Skip to content

Comment on DyLoRA: Parameter Efficient Tuning of Pre-Trained Modelsparent

Comments

Why is the only advantage at training time? I might misunderstand something but with this method you can train once, and then deploy models that use arbitrary rank (according to end-users compute requirements) and expect to have a model that performs best for that specific rank.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.