// or removed during the graph transformation. This includes feed and fetch
// nodes, keep_ops, init_ops. Fills in `num_values` and `storage_size`, they
// will be used in `TF_GetNodesToPreserveList`.
TF_CAPI_EXPORT extern void TF_GetNodesToPreserveListSize(
    const TF_GrapplerItem* item, int* num_values, size_t* storage_size,
    TF_Status* status);

// Get a set of node names that must be preserved. They can not be transformed

                                                 TF_Status* s);
void TF_DeleteExecutionContext(TF_ExecutionContext*);

// Represents a (partially-defined) shape.
typedef struct TF_Shape {
  int num_dims;  // Must be >= -1; -1 represents unknown rank.
  int64_t* dim_sizes;
} TF_Shape;

// Add a new parameter to a TensorFlow Function.
TF_AbstractTensor* TF_AddFunctionParameter(TF_ExecutionContext* func,

// Helper to un-pack `num_replicas` TFE_TensorHandles from one parallel handle.
template <std::size_t num_replicas>
void ExtractPerDeviceValues(
    TFE_Context* context, TFE_TensorHandle* input,
    std::array<TensorHandlePtr, num_replicas>* components, TF_Status* status);

// Helper to pack `num_replicas` TFE_TensorHandles into one parallel handle.
template <std::size_t num_replicas>
TensorHandlePtr CreatePerDeviceValues(

// Create a ServerDef with the given `job_name` and add `num_tasks` tasks in it.
tensorflow::ServerDef GetServerDef(const tensorflow::string& job_name,
                                   int num_tasks);

// Create a ServerDef with job name "localhost" and add `num_tasks` tasks in it.
tensorflow::ServerDef GetServerDef(int num_tasks);

// Create a multi-client ServerDef with the given `job_name`, add `num_tasks`

// Custom devices must use TFE_NewCustomDeviceTensorHandle instead.
TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_NewTensorHandleFromDeviceMemory(
    TFE_Context* ctx, const char* device_name, TF_DataType, const int64_t* dims,
    int num_dims, void* data, size_t len,
    void (*deallocator)(void* data, size_t len, void* arg),
    void* deallocator_arg, TF_Status* status);

// Retrieves the address space (i.e. job, replia, task) of the local host and

                                           const int64_t* dims, int num_dims);
// `dims` and `num_dims` must point to arrays of length `num_shapes`.
// Set `num_dims[i]` to -1 to represent "unknown rank".  Otherwise,
// `dims[i]` points to an array of length `num_dims[i]`.  `dims[i][j]`
// must be >= -1, with -1 meaning "unknown dimension".
TF_CAPI_EXPORT extern void TF_SetAttrShapeList(TF_OperationDescription* desc,

  if (values.size() != num_underlying_devices()) {
    TF_SetStatus(
        status, TF_INVALID_ARGUMENT,
        "Number of values did not match number of underlying devices.");
    return nullptr;
  }
  TF_DataType datatype_enum(
      static_cast<TF_DataType>(DataTypeToEnum<DataType>().value));
  for (int device_index = 0; device_index < num_underlying_devices();
       ++device_index) {

// 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
// passed to TFE_RegisterCustomDevice. The `device_name` arguments must match.

                           const size_t* lengths, int num_values) override;
  Status SetAttrFloatList(const char* attr_name, const float* values,
                          int num_values) override;
  Status SetAttrIntList(const char* attr_name, const int64_t* values,
                        int num_values) override;
  Status SetAttrTypeList(const char* attr_name, const DataType* values,
                         int num_values) override;

// b) ConfigProto.gpu_options.allow_growth is set to `gpu_memory_allow_growth`.
// c) ConfigProto.device_count is set to `num_cpu_devices`.
TF_CAPI_EXPORT extern TF_Buffer* TF_CreateConfig(
    unsigned char enable_xla_compilation, unsigned char gpu_memory_allow_growth,
    unsigned int num_cpu_devices);

// Create a serialized tensorflow.RunOptions proto, where RunOptions.trace_level