LoRA-Torch

This codebase reimplementes LoRA: Low-Rank Adaptation of Large Language Models (ICLR 2022) and is reconstructed based on loralib.

Features

The implementations of loratorch and loralib are very different. We take the nn.Linear as an example as follows.

1. For loralib, $h = x W_0^\top + \frac{\alpha}{r} x(BA)^\top,$

where $x\in\mathbb{R}^{k\times n}$ is the input matrix, $W_0\in\mathbb{R}^{m\times n}$ is the pre-trained weight matrix, $r$ is the predefined LoRA rank, $B\in\mathbb{R}^{m\times r}$ and $A\in \mathbb{R}^{r\times n}$ are the LoRA matrixes, and $\alpha$ is a hyper-parameter.

2. For loratorch, $h = x (W_0 + \frac{\alpha}{r} BA)^\top.$

loralib computes $xW_0^\top$ and $x(BA)^\top$ respectively and then merges the results. While loratorch merges pre-trained weight $W_0$ and its LoRA weight $BA$ and then computes the results by simply using nn.Linear.forward(). There is no difference between loralib and loratorch in the linear layers. But in some no-linear or complex layers, we are no sure whether this layer satisfies $L(x, W_0)+L(x, BA) = L(x, W_0+BA)$. Hence, it is difficult to extend LoRA to some complex layers by using loralib. On the contrary, the idea of merging weights first in loratorch is more general and extensible. You just call merge_lora_param() in loratorch to merge weights and then call forward() in the original layer to compute the results. With the help of loratorch, you can easily implement LoRA to any type of layer of torch.nn.

Supported Layers

We compare the results of loralib and loratorch in examples to demonstrate the correctness of the implementation in loratorch.

Quick Start

The usage of loratorch is the same as loralib.

1. Install loratorch.

   pip install git+https://github.com/Baijiong-Lin/LoRA-Torch
   # Alternatively for developers
   # git clone https://github.com/Baijiong-Lin/LoRA-Torch
   # cd LoRA-Torch
   # pip install -e .
   

2. Replace the layers where you would like to use LoRA by using loratorch.

   # ===== Before =====
   # layer = nn.Linear(in_features, out_features)
   
   # ===== After ======
   import loratorch as lora
   # Add a pair of low-rank adaptation matrices with rank r=16 and alpha=32
   layer = lora.Linear(in_features, out_features, r=16, lora_alpha=32)
   

3. Mark only LoRA parameters as trainable before the training loop.

   model = Model()
   # This sets requires_grad to False for all parameters without the string "lora_" in their names
   lora.mark_only_lora_as_trainable(model)
   # Training loop
   for batch in dataloader:
       model.train()
       ...
   

4. Save LoRA model (only the LoRA matrixes will be saved).

   # ===== Before =====
   # torch.save(model.state_dict(), checkpoint_path)
   # ===== After =====
   torch.save(lora.lora_state_dict(model), checkpoint_path)
   

5. Load LoRA model (need to load the pre-trained model first).

   # Load the pre-trained checkpoint first
   model.load_state_dict(torch.load('ckpt_pretrained.pt'), strict=False)
   # Then load the LoRA checkpoint
   model.load_state_dict(torch.load('ckpt_lora.pt'), strict=False)
   

Contributor

loratorch is developed and maintained by Baijiong Lin.

Contact Us

If you have any question or suggestion, please feel free to contact us by raising an issue or sending an email to [email protected].

Acknowledgements

loratorch is heavily based on loralib. We thank its authors for their wonderful and open-source codebase.

Citation

If you find loratorch useful for your research or development, please cite the following:

@inproceedings{hu2022lora,
title={Lo{RA}: Low-Rank Adaptation of Large Language Models},
author={Edward J Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen},
booktitle={International Conference on Learning Representations},
year={2022},
}

@software{lin2023loratorch,
  author = {Baijiong Lin},
  title = {{LoRA-Torch}: {PyTorch} Reimplementation of {LoRA}},
  url = {https://github.com/Baijiong-Lin/LoRA-Torch},
  year = {2023}
}

License

loratorch is released under the MIT license.