About inference phase - Githubissues

Clin0212 / HydraLoRA

[NeurIPS'24 Oral] HydraLoRA: An Asymmetric LoRA Architecture for Efficient Fine-Tuning

72 stars 5 forks source link

Closed chenyou203 closed 5 days ago

chenyou203 commented 2 weeks ago

During the inference phase, how did you merge the original parameters with the parameters obtained from HydraLoRA?

Clin0212 commented 2 weeks ago

We replace the original LoRA class (peft) with a new architecture, allowing for direct inference using the modified and fine-tuned LoRA.