// 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>) {

      %2 = "tf.Cast"(%arg0) {Truncate = false} : (tensor<1x8x2xi32>) -> tensor<1x8x2xf32>
      tf_device.return %2 : tensor<1x8x2xf32>
    // CHECK: () -> tensor<1x8x2xf32>
    }) {device = "/device:CPU:0"} : () -> tensor<*xf32>
    // CHECK: "tf.Cast"(%{{.*}}) <{Truncate = false}> : (tensor<1x8x2xf32>) -> tensor<*xf32>
    // CHECK: (tensor<i32>, tensor<1x8x2xf32>) -> (tensor<1x8x1xf32>, tensor<1x8x1xf32>)

  %1 = "tf.TensorListGetItem"(%0, %arg2, %arg1) {device = "/job:localhost/replica:0/task:0/device:GPU:0"} : (tensor<!tf_type.variant<tensor<1x32xf32>>>, tensor<i32>, tensor<2xi32>) -> tensor<1x32xf32>
  %2 = "tf.TensorListGetItem"(%0, %arg3, %arg1) {device = "/job:localhost/replica:0/task:0/device:GPU:0"} : (tensor<!tf_type.variant<tensor<1x32xf32>>>, tensor<i32>, tensor<2xi32>) -> tensor<1x32xf32>

  // tf.XlaReduceScatter legalization
  //===--------------------------------------------------------------------===//
  // CHECK-LABEL: func @xla_reduce_scatter
  func.func @xla_reduce_scatter(%arg0: tensor<128x128xf32>) -> tensor<64x128xf32> {
      %cst = "tf.Const"() {value = dense<0> : tensor<i32>} : () -> tensor<i32>
      %cst_0 = "tf.Const"() {value = dense<[[0, 4], [1, 5], [2, 6], [3, 7]]> : tensor<4x2xi32>} : () -> tensor<4x2xi32>

// dimensions into a single dimension. For example,
//
//   %shape = arith.constant dense<[1, 128, 64]> : tensor<3xi32>
//   %reshape = tfl.reshape(%input, %shape) // %input: tensor<128x64xf32>
//   %fc = tfl.fully_connected(%reshape, %filter, %bias)
//           {keep_num_dims = false, weights_format = "DEFAULT"}
//
// can be canonicalized to
//