Node* add1 =
        ops::BinaryOp("Add", unstage, b, builder.opts().WithName("add1"));
    Node* relu0 = ops::UnaryOp("Relu", add0, builder.opts().WithName("relu0"));
    ops::UnaryOp("Relu", add1, builder.opts().WithName("relu1"));
    MakeStageNode(builder, "stage", {DT_FLOAT}, {relu0});

    return GraphDefBuilderToGraph(builder, graph->get());
  };

      # Check activation functions are explicitly present.
      # If present the last op before return should be stablehlo.clamp for relu6
      # and stablehlo.maximum for relu.
      if activation_fn is nn_ops.relu6:
        self.assertRegex(module_str, r'stablehlo.clamp.*\n.*return')
      elif activation_fn is nn_ops.relu:
        self.assertRegex(module_str, r'stablehlo.maximum.*\n.*return')
    else:

     (HasRankAtMost<4> $a),
     (HasRankAtMost<4> $b)]>;
}

// We can eliminate Relu from Relu(SquaredDifference(x, y)),
// since the result of SquaredDifference is always non-negative.
// TFLite interpreter doesn't support Relu+int32 for now. So the test cases
// are failing without the following pattern to optimize Relu away fixes
// the problem.
def OptimizeReluSquaredDifference : Pat<

    %10 = "tf.Cast"(%9) {Truncate = false, device = ""} : (tensor<1x3xi32>) -> tensor<1x3xf32>
    %11 = "tf.Mul"(%10, %cst) {device = ""} : (tensor<1x3xf32>, tensor<f32>) -> tensor<1x3xf32>
    %12 = "tf.Relu"(%11) {device = ""} : (tensor<1x3xf32>) -> tensor<1x3xf32>
    return %12 : tensor<1x3xf32>
  }
// CHECK-LABEL: func @matmul_with_relu

  %1 = "tfl.mul"(%0, %arg2) {tac.device = "GPU", fused_activation_function = "RELU6", tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

// CHECK: return %[[MAX]] : tensor<?x28x28x16xf32>
// CHECK: }

// -----

// Because the operand of shape_of is other than the target conv,
// should not match conv relu dynamic pattern.

// CHECK-LABEL: @conv_with_relu_dynamic_shape_not_same_op_fn(
// CHECK-SAME:                    %[[ARG_0:.*]]: tensor<?x28x28x1xf32>

      dilations: Sequence[int] = (1, 1, 1, 1),
      padding: str = 'SAME',
  ):
    class DepthwiseConvModel(module.Module):
      """A simple model with a single depthwise conv2d, bias and relu."""

      def __init__(self):
        self.out_channel_size = filter_shape[2] * filter_shape[3]

        # This ensures filters will have different value range per out channel
        self.filters = np.stack(

  %5 = "quantfork.stats"(%4) {layerStats = dense<[-56.2916565, 122.922478]> : tensor<2xf32>} : (tensor<1x4xf32>) -> tensor<1x4xf32>
  %6 = "tfl.svdf"(%0, %1, %2, %3, %5) {fused_activation_function = "RELU", rank = 1 : i32} : (tensor<1x3xf32>, tensor<2x3xf32>, tensor<2x1xf32>, tensor<2xf32>, tensor<1x4xf32>) -> tensor<1x2xf32>

            // Others
            "AddN", "Bitcast", "Cast", "ClipByValue", "Const", "Empty",
            "Identity", "IdentityN", "Relu", "Relu6", "ReluGrad", "Relu6Grad",
            "LeakyReluGrad", "Elu", "EluGrad", "Selu", "SeluGrad", "Select",
            "SelectV2", "Transpose", "ConjugateTranspose",
            "_UnaryOpsComposition", "CollectiveReduceV2",
            "CollectiveAssignGroupV2",

 * <<#rel4.8:switch_to_publishing_plugins,Maven Publish and Ivy Publish Plugins>> that now support digital signatures with the <<signing_plugin#signing_plugin,Signing Plugin>>.
 * Use native <<#rel5.0:bom_import,BOM import>> in your builds.
 * The <<worker_api.adoc#worker_api,Worker API>> for enabling units of work to run in parallel.