// CHECK-LABEL @lowers_string_ops
  // CHECK-DAG: ml_program.global public @vars.Variable_1([]) : tensor<!tf_type.string>
  func.func @lowers_string_ops(%arg0: tensor<128xi32>, %arg1: tensor<128xi32>, %arg2: tensor<128x1xi32>, %arg3: tensor<128x90xi32>, %arg4: tensor<128x90xi32>, %arg5: tensor<128x90xi32>, %arg6: tensor<128x90x64xf32>, %arg7: tensor<128x90x64xf32>) -> tensor<!tf_type.string> {

// 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>

  %group_assignment = "tf.Const"() { value = dense<[[0],[1]]> : tensor<2x1xi32> } : () -> tensor<2x1xi32>
  // CHECK: "mhlo.all_reduce"
  // CHECK{LITERAL}: replica_groups = dense<[[0], [1]]> : tensor<2x1xi64>
  // CHECK-NOT: channel_handle
  // CHECK: mhlo.add
  %0 = "tf.XlaAllReduce"(%input, %group_assignment) {reduce_op = "Add", mode = "CrossReplica"} : (tensor<f32>, tensor<2x1xi32>) -> tensor<f32>
  func.return %0 : tensor<f32>
}

// -----

  // 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

// CHECK-LABEL: @ShardingAttr
func.func @ShardingAttr(%arg0: tensor<128x10xf32> {mhlo.sharding = "\08\03\1A\02\01\02\22\02\00\01"}, %arg1: tensor<10x1024xf32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}, %arg2: tensor<128x1024xf32> {mhlo.sharding = ""}) -> (tensor<128x10xf32>, tensor<10x1024xf32>, tensor<128x1024xf32>) {

    %11 = stablehlo.convert %2 : (tensor<1x1x1xi8>) -> tensor<1x1x1xi32>
    %12 = stablehlo.broadcast_in_dim %11, dims = [0, 1, 2] : (tensor<1x1x1xi32>) -> tensor<8x16x16xi32>
    %13 = stablehlo.subtract %10, %12 : tensor<8x16x16xi32>  // q1 - z1
    %14 = stablehlo.convert %9 : (tensor<8x16x4xi8>) -> tensor<8x16x4xi32>
    %15 = stablehlo.convert %4 : (tensor<1x1x1xi8>) -> tensor<1x1x1xi32>

// -----

func.func @constant_tensor_invalid_3(%arg0: vector<1xi32>) -> tensor<1x1xi32> {
    // expected-error@+1 {{input and output should have same shape and element type}}
  %0 = "tfr.constant_tensor"(%arg0) : (vector<1xi32>) -> tensor<1x1xi32>
  func.return %0 : tensor<1x1xi32>
}

// -----

func.func @constant_tensor_invalid_4(%arg0: i32) -> tensor<1x1xi32> {
    // expected-error@+1 {{input can not be converted to an output tensor}}

// CHECK:           %[[VAL_5:.*]] = "tf.Const"() <{value = dense<[1, 0]> : tensor<2xi64>}> : () -> tensor<2xi64>
// CHECK:           %[[VAL_6:.*]] = "tf.Transpose"(%[[VAL_4]], %[[VAL_5]]) : (tensor<1x1xi32>, tensor<2xi64>) -> tensor<1x1xi32>
// CHECK:           return %[[VAL_6]] : tensor<1x1xi32>
// CHECK:         }

  %cst = "tf.Const"() {device = "", value = dense<2> : tensor<i32>} : () -> tensor<i32>
  // idx 1 should be a compile time constant
  %0 = "tf.ExpandDims"(%arg0, %cst) {device = ""} : (tensor<?x2xi32>, tensor<i32>) -> tensor<?x2x1xi32>
  %1 = "tf.AddV2"(%cst, %cst) {device = ""} : (tensor<i32>, tensor<i32>) -> tensor<i32>

  return %0 : tensor<?x2x1xi32>
}

  %1 = "tfl.pseudo_qconst"() {qtype = tensor<128x!quant.uniform<i32:f32, 0.7>>, value = dense<0> : tensor<128xi32>} : () -> tensor<128x!quant.uniform<i32:f32, 0.7>>