device_assignment = [0, 0, 0, 0], func = @_func, host_compute_core = [], input_sharding_configuration = ["\08\01\1A\01\01\22\01\00", "\08\01\1A\01\01\22\01\00", "\08\01\1A\01\01\22\01\00", "\08\01\1A\01\01\22\01\00"], num_cores_per_replica = 1 : i64, output_sharding_configuration = ["\08\01\1A\01\01\22\01\00", "\08\01\1A\01\01\22\01\00"], step_marker_location = "STEP_MARK_AT_ENTRY", topology = "\0A\04\01\01\01\02\10\01\18\02\22\08\00\00\00\00\00\00\00\01", use_tpu = true} : (tensor<2x224x224x3xf32>, tensor<7x7x3x64xf32>,...

    ```mlir
      func @test(%arg0: tensor<*xi32>) {
        %0 = "tf_device.cluster_func"(%arg0) {
            func = @func,
            input_sharding_configuration = ["\01\02\03"],
            output_sharding_configuration = ["\08\01\1A\01\01\22\01\00"],
            step_marker_location = ""} : (tensor<*xi32>) -> tensor<*xi32>
        return
      }
      func @func(%arg0: tensor<*xi32> {mhlo.sharding = "\01\02\03"}) ->

  const auto output_sharding_attrs =
      cluster_func.getOperation()->getAttrOfType<mlir::ArrayAttr>(
          kOutputShardingAttr);
  if (!output_sharding_attrs)
    return cluster_func.emitError(
        "output_sharding_configuration missing from cluster func");

  if (output_sharding_attrs.size() != cluster_func.getNumResults())
    return cluster_func.emitError("incorrect number of output sharding");

  %0 = "tf_device.cluster_func"(%arg0) {_xla_compile_device_type = "TPU", _replication_info = "cluster0", func = @func, num_cores_per_replica = 2, input_sharding_configuration = ["\08\01\1A\01\01\22\01\00"], output_sharding_configuration = ["\08\01\1A\01\01\22\01\00"]} : (tensor<8xi32>) -> tensor<8xi32>
  return %0 : tensor<8xi32>
}
```

will be rewritten as:

```mlir