int version = TFE_CUSTOM_DEVICE_VERSION;
  // Method to copy a tensor to the custom device.
  TFE_TensorHandle* (*copy_tensor_to_device)(TFE_Context* context,
                                             TFE_TensorHandle* tensor,
                                             TF_Status* status,
                                             void* device_info);

  // Method to copy a tensor from the custom device to a target device.

#include <memory>

#include "absl/types/optional.h"
#include "absl/types/span.h"
#include "tensorflow/c/eager/abstract_operation.h"
#include "tensorflow/c/eager/immediate_execution_tensor_handle.h"
#include "tensorflow/c/tensor_interface.h"
#include "tensorflow/core/framework/cancellation.h"
#include "tensorflow/core/framework/device_attributes.pb.h"
#include "tensorflow/core/framework/op_def.pb.h"

#ifndef TENSORFLOW_C_EAGER_TFE_TENSORHANDLE_INTERNAL_H_
#define TENSORFLOW_C_EAGER_TFE_TENSORHANDLE_INTERNAL_H_

#include "tensorflow/c/conversion_macros.h"
#include "tensorflow/c/eager/immediate_execution_tensor_handle.h"

// Wraps a pointer to a tensor handle implementation.
//
// WARNING: Since the underlying object could be ref-counted a user of this
// interface cannot destruct the underlying handle object. Instead, call

int TF_OutputListNumOutputs(TF_OutputList* o);
// Return the `i`th output in the list.
TF_AbstractTensor* TF_OutputListGet(TF_OutputList* o, int i);
// Append a tensor at the end of the output list, growing its size by one.
void TF_OutputListPushBack(TF_OutputList* o, TF_AbstractTensor* tensor,
                           TF_Status*);

// TF_ExecuteOperation will, if in eager mode, execute, if in graph mode, maybe

#ifndef TENSORFLOW_C_EAGER_GRADIENT_CHECKER_H_
#define TENSORFLOW_C_EAGER_GRADIENT_CHECKER_H_

#include <memory>

#include "absl/types/span.h"
#include "tensorflow/c/eager/abstract_tensor_handle.h"
#include "tensorflow/c/eager/unified_api_testutil.h"

namespace tensorflow {
namespace gradients {

/* Returns numerical grad inside `dtheta_approx` given `forward` model and

// Return an allreduce op adding up input tensor `in` from `group_size` workers.
TFE_Op* AllReduceOp(TFE_Context* ctx, TFE_TensorHandle* in, int group_size);

// Return a SendOp op `op_name` with send input tensor `in` and attributes
// `send_device`, `recv_device`, and `send_device_incarnation` set.
TFE_Op* SendOp(TFE_Context* ctx, TFE_TensorHandle* in,

                                                       TF_Status* status) const;

  // Construct a parallel tensor consisting of the scalar values from `values`.
  template <typename DataType>
  std::unique_ptr<ParallelTensor> ScalarsFromSequence(
      absl::Span<const DataType> values, TFE_Context* context,
      TF_Status* status) const;

                                                     TF_Tensor** value,
                                                     TF_Status* status);

// Fills in `values` with the TF_Tensor values of the attribute `attr_name` of
// `oper`. `values` must point to an array of TF_Tensor* of length at least
// `max_values` (ideally set to TF_AttrMetadata.list_size from

// two component devices, running `x = TPUReplicatedInput(inputs=[a, b])` on the
// parallel device creates a parallel tensor `x` with `a` on the first of
// `underlying_devices` and `b` on the second. Running `a_unpacked, b_unpacked =
// TPUReplicatedOutput(input=x, num_replicas=2)` un-packs the parallel tensor
// into its components.
//
// The filled `device` struct and the allocated `device_info` struct may be

  // AssignSub, ...).
  void GeneralAssignment(const char* op_name, TFE_Context* context,
                         TFE_TensorHandle* value, TF_Status* status);

  // The a handle for the resource-dtype tensor pointing to the variable's
  // buffer.
  TFE_TensorHandle* handle_;
  // The dtype of the variable's buffer (input dtype for assignments, output
  // dtype of read operations).
  TF_DataType type_;
};