func (curve *CurveParams) affineFromJacobian(x, y, z *big.Int) (xOut, yOut *big.Int) {
	if z.Sign() == 0 {
		return new(big.Int), new(big.Int)
	}

	zinv := new(big.Int).ModInverse(z, curve.P)
	zinvsq := new(big.Int).Mul(zinv, zinv)

	xOut = new(big.Int).Mul(x, zinvsq)
	xOut.Mod(xOut, curve.P)
	zinvsq.Mul(zinvsq, zinv)
	yOut = new(big.Int).Mul(y, zinvsq)
	yOut.Mod(yOut, curve.P)
	return
}

// Add implements [Curve.Add].

// to the rotation schedule.
func ksRotate(in uint32) (out []uint32) {
	out = make([]uint32, 16)
	last := in
	for i := 0; i < 16; i++ {
		// 28-bit circular left shift
		left := (last << (4 + ksRotations[i])) >> 4
		right := (last << 4) >> (32 - ksRotations[i])
		out[i] = left | right
		last = out[i]
	}
	return
}

// creates 16 56-bit subkeys from the original key.

  // CHECK: %[[CUSTOM_AGGREGATOR_2:.*]], {{.*}}, {{.*}}, {{.*}} = "tf.CustomAggregator"(%[[XLA_CALL_MODULE:.*]])

  %cst = "tf.Const"() {value = dense<3.000000e-01> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
  %0 = "tf.XlaCallModule"(%arg0, %cst) {
    Sout = [#tf_type.shape<1x2x2x2>], _entry_function = @composite_conv_fn,
    _original_entry_function = "composite_conv_fn",
    _stablehlo_module_attrs = {}, _quantization_method = "weight_only_ptq { }",

  std::unordered_map<Node*, int> requested_grads;
  for (const Output& nout : inputs_) {
    if (pending_[nout.node()->id()] > 0) {
      DCHECK_GT(nout.node()->num_outputs(), 1);
      int idx = input_nodes_[nout];
      DCHECK(((*grad_outputs_)[idx].node() == nullptr));
      TF_RETURN_IF_ERROR(SumGradients(nout, &(*grad_outputs_)[idx]));
      ++requested_grads[nout.node()];
    }
  }
  for (const auto& p : requested_grads) {

### Court

Des **valeurs par défaut** sont définies pour tout, des configurations optionnelles sont présentent partout. Tous ces paramètres peuvent être ajustés afin de faire ce que vous voulez et définir l'API dont vous avez besoin.

Mais, **tout fonctionne** par défaut.

### Validation

* Validation pour la plupart (ou tous?) les **types de données** Python incluant:

}

// newSequencer returns a sequencer that allows concurrent tasks up to maxWeight
// and writes tasks' output to out and err.
func newSequencer(maxWeight int64, out, err io.Writer) *sequencer {
	sem := semaphore.NewWeighted(maxWeight)
	prev := make(chan *reporterState, 1)
	prev <- &reporterState{out: out, err: err}
	return &sequencer{
		maxWeight: maxWeight,
		sem:       sem,
		prev:      prev,
	}
}

  constexpr absl::string_view kXlaCallModuleOpWithQuantizationMethodAttr =
      R"mlir(
    func.func @main(%arg0: tensor<1x1x3xf32>, %arg1: tensor<3x4xf32>) -> tensor<1x1x4xf32> {

  %2 = "tf.XlaCallModule"(%1#0, %0) <{Sout = [#tf_type.shape<1x3>], dim_args_spec = [], disabled_checks = [], has_token_input_output = false, module = "", platforms = [], version = 5 : i64}> {_entry_function = @composite_dot_general_fn_1, _original_entry_function = "composite_dot_general_fn_1",...

    %1 = "tf.Identity"(%arg1) {device = ""} : (tensor<100352x10xf32>) -> tensor<100352x10xf32>
    %2 = "tf.XlaCallModule"(%0, %1) <{
        Sout = [#tf_type.shape<?x10>], dim_args_spec = [],
        disabled_checks = [], function_list = [],
        has_token_input_output = false, module = "", platforms = [],
        version = 5 : i64
    }> {