%3:2 = "tf.TPUReplicatedOutput"(%2#0) : (tensor<?xi32>) -> (tensor<?xi32>, tensor<?xi32>)
  %5:2 = "tf.TPUPartitionedOutputV2"(%3#0) {_XlaSharding = "", partition_dims = []} : (tensor<?xi32>) -> (tensor<?xi32>, tensor<?xi32>)
  %7:2 = "tf.TPUPartitionedOutputV2"(%3#1) {_XlaSharding = "", partition_dims = []} : (tensor<?xi32>) -> (tensor<?xi32>, tensor<?xi32>)

  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<10xf32>} : () -> tensor<10xf32>
  %cst_0 = "tf.Const"() {value = dense<[-1, 10]> : tensor<2xi32>} : () -> tensor<2xi32>
  %1 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = false, transpose_b = true
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>
  %2 = "tf.Reshape"(%1, %cst_0) : (tensor<*xf32>, tensor<2xi32>) -> tensor<*xf32>

    %s0 = "tf.Const"() {value = dense<[501, 1, 32, 1280]> : tensor<4xi32>} : () -> tensor<4xi32>
    %s1 = "tf.Const"() {value = dense<[  1, 1,  1, 1280]> : tensor<4xi32>} : () -> tensor<4xi32>

    // CHECK: "tf.BroadcastGradientArgs"
    // CHECK-NOT: _xla_outside_compilation
    %r0, %r1 = "tf.BroadcastGradientArgs"(%s0, %s1) {} : (tensor<4xi32>, tensor<4xi32>) -> (tensor<1xi32>, tensor<3xi32>)
    tf_device.return

//          (tensor<?x3x4x2xf32>) -> tensor<i32>
// %5 = stablehlo.reshape %4 :
//          (tensor<i32>) -> tensor<1xi32>
// %6 = stablehlo.concatenate %5, %0, %1, %2, dim = 0 :
//          (tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>)
//            -> tensor<4xi32>
// %7 = stablehlo.dynamic_broadcast_in_dim %arg2, %6
// %8 = stablehlo.add %3, %7
// ```
//

// CHECK-DAG:       %[[VAL_3:.*]] = "tfl.pseudo_const"{{.*}}dense<2> : tensor<1xi32>
// CHECK-DAG:       %[[VAL_4:.*]] = "tfl.pseudo_const"{{.*}}dense<[2, 1]> : tensor<2xi32>
// CHECK:           %[[VAL_5:.*]] = "tfl.reshape"(%[[VAL_0]], %[[VAL_2]]) : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:           %[[VAL_6:.*]] = "tfl.reshape"(%[[VAL_1]], %[[VAL_2]]) : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>

}

// CHECK-LABEL: QuantizeStridedSlice
func.func @QuantizeStridedSlice(tensor<12x2x2x5x!quant.uniform<u8:f32, 0.1>>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<1x2x2x5xf32> {
^bb0(%arg0: tensor<12x2x2x5x!quant.uniform<u8:f32, 0.1>>, %arg1: tensor<1xi32>, %arg2: tensor<1xi32>, %arg3: tensor<1xi32>):
  %0 = "tfl.dequantize"(%arg0) : (tensor<12x2x2x5x!quant.uniform<u8:f32, 0.1>>) -> tensor<12x2x2x5xf32>

    %8 = stablehlo.convert %7 : (tensor<1x2xi32>) -> tensor<1x2xf32>
    %9 = stablehlo.convert %2 : (tensor<2x3xi8>) -> tensor<2x3xf32>
    %10 = stablehlo.dot_general %8, %9, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    %11 = stablehlo.convert %3 : (tensor<1x3xi32>) -> tensor<1x3xf32>

    ```mlir
      %2 = "tf.A"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>
      %3 = "tf.B"(%2) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>
      %4 = "tf.C"(%2, %3) {device = "tpu0"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
      %5 = "tf.D"(%4) : (tensor<?xi32>) -> tensor<?xi32>
    ```

    After the pass, we will have:

    ```mlir
      %0 = "tf.A"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>

    %0 = "tfl.pseudo_const"() {value = dense<0.962260901> : tensor<1xf32>} : () -> tensor<1xf32>
    %1 = func.call @func_0_GPU_FLOAT(%arg0, %0) {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_0"} : (tensor<1x200x200x200xf32>, tensor<1xf32>) -> tensor<1x200x200x200xf32>
    %2 = "tfl.pseudo_const"() {value = dense<0.895973444> : tensor<1xf32>} : () -> tensor<1xf32>

// CHECK-DAG: %[[BROADCAST_DIM:.+]] = arith.constant dense<{{\[3, 2, 1, 1\]}}> : tensor<4xi32>
// CHECK-DAG: %[[EXPAND_DIM1:.+]] = arith.constant dense<3> : tensor<1xi32>
// CHECK-DAG: %[[EXPAND_DIM0:.+]] = arith.constant dense<2> : tensor<1xi32>
// CHECK: %[[EXPAND0:.+]] = "tfl.expand_dims"(%[[ARG0]], %[[EXPAND_DIM0]]) : (tensor<1x2x!quant.uniform<i8:f32, 2.000000e+00:3>>, tensor<1xi32>) -> tensor<1x2x1x!quant.uniform<i8:f32, 2.000000e+00:3>>