// Add version and compress.
	_, err := dst.Write([]byte{dataUsageCacheVerCurrent})
	if err != nil {
		return err
	}
	enc, err := zstd.NewWriter(dst,
		zstd.WithEncoderLevel(zstd.SpeedFastest),
		zstd.WithWindowSize(1<<20),
		zstd.WithEncoderConcurrency(2))
	if err != nil {
		return err
	}
	mEnc := msgp.NewWriter(enc)
	err = d.EncodeMsg(mEnc)
	if err != nil {
		return err

  absl::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.
  absl::Status SummarizeValue(std::string& summary);

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

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

 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_;
};

  vector_handles.push_back(std::move(two_vector));
  vector_handles.push_back(std::move(three_vector));
  std::unique_ptr<ParallelTensor> unknown_length_vector =
      ParallelTensor::FromTensorHandles(
          parallel_device, std::move(vector_handles), status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  const std::vector<int64_t>* shape;
  TF_ASSERT_OK(unknown_length_vector->Shape(&shape));

  }
}

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;

# By default, build TF in C++ 17 mode.
build:android --cxxopt=-std=c++17
build:android --host_cxxopt=-std=c++17
build:ios --cxxopt=-std=c++17
build:ios --host_cxxopt=-std=c++17
build:linux --cxxopt=-std=c++17
build:linux --host_cxxopt=-std=c++17
build:macos --cxxopt=-std=c++17
build:macos --host_cxxopt=-std=c++17
build:windows --cxxopt=/std:c++17
build:windows --host_cxxopt=/std:c++17

        "//tensorflow/core:session_options",
        "//tensorflow/core:test",
    ],
)

tf_cc_test(
    name = "c_test",
    srcs = ["c_test.c"],
    extra_copts = ["-std=c11"],
    deps = [
        ":c_api",
        ":c_api_experimental",
        ":env",
        ":kernels",
    ],
)

tf_cuda_cc_test(
    name = "c_api_test",
    size = "medium",

*   Build, Compilation and Packaging
    * TensorFlow is now compiled with `_GLIBCXX_USE_CXX11_ABI=1`. Downstream projects that encounter `std::__cxx11` or `[abi:cxx11]` linker errors will need to adopt this compiler option. See [the GNU C++ Library docs on Dual ABI](https://gcc.gnu.org/onlinedocs/libstdc++/manual/using_dual_abi.html).