LoRA: Low-Rank Adaptation of Large Language Models

Re:LoRA: Low-Rank Adaptation of Large Language Models by stu2b50 on Friday March 24, 2023 @12:56PM (#35291965)

I mean the derivative of a constant is 0. So if all of the original weights are considered constants, then computing their gradients is trivial, since they’re just zero.

[ Parent | on HN ]

Re:LoRA: Low-Rank Adaptation of Large Language Models by jprafael on Friday March 24, 2023 @01:17PM (#35292306)

Computing gradients is easy/cheap. What this technique solves is that you no longer need to store the computed values of the gradient until the backpropagation phase, which saves on expensive GPU RAM, allowing you to use commodity hardware.

[ Parent | on HN ]