}

void TFE_HostAddressSpace(TFE_Context* ctx, TF_Buffer* buf) {
  auto address_space = tensorflow::DeviceNameUtils::AddressSpace(
      tensorflow::unwrap(ctx)->HostCPUParsedName());
  auto str = tensorflow::DeviceNameUtils::ParsedNameToString(address_space);
  void* data = tensorflow::port::Malloc(str.length());
  str.copy(static_cast<char*>(data), str.length(), 0);
  buf->data = data;

  /// Plugins:
  ///   * Must set `status` to `TF_OK` if exactly `n` bytes have been written.
  ///   * Must set `status` to `TF_RESOURCE_EXHAUSTED` if fewer than `n` bytes
  ///     have been written, potentially due to quota/disk space.
  ///   * Might use any other error value for `status` to signal other errors.
  void (*append)(const TF_WritableFile* file, const char* buffer, size_t n,
                 TF_Status* status);

TF_CAPI_EXPORT extern void TFE_InferShapes(
    TFE_Op* op, TF_ShapeAndTypeList* input_shapes, TF_Tensor** input_tensors,
    TF_ShapeAndTypeList* input_tensor_as_shapes,
    TF_ShapeAndTypeList** input_resource_shapes_and_types,
    TF_ShapeAndTypeList** output_shapes,
    TF_ShapeAndTypeList*** output_resource_shapes_and_types, TF_Status* status);

TF_CAPI_EXPORT extern void
TF_ImportGraphDefOptionsSetValidateColocationConstraints(

void TFE_InferShapes(TFE_Op* tfe_op, TF_ShapeAndTypeList* input_shapes,
                     TF_Tensor** input_tensors,
                     TF_ShapeAndTypeList* input_tensors_as_shapes,
                     TF_ShapeAndTypeList** input_resource_shapes_and_types,
                     TF_ShapeAndTypeList** output_shapes,
                     TF_ShapeAndTypeList*** output_resource_shapes_and_types,
                     TF_Status* status) {

                               /*expected_max_outputs=*/1,
                               cancellation_manager);
  auto outputs = parallel_device.Join(
      /*expected_output_shapes=*/{PartialTensorShape({})}, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  const std::vector<std::unique_ptr<ParallelTensor>>& handles = *outputs;

class ForwardAccumulator {
 public:
  // Does not take ownership of `vspace`, which must outlive the
  // ForwardAccumulator.
  explicit ForwardAccumulator(
      const VSpace<Gradient, BackwardFunction, TapeTensor>& vspace,
      bool use_batch)
      : vspace_(vspace), use_batch_(use_batch) {
    call_state_.emplace(nullptr, false);
  }

  virtual ~ForwardAccumulator() {

// Get the OpList of all OpDefs defined in this address space.
// Returns a TF_Buffer, ownership of which is transferred to the caller
// (and can be freed using TF_DeleteBuffer).
//
// The data in the buffer will be the serialized OpList proto for ops registered
// in this address space.
TF_CAPI_EXPORT extern TF_Buffer* TF_GetAllOpList(void);

  // run and have results. If is_async=true it means that all of the
  // device-specific executors have scheduled the op.
  //
  // Accepts inferred shapes for outputs (`expected_output_shapes`), which if
  // fully defined will avoid querying the shapes of the underlying
  // TensorHandles when ParallelTensor::Shape is called. This allows async
  // computation to continue without blocking.
  //

  int num_elems = TF_TensorElementCount(theta_tensor);
  vector<float> theta_data(num_elems);
  memcpy(theta_data.data(), TF_TensorData(theta_tensor),
         TF_TensorByteSize(theta_tensor));

  // Initialize space for the numerical gradient.
  vector<float> dtheta_approx(num_elems);

  // Get theta shape and store in theta_dims.
  int num_dims = TF_NumDims(theta_tensor);
  vector<int64_t> theta_dims(num_dims);

typedef struct TF_ThreadOptions {
  // Thread stack size to use (in bytes), zero implies that the system default
  // will be used.
  size_t stack_size;

  // Guard area size to use near thread stacks to use (in bytes), zero implies
  // that the system default will be used.
  size_t guard_size;

  // The NUMA node to use, -1 implies that there should be no NUMA affinity for
  // this thread.
  int numa_node;