auto status = tensorflow::DeviceNameUtils::DeviceNameToCpuDeviceName(
      tpu0_device, cpu0_device);
  if (!status.ok())
    return op->emitError()
           << "error in converting TPU0 to CPU0. The TPU device is "
           << tpu0_device;
  return mlir::success();
}

#include "xla/stream_executor/tpu/c_api_conversions.h"
#include "xla/stream_executor/tpu/status_helper.h"
#include "xla/stream_executor/tpu/tpu_api.h"
#include "xla/stream_executor/tpu/tpu_node_context.h"
#include "xla/stream_executor/tpu/tpu_platform.h"
#include "xla/stream_executor/tpu/tpu_platform_interface.h"
#include "xla/stream_executor/tpu/tpu_stream_interface.h"

limitations under the License.
==============================================================================*/

include "mlir/Pass/PassBase.td"

def TPURewritePass : Pass<"tf-tpu-rewrite", "mlir::ModuleOp"> {
  let summary = "Rewrites a `tf_device.cluster_func` on TPUs into TPU runtime operations.";

  let description = [{
    This pass rewrites a `tf_device.cluster_func` operation into a sequence of `tf._TPUCompileMlir`

  }];

  let constructor = "tensorflow::tf2xla::internal::CreateVerifyClusteringPass()";
}

def TPUClusterFormationPass : Pass<"tf-tpu-cluster-formation", "ModuleOp"> {
  let summary = "Forms clusters from operations assigned to the same TPU computation";

  let description = [{
    TPU computations from the frontend are composed of a `tf.TPUReplicateMetadata`
    op, a subgraph of ops (TensorFlow Dialect) each with a matching

namespace tensorflow {
namespace tf2xla {
namespace v1 {

using ::tensorflow::tpu::FunctionToHloArgs;
using ::tensorflow::tpu::GuaranteedConsts;
using ::tensorflow::tpu::MlirToHloArgs;
using ::tensorflow::tpu::ShardingAndIndex;

auto* phase2_bridge_compilation_status =
    tensorflow::monitoring::Counter<1>::New(
        "/tensorflow/core/tf2xla/api/v1/"
        "phase2_compilation_status", /*metric_name*/

  if (!op->hasAttr(TF::kTpuReplicateAttr)) {
    op->emitOpError("TF2XLA TPU bridge input check: " + errormsg() +
                    "missing _tpu_replicate attr");
    return false;
  }
  auto opattr = op->getAttr(TF::kTpuReplicateAttr);
  if (opattr != attr) {
    op->emitOpError("TF2XLA TPU bridge input check: " + errormsg() +
                    "invalid _tpu_replicate attr.")

        if (!exported_model.ok()) return exported_model.status();

        // Remove the `tpu` tag from the debug quantized saved model as it is
        // for CPU. Note the 'tpu' value should be the same as `TPU` defined in
        // tensorflow/python/saved_model/tag_constants.py.
        if (quantization_options.has_debugger_config()) {
          tags.erase("tpu");
        }
        py_function_library.SaveExportedModel(

    mlir::ArrayAttr device_assignment_attr);

// Finds the TPU compilation device and execution devices from `devices` for a
// TPU computation subgraph. Compilation device is determined from looking up
// all TPU_SYSTEM:0 devices and choosing the CPU device associated to the first
// TPU_SYSTEM device sorted lexicographically by replica and task. Execution
// devices are determined by looking up all TPU devices associated with each

    // expected-warning @+1 {{TF2XLA TPU bridge input check: cluster op = tf.opA with cluster = cluster has successor as non cluster op tf.opB}}
    %out2, %c2 = tf_executor.island wraps "tf.opB"(%out) : (tensor<i32>) -> tensor<i32>
    // expected-error @+1 {{tf.TPUReplicatedOutput' op TF2XLA TPU bridge input check: non-cluster op = tf.opB has invalid successor op = tf.TPUReplicatedOutput}}

    arg->set_dtype(dtype);
    // TODO(lyandy): Support other arg kinds.
    if (dtype == tensorflow::DT_RESOURCE)
      arg->set_kind(tensorflow::tpu::TPUCompileMetadataProto::Arg::VARIABLE);
    else
      arg->set_kind(tensorflow::tpu::TPUCompileMetadataProto::Arg::PARAMETER);

    // Populate argument shapes.
    *arg->mutable_shape() = tensorflow::TensorShapeProto();
    if (auto ranked_tensor_type =