}
  if (remote) {
    TFE_OpSetDevice(matmul, task1_name, status);
    ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  } else if (!async) {
    // Set the local device to CPU to easily validate mirroring
    string cpu_device_name;
    ASSERT_TRUE(GetDeviceName(ctx, &cpu_device_name, "CPU"));
    TFE_OpSetDevice(matmul, cpu_device_name.c_str(), status);

  auto* h1_task1 =
      TFE_TensorHandleCopyToDevice(h1_task0, ctx, remote_device_name, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_Op* matmul = MatMulOp(ctx, h0_task1, h1_task1);
  TFE_OpSetDevice(matmul, remote_device_name, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);

  const TFE_OpAttrs* attributes = TFE_OpGetAttrs(original_op);

  TFE_Op* op(TFE_NewOp(context, operation_name, s));
  if (TF_GetCode(s) != TF_OK) return;
  TFE_OpAddAttrs(op, attributes);
  TFE_OpSetDevice(op, dev->underlying_device.c_str(), s);
  if (TF_GetCode(s) != TF_OK) return;
  int num_inputs = TFE_OpGetFlatInputCount(original_op, s);
  if (TF_GetCode(s) != TF_OK) return;
  for (int j = 0; j < num_inputs; ++j) {