Open renxiangnan opened 1 year ago
The LoRA in this case is just used for training and then is merged back to the model. Do you want LoRA for the plug and play adaptors? If so I can rework it to go that route instead when I can get to it
The LoRA in this case is just used for training and then is merged back to the model. Do you want LoRA for the plug and play adaptors? If so I can rework it to go that route instead when I can get to it
I appreciate your detailed explanation, which helped me gain a clearer understanding of the topic. It would be excellent if you could incorporate the plug-and-play adaptors as part of this setup. If I understand correctly (please correct me if I am wrong), given that LoRA is designed to prevent catastrophic forgetting, I believe it might be beneficial to consider not merging the weights back into the model and instead retain this option. Doing so could also lead to a less hacky when it comes to loading the model.
Hi guys, in generation.py, I noticed following code snippet as below. Looks like LORA is not used for inference at all, or is there anything I missed ? Thank you