auto allocator = std::make_unique<AsyncValueAllocator>();
  tensorflow::Tensor tensor(allocator.get(), DT_FLOAT, {1});
  TF_ASSERT_OK_AND_ASSIGN(
      auto pjrt_client,
      xla::GetTfrtCpuClient(/*asynchronous=*/true, /*cpu_device_count=*/1));
  std::vector<int32_t> data(1, 0);
  xla::Shape shape = xla::ShapeUtil::MakeShape(xla::S32, {1});
  TF_ASSERT_OK_AND_ASSIGN(
      auto buffer,
      pjrt_client->BufferFromHostBuffer(

  ExpectScalarEq<float>(result_components[1].get(), 3.);
}

TEST(PARALLEL_DEVICE, TestCollectiveSync) { TestCollective(/*async=*/false); }

// Note that ops on the parallel device currently don't execute
// asynchronously. The test is just that we don't get deadlocks.
TEST(PARALLEL_DEVICE, TestCollectiveAsync) { TestCollective(/*async=*/true); }

void RegisterCollectiveMulFunction(TFE_Context* context,