PiSSA: Principal Singular Values and Singular Vectors Adaptation of Large Language Models

To parameter-efficiently fine-tune (PEFT) large language models (LLMs), the low-rank adaptation (LoRA) method approximates the model changes $ΔW \in \mathbb{R}^{m \times n}$ through the product of two matrices $A \in \mathbb{R}^{m \times r}$ and $B \in \mathbb{R}^{r \times n}$, where $r \ll \min(m, n)$, $A$ is initialized with Gaussian noise, and $B$ with zeros. LoRA freezes the original model $W$ and updates the "Noise & Zero" adapter, which may lead to slow convergence. To overcome this limitation, we introduce Principal Singular values and Singular vectors Adaptation (PiSSA). PiSSA shares the same architecture as LoRA, but initializes the adaptor matrices $A$ and $B$ with the principal components of the original matrix $W$, and put the remaining components into a residual matrix $W^{res} \in \mathbb{R}^{m \times n}$ which is frozen during fine-tuning. Compared to LoRA, PiSSA updates the principal components while freezing the "residual" parts, allowing faster convergence and enhanced performance. Comparative experiments of PiSSA and LoRA across 12 different models, ranging from 184M to 70B, encompassing 5 NLG and 8 NLU tasks, reveal that PiSSA consistently outperforms LoRA under identical experimental setups. On the GSM8K benchmark, Mistral-7B fine-tuned with PiSSA achieves an accuracy of 72.86%, surpassing LoRA's 67.7% by 5.16%. Due to the same architecture, PiSSA is also compatible with quantization to further reduce the memory requirement of fine-tuning. Compared to QLoRA, QPiSSA exhibits smaller quantization errors in the initial stages. Fine-tuning LLaMA-3-70B on GSM8K, QPiSSA attains an accuracy of 86.05%, exceeding the performances of QLoRA at 81.73%. Leveraging a fast SVD technique, PiSSA can be initialized in only a few seconds, presenting a negligible cost for transitioning from LoRA to PiSSA. Code is available at https://github.com/GraphPKU/PiSSA.

Paper

Similar papers

Peer review

Reviewer m1t97/10 · confidence 4/52024-07-09

Summary

The paper presents a novel initialization strategy for LoRA finetuning where A and B are initialized according to the top r singular values and the pre-trained weights are initialzed with the remaining components. The authors show that this decomposition can help finetuning converge faster and behaves favorably compared to traditional quantization techniques such as NF4. The authors conduct experiments on NLG and NLU tasks show show performance improvements of PiSSA over vanilla LoRA.

Strengths

**Presentation** The theoretical motivation of the paper is clear and the method is well motivated and explained. **Empirical results** The authors provided several supporting ablations and experiments on improved quantization error compared to LoftQ and QLoRA, faster convergence etc.

Weaknesses

**Significance of results** The authors do not provide error bars in their results which questions whether the obtained results are actually statistically significant. Error bars should be provided as well as significance tests should be performed. **Unsupported claims** The authors claim that LoRA gradients point in random directions early during the fine-tuning stage which introduces wasteful update steps. Was this claim experimentally verified? If so, the authors should clarify how these results were obtained and elaborte on it in more detail in the paper. **Comparison to baselines** PiSSA is not compared to other commonly used initialization schemes, e.g., uniform gaussian [1], or kaiming init [2] for NLU and NLG experiments. Further, the authors compare PiSSA only to standard LoRA, but there have been several improvements to LoRA, such as AdaLoRA [3], or DoRA [4]. A comparison to those extensions would strengthen the paper. Also, the authors may investigate whether their initialization could be applied to these extensions to further improve perormance. Finally, some related work uses a similar initialization by considering principal components [5]. [1] Hu et al., LoRA: Low-Rank Adaptation of Large Language Models, ICLR 2022 [2] He et al., & Sun, J. Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification., ICCV 2015 [3] Zhang et al., AdaLoRA: Adaptive Budget Allocation for Parameter-Efficient Fine-Tuning, ICLR 2023 [4] Liu et al., DoRA: Weight-Decomposed Low-Rank Adaptation, ICML 2024 [5] Sheng et al., S-LoRA: Serving Thousands of Concurrent LoRA Adapters, MLSys 2024

