func.func @check_propagation_for_output_sharding_from_tf_matmul(%arg0: tensor<2x4xf32>, %arg1: tensor<4x2xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>) {
  %0 = "tf_device.cluster_func"(%arg0, %arg1) {func = @_func, use_spmd_for_xla_partitioning = true, use_tpu = true, num_cores_per_replica = 2 : i64} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>

  %3 = "tf.TensorListStack"(%1#3, %cst_0) {device = "/job:localhost/replica:0/task:0/device:CPU:0", num_elements = 2 : i64} : (tensor<!tf_type.variant<tensor<*xf32>>>, tensor<1xi32>) -> tensor<2x8xf32>
  return %3, %2 : tensor<2x8xf32>, tensor<2x8xf32>
}


// -----

// Convert a while with multiple tensor array to map_fn

      %cst_2, %cst_2, %cst_2, %cst_2) {cell_clip = 1.000000e+01 : f32, fused_activation_function = "TANH", proj_clip = 0.000000e+00 : f32, time_major = false}
    : ( tensor<1x28x28xf32>,
        tensor<20x28xf32>, tensor<20x28xf32>, tensor<20x28xf32>, tensor<20x28xf32>,
        tensor<20x20xf32>, tensor<20x20xf32>, tensor<20x20xf32>, tensor<20x20xf32>,
        none, none, none,
        tensor<20xf32>, tensor<20xf32>, tensor<20xf32>, tensor<20xf32>,

  %0 = "tf.MatMul"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", device = "/device:CPU:0", name = "MatMul", transpose_a = false, transpose_b = true} :
(tensor<40x37xf32>, tensor<40x37xf32>) -> tensor<40x40xf32>
  func.return %0 : tensor<40x40xf32>
// CHECK-LABEL: matmul_transposed_b
// CHECK: %[[CST:.*]] = "tfl.no_value"() <{value}> : () -> none

  %3 = "tf.AddV2"(%arg0, %0): (tensor<4x4xf32>, tensor<1xf32>) -> tensor<4x4xf32>
  %4 = "tf.Log"(%3) {device = "/job:localhost/replica:0/task:0/device:GPU:0"}: (tensor<4x4xf32>) -> tensor<4x4xf32>

  // CHECK: %[[ADD1:.*]] = "tf.AddV2"
  // CHECK: %[[LOG1:.*]] = "tf.Log"(%[[ADD1]])
  %5 = "tf.AddV2"(%4, %1): (tensor<4x4xf32>, tensor<1xf32>) -> tensor<4x4xf32>

@testLstmWithInvalidInputsRankMatch(%arg0: tensor<1x4xf32>, %arg1: tensor<4x2xf32>, %arg2: tensor<4x2xf32>, %arg3: tensor<4x2xf32>, %arg4: tensor<4x2xf32>, %arg5: tensor<4x4xf32>, %arg6: tensor<4x4xf32>, %arg7: tensor<4x4xf32>, %arg8: tensor<4x4xf32>, %arg9: tensor<4xf32>, %arg10: tensor<4xf32>, %arg11: tensor<4xf32>, %arg12: tensor<1x4xf32>, %arg13: tensor<4xf32>, %arg14: tensor<4xf32>, %arg15: tensor<4xf32>, %arg16: tensor<4x4xf32>, %arg17: tensor<4xf32>, %arg18: tensor<4xf32>, %arg19: tensor<4xf32>,...

  %size = "tf.Const"() {value = dense<10> : tensor<i32>} : () -> tensor<i32>
  // CHECK-NOT: tf.TensorArrayV3
  // CHECK: %[[TA_BUFFER:.*]] = "tf.MlirLocalVarOp"() : () -> tensor<!tf_type.resource<tensor<10x3xf32>>>
  // CHECK: "tf.AssignVariableOp"(%[[TA_BUFFER]]
  // CHECK-NOT: tf.TensorArrayV3

    } else {
      // Recurse on the subtypes in the variant/resource. Basically if the input
      // were:
      //   tensor<!tf_type.variant<tensor<?x8xf32>>>
      // and:
      //   tensor<!tf_type.variant<tensor<10x8xf32>>>
      // we'll try here to refine tensor<?x8xf32> with tensor<10x8xf32>.
      auto refined_subtype = mlir::cast<TensorType>(
          TypeMeet(lhs_element_type_with_subtype.GetSubtypes().front(),

    data and model parallelism in an easier manner.

    For example, the following:

    ```mlir
    !rtype = type tensor<!tf_type.resource<tensor<10x3xf32>>>
    func @data_and_model_parallelism(%arg0: !rtype, %arg1: !rtype, %arg2: !rtype, %arg3: !rtype) -> !rtype {

a manner will allow subsequent cluster formation pass to handle IR with both
data and model parallelism in an easier manner.

For example, the following:

```mlir
!rtype = type tensor<!tf_type.resource<tensor<10x3xf32>>>
func @data_and_model_parallelism(%arg0: !rtype, %arg1: !rtype, %arg2: !rtype, %arg3: !rtype) -> !rtype {