ASSERT_FALSE(executed);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> matmul(
      MatMulOp(context, hcpu, hdevice), TFE_DeleteOp);
  TFE_OpSetDevice(matmul.get(), name, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_TensorHandle* retval;
  int num_retvals = 1;

  // API.
  std::string no_sharing = "cd2c89b7-88b7-44c8-ad83-06c2a9158347";
  TFE_OpSetAttrString(op.get(), "shared_name", no_sharing.c_str(),
                      no_sharing.length());
  TFE_OpSetDevice(op.get(), device, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_TensorHandle* var_handle = nullptr;
  int num_retvals = 1;
  TFE_Execute(op.get(), &var_handle, &num_retvals, status);

                                 const char* local_device_name) {
  TF_Status* status = TF_NewStatus();
  TFE_TensorHandle* h0_task0 = TestMatrixTensorHandle(ctx);

  TFE_Op* matmul = MatMulOp(ctx, h0_task0, h0_task0);
  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);

  }
  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);

    TFE_Op* matmul = MatMulOp(ctx, hcpu, hgpu);
    if (cpu_op) {
      string cpu_device_name;
      ASSERT_TRUE(GetDeviceName(ctx, &cpu_device_name, "CPU"));
      TFE_OpSetDevice(matmul, cpu_device_name.c_str(), status.get());
    } else {
      TFE_OpSetDevice(matmul, gpu_device_name.c_str(), status.get());
    }
    ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

    TFE_TensorHandle* retvals[1];

    if (TF_GetCode(status.get()) != TF_OK) return;
    TFE_OpSetAttrInt(op.get(), "num_replicas", 1);
    TFE_OpAddInput(op.get(), value_one.get(), status.get());
    if (TF_GetCode(status.get()) != TF_OK) return;
    TFE_OpSetDevice(op.get(), device_name, status.get());
    if (TF_GetCode(status.get()) != TF_OK) return;

    TFE_TensorHandle* result_handles;
    int num_retvals = 1;

  TFE_OpSetAttrInt(op.get(), "N", num_replicas);
  for (int i = 0; i < num_replicas; ++i) {
    TFE_OpAddInput(op.get(), components[i], status);
    if (TF_GetCode(status) != TF_OK) return nullptr;
  }
  TFE_OpSetDevice(op.get(), device, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;

  TFE_TensorHandle* result_handle;
  int num_retvals = 1;
  TFE_Execute(op.get(), &result_handle, &num_retvals, status);

  TFE_OpSetAttrShape(op, "shape", {}, 0, status);
  TFE_OpSetAttrString(op, "container", "localhost", 0);
  TFE_OpSetAttrString(op, "shared_name", "", 0);
  if (!device_name.empty()) {
    TFE_OpSetDevice(op, device_name.c_str(), status);
  }
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_TensorHandle* var_handle = nullptr;
  int num_retvals = 1;
  TFE_Execute(op, &var_handle, &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) {