:py:mod:`kosmos.ml.models.vqc.circuit.qiskit_circuit.autograd_function`
=======================================================================

.. py:module:: kosmos.ml.models.vqc.circuit.qiskit_circuit.autograd_function







Classes
-------

.. py:class:: AutogradCtx

   Bases: :py:class:`Protocol`

   Autograd context for the QiskitAutogradFunction.







----

.. py:class:: QiskitAutogradFunction(*args, **kwargs)

   Bases: :py:class:`torch.autograd.Function`

   Custom autograd bridge between PyTorch and Qiskit for variational quantum circuits.





   |

   .. rubric:: Methods

   .. py:method:: forward(ctx: torch.autograd.function.FunctionCtx, x: torch.Tensor, weights: torch.Tensor, evaluator: collections.abc.Callable, gradient_fn: collections.abc.Callable) -> torch.Tensor

      Perform the forward pass by evaluating the quantum circuit via the provided evaluator.

      :param ctx: Autograd context to save information for the backward pass.
      :type ctx: FunctionCtx
      :param x: Input batch of shape (B, input_dim).
      :type x: torch.Tensor
      :param weights: Trainable weights of the circuit.
      :type weights: torch.Tensor
      :param evaluator: Callable that evaluates the quantum circuit and returns
                        the output values for the given inputs and weights.
      :type evaluator: Callable
      :param gradient_fn: Callable computing gradients of the circuit output with
                          respect to its parameters, e.g., via the parameter-shift rule.
      :type gradient_fn: Callable

      :returns: Output batch of shape (B, output_dim).
      :rtype: torch.Tensor


   .. py:method:: backward(ctx: AutogradCtx, *grad_outputs: torch.Tensor) -> tuple[None, torch.Tensor, None, None]

      Compute gradients using the provided quantum gradient function.

      :param ctx: Autograd context with saved tensors from the forward pass.
      :type ctx: AutogradCtx
      :param \*grad_outputs: Gradient of the loss with respect to
                             the forward output, shape (B, output_dim).
      :type \*grad_outputs: torch.Tensor

      :returns:

                Gradients for each forward input. Only the
                    gradient with respect to `weights` is returned; gradients for `model` and `x`
                    are None.
      :rtype: tuple[None, torch.Tensor, None, None]