const char task2_name[] = "/job:localhost/replica:0/task:2/device:CPU:0";

  // Create one variable per task.
  TFE_TensorHandle* h0 = TestVariable(ctx, 1.0, task1_name);
  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);
  TFE_TensorHandle* h2 = TestVariable(ctx, 3.0, task0_name);

  // Add a sync point to make sure that variables have been initialized
  // before the function execution starts.

  {
    // Try to pass a ParallelTensor to TPUReplicatedInput
    std::array<TFE_TensorHandle*, 1> correct_components{value_one.get()};
    TensorHandlePtr combined_value = CreatePerDeviceValues(
        context.get(), correct_components, device_name, status.get());
    ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

    std::array<TFE_TensorHandle*, 1> incorrect_components{combined_value.get()};

      /*is_async=*/async, /*enable_streaming_enqueue=*/true,
      /*in_flight_nodes_limit=*/0);
  TFE_ContextSetExecutorForThread(ctx, executor);

  TFE_TensorHandle* m = TestMatrixTensorHandle(ctx);
  TFE_Op* matmul = MatMulOp(ctx, m, m);
  TFE_TensorHandle* retvals[2] = {nullptr, nullptr};
  int num_retvals = 2;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(1, num_retvals);

  TFE_TensorHandle* (*copy_tensor_to_device)(TFE_Context* context,
                                             TFE_TensorHandle* tensor,
                                             TF_Status* status,
                                             void* device_info);

  // Method to copy a tensor from the custom device to a target device.
  TFE_TensorHandle* (*copy_tensor_from_device)(TFE_Context* context,

  const char dev1_name[] = "/job:localhost/replica:0/task:1/device:CPU:0";

  TFE_TensorHandle* var_handle0 = TestVariable(ctx, 1.0, dev0_name);
  EXPECT_NE(var_handle0, nullptr);
  TFE_TensorHandle* var_handle1 = TestVariable(ctx, 2.0, dev1_name);
  EXPECT_NE(var_handle1, nullptr);

  TFE_TensorHandle* value_handle = nullptr;
  ReadVariable(ctx, var_handle1, &value_handle);

  TFE_ContextOptions* opts = TFE_NewContextOptions();
  TFE_Context* ctx = TFE_NewContext(opts, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteContextOptions(opts);

  TFE_TensorHandle* h = TestScalarTensorHandle(ctx, 1.0f);
  TFE_TensorDebugInfo* debug_info = TFE_TensorHandleTensorDebugInfo(h, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Build an abstract input tensor.
  TFE_Context* eager_ctx = TF_ExecutionContextGetTFEContext(ctx, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
  TFE_TensorHandle* t = TestScalarTensorHandle(eager_ctx, 2.0f);
  TF_AbstractTensor* at =
      TF_CreateAbstractTensorFromEagerTensor(t, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

// void* for further PyCapsule construction.
TF_CAPI_EXPORT extern void* TFE_HandleToDLPack(TFE_TensorHandle* h,
                                               TF_Status* status);

// Converts DLPack (DLManagedTensor*) to eager tensor handle.
TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_HandleFromDLPack(void* dlm,
                                                             TF_Status* status,

  TFE_DeleteContextOptions(opts);

  TFE_ContextSetServerDef(ctx, 0, serialized.data(), serialized.size(), status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* h0_task0 = TestMatrixTensorHandle(ctx);
  TFE_TensorHandle* h1_task0 = TestMatrixTensorHandle(ctx);
  const char remote_device_name[] =
      "/job:localhost/replica:0/task:1/device:CPU:0";
  auto* h0_task1 =

  DLManagedTensor tensor;

  explicit TfDlManagedTensorCtx(const TensorReference& ref) : reference(ref) {}
};

// Gets tensor from eager tensor handle.
const Tensor* GetTensorFromHandle(TFE_TensorHandle* h, TF_Status* status) {
  if (h == nullptr) {
    status->status = tensorflow::errors::InvalidArgument("Invalid handle");
    return nullptr;
  }
  tensorflow::TensorHandle* handle =