func.return %1 : tensor<128x32xf32>
  // CHECK: %[[FULLY_CONNECTED:.*]] = "tfl.fully_connected"(%arg0, %arg1, %arg2) <{fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"}> : (tensor<128x64xf32>, tensor<32x64xf32>, tensor<32xf32>) -> tensor<128x32xf32>
  // CHECK: return %[[FULLY_CONNECTED]] : tensor<128x32xf32>
}

  %0 = "tf.CollectivePermute"(%arg0, %source_target_pairs) {
  } : (tensor<128x32xf32>, tensor<3x2xi32>) -> tensor<128x32xf32>

  func.return %0 : tensor<128x32xf32>
}

// -----

// CHECK-LABEL: @cross_replica_sum
func.func @cross_replica_sum(%input: tensor<10xf32>) -> tensor<10xf32> {
  %replica_groups = "tf.Const" () {

}

// -----

module {
func.func private @func_20_GPU_FLOAT(%arg0: tensor<128x128xf32>, %arg1: tensor<3xi32>) -> tensor<1x128x128xf32> attributes {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_20"} {
  %0 = "tfl.reshape"(%arg0, %arg1) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<128x128xf32>, tensor<3xi32>) -> tensor<1x128x128xf32>
  func.return %0 : tensor<1x128x128xf32>
}

func.func @Int64SliceBeginSize(%arg0: tensor<4x128x32xf32>) -> tensor<1x128x32xf32> {
  %0 = "tfl.pseudo_const"() {value = dense<0> : tensor<3xi64>} : () -> tensor<3xi64>
  %1 = "tfl.pseudo_const"() {value = dense<[1, 128, 32]> : tensor<3xi64>} : () -> tensor<3xi64>
  %2 = "tfl.slice"(%arg0, %0, %1) : (tensor<4x128x32xf32>, tensor<3xi64>, tensor<3xi64>) -> tensor<1x128x32xf32>
  func.return %2 : tensor<1x128x32xf32>

// CHECK-NEXT:  %[[cst:.*]] = arith.constant dense<1> : tensor<i32>
  %cst = arith.constant dense<1> : tensor<i32>
// CHECK-NEXT:  %[[softmax:.*]] = "tfl.softmax"(%arg0) <{beta = 1.000000e+00 : f32}> : (tensor<128x16xf32>) -> tensor<128x16xf32>
  %0 = "tfl.softmax"(%arg0) {beta = 1.000000e+00 : f32} : (tensor<128x16xf32>) -> tensor<128x16xf32>

   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

  // CHECK-LABEL: func @parallel_execute_with_tiled_input
  // CHECK-SAME: (%[[ARG_0:[a-z0-9]*]]: tensor<128x10xf32>, %[[ARG_1:[a-z0-9]*]]: tensor<128x10xf32>, %[[ARG_2:[a-z0-9]*]]: tensor<*xi32>, %[[ARG_3:[a-z0-9]*]]: tensor<*xi32>)
  func.func @parallel_execute_with_tiled_input(%arg0: tensor<128x10xf32>, %arg1: tensor<128x10xf32>, %arg2: tensor<*xi32>, %arg3: tensor<*xi32>) -> (tensor<*xi32>, tensor<*xi1>) {
    // CHECK: tf_device.replicate

  %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>