Adapters v1.3.0

This version is built for Hugging Face Transformers v4.57.x.

New

Add support for DoRA adapters (@julian-fong via #790):

DoRA (Liu et al., 2024) is an enhanced LoRA variant. See https://docs.adapterhub.ml/methods.html#dora.

More

• Add EAT model interface for LoRA adapter support (@mgustineli via #832)

Changed

• Upgrade supported Transformers version to v4.57.x (@lenglaender via #814, @calpt via #829)

Fixed

• Fix Prefix Finetuning for Group Query Attention (GQA) (@ha405 via #825)
• Handle NotImplementedError from get_input_embeddings() during init (@mgustineli via #831)
• Fix LoRAMergedLinear.get_n_heads() shape mismatch with fused QKV + "o" matrix (@mgustineli via #830)
• Fix path traversal issue for downloading adapter archives (@calpt via #833)