func.func @convert_argmax(%arg0: tensor<4x32x256xf32>) -> (tensor<4x32xf32>, tensor<4x32xi32>) {
  %0 = mhlo.constant dense<0xFF800000> : tensor<f32>
  %1 = mhlo.constant dense<0> : tensor<i32>
  %2 = "mhlo.iota"() <{iota_dimension = 0 : i64}> : () -> tensor<256xi32>
  %3 = "mhlo.broadcast_in_dim"(%2) <{broadcast_dimensions = dense<2> : tensor<1xi64>}> : (tensor<256xi32>) -> tensor<4x32x256xi32>
  %4:2 = "mhlo.reduce"(%arg0, %3, %0, %1) ({