%0 = "tf.ReadVariableOp"(%arg1) {device = "/CPU:0"} : (tensor<*x!tf_type.resource>) -> tensor<*xi32>
  %1 = "tf.Shape"(%arg0) {device = "/CPU:0"} : (tensor<*xi32>) -> tensor<?xi64>
  %2 = "tf.Shape"(%0) {device = "/CPU:0"} : (tensor<*xi32>) -> tensor<?xi64>

    %5 = "tf.Identity"(%arg4) {device = ""} : (tensor<1024x1024xf32>) -> tensor<1024x1024xf32>
    %6 = "tf.MatMul"(%arg3, %5) {device = "", transpose_a = false, transpose_b = false} : (tensor<1x1024xf32>, tensor<1024x1024xf32>) -> tensor<1x1024xf32>
    %7 = "tf.AddV2"(%2, %6) {device = ""} : (tensor<1x1024xf32>, tensor<1x1024xf32>) -> tensor<1x1024xf32>

auto* auto_outside_compilation_gauge =
    tensorflow::monitoring::Gauge<bool, 0>::New(
        "/tensorflow/core/use_auto_outside_compilation",
        "Tracks if auto outside compilation is enabled");

#define GEN_PASS_DEF_MARKOPSFOROUTSIDECOMPILATIONPASS
#include "tensorflow/compiler/mlir/tf2xla/internal/passes/clustering_passes.h.inc"

struct MarkOpsForOutsideCompilation
    : public impl::MarkOpsForOutsideCompilationPassBase<

  #define CU_TRUE 1
#endif

#ifndef CU_FALSE
  /** Boolean FALSE for CUnit use. */
  #define CU_FALSE 0
#endif

#ifndef CU_UNREFERENCED_PARAMETER
  /** Consistent approach to referencing unused parameters. */
  #define CU_UNREFERENCED_PARAMETER(x) (void)x
#endif

#ifndef CU_MAX
#  define CU_MAX(a,b) (((a) >= (b)) ? (a) : (b))
#endif

#ifndef CU_MIN

    func.return %0 : tensor<2x100xf32>
  }
  func.func @func_0_GPU_FLOAT(%arg0: tensor<100xf32>, %arg1: tensor<100xf32>, %arg2: tensor<100xf32>) -> tensor<100xf32> attributes {tac.cost = 4.000000e+01 : f32, tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_0"} {

// Returns the Device ID of the device that the context possesses. Returns the
// PlatformDeviceId if a mapping between between TfDeviceId and PlatformDeviceId
// is set; otherwise returns the id in the device name. Please refer to
// tensorflow/tsl/framework/device_id.h for more details.
// For mobile or slim build, returns the id in the device name.

// CHECK-SAME: ({{%.+}}: tensor<i64> {tf.device = "/job:localhost/replica:0/task:0/device:TPU:0"}, {{%.+}}: tensor<i32> {tf.device = "/job:localhost/replica:0/task:0/device:TPU:0"})
// CHECK-SAME: -> (tensor<i64> {tf.device = "/job:localhost/replica:0/task:0/device:TPU:0"}, tensor<i32> {tf.device = "/job:localhost/replica:0/task:0/device:TPU:0"})

      task_and_device = {task, device};
    }
  }

  return topology;
}

// Determine execution devices when topology and device assignment are defined.
// With a topology device coordinate to task and device mapping, device
// assignment device coordinates can then be mapped to task and device for TPU
// devices. The device assignment array is also validated.
//

    %0 = tfl.add %arg0, %arg1 {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
    %1 = tfl.mul %0, %arg2 {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
    func.return %1 : tensor<1xf32>

  device_setup_.AddDevicesAndSetUp({DEVICE_GPU});
  Device* device = device_setup_.GetDevice(DEVICE_GPU);
  XlaPlatformInfo platform_info = XlaPlatformInfoFromDevice(device);

  ResourceMgr resource_mgr("");
  OpKernelContext::Params params;
  params.resource_manager = &resource_mgr;
  params.device = device;
  OpKernelContext ctx(&params, 0);

  PjRtDeviceCompiler* pjrt_device_compiler = nullptr;