huggingface--peft/docs/source/package_reference/road.md

RoAd

RoAd is a parameter‑efficient fine‑tuning technique that adapts large language models by learning a small set of 2×2 rotation matrices (and optional scaling factors) applied to pairs of hidden dimensions. RoAd achieves competitive or superior performance compared to other PEFT methods with under 0.1% trainable parameters. Unlike LoRA’s batched low‑rank updates, RoAd’s sparse rotations reformulate to simple element‑wise operations, yielding significantly higher serving throughput when handling heterogeneous requests in the same batch, i.e. serving multiple adapters simultaneously. Moreover, RoAd integrates seamlessly into a distributed interchange intervention framework, interpreting its sparse 2D rotations as task-specific interventions within learned subspaces of hidden representations. These orthogonal subspaces can be composed to merge multiple task-specific behaviors—like multilingual capabilities or instruction following—without additional fine-tuning, enabling modular, interpretable adaptations in LLMs.

Finetuning with RoAd typically requires higher learning rate compared to LoRA or similar methods, around 1e-3. Currently RoAd only supports linear layers and it can be used on models quantized with bitsandbytes (4-bit or 8-bit).

For running inference with different RoAd adapters in the same batch see Inference with different LoRA adapters in the same batch.

Benchmark overview

API

RoadConfig

autodoc tuners.road.config.RoadConfig

RoadModel

autodoc tuners.road.model.RoadModel