Closed JaheimLee closed 6 days ago
Yes, this is expected. Methods like OLoRA modify the base weights too. When you want to convert the OLoRA weights to LoRA weights, it needs to be ensured that the original base weights can be used. This is only possible by performing some changes on the OLoRA weights, which involves doubling their size. The reason is not quite straightforward to understand but it's explained here (this is for LoftQ but the same idea applies to OLoRA).
Ping @tokenizer-decode for info.
Yes, this is expected. Methods like OLoRA modify the base weights too. When you want to convert the OLoRA weights to LoRA weights, it needs to be ensured that the original base weights can be used. This is only possible by performing some changes on the OLoRA weights, which involves doubling their size. The reason is not quite straightforward to understand but it's explained here (this is for LoftQ but the same idea applies to OLoRA).
Ping @tokenizer-decode for info.
Got it, thanks for your reply
Yes, this is expected. Methods like OLoRA modify the base weights too. When you want to convert the OLoRA weights to LoRA weights, it needs to be ensured that the original base weights can be used. This is only possible by performing some changes on the OLoRA weights, which involves doubling their size. The reason is not quite straightforward to understand but it's explained here (this is for LoftQ but the same idea applies to OLoRA).
Ping @tokenizer-decode for info.
Found a new problem. After converting to lora model, r and alpha of the base model will be 2r and 2alpha. So maybe it's better to reset them to r and alpha after saving is finished.
Good point @JaheimLee, I created a PR to address that: #2077.
System Info
transformers
version: 4.44.2Who can help?
No response
Information
Tasks
examples
folderReproduction
Expected behavior
The lora_r of init adapter is 16.
But the converted one is 32.
Model size is also double. Is it as expected?