// Warmup the persistent cache(s) with multiple runs. 4 is a magic number to
  // detect non-determinism in TF when running the test.
  listener()->ClearListenerHistory();
  for (int b = 1; b < 4; ++b) {
    TF_ASSERT_OK(ExecuteWithBatch(graph, b));
  }
  TF_ASSERT_OK(listener()->VerifyPersistentCacheUseListenerHistory(
      /*expect_persistent_cache_use=*/false));

  // Reset the cluster numbering between sessions so we can get the same

  // Warmup the persistent cache(s) with multiple runs. 4 is a magic number to
  // detect non-determinism in TF when running the test.
  listener()->ClearListenerHistory();
  for (int b = 1; b < 4; ++b) {
    TF_ASSERT_OK(ExecuteWithBatch(graph, b));
  }
  TF_ASSERT_OK(listener()->VerifyPersistentCacheUseListenerHistory(
      /*expect_persistent_cache_use=*/false));

  // Reset the cluster numbering between sessions so we can get the same

  // Runs the graph using specified batch size both with and without XLA JIT
  // compilation. Returns an error if the results between the two do not match.
  Status ExecuteWithBatch(const GraphDef& graph, int batch);

  // Adds the suffix "_altered" to the HLO module names of all of the persistent
  // XLA compilation cache entries found at the specified directory. If none are

                               {{"dtype", DT_FLOAT}, {"shape", input_shape}});
  *graph.add_node() = MakeNode("m", "TestFn", {"a", "b", "c"}, {});
  return graph;
}

Status DeviceCompilerSerializeTest::ExecuteWithBatch(const GraphDef& graph,
                                                     int batch) {
  const TensorShape shape({batch, 4});

  // Compute the golden output tensor
  std::vector<Tensor> golden_output_tensors;
  {