Questions

- In line 174 the authors mention that $\alpha=r$, why was this choice made? In [1] $\alpha$ is set to a constant. Setting $\alpha=r$ ultimately results in different learning rates for different r. - The advantage of of quantized PiSSA over QLoRA might come from outlier removal, was there any experiments done comparing QPiSSA to efficient quantization techniques such as GPTQ [2] or llm.int8 [3]? - In line 188 the authors mention that training for NLG experiments was only done for 1 epoch. It would be interesting to see if LoRA ultimately converges to a similar optimum after longer training, or whether the improved initialization also leads to a better optimum. - What do the entries in the legend of Figure 2b mean exactly? [1] Hu et al., LoRA: Low-Rank Adaptation of Large Language Models, ICLR 2022 [2] Frantar et al., GPTQ: Accurate post-training quantization for generative pre-trained transformers., ICLR 2023 [3] Dettmers et al., LLM.int8(): 8-bit matrix multiplication for transformers at scale., NeurIPS 2022

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The fact that PiSSA needs to store 2r ranks compared to r ranks for LoRA which is mentioned in Appendix C should be also mentioned under Limitations.

Authorsrebuttal2024-08-10

Completing the Remaining Experiments

**Dear Reviewer M1T9**, In your questions Q4 and Q5, you suggested integrating PiSSA with both AdaLoRA and DoRA, as well as combining PiSSA with llm.int8 and GPTQ. During the rebuttal period, due to time constraints, we were only able to combine PiSSA with AdaLoRA and llm.int8, and the experimental results have shown the advantages of PiSSA when integrated. Over the past few days, we have completed the experiments combining PiSSA with DoRA and GPTQ, and the results are presented in the table below: | Method | Run 1 | Run 2 | Run 3 | GSM8K Average | | --- | --- | --- | --- | --- | | Full FT | 74.89 | 74.22 | 74.07 | 74.39±0.356 | | LoRA(gaussian-init) | 71.11 | 71.19 | 70.74 | 71.01±0.199 | | LoRA(kaiming init) | 72.25 | 71.57 | 71.95 | 71.92±0.279 | | **PiSSA** | **76.72** | **76.72** | **76.80** | **76.75±0.036** | | AdaLoRA | 72.48 | 72.42 | 72.02 | 72.31±0.202 | | **PiSSA+AdaLoRA** | **78.77** | **78.32** | **78.69** | **78.59±0.199** | | DoRA | 72.18 | 72.33 | 72.63 | 72.38±0.189 | | **PiSSA+DoRA** | **77.86** | **77.41** | **77.26** | **77.51±0.257** | | QLoRA+int8 | 71.78 | 71.48 | 71.78 | 71.68±0.143 | | **QPiSSA+int8** | **76.18** | **76.48** | **76.95** | **76.54±0.318** | | QLoRA+GPTQ-4bit | 70.51 | 70.43 | 69.60 | 70.18±0.419 | | **QPiSSA+GPTQ-4bit** | **74.60** | **74.30** | **74.83** | **74.58±0.217** | **Q4: Combining AdaLoRA or DoRA with PiSSA** **A4 part 2:** From lines 4, 7, and 8 of the table, it is evident that the performance of PiSSA combined with DoRA significantly surpasses that of DoRA alone and also exceeds the performance of PiSSA alone. Taking into account the combination experiments of PiSSA with AdaLoRA, it can be inferred that PiSSA benefits from the enhancement techniques of LoRA, demonstrating the potential of PiSSA when integrated with other methods. **Q5: Comparing QPiSSA to GPTQ or int8** **A5 part 2**: GPTQ handles each row $w$ independently, quantizing one weight at a time while updating all not-yet-quantized weights. Therefore, the method of using nuclear norm to calculate quantization error, as used in our paper, is not applicable for GPTQ. Thus, we turn to use Perplexity on WikiText-2 to calculate quantization error, where a lower Perplexity indicates reduced quantization error. The Perplexity of LLaMA-3-8B used in our experiments is **5.14**. After quantization with GPTQ-4bit, the Perplexity increased to **20.79**. However, using PiSSA, we were able to reduce the Perplexity to **6.23**. These results validate the effectiveness of PiSSA in reducing quantization error as discussed in our paper. From lines 11-12 of the table, it is evident that the performance of QPiSSA combined with GPTQ-4bit significantly surpasses that of QLoRA combined with GPTQ-4bit. This confirms the advantages mentioned in our paper, where QPiSSA retains the rapid convergence and superior performance characteristics of PiSSA. We believe we have addressed all of your questions comprehensively. If you have any additional queries or require further clarification, please do not hesitate to contact us. We are committed to providing any necessary information. If you find the contributions of our paper and the supplementary experiments we have provided to be valuable, might we kindly ask you to consider adjusting your score accordingly?

