device = "/job:tpu_host_worker/replica:0/task:0/device:CPU:0",
      finalize_func = @finalize,
      init_func = @init,
      next_func = @next,
      operandSegmentSizes = array<i32: 1, 2, 1>,
      output_shapes = [#tf_type.shape<>],
      output_types = [!tf_type.string],
      metadata = ""} : (
              tensor<4xf32>,
              tensor<3xf32>,
              tensor<!tf_type.resource>,

  %2 = "tf.Const"() {__op_key = 2: i32, device = "/device:CPU:0", value = dense<1> : tensor<i64>} : () -> tensor<i64>
  %3 = "tf.RangeDataset"(%0, %1, %2) {__op_key = 3: i32, device = "/device:CPU:0", output_shapes = [#tf_type.shape<>], output_types = [i64], metadata = ""} : (tensor<i64>, tensor<i64>, tensor<i64>) -> tensor<!tf_type.variant>
  // CHECK: tf_mlrt.executeop{{.*}}op: \22FlatMapDataset\22

      continue;
    }
    shape.dims = new int64_t[shape.num_dims];
    for (size_t j = 0; j < shape.num_dims; ++j) {
      shape.dims[j] = c.Value(c.Dim(shape_handle, j));
    }
  }
  if (output_shapes != nullptr) *output_shapes = output_shapes_result;

  // TODO(bgogul): Set output_resource_shapes_and_types.
}

void TF_ImportGraphDefOptionsSetValidateColocationConstraints(

//   - The types need not be set in `input_shapes` as it is not used.
//   - The number of `input_tensors` should be the same as the number of items
//     in `input_shapes`.
//
// The results are returned in `output_shapes` and
// `output_resource_shapes_and_types`. The caller is responsible for freeing the
// memory in these buffers by calling `TF_DeleteShapeAndTypeList`.
TF_CAPI_EXPORT extern void TFE_InferShapes(

    return %0 : tensor<200x10xf32>
  }

  // CHECK-LABEL: @main
  func.func @main()
  attributes {tf_saved_model.exported_names = ["main"]} {
    %0 = "tf.Iterator"() {container = "", output_shapes = [#tf_type.shape<200x10>], output_types = [f32], shared_name = "foo_iterator"} : () -> tensor<!tf_type.resource>
    %1 = func.call @get_next(%0) : (tensor<!tf_type.resource>) -> tensor<200x10xf32>
    return
  }
}

"tf.While"(%cst_0, %cst, %cst_0, %arg0, %cst_1) {T = [i32, i32, i32, f32, f32],_lower_using_switch_merge = true, _num_original_outputs = 5 : i64, _read_only_resource_inputs = [], body = @while_body, cond = @while_cond, device = "", is_stateless = true, output_shapes = [#tf_type.shape<>, #tf_type.shape<>, #tf_type.shape<>, #tf_type.shape<1x1024>, #tf_type.shape<1024x1024>], parallel_iterations = 10 : i64, shape_invariant} : (tensor<i32>, tensor<i32>, tensor<i32>, tensor<1x1024xf32>, tensor<1024x1024xf32>) ->...

    const xla::HloInputOutputAliasConfig& input_output_alias,
    absl::Span<const int> input_mapping,
    const std::map<int, const Tensor*>& resource_vars_snapshots,
    DataType output_dtype, const TensorShape& output_shape,
    Allocator* output_allocator, bool allocate_xla_tensors, se::Stream* stream,
    bool use_multiple_streams, std::shared_ptr<se::Event> definition_event) {

  return ConvertTensor(t, builder);
}

void ConvertToTensorShapeProto(ArrayRef<int64_t> shape,
                               TensorShapeProto* output_shape) {
  for (auto d : shape) {
    output_shape->add_dim()->set_size(ShapedType::isDynamic(d) ? kTFDynamicSize
                                                               : d);
  }
}

  %output_shape = arith.constant dense<[2, 3, 2, 2048]> : tensor<4xi32>
  %f16_weights = "tfl.pseudo_const"() {value = dense<1.0> : tensor<4x2x2x2048xf16>} : () -> tensor<4x2x2x2048xf16>
  %dq_weights = "tfl.dequantize"(%f16_weights) : (tensor<4x2x2x2048xf16>) -> tensor<4x2x2x2048xf32>
  %bias = "tfl.no_value"() {value} : () -> none

        init_func = @__init_func_530, next_func = @__next_func_680,
        next_func.experimental_ints_on_device = true,
        operandSegmentSizes = array<i32: 2, 2, 1>,
        output_shapes = [#tf_type.shape<?>],
        output_types = [f32],
        metadata = ""
      } :