&build_ops_flags->tf_xla_check_cluster_output_numerics,
            "If true then insert CheckNumerics nodes to check all cluster "
            "outputs."),
       Flag("tf_xla_disable_constant_folding",
            &build_ops_flags->tf_xla_disable_constant_folding,
            "If true then disables constant folding on TF graph before XLA "
            "compilation."),
       Flag("tf_xla_disable_full_embedding_pipelining",

unsigned char TF_GetXlaConstantFoldingDisabled() {
  return static_cast<unsigned char>(
      tensorflow::GetBuildXlaOpsPassFlags()->tf_xla_disable_constant_folding);
}

void TF_SetXlaConstantFoldingDisabled(unsigned char should_enable) {
  tensorflow::GetBuildXlaOpsPassFlags()->tf_xla_disable_constant_folding =
      static_cast<bool>(should_enable);
}

void TF_SetXlaMinClusterSize(int size) {

  bool tf_xla_check_cluster_output_numerics;

  // Disables all constant folding. The primary use for this is for testing to
  // guarantee that tests are run on XLA and not on TF's CPU implementation.
  bool tf_xla_disable_constant_folding;

  // Disables full embedding pipelining when true. Instead, strict SparseCore
  // TensorCore sequencing will be used.
  bool tf_xla_disable_full_embedding_pipelining;

        // disallow any sort of constant folding on Variant nodes for now.
        bool disable_constant_folding =
            GetBuildXlaOpsPassFlags()->tf_xla_disable_constant_folding;
        auto cf_consider_fn = [disable_constant_folding](const Node* n) {
          if (disable_constant_folding) return false;
          for (const auto& output_arg : n->op_def().output_arg()) {