Reviewer m1t92024-08-10

I would like to thank the authors for the detailed response and commend them for the amount of experiments delivered during the rebuttal. I would like to highlight that claiming statistical significance based on the outcome of different experiments without verification is not convincing. Verifying statistical significance should be done for each experiment by providing variance estimates. Having said that, I greatly appreciate the reported error bars for GSM8K. I strongly recommend to the authors to report variance estimates for the remaining experiments as well and to verify statistical significance of their results. Further, the additional findings for applying PiSSA to other PEFT methods such as DoRA and AdaLoRA look very promising. Therefore I have decided to increase my score.

Authorsrebuttal2024-08-10

Thank you for your suggestions and recognition

In the camera-ready version, we will repeat the key experiments multiple times, include error bars, and verify statistical significance. Your professional insights have been invaluable in improving the quality of the PiSSA paper. We will also incorporate the additional experiments conducted during the rebuttal period into the final version of the paper. Thank you again for your suggestion towards PiSSA.

Reviewer Bb8M7/10 · confidence 4/52024-07-10

Summary

The authors present PiSSA, a relatively simple change to the LoRA framework leading to a large amount of demonstrated benefits. PiSSA proceeds by adjusting the initialization step of standard LoRA; rather than freezing the original weight matrix W and learning a low-rank perturbation W' = W + BA, PiSSA instead considers the SVD W = USV^T = U[:,r:]S[r:,r:]V[:,r:]^T + U[:,:r]S[:r,:r]V[:,:r]^T = W^{res} + AB, s.t. A=U[:,:r]S^{0.5}[:r,:r], B = S^{0.5}[:r,:r]V[:,:r]^T. Due to this backwards compatibility, PiSSA enjoys the fine-tuning speed/memory savings of LoRA while also being easily implementable to existing PEFT packages. Most importantly, the authors demonstrate across a large number of models, fine-tuning recipes, and benchmarks that PiSSA greatly improves on downstream accuracy compared to LoRA. Furthermore, the authors show that this initialization + fine-tuning strategy also leads to significant decreases in quantization error, outperforming both QLoRA and the recently proposed LoftQ both in terms of training loss (e.g., faster loss reduction) and performance for GSM8K and MATH tasks.

Strengths

**Originality** - the specific algorithm is an original, lightweight alteration to the LoRA framework. In particular, the fact that there are no additional hyperparameters and, after the initialization phase, PiSSA is essentially swappable with LoRA implementations are all significant benefits of the method. Furthermore, PiSSA's use of only fine-tuning the principle components in a LoRA fashion lends itself to strong intuition as to why the method leads to such strong downstream fine-tuning performance, and why LoRA is less successful in comparison. The latter point is also true when considering quantization. **Quality** - the extensive experiments across models and fine-tuning evaluations well convey the benefits of this method over standard LoRA. This is also true of the extensive quantization experiments. **Clarity** - for the most part, the paper is well written. However, there are specific points where the paper could improve clarity (see weaknesses below). **Significance** - the performance benefits and backwards compatibility lead to significant opportunity for future impact of this work.

