%0 = "tfl.pseudo_const"() {value = dense<0.962260901> : tensor<1xf32>} : () -> tensor<1xf32>
    %1 = func.call @func_0_GPU_FLOAT(%arg0, %0) {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_0"} : (tensor<1x200x200x200xf32>, tensor<1xf32>) -> tensor<1x200x200x200xf32>
    %2 = "tfl.pseudo_const"() {value = dense<0.895973444> : tensor<1xf32>} : () -> tensor<1xf32>

      %1 = "tf.Const"() {value = dense<1.0> : tensor<2xf32>} : () -> tensor<2xf32>
      %2 = "tf.Const"() {value = dense<2> : tensor<2xi32>} : () -> tensor<2xi32>
      %3 = "tf.Const"() {value = dense<1.0> : tensor<2xf32>} : () -> tensor<2xf32>
      %4 = "tf.Const"() {value = dense<0> : tensor<2xi32>} : () -> tensor<2xi32>

// CHECK-NEXT:      %12 = mhlo.dynamic_reshape %2, %11 : (tensor<4x256x?xf32>, tensor<3xi32>) -> tensor<4x256x?xf32>
// CHECK-NEXT:      %13 = "tfl.batch_matmul"(%3, %12) <{adj_x = false, adj_y = false, asymmetric_quantize_inputs = false}> : (tensor<4x4x256xf32>, tensor<4x256x?xf32>) -> tensor<4x4x?xf32>
// CHECK-NEXT:      %14 = "tfl.shape"(%arg0) : (tensor<4x4x256xf32>) -> tensor<3xi32>

        %6:2 = tf_executor.island wraps "tf.opC"() : () -> tensor<1xf32>
        %8:2 = tf_executor.island(%4#1) wraps "tf.opD"(%6#0) : (tensor<1xf32>) -> tensor<1xf32>
        tf_executor.fetch %8#0 : tensor<1xf32>
      }
      tf_executor.yield %3 : tensor<1xf32>
    }
    tf_executor.fetch
  }
  func.return
}

// CHECK:        tf_executor.island wraps "tf.opA"

    %3 = "tf.XlaCallModule"(%2, %1, %0) <{Sout = [#tf_type.shape<?x2>], module = "", version = 9 : i64}> {_entry_function = @composite_dot_general_with_bias_and_relu6_dynamic_fn_1, _original_entry_function = "composite_dot_general_with_bias_and_relu6_dynamic_fn_1", _tfl_quant_trait = "fully_quantizable"} : (tensor<?x2xf32>, tensor<2x2xf32>, tensor<2xf32>) -> tensor<?x2xf32>

  } : (tensor<1x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<1xf32>

  // CHECK: return %[[y]] : tensor<1xf32>
  func.return %0 : tensor<1xf32>
}

// -----

// CHECK-LABEL: func @uniform_quantize_dequantize
func.func @uniform_quantize_dequantize(%arg0: tensor<1xf32>) -> tensor<1xf32>
{
  %scales = "tf.Const"() { value = dense<1.0> : tensor<f32> } : () -> tensor<f32>

// MinElement-LABEL: QuantizeCustomOp
func.func @QuantizeCustomOp(%arg0: tensor<1x1x1x1xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) attributes {tf.entry_function = {inputs = "input", outputs = "custom_op"}} {
  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>} : (tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>
  %w_1 = arith.constant dense<127.0> : tensor<4096x1x1x1xf32>

    module {
      func.func @test_fold_constant(%arg0: tensor<*xf32>) -> (tensor<?x?x?x3xf32>) {
        %cst = "tf.Const"() {value = dense<2.000000e+00> : tensor<2x3x3x1xf32>} : () -> tensor<2x3x3x1xf32>
        %cst_0 = "tf.Const"() {value = dense<0.400000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
        %cst_1 = "tf.Const"() {value = dense<0.500000e+00> : tensor<3xf32>} : () -> tensor<3xf32>

     %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>

func.func @func(%arg0: tensor<4xf32>) -> tensor<4xf32> {
  %cst = "tf.Const"() {value = dense<23.0> : tensor<4xf32>} : () -> tensor<4xf32>
  %0 = "tf.XlaSharding"(%arg0) { _XlaSharding = "\01\02\03"} : (tensor<4xf32>) -> tensor<4xf32>
  %1 = "tf.AddV2"(%0, %cst) : (tensor<4xf32>, tensor<4xf32>) -> (tensor<4xf32>)
  %2 = "tf.AddV2"(%cst, %1) : (tensor<4xf32>, tensor<4xf32>) -> (tensor<4xf32>)