absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    /* Given upstream grad U and a BiasAdd: A + bias, the gradients are:
     *
     *    dA = U
     *    dbias = reduceSum(U, dims = channel_dim)
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];
    DCHECK(upstream_grad);

  return nullptr;
}

}  // end namespace tracing
}  // end namespace tensorflow

// =============================================================================
// Public C API entry points
//
// These are only the generic entry points for the C API. This file does not
// have any visibility into the graph/eager implementation and is only providing
// C bindings to the abstract classes defined in the

// `src_inputs[i]` will have that input replaced with `dst_inputs[i]`.  `prefix`
// will be prepended to copied node names. `control_deps` are nodes in
// `dst_graph` that the copied `src_graph` nodes will have control dependencies
// on. `return_nodes` are nodes in `src_graph`, and the new corresponding nodes
// in `dst_graph` will be returned. `return_nodes` must be non-null.

    for (int i = 0; i < grad_outputs.size(); i++) {
      auto grad_input = grad_outputs[i];
      // TODO(srbs): Should we add a copy contructor to AbstractTensorHandle
      // that takes care of this similar to `Tensor`?
      if (grad_input) {
        grad_input->Ref();
      }
      grad_inputs[i] = grad_input;
    }
    return absl::OkStatus();
  }
  ~IdentityNGradientFunction() override {}
};

  auto* optimizer_options =
      config.mutable_graph_options()->mutable_optimizer_options();
  if (enable) {
    optimizer_options->set_global_jit_level(tensorflow::OptimizerOptions::ON_1);

    // These XLA flags are needed to trigger XLA properly from C (more generally
    // non-Python) clients. If this API is called again with `enable` set to
    // false, it is safe to keep these flag values as is.

    return;
  } else if (S_ISDIR(st.st_mode)) {
    TF_SetStatus(status, TF_FAILED_PRECONDITION, "source path is a directory");
    return;
  }

  // Do the actual rename. Here both arguments are filenames.
  if (rename(src, dst) != 0)
    TF_SetStatusFromIOError(status, errno, dst);
  else
    TF_SetStatus(status, TF_OK, "");
}

static void CopyFile(const TF_Filesystem* filesystem, const char* src,

  tensorflow::condition_variable finished_join_;

  // Temporary state between `StartExecute` and `Join`.
  //
  //   Inputs; pointers are to objects not owned by the DeviceThread, but which
  //   are expected to live at least until `Join` finishes:
  TFE_Context* context_ TF_GUARDED_BY(execution_mutex_);
  const char* operation_name_ TF_GUARDED_BY(execution_mutex_);

  for (int i{}; i < inputs.size(); ++i) {
    theta_inputs[i] = inputs[i];
  }

  AbstractTensorHandle* theta =
      theta_inputs[input_index];  // parameter we are grad checking

  // Convert from AbstractTensor to TF_Tensor.
  TF_Tensor* theta_tensor;
  TF_RETURN_IF_ERROR(GetValue(theta, &theta_tensor));

  // Get number of elements and fill data.

  // Pack 3 variable handles into one TFE_TensorHandle.
  // When remote is false, function device is placed on task0. Handle types are
  // REMOTE, REMOTE, LOCAL on task0. When remote is true, function device is
  // placed on task1, Handle types are LOCAL, REMOTE, LOCAL on task1.
  int num_replicas = 3;
  std::vector<TFE_TensorHandle*> handles = {h0, h1, h2};
  TFE_TensorHandle* packed_handle =

  EXPECT_TRUE(absl::StrContains(s, expected))
      << "'" << s << "' does not contain '" << expected << "'";
}

// Returns the GPU device name if there is one (with arbitrary tie breaking if
// there are more than one), or "" otherwise.
string GPUDeviceName(TF_Session* session) {
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  TF_Status* s = status.get();