version = 5 : i64
  } : (tensor<2x2xf32>, tensor<2x3xf32>) -> tensor<2x3xf32>
  %2 = stablehlo.add %cst, %0 : tensor<2x3xf32>
  return %2 : tensor<2x3xf32>
}

// CHECK-LABEL: func.func @no_qdq_for_multiple_users
// CHECK-NOT: quantfork.qcast

}
func.func @_func(%arg0: tensor<2x4xf32>, %arg1: tensor<4x2xf32>) -> tensor<2x2xf32> {
  %0 = "tf.MatMul"(%arg0, %arg1) {_XlaSharding = "\08\03\1A\02\02\01\22\02\00\01"} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>
  %1 = "tf.Identity"(%0) : (tensor<2x2xf32>) -> tensor<2x2xf32>
  return %1 : tensor<2x2xf32>
}

// -----
// The following op sharding is used in the following test case:

    %26 = "tf.Cast"(%25) {Truncate = false} : (tensor<2xi64>) -> tensor<2xf32>
    %27 = "tf.Equal"(%14, %26) {incompatible_shape_error = true} : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xi1>
    %28 = "tf.Cast"(%27) {Truncate = false} : (tensor<2xi1>) -> tensor<2xf32>
    %29 = "tf.Sum"(%28, %6) {keep_dims = false} : (tensor<2xf32>, tensor<1xi32>) -> tensor<f32>

  %noncon3 = "tf.Square"(%noncon2) : (tensor<2xf32>) -> tensor<2xf32>

  %res1 = "tf.DivNoNan"(%arg0, %arg1) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
  %res2 = "tf.MulNoNan"(%arg0, %noncon2) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
  %res3 = "tf.DivNoNan"(%arg0, %noncon3) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
  func.return %res1, %res2, %res3 : tensor<2xf32>, tensor<2xf32>, tensor<2xf32>

func.func @add_dense_dense_float_mixfng_1_n() -> tensor<2x2xf32> {
  %cst_0 = arith.constant dense<[[1.5, -2.5]]> : tensor<1x2xf32>
  %cst_1 = arith.constant dense<[[-3.], [4.]]> : tensor<2x1xf32>

  %0 = "tfl.add"(%cst_0, %cst_1) {fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<2x1xf32>) -> tensor<2x2xf32>

  func.return %0 : tensor<2x2xf32>

    %input:2 = tf_executor.island wraps "tf.Const"() {dtype = "tftype$DT_INT32", value = dense<1.0> : tensor<4x6xf32>} : () -> tensor<4x6xf32>
    %split:3 = tf_executor.island wraps "tf.Split"(%dim#0, %input#0) : (tensor<i32>, tensor<4x6xf32>) -> (tensor<2x6xf32>, tensor<2x6xf32>)
    tf_executor.fetch
  }
  func.return

  %outputs_1, %outputs_2 = "tf.IdentityN"(%outputs, %outputs_0) {device = "TPU"} : (tensor<2x2xf32>, tensor<2x2xf32>) -> (tensor<2x2xf32>, tensor<2x2xf32>)
  func.return %outputs_0, %outputs_1 : tensor<2x2xf32>, tensor<2x2xf32>
}

// CHECK-LABEL: func @keep_identity_n_test
func.func @keep_identity_n_test(%arg0: tensor<2x2xf32>, %arg1: tensor<2x2xf32>) -> (tensor<2x2xf32>, tensor<2x2xf32>) {
    // CHECK: tf.MatMul

// WRAPPED-NEXT:   tf_executor.island wraps "tfl.custom_tf"
// WRAPPED-NEXT:     ^bb0(%arg1: tensor<*xf32>, %arg2: tensor<186xf32>, %arg3: tensor<186xf32>):
// WRAPPED-NEXT:   %[[fq:.*]] = "tf.FakeQuantWithMinMaxVarsPerChannel"(%arg1, %arg2, %arg3) <{narrow_range = true, num_bits = 8 : i64}> {device = ""} : (tensor<*xf32>, tensor<186xf32>, tensor<186xf32>) -> tensor<*xf32>

    %0 = "tf.MatMul"(%arg0, %arg1) {attr_map = "0:transpose_a,1:transpose_b", device = "", transpose_a = false, transpose_b = false} : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
    return %0 : tensor<2x2xf32>
  }

}

// CHECK-LABEL: pack_multiple
func.func @pack_multiple(%arg0: tensor<2x3xf32>,
                    %arg1: tensor<2x3xf32>,
                    %arg2: tensor<2x3xf32>) -> tensor<3x2x3xf32> {
  %0 = "tf.MyPack"(%arg0, %arg1, %arg2) {N=3:i32, axis=0:i32} : (tensor<2x3xf32>, tensor<2x3xf32>, tensor<2x3xf32>) -> tensor<3x2x3xf32>
  func.return %0 : tensor<3x2x3xf32>

// CHECK-NEXT: %[[AXIS:.*]] = arith.constant 0 : i32