%4 = "tfl.relu"(%3) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<128x128xf32>) -> tensor<128x128xf32>
  %5 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<128x128xf32>} : () -> tensor<128x128xf32>
  %6 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<128xf32>} : () -> tensor<128xf32>

    %0 = "tf.Div"(%arg0, %cst_3) {device = ""} : (tensor<1x1024xf32>, tensor<f32>) -> tensor<1x1024xf32>
    %1 = "tf.AddV2"(%0, %cst_0) {device = ""} : (tensor<1x1024xf32>, tensor<f32>) -> tensor<1x1024xf32>
    %2 = "tf.Floor"(%1) {device = ""} : (tensor<1x1024xf32>) -> tensor<1x1024xf32>
    %3 = "tf.ClipByValue"(%2, %cst_1, %cst_5) {device = ""} : (tensor<1x1024xf32>, tensor<f32>, tensor<f32>) -> tensor<1x1024xf32>

}

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>)

  // CHECK-DAG: %[[ITEMS0_0:.*]] = "tf.ExpandDims"(%[[ITEMS0]], %[[AXIS]])
  // CHECK-DAG: "tf.ConcatV2"(%[[ITEMS1_3]], %[[ITEMS1_2]], %[[ITEMS1_1]], %[[ITEMS1_0]], %[[ITEMS0_0]], %[[AXIS]]) : (tensor<1x2xf32>, tensor<1x2xf32>, tensor<1x2xf32>, tensor<1x2xf32>, tensor<1x2xf32>, tensor<i64>) -> tensor<5x2xf32>

  %indices0 = "tf.Const"() {value = dense<4> : tensor<i32>} : () -> tensor<i32>

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