session_options.config.intra_op_parallelism_threads()),
      use_multiple_streams_(options.use_multiple_streams),
      shape_determination_fns_(options.shape_determination_fns),
      allowed_devices_(options.allowed_devices),
      use_global_compute_stream_(options.use_global_compute_stream) {
  if (options.shape_determination_fns.empty()) {
    LOG(ERROR) << "shape_representation_fns must be non-empty.";
  }

  // platform will have resources allocated. For GPUs this will be
  // filled from visible_gpu_devices list from session configuration.
  std::optional<std::set<int>> allowed_devices_;

  const bool use_global_compute_stream_;

  // A static vector with device_ordinal as its index, describing the global
  // compute streams used in each XLA device. It is only used if

    options.device_name = DEVICE_XLA_GPU;
    options.device_ordinal = i;
    options.compilation_device_name = DEVICE_GPU_XLA_JIT;
    options.use_multiple_streams = true;
    options.allowed_devices = gpu_ids;
    XlaShapeLayoutHelpers::ShapeDeterminationFns shape_representation_fns{
        UseNoPreferenceLayoutFn(), IdentityShapeRepresentationFn()};
    options.shape_determination_fns = {shape_representation_fns};

  "tf_device.cluster"() ({
    // CHECK:      "tf.VarHandleOp"
    // CHECK-SAME: _xla_outside_compilation
    "tf.VarHandleOp"() {allowed_devices = [], container = "", device = "", shared_name = ""} : () -> tensor<!tf_type.resource<tensor<!tf_type.string>>>
    tf_device.return

  // CHECK: [[elems:%.*]] = "tf.VarHandleOp"
  %outputs_10  =  "tf.VarHandleOp"() {_xla_inferred_shapes = [#tf_type.shape<>], allowed_devices = [], container = "", device = "/job:localhost/replica:0/task:0/device:CPU:0", shared_name = "w"} : () -> tensor<!tf_type.resource<tensor<3x1xf32>>>