elixir-nx / bumblebee

Pre-trained Neural Network models in Axon (+ 🤗 Models integration)
Apache License 2.0
1.26k stars 90 forks source link

Migrate optional outputs to use global layer options #360

Closed jonatanklosko closed 4 months ago

jonatanklosko commented 4 months ago

Moves :output_hidden_states and :output_attentions from spec attributes to global layer options (https://github.com/elixir-nx/axon/pull/563).

This fixes contrastive search to respect the model, which is currently ignores because we recreate the model.