template <typename Gradient>
class ForwardFunction
    : public std::function<Status(const std::vector<Gradient*>&,
                                  std::vector<Gradient*>*, bool)> {
 public:
  template <typename lambda_type>
  explicit ForwardFunction(lambda_type lambda)
      : std::function<Status(const std::vector<Gradient*>&,
                             std::vector<Gradient*>*, bool)>(lambda) {}
};

namespace tensorflow {
namespace gradients {
namespace internal {
namespace {
using std::vector;

class CustomGradientTest
    : public ::testing::TestWithParam<std::tuple<const char*, bool, bool>> {
 protected:
  void SetUp() override {
    TF_StatusPtr status(TF_NewStatus());
    TF_SetTracingImplementation(std::get<0>(GetParam()), status.get());
    Status s = StatusFromTF_Status(status.get());

namespace tensorflow {
namespace parallel_device {

TEST(PARALLEL_DEVICE, TestRemoteBasic) {
  std::unique_ptr<TFE_ContextOptions, decltype(&TFE_DeleteContextOptions)> opts(
      TFE_NewContextOptions(), TFE_DeleteContextOptions);
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  std::unique_ptr<TFE_Context, decltype(&TFE_DeleteContext)> context(

  InferenceContext c(TF_GRAPH_DEF_VERSION, node_def, op_reg_data->op_def,
                     std::vector<ShapeHandle>(num_inputs), input_tensors_vector,
                     {},
                     std::vector<std::unique_ptr<std::vector<ShapeAndType>>>());

  // Set input_shapes.
  for (int i = 0; i < num_inputs; ++i) {
    std::vector<DimensionHandle> dims;
    const TF_ShapeAndType& input_shape = input_shapes->items[i];

Vendoring in std and cmd
========================

The Go command maintains copies of external packages needed by the
standard library in the src/vendor and src/cmd/vendor directories.

There are two modules, std and cmd, defined in src/go.mod and
src/cmd/go.mod. When a package outside std or cmd is imported
by a package inside std or cmd, the import path is interpreted
as if it had a "vendor/" prefix. For example, within "crypto/tls",

                           const string& op_name) {
  std::vector<int64_t> input_ids(inputs.size());
  std::vector<tensorflow::DataType> input_dtypes(inputs.size());
  for (int i = 0; i < inputs.size(); i++) {
    input_ids[i] = ToId(inputs[i]);
    input_dtypes[i] = inputs[i]->DataType();
  }
  std::vector<TapeTensor> tape_tensors;
  tape_tensors.reserve(outputs.size());
  for (auto t : outputs) {

  Status Shape(const std::vector<int64_t>** shape) const;
  TF_DataType dtype() const { return dtype_; }

  // Sets its output argument to a summary of the values of this tensor on every
  // component device.
  Status SummarizeValue(std::string& summary);

  std::vector<TensorHandlePtr> release_tensors() { return std::move(tensors_); }

  std::vector<TFE_TensorHandle*> tensors() const {

  }
}

absl::optional<std::vector<std::unique_ptr<ParallelTensor>>>
ParallelDevice::Join(
    const std::vector<PartialTensorShape>& expected_output_shapes,
    TF_Status* status) const {
  absl::optional<std::vector<std::unique_ptr<ParallelTensor>>> result;
  // Compute per-device per-output tensors
  std::vector<std::vector<TensorHandlePtr>> per_device_output_tensors;

 public:
  NamedParallelDevice(const std::string& name,
                      std::unique_ptr<ParallelDevice> parallel_device)
      : device_name_(name), parallel_device_(std::move(parallel_device)) {}
  const std::string& name() const { return device_name_; }
  const ParallelDevice& device() const { return *parallel_device_; }

 private:
  std::string device_name_;
  std::unique_ptr<ParallelDevice> parallel_device_;
};

    // Input tensors
    const std::vector<std::pair<string, Tensor>>& input_pairs,
    // Output tensors
    const std::vector<string>& output_tensor_names, TF_Tensor** c_outputs,
    // Target nodes
    const std::vector<string>& target_oper_names, TF_Buffer* run_metadata,
    TF_Status* status) {
  const int noutputs = output_tensor_names.size();
  std::vector<Tensor> outputs(noutputs);
  Status result;