We would like to additionally save the trainable part of the model (e.g. head, decoder), so that it would be easy to use afterwords.

We can achieve this by including an io torch save plugin, which saves the part of the desired model along with the full one.