%1 = "tf.Relu"(%0) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>
    %2 = "tf.Add"(%0, %1) {device = "tpu0"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
    %3 = "tf.Relu"(%arg1) {device = "tpu1"} : (tensor<?xi32>) -> tensor<?xi32>
    %4 = "tf.Add"(%3, %arg1) {device = "tpu1"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
    %5 = "tf.Relu"(%4) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>

    // CHECK-SAME: <{device = "tpu0"}>
    // CHECK: %[[B_OUTPUT:[0-9]*]] = "tf.B"(%[[A_OUTPUT]]) : (tensor<?xi32>) -> tensor<?xi32>
    %3 = "tf.B"(%2) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>

    // CHECK: %[[C_OUTPUT:[0-9]*]] = "tf.C"(%[[A_OUTPUT]], %[[B_OUTPUT]]) : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
    %4 = "tf.C"(%2, %3) {device = "tpu0"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>

  const std::string cpu0 = "/job:worker/replica:0/task:0/device:CPU:0";
  const std::string gpu0 = "/job:worker/replica:1/task:2/device:GPU:0";
  const std::string gpu1 = "/job:worker/replica:1/task:2/device:GPU:1";

  llvm::SmallVector<std::unique_ptr<Device>, 2> devices;
  devices.push_back(FakeDevice::Make(cpu0));
  devices.push_back(FakeDevice::Make(gpu0, "compute capability: 7.0"));

  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();
}

}

def TPUPartitionedOpConversionPass : Pass<"tf-tpu-partitioned-op-conversion", "mlir::func::FuncOp"> {
  let summary = "Rewrite all TPU Partitioned ops into their V2 counterparts.";

  let constructor = "TFTPU::CreateTPUPartitionedOpConversionPass()";
}

def TPUValidateInputsPass : Pass<"tf-tpu-validate-inputs", "ModuleOp"> {
  let summary = "Validates inputs to the TPU TF/XLA bridge";

  let description = [{

This pass expands outside compilation attributes to Identity/Cast ops
at the head of TPU computation if it's only used by outside compiled ops.
### `-tf-tpu-identity-pruning`

_Removes Identity/IdentityN ops from the TPU computation_

### `-tf-tpu-merge-variables-with-execute`

_Merges device variable reads and updates into TPU execute ops_

      }) {is_stateless = true, _xla_compile_device_type = "TPU", _replication_info = "x" } : (tensor<i1>) -> (tensor<f32>)
    tf_device.return %3 : tensor<f32>
  // CHECK: {_replication_info = "x", _xla_compile_device_type = "TPU", cluster_attr = "cluster_attr", device = "y"}
  }) {cluster_attr = "cluster_attr", _xla_compile_device_type = "TPU", _replication_info = "x", device = "y"} : () -> tensor<f32>
  // CHECK: "tf.Add"

#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"

  // CHECK-SAME: _replication_info = "write", _xla_compile_device_type = "TPU"
  %0:2 = "tf_device.cluster_func"(%arg0, %arg1) {_replication_info = "write", _xla_compile_device_type = "TPU", func = @write_func} : (tensor<i32>, tensor<f32>) -> (tensor<f32>, tensor<i32>)
  // CHECK-NEXT: "tf.AssignVariableOp"([[ARG2]], [[CLUSTER]]#1)
  "tf.AssignVariableOp"(%arg2, %0#1) : (tensor<*x!tf_type.resource<tensor<i32>>>, tensor<i32>) -> ()

// RUN: tf-opt %s -tf-replicated-clustering-bridge-v2 -tfrt-lower-cluster-to-runtime-ops-tpu 2>&1 | FileCheck %s
// RUN: tf-opt %s -tf-replicated-clustering-bridge-v1 -tfrt-lower-cluster-to-runtime-ops-tpu -tf-dialect-to-executor-v1 2>&1 | FileCheck %s

// This test verifies there is no warning about shape inference failure in TPU
// bridge in handling multiple usage of the same function.

// Since it is possible that this warning may become an error in the future,