* 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>

// 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>

include "mlir/IR/OpBase.td"
include "mlir/IR/PatternBase.td"
include "mlir/Dialect/Func/IR/FuncOps.td"
include "tensorflow/compiler/mlir/lite/ir/tfl_ops.td"

// Both Quantize and Dequantize ops have side effects, so we have to define
// patterns to remove dead ones after the quantization rewrite.
def : Pat<(TFL_QuantizeOp:$op $in, $qt), (replaceWithValue $in), [(HasNoUseOf:$op)]>;

}

def QuantConvertSimulatedQuant
    : Pass<"quant-convert-simulated-quantization", "func::FuncOp"> {
  let summary = "Converts training-time simulated quantization ops to "
                "corresponding quantize/dequantize casts";
  let constructor = "mlir::quantfork::createConvertSimulatedQuantPass()";
}

              new_q_op.getResult());
          assign_variable_op->replaceAllUsesWith(new_assign_variable_op);
        }
        assign_variable_op.erase();
        dq_op.erase();
      } else {
        // Add quantize op.
        builder.setInsertionPoint(assign_variable_op);
        auto new_q_op = builder.create<QuantizeOp>(
            assign_variable_op.getLoc(), ref_qtype,

  StringRef getArgument() const final {
    // This is the argument used to refer to the pass in
    // the textual format (on the commandline for example).
    return "quant-propagate-quantize-type";
  }
  StringRef getDescription() const final {
    // This is a brief description of the pass.
    return "Propagate quantized type through allowed ops.";
  }

  void runOnOperation() override;
};

  (TFL_AddOp $lhs, (TFL_MulOp $rhs,
                  (Arith_ConstantOp ConstantAttr<RankedF32ElementsAttr<[]>,
                   "-1.0f">), TFL_AF_None), $act)>;

// Squash tfl.dequantize and tfl.quantize pairs.
// TODO(b/185915462): Compare the scale of input and output. This can also be
// squashed to a requantize op if the scales are different.