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);

    // Finally, use the quantization parameter to create the quantize and
    // dequantize ops, and insert them between the tf.FakeQuantWithMinMaxVarsOp
    // and its users.
    auto quantize = rewriter.create<quantfork::QuantizeCastOp>(
        tf_op.getLoc(), qtype.getValue(), input);
    auto dequantize = rewriter.create<quantfork::DequantizeCastOp>(
        tf_op.getLoc(), res_type, quantize.getResult());

    func.return %dot_out : tensor<*x!tf_type.qint32>
  }

  // Quantize initial input at the start of the graph. Output is qint8.
  func.func @quantize_i8(%input : tensor<*xf32>, %input_scale : tensor<*xf32>, %input_zp : tensor<*xi32>) -> tensor<*x!tf_type.qint8> {
    %quantize = "tf.UniformQuantize"(%input, %input_scale, %input_zp) {
      Tin = "tfdtype$DT_FLOAT",
      Tout = "tfdtype$DT_QINT8",

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

      // The input of the quantize op has already been quantized, i.e.
      // rescale.
      return failure();
    }

    Operation* operand_op = operand.getDefiningOp();
    if (operand_op == nullptr) {
      // When `QuantizeOpT`'s operand does not have a defining op, it means it
      // is a `BlockArgument`. The pattern does not match if there is no op to
      // quantize.
      return failure();
    }

                  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,

    * Si on compare Uvicorn, il faut le comparer à d'autre applications de serveurs comme Daphne, Hypercorn, uWSGI, etc.
* **Starlette** :

 private:
  void runOnOperation() override;
};

// TODO: b/305815328 - Consider preserving leading and trailing QDQs for
// ModifyIONodesPass in TFLite use cases.
// Removes the back-to-back quantize and dequantize ops with volatile attribute.
class RemoveVolatileQdqPattern
    : public OpRewritePattern<quantfork::DequantizeCastOp> {
 public:
  explicit RemoveVolatileQdqPattern(MLIRContext* context)

// and tfl.dequantize pairs before tf.FakeQuant* being foled.
LogicalResult ConvertFakeQuantOps(func::FuncOp func, MLIRContext* ctx,
                                  bool use_fake_quant_num_bits) {
  OpBuilder builder(func);
  if (failed(UnwrapTFCustomOps(func, builder))) {
    return failure();
  }

  // Insert the tfl.quantize/tfl.dequantize ops after the tf.FakeQuant* ops to

// 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>