assert(scales_.size() == zero_points_.size());
  }

  // Quantize an Attribute by the quantization parameters. Return nullptr if
  // the conversion fails or the input array isn't an ElementsAttr.
  ElementsAttr convert(Attribute real_value);

 private:
  // Quantize an DenseFPElementsAttr by the quantization parameters.
  DenseElementsAttr convert(DenseFPElementsAttr attr);

A. Unique TensorFlower <******@****.***> 1713119208 -0700

                  const tensorflow::quantization::PyFunctionLibrary*
                      quantization_py_function_library = nullptr);

// Quantize the model with calibration data. Throw errors if `fully_quantize`
// is specified by the calibration data are not sufficient to quantize the
// model.
PyObject* MlirQuantizeModel(PyObject* data, bool disable_per_channel,
                            bool fully_quantize, int inference_type,

bool NotFromDequant(mlir::Value value) {
  auto dequant_op = value.getDefiningOp<DequantizeOp>();
  if (dequant_op) {
    return false;
  }
  auto split_op = value.getDefiningOp<SplitOp>();
  if (!split_op) {
    return true;
  }
  return !split_op.getValue().getDefiningOp<DequantizeOp>();
}

// Optimize TFLite operations in functions.
class OptimizeBatchMatmulPass

// RUN: stablehlo-quant-opt %s -split-input-file -stablehlo-prepare-quantize=bit-width=4 -verify-diagnostics | FileCheck %s

// CHECK-LABEL: func @dot_int4
// CHECK-SAME: (%[[ARG_0:.*]]: tensor<?x3xf32>) -> tensor<?x2xf32>
func.func @dot_int4(%arg0: tensor<?x3xf32>) -> tensor<?x2xf32> {
  // CHECK: %[[cst:.*]] = stablehlo.constant
  // CHECK: %[[q1:.*]] = "quantfork.qcast"(%[[cst]])
  // CHECK-SAME: quant.uniform<i8:f32, 0.0040316890267764818:127>

  EXPECT_FALSE(IsOpFullyQuantized(*add_op_itr));
}

TEST_F(IsOpFullyQuantizedTest, FalseIfOpPartiallyQuantized) {
  constexpr absl::string_view kQuantizeOp = R"mlir(
    func.func @quantize(%arg0: tensor<2xf32>) -> tensor<2x!quant.uniform<i8:f32, 1.000000e+00:0>> {
      %0 = stablehlo.uniform_quantize %arg0 : (tensor<2xf32>) -> tensor<2x!quant.uniform<i8:f32, 1.000000e+00:0>>

// RUN: tf-quant-opt %s -split-input-file -quant-prepare-quantize='post-training-quantize=true enable-per-channel-quantization=true' | FileCheck %s

module {
  func.func private @conv_with_bias_and_relu(%arg0: tensor<1x3x4x3xf32>) -> tensor<*xf32> {
    %cst = "tf.Const"() {device = "", value = dense<[7.11401462, 7.05456924]> : tensor<2xf32>} : () -> tensor<2xf32>

    // 1. Collect quantizable ops.
    QuantizationUnits quantizable_ops = GetQuantizableOps(op);
    if (quantizable_ops.empty()) {
      return failure();
    }

    // 2. Quantize collected ops.
    if (!QuantizeOps(rewriter, op, quantizable_ops)) {
      return failure();
    }

    // 3. Complete the Q-DQ pair for each inference type.
    if (!ConvertToFloat16Constant(rewriter, op)) {

#include "tensorflow/compiler/mlir/lite/utils/utils.h"

namespace mlir {
namespace TFL {
namespace tac {

bool NotTFLQuantDequantizeOp(Operation* op) {
  if (!op) return false;
  if (llvm::isa<TFL::QuantizeOp, TFL::DequantizeOp>(op)) return false;
  return true;
}

bool IsTerminatorOp(Operation* op) {
  if (!op) return false;
  return op->hasTrait<OpTrait::IsTerminator>();
}

// Try to guess the inference type of the op.

// RUN: tf-quant-opt %s -split-input-file -quant-prepare-quantize | FileCheck %s

module {
  func.func @same_scale_test(%arg0: tensor<*xf32>) -> tensor<*xf32> {
    %cst = arith.constant dense<[-1, 144]> : tensor<2xi32>
    %cst_1 = arith.constant dense<1.0> : tensor<144x10xf32>
    %cst_2 = arith.constant dense<0.1> : tensor<10xf32>
    %0 = "quantfork.qcast"(%arg0) : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 0.05:-10>>