%2 = "tfl.dequantize"(%1#0) : (tensor<2x!quant.uniform<u8:f32, 1.0>>) -> tensor<2xf32>
  %3 = "tfl.dequantize"(%1#1) : (tensor<2x!quant.uniform<u8:f32, 1.0>>) -> tensor<2xf32>

  // unused quantization ops should be removed as well.
  %4 = "tfl.dequantize"(%1#2) : (tensor<2x!quant.uniform<u8:f32, 1.0>>) -> tensor<2xf32>

// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range="enable-custom-op-quantization=CustomTestOp=1" -tfl-quantize="enable-dynamic-range-quantization=true enable-custom-op-weight-only=CustomTestOp=false" -tfl-post-quantize="enable-no-side-effect=CustomTestOp=false" | FileCheck --check-prefix=NotPrune %s

  // CHECK-DAG: %[[VAL2:.+]] = "tfl.dequantize"(%[[VAL0]])
  // CHECK-DAG: %[[VAL3:.+]] = "tfl.dequantize"(%[[VAL1]])
  // CHECK-DAG: %[[VAL4:.+]] = "tfl.conv_2d"(%arg0, %[[VAL2]], %[[VAL3]]) <{dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32}>

    %0 = "tfl.dequantize"(%arg0) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<100x!quant.uniform<i8:f32, 2.000000e-01:-3>>) -> tensor<100xf32>
    %1 = "tfl.dequantize"(%arg1) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<100x!quant.uniform<i8:f32, 2.000000e-01:-3>>) -> tensor<100xf32>

// CHECK: %[[dequantize:.*]] = "tf.PartitionedCall"(%[[conv_quant]], %[[out_scale]], %[[out_zp]])
// CHECK-SAME: f = @dequantize_i8

// CHECK: %[[conv_float:.*]] = "tf.PartitionedCall"(%arg0, %[[w_float]], %[[b_float]])
// CHECK-SAME: f = @composite_conv2d_with_bias_and_relu6_fn_1

// CHECK: return %[[dequantize]], %[[conv_float]]

  // Inserts the Quantize ops for requantizing a block argument.
  void RequantizeArg(BlockArgument arg, RequantizeStates& states);

  // Inserts the Quantize and Dequantize ops to quantize the value and returns
  // the Quantize op.
  void RequantizeValue(Value value, RequantizeStates& states, Location loc);

  // Returns the quantization parameter satisfies the same scale

            op->user_begin()->hasTrait<OpTrait::IsTerminator>())
          return failure();
      }
      // If the quantize op is a requantize op, it is being used in other scale
      // adjustments and should be kept. Instead, moving dequantize op before
      // the requantize op to remove the unnecessary requantize op.
      if (auto qtype = quant::QuantizedType::getQuantizedElementType(
              q.getInput().getType())) {

    const auto float_graph = model_->subgraphs()->Get(subgraph_idx);
    // The output graph should have an extra tensor from the added dequantize
    // op.
    ASSERT_EQ(quantized_graph->tensors()->size(),
              float_graph->tensors()->size() + 1);
    // Check that a dequantize op exists.
    int32_t dequant_input_idx = -1;
    int32_t dequant_output_idx = -1;

// CHECK: %[[dequantize:.*]] = "tf.PartitionedCall"(%[[maxpool]]
// CHECK-SAME: f = @dequantize_i8
// CHECK: return %[[dequantize]]

// CHECK: -------- Quantization Summary --------
// CHECK: Number of quantized layers in the model
// CHECK: --------------------------------

// CHECK-DAG:       %[[VAL_8:.*]] = "tfl.pseudo_const"(){{.*}}dense<[384, 128]> : tensor<2xi32>
// CHECK:           %[[VAL_9:.*]] = "tfl.dequantize"(%[[VAL_0]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<384x512x!quant.uniform<i8:f32, 1.000000e-01>>) -> tensor<384x512xf32>