//
// The caller of c++ interfaces should make sure to pass valid serialized
// `ProtoT` objects as arguments. Failing to do so results in raising a
// `ValueError`. Similarly, the python implementation of a c++ virtual member
// function that return an `ProtoT` should return a valid serialized `ProtoT`.
//
// See https://pybind11.readthedocs.io/en/stable/advanced/cast/custom.html

  }
  DeviceCompilerClient<ExecutableType, ClientType>* compiler_client() {
    return compiler_client_.get();
  }

  string DebugString() const override;

 private:
  // Common implementation of Compile and CompileSingleOp. The `OpKernelContext`
  // parameter is always null for the former.
  Status CompileImpl(
      const XlaCompiler::CompileOptions& compile_options,

                                              TF_Operation* input);

// Request that `desc` be co-located on the device where `op`
// is placed.
//
// Use of this is discouraged since the implementation of device placement is
// subject to change. Primarily intended for internal libraries
TF_CAPI_EXPORT extern void TF_ColocateWith(TF_OperationDescription* desc,
                                           TF_Operation* op);

 *                Removed CUNIT_MAX_STRING_LENGTH - dangerous since not enforced.
 *                Fixed off-by-1 error in CU_translate_special_characters(), 
 *                modifying implementation & results in some cases.  User can 
 *                now tell if conversion failed. (JDS)
 */

/** @file
 *  Utility functions (user interface).
 */
/** @addtogroup Framework
 * @{
 */

// (1) TF_Output or TFE_TensorHandle* as arguments to TF_AddInput,
//     TF_AddInputList
// (2) TF_ColocateWith, TF_AddControlInput etc. do not make sense.
// (3) Implementation detail: Avoid use of NodeBuilder/NodeDefBuilder since
//     the additional sanity checks there seem unnecessary;
typedef struct TFE_Op TFE_Op;

TF_CAPI_EXPORT extern TFE_Op* TFE_NewOp(TFE_Context* ctx,

 private:
  xla::LocalClient* client_;
  se::DeviceMemoryAllocator* xla_allocator_;
  bool allocate_xla_tensors_;
  bool use_multiple_streams_;
  int device_ordinal_;
};

// A simple TensorBuffer implementation that allows us to create Tensors that
// take ownership of pre-allocated memory.
class XlaTensorBuffer : public TensorBuffer {
 public:
  XlaTensorBuffer(const void* ptr, size_t expected_size, size_t actual_size,

// *containing* linked_ptr<> must have a constructor and destructor (even
// if they do nothing!).
//
// Bill Gibbons suggested we use something like this.
//
// Thread Safety:
//   Unlike other linked_ptr implementations, in this implementation
//   a linked_ptr object is thread-safe in the sense that:
//     - it's safe to copy linked_ptr objects concurrently,
//     - it's safe to copy *from* a linked_ptr and read its underlying

  // from an XLA cluster.
  bool tf_xla_check_cluster_output_numerics;

  // Disables all constant folding. The primary use for this is for testing to
  // guarantee that tests are run on XLA and not on TF's CPU implementation.
  bool tf_xla_disable_constant_folding;

  // Disables full embedding pipelining when true. Instead, strict SparseCore
  // TensorCore sequencing will be used.
  bool tf_xla_disable_full_embedding_pipelining;

    typename T6 = void, typename T7 = void, typename T8 = void,
    typename T9 = void>
class tuple;

// Anything in namespace gtest_internal is Google Test's INTERNAL
// IMPLEMENTATION DETAIL and MUST NOT BE USED DIRECTLY in user code.
namespace gtest_internal {

// ByRef<T>::type is T if T is a reference; otherwise it's const T&.
template <typename T>

// A user can override this behavior for a class type Foo by defining
// a << operator in the namespace where Foo is defined.
//
// We put this operator in namespace 'internal2' instead of 'internal'
// to simplify the implementation, as much code in 'internal' needs to
// use << in STL, which would conflict with our own << were it defined
// in 'internal'.
//
// Note that this operator<< takes a generic std::basic_ostream<Char,