TF_Status* status);

// Get GraphDebugInfo containing stack traces mapping to node names
TF_CAPI_EXPORT extern void TFE_ContextGetGraphDebugInfo(
    TFE_Context* ctx, const char* function_name, TF_Buffer* buf,
    TF_Status* status);

// Extracts a TF_Function from the context.
// Must call TF_DeleteFunction on the returned value.

TF_CAPI_EXPORT extern int TF_OperationNumOutputs(TF_Operation* oper);
TF_CAPI_EXPORT extern TF_DataType TF_OperationOutputType(TF_Output oper_out);
TF_CAPI_EXPORT extern int TF_OperationOutputListLength(TF_Operation* oper,
                                                       const char* arg_name,
                                                       TF_Status* status);

template <class T, TF_DataType datatype>
TFE_TensorHandle* TestTensorHandleWithDims(TFE_Context* ctx, const T* data,
                                           const int64_t* dims, int num_dims) {
  TF_Status* status = TF_NewStatus();
  TF_Tensor* t = TFE_AllocateHostTensor(ctx, datatype, dims, num_dims, status);
  memcpy(TF_TensorData(t), data, TF_TensorByteSize(t));

  TF_WritableFileHandle* h;
  TF_Status* status = TF_NewStatus();
  TF_NewWritableFile(full_path, &h, status);
  if (TF_GetCode(status) != TF_OK) {
    fprintf(stderr, "TF_NewWritableFile failed: %s\n", TF_Message(status));
    return 1;
  }
  fprintf(stderr, "wrote %s\n", full_path);
  free(full_path);
  TF_CloseWritableFile(h, status);
  if (TF_GetCode(status) != TF_OK) {

    const std::vector<Gradient*>& in_grads, absl::Span<Gradient*> out_grads) {
  /* This function is approximately equivalent to this Python code:

  forwardprop_aids = tf.ones_like(output_tensors)
  with tf.GradientTape() as g:
    g.watch(forwardprop_aids)
    grad = backward_function(forwardprop_aids)
  forward_grads = g.gradient(grad, forwardprop_aids, output_gradients=in_grads)

TF_CAPI_EXPORT extern TF_VariableInfo* TF_CreateVariableInfoFromContext(
    TF_OpKernelContext* ctx, int index, TF_Status* status);

TF_CAPI_EXPORT extern void TF_LockVariableInfos(TF_VariableInfo** vars,
                                                int num_vars,
                                                TF_Status* status);

TF_CAPI_EXPORT extern void TF_AllocateTempForVariableInfo(

class ModularFileSystem final : public FileSystem {
 public:
  ModularFileSystem(
      std::unique_ptr<TF_Filesystem> filesystem,
      std::unique_ptr<const TF_FilesystemOps> filesystem_ops,
      std::unique_ptr<const TF_RandomAccessFileOps> random_access_file_ops,
      std::unique_ptr<const TF_WritableFileOps> writable_file_ops,
      std::unique_ptr<const TF_ReadOnlyMemoryRegionOps>
          read_only_memory_region_ops,

  // in current TF runtime. For tfrt, it is used by fallback op handler.
  virtual void StartStep() = 0;
  // Destroy the step resource container for a training step.
  virtual void EndStep() = 0;

  // Return the Eager Executor for current thread. Please note that Eager
  // Executor is only used in current TF but not in TFRT.
  virtual EagerExecutor& Executor() = 0;

  // Set up a multi-client distributed execution environment. Must be called
  // on all tasks in the cluster. This call internally coordinates with other
  // tasks to initialize the eager context and TF server for multi-client
  // execution.
  virtual Status EnableCollectiveOps(const ServerDef& server_def) = 0;

  // Check if the remote task is alive.
  virtual Status CheckRemoteAlive(const std::string& remote_task_name,

  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  TFE_Context* eager_ctx =
      TF_ExecutionContextGetTFEContext(wrap(ctx), status.get());
  TF_RETURN_IF_ERROR(StatusFromTF_Status(status.get()));
  TFE_TensorHandle* input_eager =
      TestScalarTensorHandle<T, datatype>(eager_ctx, value);
  *tensor =