const int in_flight_nodes_limit)
      : status_(TF_NewStatus()),
        // If the context's default exector is set to async, re-using that in
        // each thread would cause collectives to deadlock. For consistency we
        // create a new sync executor for every thread.
        //
        // TODO(allenl): We should have an async API that works with the
        // parallel device.

    if len(testsuite) == 0:  # pylint: disable=g-explicit-length-test
      r._elem.remove(testsuite._elem)
  if len(r) > 0:  # pylint: disable=g-explicit-length-test
    result += r

# Insert the number of failures for each test to help identify flakes
# need to clarify for shard
for p in result._elem.xpath(".//error | .//failure"):
  key = re.sub(r"0x\w+", "", p.getparent().get("name", "")) + p.text

// of fetched mappings is returned in `num_missing_unused_input_mappings`. The
// array of each mapping's source node name is returned in `src_names`, and the
// array of each mapping's source index is returned in `src_indexes`.
//
// `*src_names`, `*src_indexes`, and the memory backing each string in
// `src_names` are owned by and have the lifetime of `results`.

    status_ = TF_NewStatus();
    filesystem_ = new TF_Filesystem;
    filesystem_->plugin_filesystem = nullptr;
    // Because different tests requires different setup for filesystem. We
    // initialize filesystem in each testcase.
  }
  void TearDown() override {
    TF_DeleteStatus(status_);
    if (filesystem_->plugin_filesystem != nullptr)
      tf_gcs_filesystem::Cleanup(filesystem_);
    delete filesystem_;
  }

  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());