Improve links to autogram engine

Author: ValerianRey

docs/source/examples/iwrm.rst

@@ -18,7 +18,7 @@ backpropagates the Jacobian of each sample's loss. It uses an
 ``.grad`` fields of the model's parameters. Because it has to store the full Jacobian, this approach
 uses a lot of memory.
 
-The recommended approach, called the :doc:`autogram <../docs/autogram/index>` engine, works by
+The recommended approach, called the :doc:`autogram engine <../docs/autogram/engine>`, works by
 backpropagating the Gramian of the Jacobian of each sample's loss with respect to the model's
 parameters. This method is more memory-efficient and generally much faster because it avoids
 storing the full Jacobians. A vector of weights is then computed by applying a

docs/source/index.rst

@@ -46,7 +46,7 @@ the gradient of the obtained weighted loss. The iterative computation of the Gra
 Algorithm 3 of
 `Jacobian Descent For Multi-Objective Optimization <https://arxiv.org/pdf/2406.16232>`_. The
 documentation and usage example of this algorithm is provided in
-:doc:`Engine <docs/autogram/engine>`.
+:doc:`autogram.Engine <docs/autogram/engine>`.
 
 TorchJD is open-source, under MIT License. The source code is available on
 `GitHub <https://github.com/TorchJD/torchjd>`_.