attrs=['act: {"", "RELU", "RELU6", "TANH"} = ""'],
    derived_attrs=['T: {float, int8}'],
    outputs=['o: T'])
def _composite_fully_connected(input_, filter_, bias, act):
  res = tf.raw_ops.MatMul(
      a=input_, b=filter_, transpose_a=False, transpose_b=True)
  res = tf.raw_ops.Add(x=res, y=bias)
  if act == 'RELU':
    return tf.raw_ops.Relu(features=res)
  elif act == 'RELU6':

 Conv2D 53
 Conv2DBackpropFilter 53
 Conv2DBackpropInput 52
 Equal 1
 FusedBatchNormGradV2 53
 FusedBatchNormV2 53
 MatMul 3
 MaxPool 1
 MaxPoolGrad 1
 Mean 1
 Mul 218
 Pad 2
 ReadVariableOp 538
 Relu 49
 ReluGrad 49
 Reshape 2
 ResourceApplyKerasMomentum 161
 Slice 1
 Softmax 1
 SparseSoftmaxCrossEntropyWithLogits 1
 Squeeze 1
 Sum 1
 Tile 1
 Transpose 1
cluster 1 size 815
 AddN 1

  // CHECK: tfrt_fallback_async.executeop key(0) cost(100) device("cpu") "tf.Relu"(%{{.*}}) {T = f32} : 1
  %2 = tfrt_fallback_async.executeop key(0) cost(100) device("cpu") "tf.Relu"(%1) {T = f32} : 1
  // CHECK: tfrt_fallback_async.fallback_tensor_to_corert_tensorhandle

// CHECK-NEXT:      outputs: [ 5, 6, 7, 8 ],
// CHECK-NEXT:      builtin_options_type: LSTMOptions,
// CHECK-NEXT:      builtin_options: {
// CHECK-NEXT:        fused_activation_function: RELU,
// CHECK-NEXT:        cell_clip: 1.0,
// CHECK-NEXT:        proj_clip: 2.0,
// CHECK-NEXT:        kernel_type: BASIC
// CHECK-NEXT:      },
// CHECK-NEXT:      intermediates: [ ]
// CHECK-NEXT:    } ],

     (TFR_ConstantTensorOp (Arith_ConstantOp ConstantAttr<I32Attr, "127">))]>;

def QuantActRangeReluPattern :
  Pattern<
    (TFR_TFRQuantActRangeOp
     (TFR_ConstOp HasStringAttr<"RELU">:$act),
     (ConstantLikeMatcher F32Attr:$scale),
     (ConstantLikeMatcher I64Attr:$zp)),
    [(TFR_ConstantTensorOp (Arith_ConstantOp (Quantize<"0.0f"> $scale, $zp))),

   // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %1 = "tfl.add"(%arg0, %0) {fused_activation_function = "RELU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
   // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %2 = "tfl.relu"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
  // CHECK: tac.device = "CPU", tac.inference_type = "FLOAT"

  MatMul \
  Neg \
  Sum \
  Sub \
  Div \
  DivNoNan \
  Exp \
  Sqrt \
  SqrtGrad \
  Log1p

${generate} \
  --category=nn \
  SparseSoftmaxCrossEntropyWithLogits \
  ReluGrad \
  Relu \
  BiasAdd \
  BiasAddGrad

${generate} \
  --category=resource_variable \
  VarHandleOp \
  ReadVariableOp \
  AssignVariableOp \
  DestroyResourceOp

${generate} \
  --category=io \

def mnist_model():
  """Creates a MNIST model."""
  model = tf.keras.models.Sequential()
  model.add(tf.keras.layers.Flatten())
  model.add(tf.keras.layers.Dense(128, activation='relu'))
  model.add(tf.keras.layers.Dense(10, activation='softmax'))
  return model


class TestModule(tf.Module):

  def __init__(self):
    super(TestModule, self).__init__()
    self.model = mnist_model()

  // CHECK-NOT: tfrt_fallback_async.copy_if_small
  %0 = tfrt_fallback_async.executeop key(0) cost(1024) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.Relu"(%arg) {T = f32} : 1
  %1 = tfrt_fallback_async.executeop key(0) cost(1024) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.Relu"(%arg) {T = f32} : 1
  tfrt.return %0, %1 : !tfrt_fallback.tf_tensor, !tfrt_fallback.tf_tensor

      %0 = "tfl.conv_2d"(%arg0, %arg1, %arg2) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "RELU", padding = "VALID", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<1x4x4x3xf32>, tensor<3x1x1x3xf32>, tensor<3xf32>) -> tensor<1x4x4x3xf32>
      return %0 : tensor<1x4x4x3xf32>
    }