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));

// XlaLaunch nodes.
//
// xla.compile() does two main things:
// a) marks operators that make up an XLA computation with the attribute
//    _xla_compile_id=XYZ, where XYZ is a unique key.
// b) adds XlaClusterOutput nodes to represent outputs of the computation.
//    These nodes are not marked with the _xla_compile_id attribute.

#ifndef TENSORFLOW_COMPILER_JIT_ENCAPSULATE_XLA_COMPUTATIONS_PASS_H_

# limitations under the License.
# ==============================================================================

from tensorflow.python.framework import ops


@ops.RegisterGradient("XlaClusterOutput")
def _XlaClusterOutputGrad(_, grad):
  del grad  # unused
  raise RuntimeError("Gradient computation of graph in xla.compile() is "
                     "prohibited because it can cause performance degradation."

    .Attr("Tresults: list(type) >= 0")
    .Attr("constants: list(int) >= 0")
    .Attr("resources: list(int) >= 0")
    .Attr("function: func")
    .Doc("XLA Launch Op. For use by the XLA JIT only.");

REGISTER_OP("XlaClusterOutput")
    .Input("input: T")
    // Note: when replication is supported, this op will have N outputs.
    .Output("outputs: T")
    .Attr("T: type")
    .SetShapeFn([](InferenceContext* c) {

#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;

  );

  TF_DerivedOperandTypeListAttr Tin = TF_DerivedOperandTypeListAttr<0>;
  TF_DerivedResultTypeListAttr Tout = TF_DerivedResultTypeListAttr<0>;
}

def TF_XlaClusterOutputOp : TF_Op<"XlaClusterOutput", [Pure, TF_NoConstantFold]> {
  let summary = [{
Operator that connects the output of an XLA computation to other consumer graph nodes.
  }];

  let arguments = (ins
    TF_Tensor:$input