Weaknesses

# Clarity There are several areas the paper could improve in clarify. In particular, the literature review requires more granularity, e.g., "AdaLoRA [42, 41, 43]," only one of these works is AdaLoRA. Furthermore, there is overlap between AdaLoRA and PiSSA in both their reliance on the SVD. It is necessary to discuss exactly how AdaLoRA leverages SVDs, and how this differs from PiSSA. There are several experimental details missing, e.g., are the GSM8K evals 8-shot? What are the shots used for the other benchmarks? Were the benchmarks run using the Eleuther eval harness? Please include these details in the paper. Paper at times states it is similar to LoRA and, at other times, seeks to differentiate itself: - "Unlike LoRA and its successors, which focus on learning low-rank approximations of weight updates, our PiSSA approach directly tunes the essential but low-rank parts of the model while keeping the noisier, high-rank, and nonessential parts frozen. Although our approach differs in philosophy from LoRA, it shares most of LoRA’s structural benefits and can be extended by these methods to enhance its performance" - "Since PiSSA shares the identical architecture with LoRA, it inherits most of LoRA’s benefits." This was confusing on an initial pass of the paper. A short, succinct list enumerating the similarities and differences early on may help readers understand exactly how PiSSA distinguishes itself from LoRA, but how it is able to inherit all of LoRA's PEFT benefits. # Quality The authors consider the weight matrix of a single attention layer in Llama-2 in Figure 2: "Since the residual model has removed the large-singular-value components, Wres has a narrower distribution than that of W, as can be seen in Figures 3a and 3b (comparing the singular value distributions of W and Wres), as well as Figures 3c and 3f (comparing the value distributions of W and Wres), which is beneficial for reducing the quantization error" <- However, since this only looks at a single layer of a single model, it hard to accept this as evidence for a general claim. Can you include a comparison of the distributions of mu/sigma for W and W^{res} (e.g., mean absolute differences between means and ratio between sigmas) across all layers for Llama2 and another model (e.g., Mistral) to show that, in general, W^{res} has a narrow distribution than W?

Questions

See weaknesses. Minor comment: "loftQ" in Table 3 <- "LoftQ"

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

4

Limitations

The authors have addressed potential limitations of the demonstrated method.

Reviewer XzvL8/10 · confidence 4/52024-07-14

Summary

In this paper, the authors proposed a modified low-rank adaptation (LoRA) method for fine-tuning large pretrained models. Specifically, the proposed method initializes the A and B matrices with singular matrices and freezes the weight matrix to be the residual. Further, the authors provide a quantization step to reduce the memory consumption. The authors demonstrate the effectiveness of this proposed method on a lot of models and tasks, showing that this method converges faster than the conventional LoRA method.

Strengths

- Overall, this paper is well-written. The author did a great job of presenting the new idea with details. The contents are organized well. - The authors comprehensively evaluated the proposed method, both quantitatively and qualitatively. - The proposed method is simple, yet effective.

Weaknesses

- It would be great if the authors could provide more insight into the proposed method. For example, is it possible to compare the subspace of the gradient of conventional LoRA? - The proposed method utilizes the first r eigenvectors. How about other dimensions in the subspace?

Questions

Please see the comments in the "limitation" block.

Rating

8

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

There is no negative societal impact as far as I can tell.

