"MLIR dump file name.">,
    Option<"merge_fusion_with_dequantize_",
        "merge-fusion-with-dequantize",
        "bool", /*default=*/"false",
        "Whether to merge quantized conv/dot_general fusion with subsequent dequantize.">,
  ];
  let dependentDialects = [
    "mlir::arith::ArithDialect",
    "mlir::stablehlo::StablehloDialect",
    "mlir::quant::QuantizationDialect",

                           TFDynamicRangeQuantization>(ctx, quant_params) {}
};

// Removes quantize-dequantize pairs that are not used in the quantization.
// The benefit of this pattern is set to lower value than other patterns, so
// that the other patterns can work on quantize/dequantize ops first.
class RemoveUnusedQdqPattern
    : public OpRewritePattern<quantfork::DequantizeCastOp> {
 public:

  // CHECK: %[[DEQUANTIZE:.*]] = mhlo.uniform_dequantize %[[CONVERT_2]] : (tensor<2x!quant.uniform<i8:f32, 1.000000e+00:3>>) -> tensor<2xf32>
  // CHECK: return %[[DEQUANTIZE]] : tensor<2xf32>

  %0 = "tf.UniformQuantize"(%arg0, %scales, %zps) {
    quantization_axis = -1 : i64, quantization_min_val = -128 : i64, quantization_max_val = 127 : i64

      } : (tensor<i8>, tensor<*xf32>, tensor<*xi32>) -> tensor<*xf32>

    %clamp_max = "tf.Maximum"(%dequantize, %clip_min) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    %clamp_min = "tf.Minimum"(%clamp_max, %clip_max) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>

    func.return %clamp_min : tensor<*xf32>
  }

  // Dequantizes and applies quantized Relu by clipping.

        // correct float op should be the user of the last DequantizeOp.
        if (llvm::isa<QuantizeOpT>(user)) {
          user = *user->getResult(0).getUsers().begin();
        }
        if (auto dequantize = llvm::dyn_cast<DequantizeOpT>(user)) {
          // Replace all uses, except not quantizable ops that are being used in
          // the float backbone.

  %6 = "tfl.dequantize"(%5) : (tensor<1x401408x!quant.uniform<i8:f32, 3.906250e-03>>) -> tensor<1x401408xf32>
  func.return %6 : tensor<1x401408xf32>

// CHECK-LABEL: func @modified(%arg0: tensor<1x224x224x3xf32>) -> tensor<1x401408xf32>
// CHECK-NEXT: %[[shape:.*]] = arith.constant dense<[1, 401408]> : tensor<2xi32>

  // TODO: b/323478683 - Make the attribute being part of op definition.
  quantize->setAttr(kVolatileOpAttrName, builder_.getUnitAttr());

  // `original_result` has a use to `quantize`, so this will replace that use
  // by the result of `dequantize`. Remember to reset that use afterwards
  value.replaceAllUsesWith(dequantize);
  quantize.getOperation()->replaceUsesOfWith(dequantize, value);
}

    } : (tensor<*xf32>, tensor<*xf32>, tensor<*xi32>) -> tensor<*x!tf_type.qint32>
    func.return %quantize : tensor<*x!tf_type.qint32>
  }

  // Dequantize final graph output back to f32. Input is qint8.
  func.func @dequantize_i8(%input : tensor<*x!tf_type.qint8>, %input_scale : tensor<*xf32>, %input_zp : tensor<*xi32>) -> tensor<*xf32> {
    %dequantize = "tf.UniformDequantize"(%input, %input_scale, %input_zp) {
      Tin = "tfdtype$DT_QINT8",

    if (failed(candidate_ops) || candidate_ops->empty()) return failure();

    // Rewrite the floating-point ops to the quantized version, by fusing
    // preceding dequantize ops and succeding quantize ops.
    for (Operation* candidate_op : *candidate_ops) {
      // If it is requantize op, we shouldn't rewrite this op.
      if (isa<QuantizeOpT, DequantizeOpT>(candidate_op)) {
        return failure();
      }

  quant::QuantizationSpecs quant_specs_;
};

#include "tensorflow/compiler/mlir/lite/utils/generated_op_quant_spec_getters.inc"

// If the weight is applicable to dynamic range quantization, insert Quantize
// and Dequantize ops with either per-axis or per-tensor scale.
class PrepareDynamicRangeQuantizableOp
    : public OpRewritePattern<arith::ConstantOp> {
 public:
  explicit PrepareDynamicRangeQuantizableOp(