scope.graph()->AddEdge(w.node(), 0, launch, 6);

  auto out0 =
      ops::XlaClusterOutput(scope.WithOpName("Out0"), Output(launch, 0));
  auto out1 =
      ops::XlaClusterOutput(scope.WithOpName("Out1"), Output(launch, 1));
  auto out2 =
      ops::XlaClusterOutput(scope.WithOpName("Out2"), Output(launch, 2));
  auto out3 =
      ops::XlaClusterOutput(scope.WithOpName("Out3"), Output(launch, 3));

#include "tensorflow/core/platform/statusor.h"
#include "tensorflow/core/util/dump_graph.h"

namespace tensorflow {

namespace {

const char* const kXlaClusterOutput = "XlaClusterOutput";

bool IsCpuGpuCompile(const Graph* graph) {
  for (Node* n : graph->nodes()) {
    string name;
    // Only consider nodes being compiled.
    if (!TryGetNodeAttr(n->attrs(), kXlaClusterIdAttr, &name)) continue;