Authorsrebuttal2024-08-11

Official Comment by Authors

**Dear Reviewer XzvL**, The discussion period has passed the **halfway** point, and we have addressed the questions you previously raised. Do you have any additional questions or concerns that you would like to discuss? Our paper investigates the **gradient behavior of LoRA**, revealing that $A$ and $B$ initially have **zero gradients** and **random gradient** directions. This leads to **slow convergence** and potentially **suboptimal local minima**. To address these issues, we propose the PiSSA initialization method. **PiSSA approximates the optimization direction of full fine-tuning** by fine-tuning the principal components of a model. To our knowledge, PiSSA is the **first** method to use SVD on the original model, employing the principal singular values and vectors to initialize the adapter and fine-tune it, while fixing the remaining parts of the model during training. Our experiments demonstrate that PiSSA not only **converges faster** but also achieves **better final results** compared to LoRA. The initialization process is **efficient**, taking only seconds to minutes due to the use of fast singular value decomposition. We extend PiSSA by integrating it with **NF4**, **llm.int8**, and **GPTQ** quantization to create QPiSSA. This approach significantly **reduces quantization error** compared to QLoRA while retaining the **fast convergence** and **strong performance** of PiSSA. PiSSA modifies only the initialization method used by LoRA, making it compatible with various LoRA enhancements. We **combined PiSSA** with **LoftQ**, **AdaLoRA**, and **DoRA**, and our results show that PiSSA+ outperforms these methods and PiSSA, demonstrating its potential for further improvements. Our extensive experiments include **5 NLG** and **8 NLU** tasks using **12** different models ranging from **184M to 70B** parameters. We compared performance across **1k-9k+** training steps and ranks from **1 to 128**, evaluating against methods such as **LoRA**, **QLoRA**, **LoftQ**, **DoRA**, **AdaLoRA**, and **full-parameter fine-tuning**. We also examined the effects of initializing with **principal**, **medium**, and **minor** singular values and vectors. If you find the content in the main paper and our rebuttal satisfactory, would you consider revising your score?

Authorsrebuttal2024-08-12

Looking forward to your response

Dear Reviewer XzvL, We have addressed the concerns you raised in your initial review and have submitted a detailed rebuttal. We are writing to confirm whether you have had the opportunity to review our responses. We hope that our rebuttal has addressed your questions satisfactorily and would appreciate if you could reconsider the contributions of our manuscript, PiSSA, along with the efforts we have made during the rebuttal process. We are eager to hear your feedback and are open to any further discussion that might help clarify any remaining issues. Looking forward to your response. Best regards, Submission 5120 Authors

Reviewer XzvL2024-08-12

Response

I would like thank the authors for detailed response and additional experiments. My concerns are mostly addressed and I am happy to raise the scores.

Reviewer x9Pi8/10 · confidence 4/52024-07-14

Summary

The paper presents a method for creating lora-like fine-tuning adapters for base models. The idea is to initialize W, A and B matrices in LoRA from SVD decomposition. W is consists of less important ranks while A and B consist of the most important ranks. This implies that less important principle components of the model remain frozen while the most important ones get updated. Another benefit of this scheme is that A and B can be in high precision while W is in low precision, thus ensuring a reduced quantization error.

Strengths

I think this is a good paper with good results and a potential for significant impact as adapter tuning becomes more mainstream. The paper is written well, the method is intuitive, and the experimental results are thorough.

Weaknesses

As authors note in the limitations section there are many more experiments they could have conducted, but those are future scope. I am happy with the current scope and experiments. Having said, that I wish the authors had included at least a few experiments comparing advanced LoRA adapter methods. DoRA is mentioned in related works, but I couldn't find any comparison with it.

Questions

Will the authors be releasing their code? I believe it is very important that they do as our community could greatly benefit from being able to reproduce the experiments here and build on top of it.

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

4

Limitations

The limitations section in the paper is good enough for me.

Reviewer Bb8M2024-08-11

Acknowledgement of rebuttal

I thank the authors for their response to my original questions. I have read the other reviews and respective authors responses, and have no further questions or concerns.

Authorsrebuttal2024-08-11

To Reviewer Bb8M

We sincerely appreciate the professionalism and sense of responsibility you demonstrated throughout the rebuttal period. Your valuable suggestions have been instrumental, and we will incorporate them into the camera-ready version. We believe your recommendations will greatly enhance PiSSA's contribution to the field of efficient fine-tuning of LLMs.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC