TF_CAPI_EXPORT extern void TF_DeleteGraph(TF_Graph*);

// Operation being built. The underlying graph must outlive this.
typedef struct TF_OperationDescription TF_OperationDescription;

// Operation that has been added to the graph. Valid until the graph is
// deleted -- in particular adding a new operation to the graph does not
// invalidate old TF_Operation* pointers.
typedef struct TF_Operation TF_Operation;

  // The number of devices operations run on.
  size_t num_underlying_devices() const { return underlying_devices_.size(); }

  // The devices operations run on.
  const std::vector<std::string>& underlying_devices() const {
    return underlying_devices_;
  }

  // Takes a description of a single operation being executed on the
  // ParallelDevice, and in turn runs one operation per component device with

  ~DeviceThread();

  // Requests that the worker thread execute the specified operation. Blocks
  // until the previously pending operation (a StartExecute without a Join) has
  // finished, if any.
  //
  // `cancellation_manager` must live until after `Join` finishes and pending
  // `is_async` operations finish. In addition to allowing the caller to cancel
  // the operation, its `StartCancel` method will be called if op execution

#include <memory>

#include "tensorflow/c/eager/abstract_function.h"
#include "tensorflow/c/eager/abstract_operation.h"

namespace tensorflow {

// Abstract interface to a context.
//
// This serves as a factory for creating `AbstractOperation`s and for
// registering traced functions.
// Operations creation within a context can only be executed in that context
// (for now at least).

#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"

namespace tensorflow {
namespace parallel_device {

// Allocate a parallel device named `device_name` which forwards operations to
// `underlying_devices`, maintaining "parallel tensors" with components placed
// on each underlying device.
//
// For example if `device_name` is
//   "/job:localhost/replica:0/task:0/device:CUSTOM:0"

                   final_op.length());
  TF_SetAttrIntList(reduce_desc, "subdiv_offsets", nullptr, 0);
  TF_AddInput(reduce_desc, x);
  TF_Operation* reduce_op = TF_FinishOperation(reduce_desc, status);
  if (TF_GetCode(status) != TF_OK) return;
  TF_Operation* operations[]{placeholder_op, reduce_op};
  TF_Output y{reduce_op, 0};
  const char* output_name = "y";
  std::unique_ptr<TF_Function, decltype(&TF_DeleteFunction)> function(

    result.emplace(std::move(result_content));
    return result;
  } else if (operation_name == std::string("TPUReplicatedOutput")) {
    // Special-cased operation for un-packing one parallel tensor into
    // per-device tensors.
    OpPtr op(TFE_NewOp(context, operation_name, status));
    TFE_OpAddAttrs(op.get(), attributes);

#include "tensorflow/c/eager/parallel_device/parallel_device_lib.h"
#include "tensorflow/core/platform/test.h"

namespace tensorflow {
namespace parallel_device {

// A helper for performing common operations on variables. A much more
// restricted stand-in for tf.Variable in Python.
class Variable {
 public:
  // Construct a Variable from a resource-dtype TFE_TensorHandle and an

  }
  TFE_Context* context = TFE_OpGetContext(original_op, s);
  if (TF_GetCode(s) != TF_OK) return;
  const char* operation_name = TFE_OpGetName(original_op, s);
  if (TF_GetCode(s) != TF_OK) return;
  const TFE_OpAttrs* attributes = TFE_OpGetAttrs(original_op);

  TFE_Op* op(TFE_NewOp(context, operation_name, s));
  if (TF_GetCode(s) != TF_OK) return;
  TFE_OpAddAttrs(op, attributes);

    const TF_Graph* fn_body, const char* fn_name,
    unsigned char append_hash_to_fn_name, int num_opers,
    const TF_Operation* const* opers, int ninputs, const TF_Output* inputs,
    int noutputs, const TF_Output* outputs, const char* const* output_names,
    int ncontrol_outputs, const TF_Operation* const* control_outputs,
    const char* const* control_output_names, const TF_FunctionOptions* opts,