Emergent Mind

LoRA: Low-Rank Adaptation of Large Language Models

Published Jun 17, 2021 in cs.CL , cs.AI , and cs.LG


An important paradigm of natural language processing consists of large-scale pre-training on general domain data and adaptation to particular tasks or domains. As we pre-train larger models, full fine-tuning, which retrains all model parameters, becomes less feasible. Using GPT-3 175B as an example -- deploying independent instances of fine-tuned models, each with 175B parameters, is prohibitively expensive. We propose Low-Rank Adaptation, or LoRA, which freezes the pre-trained model weights and injects trainable rank decomposition matrices into each layer of the Transformer architecture, greatly reducing the number of trainable parameters for downstream tasks. Compared to GPT-3 175B fine-tuned with Adam, LoRA can reduce the number of trainable parameters by 10,000 times and the GPU memory requirement by 3 times. LoRA performs on-par or better than fine-tuning in model quality on RoBERTa, DeBERTa, GPT-2, and GPT-3, despite having fewer trainable parameters, a higher training throughput, and, unlike adapters, no additional inference latency. We also provide an empirical investigation into rank-deficiency in language model adaptation, which sheds light on the efficacy of LoRA. We release a package that facilitates the integration of LoRA with PyTorch models and provide our implementations and model checkpoints for RoBERTa, DeBERTa, and GPT-2 at https://github.com/microsoft/LoRA.
Comparison of GPT-3's validation accuracy with varying parameters shows LoRA's superior scalability and performance.


  • LoRA introduces a strategy for fine-tuning LLMs like GPT-3 with reduced computational cost by incorporating low-rank matrices and freezing pre-trained weights.

  • LoRA reduces storage requirements and simplifies task-switching with minimal overhead, making it an efficient solution for model adaptation.

  • Empirical studies validate that LoRA maintains or improves performance on natural language understanding and generation tasks with significantly fewer trainable parameters.

  • LoRA's approach indicates potential redundancies in LLM parameter space, offering insights into more resource-efficient training methodologies and setting the stage for further research.

Exploring Efficient Fine-Tuning of LLMs with Low-Rank Adaptation

Introduction to Low-Rank Adaptation (LoRA)

With the growing sizes of LLMs like GPT-3, the conventional approach of fine-tuning such models for specific tasks or domains has faced increasing challenges, mainly due to its computational and financial impracticalities. In this context, Low-Rank Adaptation (LoRA) emerges as a strategy that addresses these limitations by freezing the pretrained model weights and incorporating trainable rank decomposition matrices for adaptation, thereby significantly reducing the number of parameters that need to be trained.

Key Advantages of LoRA

LoRA brings forth several advantages worth noting:

  • Efficiency in Storage and Task-switching: By maintaining most of the pre-trained model's parameters frozen and only modifying low-rank matrices for task adaptation, LoRA drastically reduces the storage requirements and simplifies the process of switching between tasks without incurring a high overhead.
  • Reduced Training Resources: Since only a small fraction of the parameters are trained, there's a substantial decrease in the necessary computational resources, opening up the possibility of fine-tuning LLMs with a much lower barrier to entry.
  • No Increase in Inference Latency: A critical advantage of LoRA is that it does not add any latency to the model inference time, positioning it as a practical solution for real-world applications where response time is crucial.

Empirical Validation

Empirical studies have demonstrated LoRA's effectiveness across various models such as RoBERTa, DeBERTa, GPT-2, and GPT-3. Remarkably, despite the significant reduction in trainable parameters — by up to 10,000 times compared to full fine-tuning — LoRA achieves on-par or even superior performance in tasks encompassing both natural language understanding (NLU) and generation (NLG).

Practical Implications and Theoretical Insights

LoRA's efficiency does not merely lie in its practical benefits but also offers valuable theoretical insights. For instance, the empirical investigation sheds light on the intrinsic rank of model adaptations, revealing that the modification of weights necessary for task-specific adjustments often resides in a surprisingly low-dimensional subspace. This discovery opens new avenues for understanding the underpinnings of LLM adaptability and efficiency.

Moreover, the ability of LoRA to achieve significant reduction in training parameters without compromising model quality speaks volumes about the potential redundancies in the parameter space of LLMs, inviting further exploration into more resource-efficient training methodologies.

Future Directions

While LoRA represents a significant stride towards efficient model adaptation, it also sets the stage for numerous research opportunities. Among these, exploring the synergy between LoRA and other fine-tuning methods, delving deeper into the theoretical aspects of parameter efficiency in LLMs, and fine-tuning the approach based on the specific characteristics of different LLM architectures are areas ripe for investigation.


LoRA offers a compelling solution to the growing challenge of fine-tuning LLMs efficiently. By leveraging low-rank adaptations, it presents a way forward that balances the need for model customization with the practical constraints of computational resources and inference efficiency. As the field continues to evolve, the principles laid down by LoRA will undoubtedly influence future advancements in LLM fine-tuning methodologies.

Get summaries of Trending computer science papers delivered straight to your inbox

Unsubscribe anytime.