calpt

This PR adds the (IA)^3 adapter architecture proposed in _"Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning"_ by Liu et al. (2022). - Paper: https://arxiv.org/pdf/2205.05638.pdf - Reference implementation:...

Fix Deberta self attention LoRA linear

1

Fixes #398. Depends on #396.

reduction_factor -> bottleneck_size

11

closes Adapter-Hub/website#25

Closes #77. This PR replaces the adapter-transformers code currently used in trankit by the newly released adapters package. Changes: - Add `adapters` & `transformers` as dependency - Small adaptations regarding...

Upgrade to Adapters library and recent Transformers versions

1

Hey Trankit team, First of all, thanks for this great toolkit! I'm one of the authors and maintainers of AdapterHub, which you use for your adapter implementations. As you might...

This PR: - introduces new methods `adapter_to()` & `adapter_fusion_to()` to move only adapter weights to device or convert dtype - avoids moving full model in `AdapterTrainer` when loading best model....

Resolves #661. Methods for saving & loading adapters, prediction heads & adapter fusions get a new parameter `use_safetensors` to save & load via Safetensors.

Changes in this PR: - Remove deprecated add_fusion() & train_fusion() methods - Remove deprecated support for passing lists to adapter activation - Add "adapters." prefix to version identifier when saving...