parallel_device.Execute(context.get(), handle_inputs, "ReadVariableOp",
                          TFE_OpGetAttrs(read_op.get()),
                          /*expected_max_outputs=*/1, status.get());
  ASSERT_FALSE(TF_GetCode(status.get()) == TF_OK);
  TF_SetStatus(status.get(), TF_OK, "");

  // Check that ops still run successfully on the device.
  parallel_device.Execute(context.get(), std::vector<ParallelTensor*>(),

  TFE_DeleteTensorHandle(h0_task0);
  TFE_DeleteTensorHandle(retvals[0]);

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TF_DeleteStatus(status);
}

// Read the value of variable `var` and save it into `out_value`.

  TFE_DeleteTensorHandle(h1_task1);
  TFE_DeleteTensorHandle(retvals[0]);

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TFE_DeleteContext(ctx);

  TF_DeleteStatus(status);

    if (TF_GetCode(s) != TF_OK) {
      return;
    }
  }
  int num_outputs = unwrap(o)->expected_num_outputs;
  tsl::Set_TF_Status_from_Status(
      s, unwrap(op)->Execute(
             absl::MakeSpan(reinterpret_cast<AbstractTensorHandle**>(
                                unwrap(o)->outputs.data()),
                            unwrap(o)->outputs.size()),
             &num_outputs));
}

  TFE_TensorHandle* var_handle = nullptr;
  int num_retvals = 1;
  executed = false;
  TFE_Execute(op.get(), &var_handle, &num_retvals, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  ASSERT_TRUE(executed);
  auto handle_cleaner = tensorflow::gtl::MakeCleanup(
      [var_handle]() { TFE_DeleteTensorHandle(var_handle); });

      "SetAttrFunctionList has not been "
      "implemented yet.");
}
Status Execute(AbstractOperation* op_, AbstractContext* ctx,
               absl::Span<AbstractTensorHandle*> retvals, int* num_retvals,
               ForwardOperation* forward_op_, Tape* tape,
               const GradientRegistry& registry) {
  TF_RETURN_IF_ERROR(op_->Execute(retvals, num_retvals));
  for (int i = 0; i < *num_retvals; i++) {

  // Block until the previous `StartExecute` operation has executed. Forwards
  // the status from `TFE_Execute` and returns outputs if the status is OK.
  std::vector<TensorHandlePtr> Join(TF_Status* status);

  // Block until all Ops finished running on the thread.
  void AsyncWait(TF_Status* status);

 private:
  void Run();

  void Execute(TFE_Context* context, const char* operation_name,

#include "tensorflow/c/c_api_experimental.h"
#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {

  custom_device.copy_tensor_from_device = &CopyTensorFromLoggingDevice;
  custom_device.delete_device = &DeleteLoggingDevice;
  custom_device.execute = &LoggingDeviceExecute;
  LoggingDevice* device = new LoggingDevice;
  device->arrived_flag = arrived_flag;
  device->executed_flag = executed_flag;
  device->device_name = name;
  device->underlying_device = "/job:localhost/replica:0/task:0/device:CPU:0";

  TFE_DeleteTensorHandle(h0);
  TFE_DeleteTensorHandle(h1);
  TFE_DeleteTensorHandle(h2);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TFE_ContextRemoveFunction(ctx, "AddVariablesFunction", status);
  TFE_DeleteContext(ctx);