_xla_compile_device_type = "TPU", _replication_info = "cluster",
    device = "/task:0/device:TPU:0", dtype = f32
  } : (tensor<1x80xf32>, tensor<1x80xf32>) -> tensor<1x80xf32>
  %3 = "tf.ResourceGatherNd"(%arg0, %0) {
    Tindices = i32
  } : (tensor<*x!tf_type.resource<tensor<80xf32>>>, tensor<i32>) -> tensor<1x80xf32>

}

func.func @QDQsFollowedByTranspose(tensor<1x2xf32>) -> (tensor<2x1xf32>) {
^bb0(%arg0: tensor<1x2xf32>):
  %cst_0 = arith.constant dense<[1, 0]> : tensor<2xi32>
  %0 = "tfl.quantize"(%arg0){qtype = tensor<1x2x!quant.uniform<u8:f32, 1.0>>}: (tensor<1x2xf32>) -> (tensor<1x2x!quant.uniform<u8:f32, 1.0>>)
  %1 = "tfl.dequantize"(%0): (tensor<1x2x!quant.uniform<u8:f32, 1.0>>) -> (tensor<1x2xf32>)

    %recurrent_stats = "quantfork.stats"(%recurrent_input) {layerStats = dense<[-2.0, 1.0]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>
    %cell_input = arith.constant dense<1.0> : tensor<1x20xf32>
    %cell_stats = "quantfork.stats"(%cell_input) {layerStats = dense<[-2.73090601, 7.94872093]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>
    %0 = "tfl.unidirectional_sequence_lstm"(%arg0,

    // CHECK-NEXT:  %[[RES:.*]] = "tf.SelectV2"(%[[PRED]], %[[SCALED_GRADIENTS]], %[[SELU_GRAD_VALUE]]) : (tensor<4x8xi1>, tensor<4x8xf32>, tensor<4x8xf32>) -> tensor<4x8xf32>
    // CHECK-NEXT:  return %[[RES]] : tensor<4x8xf32>
    %2 = "tf.SeluGrad"(%gradients, %features) : (tensor<4x8xf32>, tensor<4x8xf32>) -> tensor<4x8xf32>
    func.return %2 : tensor<4x8xf32>
}

For example, if we have the code

```mlir
  %0 = "tf.Const"() {value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %1 = "tf.Const"() {device = "", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %2 = "tf.Const"() {device = "baz", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
```

then running this pass with 'default-device=foobar', we get:

```mlir