Targs = [tensor<32xf32>, tensor<64xf32>, tensor<8xf32>, tensor<2xf32>],
      Tresults = [tensor<32xf32>, tensor<64xf32>, tensor<8xf32>]}
      : (tensor<32xf32>, tensor<64xf32>, tensor<8xf32>, tensor<2xf32>, tensor<2x!tf_type.string>)
        -> (tensor<32xf32>, tensor<64xf32>, tensor<8xf32>)
    tf_device.return %0#0, %0#1, %0#2 : tensor<32xf32>, tensor<64xf32>, tensor<8xf32>

  // Use other output
  %3:6 = "tf.FusedBatchNormV3"( %2#0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", U = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>)

    func.return %0#0 : tensor<8x8x8x8xf32>
  }

  // CHECK-LABEL: fusedBatchNormV3_training

func.func @BiasAddGrad_unranked(%arg0: tensor<*xf32>) -> tensor<?xf32> {
  // CHECK: tf.BiasAddGrad
  %0 = "tf.BiasAddGrad"(%arg0) {data_format = "NCHW"} : (tensor<*xf32>) -> tensor<?xf32>
  func.return %0 : tensor<?xf32>
}

// CHECK-LABEL: func @rsqrt_grad
// CHECK-SAME: (%[[ARG0:.*]]: tensor<2xf32>, %[[ARG1:.*]]: tensor<2xf32>)
func.func @rsqrt_grad(%arg0: tensor<2xf32>, %arg1: tensor<2xf32>) -> tensor<2xf32> {

// CHECK: return %[[DEQUANTIZE]] : tensor<8xf32>
}

// CHECK-LABEL: fakeQuantNotFolded
func.func @fakeQuantNotFolded(tensor<8xf32>, tensor<f32>, tensor<f32>) -> (tensor<8xf32>) {
^bb0(%arg0: tensor<8xf32>, %arg3: tensor<f32>, %arg4: tensor<f32>):
  %1 = "tf.FakeQuantWithMinMaxVars"(%arg0, %arg3, %arg4) {num_bits = 5, narrow_range = false} : (tensor<8xf32>, tensor<f32>, tensor<f32>) -> tensor<8xf32>
  func.return %1 : tensor<8xf32>

        tensor<20xf32>, tensor<20xf32>, tensor<20xf32>,
        tensor<20xf32>, tensor<20xf32>, tensor<20xf32>, tensor<20xf32>,
        tensor<20x20xf32>, none,
        tensor<1x20xf32>, tensor<1x20xf32>,
        none, none, none, none) -> tensor<1x28x20xf32>
    %1 = "quantfork.stats"(%0) {layerStats = dense<[-1.0, 2.0]> : tensor<2xf32>} : (tensor<1x28x20xf32>) -> tensor<1x28x20xf32>

     %2 = "tf.ReadVariableOp"(%arg2) {device = ""} : (tensor<*x!tf_type.resource<tensor<8xf32>>>) -> tensor<8xf32>
     "tf.LoadTPUEmbeddingAdagradParameters"(%0, %1) {config = "", num_shards = 1 : i64, shard_id = 0 : i64, table_id = -1 : i64, table_name = "table1"} : (tensor<8xf32>, tensor<8xf32>) -> ()
     %3 = "tf.ReadVariableOp"(%arg3) {device = ""} : (tensor<*x!tf_type.resource<tensor<8xf32>>>) -> tensor<8xf32>

// CHECK:   tf_executor.fetch %[[outputs]], %arg0 : tensor<2048xf32>, tensor<?x?x?x?xbf16>

func.func @tpu_load_embedding_ops_sink_controls(%arg0: tensor<*x!tf_type.resource<tensor<8xf32>>>, %arg1: tensor<*x!tf_type.resource<tensor<8xf32>>>, %arg2: tensor<*x!tf_type.resource<tensor<8xf32>>>, %arg3: tensor<*x!tf_type.resource<tensor<8xf32>>>) {
  tf_executor.graph {

}

// CHECK-LABEL: func @merge_with_dynamic_shape
func.func @merge_with_dynamic_shape(%arg0: tensor<2xf32>, %arg1: tensor<3xf32>) -> tensor<?xf32> {
  %result = tf_executor.graph {

// CHECK: tf_executor.Merge{{.*}}(tensor<2xf32>, tensor<3xf32>) -> (tensor<?xf32>, tensor<i32>, !tf_executor.control)

module {
  func.func @simpleTest(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>) -> tensor<2x1xf32> {
    %0 = func.call @func_0_GPU_FLOAT(%arg0, %arg1, %arg2) {tac.interface_name = "func_0"} : (tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
    %1 = func.call @func_1_GPU_FLOAT(%arg0, %arg3) {tac.interface_name = "func_1"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>