// re-use a thread.
  std::vector<std::unique_ptr<DeviceThread>> device_threads_;
  // A cancellation manager to use if the caller does not provide one. When ops
  // are executed asynchronously this must outlive the queued op, so it can't be
  // function-local to Execute.
  mutable std::unique_ptr<CancellationManager> default_cancellation_manager_;
};

  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,

*   Introduce TFDecorator.
*   Added an Mfcc op for speech feature generation.
*   Improved DirectSession::Run() overhead and error checking. Feeding a value
    of the wrong type will now synchronously raise an INVALID_ARGUMENT error
    instead of asynchronously raising an INTERNAL error. Code that depends on
    the (undefined) behavior when feeding a tensor of the wrong type may need to
    be updated.

  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(