func.return %arg0, %arg1, %arg2 : tensor<128x10xf32>, tensor<10x1024xf32>, tensor<128x1024xf32>
  }
}

// The following xla::OpSharding protos are used:
//  Serialized string:
//   "\08\03\1A\02\01\02\22\02\00\01"
//  Proto debug string:
//   type: OTHER

        Stage.from(1, "9") < Stage.from(2, "1")
        Stage.from(2, "1") > Stage.from(1, "9")

        Stage.from(1, "1") < Stage.from(1, "1a")
        Stage.from(1, "1a") > Stage.from(1, "1")

        Stage.from(1, "1a") < Stage.from(1, "1b")
        Stage.from(1, "1b") > Stage.from(1, "1a")

        Stage.from(1, "1c") < Stage.from(1, "2b")
        Stage.from(1, "2b") > Stage.from(1, "1c")
    }

00000110  b6 d7 49 a6 b5 68 1a 41  03 56 6b dc 5a 89        |..I..h.A.Vk.Z.|
>>> Flow 2 (server to client)
00000000  16 03 03 00 5d 02 00 00  59 03 03 1c ad ae e9 64  |....]...Y......d|
00000010  35 80 c6 04 f5 69 05 bb  5c 38 78 d3 99 2f ad e6  |5....i..\8x../..|
00000020  56 20 d4 79 89 1a 51 54  14 ac 6d 20 06 6d 1a bd  |V .y..QT..m .m..|

  func.func @main(%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 = ""}) {
    func.return
  }
}

// The following xla::OpSharding protos are used:
//  Serialized string:
//   "\08\03\1A\02\01\02\22\02\00\01"
//  Proto debug string:
//   type: OTHER

00000220  6f 6c 61 6e 67 30 0d 06  09 2a 86 48 86 f7 0d 01  |olang0...*.H....|
00000230  01 0b 05 00 03 81 81 00  9d 30 cc 40 2b 5b 50 a0  |.........0.@+[P.|
00000240  61 cb ba e5 53 58 e1 ed  83 28 a9 58 1a a9 38 a4  |a...SX...(.X..8.|
00000250  95 a1 ac 31 5a 1a 84 66  3d 43 d3 2d d9 0b f2 97  |...1Z..f=C.-....|
00000260  df d3 20 64 38 92 24 3a  00 bc cf 9c 7d b7 40 20  |.. d8.$:....}.@ |
00000270  01 5f aa d3 16 61 09 a2  76 fd 13 c3 cc e1 0c 5c  |._...a..v......\|

00000000  16 03 02 00 5d 02 00 00  59 03 02 79 2f de f0 f3  |....]...Y..y/...|
00000010  97 6a 5e e3 07 66 d9 f2  8b e9 be 1a 97 9f 6d c8  |.j^..f........m.|
00000020  9d b5 88 57 fc 8c 5d 7f  d7 3e 34 20 e5 4c 51 33  |...W..]..>4 .LQ3|
00000030  52 23 a0 52 77 ca 41 4a  1b e4 22 fb 1c 82 f8 e4  |R#.Rw.AJ..".....|
00000040  aa 13 c2 1a ce cd d5 42  de b0 e9 10 c0 13 00 00  |.......B........|

00000200  6f 6c 61 6e 67 30 0d 06  09 2a 86 48 86 f7 0d 01  |olang0...*.H....|
00000210  01 0b 05 00 03 81 81 00  9d 30 cc 40 2b 5b 50 a0  |.........0.@+[P.|
00000220  61 cb ba e5 53 58 e1 ed  83 28 a9 58 1a a9 38 a4  |a...SX...(.X..8.|
00000230  95 a1 ac 31 5a 1a 84 66  3d 43 d3 2d d9 0b f2 97  |...1Z..f=C.-....|
00000240  df d3 20 64 38 92 24 3a  00 bc cf 9c 7d b7 40 20  |.. d8.$:....}.@ |
00000250  01 5f aa d3 16 61 09 a2  76 fd 13 c3 cc e1 0c 5c  |._...a..v......\|

00000200  6f 6c 61 6e 67 30 0d 06  09 2a 86 48 86 f7 0d 01  |olang0...*.H....|
00000210  01 0b 05 00 03 81 81 00  9d 30 cc 40 2b 5b 50 a0  |.........0.@+[P.|
00000220  61 cb ba e5 53 58 e1 ed  83 28 a9 58 1a a9 38 a4  |a...SX...(.X..8.|
00000230  95 a1 ac 31 5a 1a 84 66  3d 43 d3 2d d9 0b f2 97  |...1Z..f=C.-....|
00000240  df d3 20 64 38 92 24 3a  00 bc cf 9c 7d b7 40 20  |.. d8.$:....}.@ |
00000250  01 5f aa d3 16 61 09 a2  76 fd 13 c3 cc e1 0c 5c  |._...a..v......\|

00000210  0d 06 09 2a 86 48 86 f7  0d 01 01 0b 05 00 03 81  |...*.H..........|
00000220  81 00 9d 30 cc 40 2b 5b  50 a0 61 cb ba e5 53 58  |...0.@+[P.a...SX|
00000230  e1 ed 83 28 a9 58 1a a9  38 a4 95 a1 ac 31 5a 1a  |...(.X..8....1Z.|
00000240  84 66 3d 43 d3 2d d9 0b  f2 97 df d3 20 64 38 92  |.f=C.-...... d8.|
00000250  24 3a 00 bc cf 9c 7d b7  40 20 01 5f aa d3 16 61  |$:....}.@ ._...a|

   }
   // CHECK-LABEL: func @tpu_func
   // CHECK: mhlo.type_extensions
   func.func @tpu_func (
      %arg0: tensor<2048xi32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}, %arg1: tensor<2048xi32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}) -> (tensor<2048xi32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}) {
  // TODO(b/292540052): Below tf.addV2 instruction is replaced with just
  // returning arg0 due to the workaround mentioned in the above bug. Revert