const void** buf, size_t* len);

// TODO(josh11b,mrry): Change Session to be able to use a Graph*
// directly, instead of requiring us to serialize to a GraphDef and
// call Session::Extend().
bool ExtendSessionGraphHelper(TF_Session* session, TF_Status* status) {
  if (session->graph != nullptr) {
    // Take the graph lock before the session lock to avoid deadlock. This is

    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 {}
};

           /*dep=*/{"assert"}},
      });
  TF_ASSERT_OK(ContextFromInterface(unwrap(context.get()))
                   ->AddFunctionDef(assert_and_collective));

  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> call_op(
      TFE_NewOp(context.get(), "AssertAndCollective", status.get()),
      TFE_DeleteOp);
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  std::unique_ptr<ParallelTensor> reduced_values =

  CheckRemoteMatMulExecutesOK(ctx, remote_device_name, local_device_name);

  // Adding a non-existent remote worker to cluster def. This should cause the
  // UpdateServerDef call to fail.
  tensorflow::ClusterDef* cluster_def = server_def.mutable_cluster();
  tensorflow::JobDef* job_def = cluster_def->mutable_job(0);
  int port = tensorflow::testing::PickUnusedPortOrDie();

  using tensorflow::AttrBuilder::AttrBuilder;
  // The string buffers to make sure that any `attr_name` we pass into
  // `builder->Set()` will outlive the subsequent
  // `TF_AttrBuilderCheckCanRunOnDevice()` call(s) on the same `builder`.
  std::set<std::string> attr_names;
};

TF_AttrBuilder* TF_NewAttrBuilder(const char* op_name) {
  return new TF_AttrBuilder(op_name);
}

                                    TF_Graph* graph, TF_Operation* ctrl_op,
                                    TF_Status* s, const char* name) {
  TF_OperationDescription* desc = TF_NewOperation(graph, "AddN", name);
  TF_Output add_inputs[2] = {{l, 0}, {r, 0}};
  TF_AddInputList(desc, add_inputs, 2);
  TF_AddControlInput(desc, ctrl_op);
  return TF_FinishOperation(desc, s);
}

  auto tf_dlm_type = GetDlDataType(data_type, status);
  if (!status->status.ok()) {
    return nullptr;
  }

  TensorReference tensor_ref(*tensor);  // This will call buf_->Ref()
  auto* tf_dlm_tensor_ctx = new TfDlManagedTensorCtx(tensor_ref);
  tf_dlm_tensor_ctx->reference = tensor_ref;

  DLManagedTensor* dlm_tensor = &tf_dlm_tensor_ctx->tensor;

      output_gradients, result, /*build_default_zeros_grads*/ false));
  return absl::OkStatus();
}

// Helper functions which delegate to `AbstractOperation`, update
// the state of the ForwardOperation and call the tape as appropriate.
// These APIs are mainly to facilitate testing and are subject to change.
namespace internal {
Status Reset(AbstractOperation* op_, const char* op,

    if (!enabled_) {
      return absl::OkStatus();
    }
    if (first_call_) {
      first_call_ = false;
      return absl::OkStatus();
    }
    return tensorflow::errors::Internal("Graph pass runs for more than once!");
  }

 private:
  bool first_call_ = true;
};

// After the graph pass is registered, it takes effect globally and can affect

  // The resulting TensorHandle owns an opaque pointer to "device memory", which
  // for a ParallelDevice is really a ParallelTensor. When the TensorHandle is
  // deleted, it will call ParallelTensorDeallocator to free the struct.
  ParallelTensor* t_released = t.release();
  TFE_CustomDeviceTensorHandleMethods handle_methods;
  handle_methods.num_dims = &ParallelTensorNumDims;