CA-LoRA: Adapting Existing LoRA for Compressed LLMs to Enable Efficient Multi-Tasking on Personal Devices

Recently, there has been a demand to deploy Large Language Models (LLMs) on personal devices such as laptops and smartphones. These LLMs have different model variants when handling different tasks. However, personal devices have limited resources and require reduced storage overhead. To address this, there are two key methods available: the first is model compression, which compresses LLMs into smaller sizes; the second is LoRA, which can transfer an LLM to other tasks with very few parameters, avoiding the storage of multiple model variants in multi-task scenarios by only preserving LoRAs. However, our experiments show that directly combining these two methods yields sub-optimal performance. Considering that the open-source community has already contributed many LoRAs to LLMs, we propose to adapt these existing LoRAs from the LLMs to their compressed version and introduce a Compression-Aware LoRA (CA-LoRA) framework. We incorporate knowledge inheritance and recovery strategies to recover the lost knowledge caused by model compression. Experiment results demonstrate that CA-LoRA outperforms the vanilla LoRA methods applied to a compressed LLM and achieves comparable performance to the non-compressed LLM with existing LoRA modules. The source code of CA-LoRA is available at https://github.com/thunlp/CA-LoRA.

Paper

Similar papers

Reviewer z39k7/10 · confidence 2/52024-05-10

Summary

This paper proposes a method to adapt a trained LoRA module for an LLM to create a LoRA module for it's compressed version by initializing the latter with the former before training, and learning an additional a task specific low rank recovery module to recover some of the knowledge lost from the base LM in the compression process. The technique is tested both for task specific adaptation and instruction tuning with T5-3B and adapters trained with the proposed method are shown to improve over standard LoRA for compressed versions of T5-3B obtained by quantization, MoEfication, structured pruning and unstructured pruning.

Rating

7

Confidence

2

Ethics flag

1

Reasons to accept

There is demand for efficient task specific models in many applications and this method shows promise as being an effective to perform task specific tuning of compressed LMs. The paper includes a result where on instruction tuning, compressed T5-3B + CA-LoRA outperforms T5-base + LoRA.

Reasons to reject

The experiments on this paper are restricted to only one model architecture - T5. This should be acknowledged as a limitation and potential direction for future work. It is also unclear why the paper does not compare to QLoRA.

Reviewer rQV54/10 · confidence 4/52024-05-13

Summary

The main motivation is on device personalization where having the big LLM on in device memory is expensive. Hence the authors try to understand ways this requirement can be reduced. They propose using some existing compression techniques on the base LLM and then learning Lora modules on top of it. They propose two types of Lora modules one initialized from a task-specific Lora module and another that learns how to recover the information loss that happens due to compression. However, this part if very confusing to me.

Rating

4

Confidence

4

Ethics flag

1

Reasons to accept

Problem setting kind-off makes sense.

Reasons to reject

I have some issues with the experimental setting and the method that are listed below.

Questions to authors

1. In CLM + LORA are the LORA modules initialized from the trained checkpoint as in CLM + CA-LORA. 2. The QA-Lora method is the strongest on MMLU and authors have omitted the results on it saying they took the results from the original paper. I feel like Table-3 should be completed with the results of QA-Lora on other benchmarks. 3. How are the recovery Lora modules initialized? 4. Can you provide training time comparisons CA-LORA with further training an inherited lora on top of a CLM? Because it seems like the cost overhead might be too high as distillation requires forward passes on two models. 5. Looking at Table 4, it seems like distillation is the least useful component, then it would be good to compare with inheret + recover at all the places in other tables? As it is not very clear how this combination performs at other places and if the additional training cost of distillation is worth it. 6. It does not conceptually make sense to me why do we need two lora modules in parallel to each other as the final output is just a linear combination of the two lora's anyways and both the lora's are trained on the same joint objective. Then why do we even need to recover lora. Can you just not have a single lora and use the objective in eqn-6 to obtain similar results? I feel like this would work equally well as the CA+LORA method.

Reviewer Zo747/10 · confidence 4/52024-05-17

Summary

The authors present CA-LoRA, a compression-aware parameter-efficient finetuning method based on LoRA. They train LoRA modules on an uncompressed LLM and then adapt them downstream to be used with a compressed version of the same LLM, with a relatively small number of additional parameters introduced in addition to the LoRA modules. They explore two model families -- T5 and Llama 2 -- across both SuperGLUE and instruction tuning tasks, and see noticeable improvements in nearly all scenarios they tested. The experiments are reasonable, the idea is straightforward yet novel, and motivation is sound. The significance of this work lies in their stated motivation: enabling strong, static (or rarely updated) task-specific models to be deployed in resource-constrained environments, such as mobile devices. As far as I am aware, this is a novel method for adapting compressed language models to downstream tasks. They are also clear about limitations: this method requires extra compute to adapt the full sized model before being compressed. Whether this tradeoff is worth it depends on the specific training and deployment scenario. The structure of the paper is good, but the clarity could be improved. For example, I am not sure why the extra parameters introduced are called “knowledge recovery modules” – based on my reading, they are essentially extra parameters (+ a nonlinear function) that allow the compressed model + LoRA weights extra freedom during finetuning, but they do not explicitly encode any knowledge from the original LLM. Additionally, I believe that many of the equations written are not vital to understanding the work, and replacing them with prose and/or diagrams would improve the overall readability of the paper.

Rating

7

Confidence

4

Ethics flag

1

Reasons to accept

This paper demonstrates an interesting and useful strategy for building small but performant task-specific LLMs. They show applicability to real world use cases, and ground their work in an important motivation: deploying strong, task-specific models in resource constrained environments. Additionally, their method is theoretically straightforward, and adds little latency to inference. Overall, the method described seems like a very useful way to adapt compressed LLMs in certain scenarios.

Reasons to reject

The first experiments focus on older models and tasks, which are less relevant for real world scenarios as described in the motivation. SuperGLUE tasks are fairly outdated, and I would prefer to see the instruction tuning experiments expanded and highlighted, because those results are also promising. The paper could also be written more clearly. The amount of equations is – in my opinion – unnecessary, and the paper would be more straightforward to read if some of these were changed to prose (and/or diagrams) instead.

Questions to authors

While the T5 experiments show value in CA-LoRA for task-specific finetuning, I believe the strength of the method lies in the instruction tuning experiments; I think they’re the most interesting and applicable to real world scenarios. If compute allows, I have a few suggestions for extending the IFT experiments: * Adding more evaluation domains, to cover other important properties of instruction-tuned LLMs: * Safety (XSTest, ToxiGen, etc) * Instruction-following ability (AlpacaEval 1 and/or 2, IFEval) * Math (GSM8K) * If your compute budget allows, I would recommend adding another baseline: * Llama 2 13B finetuned (with LoRA, most likely), on Alpaca. This would be a fairer baseline vs your method than Llama 2 13B without instruction tuning. * Try instruction tuning on a domain-specific IFT set, such as CodeAlpaca-20k, and then evaluating on domain-specific evaluations (e.g. for coding, HumanEval+ and MBPP+). I don’t think any of these are strictly necessary, but I think they’d help strengthen your argument (especially the evaluations I mentioned in the first bullet point). Additionally, if you have the space, I’d recommend moving Table 6 from the appendix to section 4. Overall, I really like your work!

Reviewer Zo742024-06-04

Great, thank you for your detailed response! I believe the changes you've mentioned will make the paper stronger, and I'm very interested to see the additional evaluation results in the camera ready version.

Program Chairsdecision2024-07-10

Decision

Accept

© 2026 NYSGPT2525 LLC