LoRA: Low Rank Adaptation of Large Language Models скачать в хорошем качестве

LoRA: Low Rank Adaptation of Large Language Models

LoRA stands for Low-Rank Adaptation. It is an efficient adaptation strategy that neither introduces inference latency nor reduces input sequence length while retaining high model quality. It allows for quick task-switching when deployed as a service by sharing the vast majority of the model parameters. The idea is to freeze the pretrained model weights and inject trainable rank decomposition matrices into each layer. For GPT-3, LoRA reduces number of trainable parameters by 10,000 times and the GPU memory requirement by 3 times. It is very popular for finetuning LLMs like instruction-based models including Alpaca and Vicuna. It is also used to tune stable diffusion to adapt the style of generated images. Here is the agenda for this video: 00:00:00 Why finetuning LLMs is difficult? 00:04:24 What is LoRA? 00:07:18 What are the main advantages of LoRA? 00:08:31 How can we apply LoRA to Transformers? 00:11:41 How does LoRA perform for RoBERTa, DeBERTa and GPT3? 00:12:59 Understanding Low-rank updates done by LoRA For more details, please look at https://arxiv.org/pdf/2106.09685.pdf and https://github.com/microsoft/LoRA Hu, Edward J., Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. "LoRA: Low-Rank Adaptation of Large Language Models." In International Conference on Learning Representations. 2022.