->points.at(0)
                     ->histogram_value.sum());

  TFE_MonitoringSamplerCellAdd(cell, 5.0);
  metrics = collection_registry->CollectMetrics(options);
  EXPECT_EQ(6.0, metrics->point_set_map.at("test/sampler")
                     ->points.at(0)
                     ->histogram_value.sum());
  TFE_MonitoringDeleteBuckets(buckets);
  TFE_MonitoringDeleteSampler0(sampler);

  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteTensorHandle(retvals[0]);
  float sum = 0;
  EXPECT_EQ(sizeof(sum), TF_TensorByteSize(t));
  memcpy(&sum, TF_TensorData(t), TF_TensorByteSize(t));
  TF_DeleteTensor(t);
  EXPECT_EQ(sum, 6.0);

  TFE_DeleteTensorHandle(h0);
  TFE_DeleteTensorHandle(h1);
  TFE_DeleteTensorHandle(h2);

  }

 private:
  /// The size of the blocks stored in the LRU cache, as well as the size of the
  /// reads from the underlying filesystem.
  const size_t block_size_;
  /// The maximum number of bytes (sum of block sizes) allowed in the LRU cache.
  const size_t max_bytes_;
  /// The maximum staleness of any block in the LRU cache, in seconds.
  const uint64_t max_staleness_;

        // allows code like "tf.constant(1.)" or "tf.reduce_sum(..., axis=1)"
        // (where the value starts on the host), without allowing other implicit
        // copies/broadcasts. Other implicit copies may be supported eventually,
        // but need special handling for gradients (gradient of copy-on is not
        // just copy-off but includes a sum) and consideration of performance.
        //

  TensorHandlePtr parallel_value = CreatePerDeviceValues(
      context.get(), components, device_name, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Run a collective sum, so each component should now be the same.
  TensorHandlePtr reduced(
      CollectiveSum(context.get(), parallel_value.get(), 2, status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());