}

  if (output_to_input_alias.empty()) return absl::OkStatus();

  xla::HloModuleProto* module_proto =
      compilation_result->computation->mutable_proto();
  absl::StatusOr<xla::ProgramShape> program_shape_or_status =
      compilation_result->computation->GetProgramShape();
  TF_RET_CHECK(program_shape_or_status.ok());

  xla::ProgramShape& program_shape = program_shape_or_status.value();

// Creates a pass that lifts operations on external resource variables from
// device computation nested in `tf_device::LaunchOp` out so that resource
// variable load operations are all before device computation while resource
// variable store operations are all after device computation. After this pass,
// device computation no longer interacts with external resource variables.

      %computation = "tf_device.cluster_func"(%read) {func = @computation, use_spmd_for_xla_partitioning = true} : (tensor<i32>) -> tensor<i32>
      "tf.AssignVariableOp"(%partitioned_variable, %computation) : (tensor<!tf_type.resource<tensor<i32>>>, tensor<i32>) -> ()
      return
    }

    func @computation(%arg0: tensor<i32>) -> tensor<i32> {
      return %arg0: tensor<i32>

  let description = [{
This operation holds a replicated output from a `tpu.replicate()` computation subgraph.
Each replicated output has the same shape and type alongside the input.

For example:
```
%computation = "tf.Computation"()
%replicated_output:2 = "tf.TPUReplicatedOutput"(%computation)
```

    output_dims.insert(output_dims.begin() + axis, depth);

    Location loc = op.getLoc();

    // The iota result is the effective output shape of the computation,
    // and indices must be broadcast into it. At this point, this computation
    // would need to be reworked quite a bit to support dynamic shapes, so
    // just using static broadcasting.
    auto index_type =

  }
  return DeviceNameUtils::ParsedNameToString(tpu_device);
}

// Determine execution devices when topology and device assignment are not
// defined. This is a special case where a single core computation is replicated
// to every core in the mesh. TPU devices are simply added to
// `execution_devices` of one replica. `num_replicas` must be 1 or the total
// number of TPU devices available, and `num_cores_per_replica` must be 1.

 *       SMALL_DELAY_MS}, {@code MEDIUM_DELAY_MS}, {@code LONG_DELAY_MS}. The idea here is that a
 *       SHORT is always discriminable from zero time, and always allows enough time for the small
 *       amounts of computation (creating a thread, calling a few methods, etc) needed to reach a
 *       timeout point. Similarly, a SMALL is always discriminable as larger than SHORT and smaller

    *   oneDNN optimizations can yield slightly different numerical results compared to when oneDNN optimizations are disabled due to floating-point round-off errors from
 different computation approaches and orders.
    *   To verify if oneDNN optimizations are on, look for a message with *"oneDNN custom operations are on"* in the log. If the exact phrase is not there, it means they are off.

  // Returns true if a return type was changed.
  bool InferShapeForXlaCallModule(XlaCallModuleOp op);

  // Infers the shape of _XlaHostComputeMlir based on the host computation
  // module.  Returns true if a return type was changed.
  bool InferShapeForXlaHostComputeMlir(_XlaHostComputeMlirOp op);

  // Infers the shape of function attached to XlaHostCompute.

	}

	gw, gwMap, nsReferences := convertGateways(ctx)
	ctx.GatewayReferences = gwMap
	result.Gateway = gw

	result.VirtualService = convertVirtualService(ctx)

	// Once we have gone through all route computation, we will know how many routes bound to each gateway.
	// Report this in the status.
	for _, dm := range gwMap {
		for _, pri := range dm {
			if pri.ReportAttachedRoutes != nil {
				pri.ReportAttachedRoutes()
			}