}

  // Populate return_nodes
  DCHECK(tf_results->return_nodes.empty());
  tf_results->return_nodes.resize(results.return_nodes.size());
  for (int i = 0; i < results.return_nodes.size(); ++i) {
    tf_results->return_nodes[i] = ToOperation(results.return_nodes[i]);
  }

  // Populate missing unused map keys
  DCHECK(tf_results->missing_unused_key_names.empty());

    float grad_data[1];
    memcpy(&grad_data[0], TF_TensorData(grad_tensor),
           TF_TensorByteSize(grad_tensor));
    TF_DeleteTensor(grad_tensor);
    dtheta_approx[i] = grad_data[0];
  }

  // Populate *numerical_grad with the data from dtheta_approx.
  TF_RETURN_IF_ERROR(TestTensorHandleWithDims<float, TF_FLOAT>(
      ctx, dtheta_approx.data(), theta_dims.data(), num_dims, numerical_grad));
  TF_DeleteTensor(theta_tensor);

TF_CAPI_EXPORT extern void TFE_ContextEnableRunMetadata(TFE_Context* ctx);

// Disables tracing of RunMetadata on the ops executed from this context.
TF_CAPI_EXPORT extern void TFE_ContextDisableRunMetadata(TFE_Context* ctx);

// Populates the passed-in buffer with a serialized RunMetadata protocol buffer
// containing any run metadata information accumulated so far and clears this
// information.

  // Consider getting rid of this and making the behavior between number types
  // and string consistent.
  forward_op_.attrs.BuildNodeDef();
  // TODO(b/170307493): Populate skip_input_indices here.
  std::unique_ptr<GradientFunction> backward_fn;
  TF_RETURN_IF_ERROR(registry_.Lookup(forward_op_, &backward_fn));
  tape_->RecordOperation(forward_op_.inputs, forward_op_.outputs,

  // Finalize the function by providing the returned values.
  TF_AbstractFunction* func;
  {
    // We want to return the output of both add operations, create a new list
    // and populate it.
    TF_OutputList* func_outputs = TF_NewOutputList();
    TF_OutputListPushBack(func_outputs, add_output1, s);
    ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

TF_CAPI_EXPORT extern uint64_t TF_NowMicros(void);

// Returns the number of seconds since the Unix epoch.
TF_CAPI_EXPORT extern uint64_t TF_NowSeconds(void);

// Populates a TF_ThreadOptions struct with system-default values.
TF_CAPI_EXPORT extern void TF_DefaultThreadOptions(TF_ThreadOptions* options);

// Returns a new thread that is running work_func and is identified

                                    "Encountered while creating function '",
                                    fn_name, "'");
    output_tensors->emplace_back(node, idx);
  }
  return absl::OkStatus();
}

// Populates `body_nodes` with the nodes that will become function's body.
// Performs various checks.
Status ComputeBodyNodes(
    const TF_Graph* fn_body, const char* fn_name, int num_opers,
    const TF_Operation* const* opers,

//
// Design decisions made:
// * Backing store for tensor memory has an associated deallocation
//   function.  This deallocation function will point to client code
//   for tensors populated by the client.  So the client can do things
//   like shadowing a numpy array.
// * We do not provide TF_OK since it is not strictly necessary and we
//   are not optimizing for convenience.

  // once) and produces the gradient of the target tensors with respect to the
  // source tensors. The output gradients are used if not empty and not
  // null. The result is populated with one tensor per target element.
  // When running backward functions, builds zeros-like tensors for
  // incoming grads which are nullptrs, unless `build_default_zeros_grads`
  // is set to false.

//     core and the plugin.
//     * For example, `TF_InitGraph` function receives
//       `TP_OptimizerRegistrationParams*` as input with `struct_size`
//       populated by core. The plugin is responsible for setting
//       `struct_size` as well, along with all other fields.
//     * Refer to "TensorFlow Versioning Strategy" section at