[docs] Lora-like guides #1371
[docs] Lora-like guides #1371
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
There was a problem hiding this comment.
Thank you @stevhliu for consolidating all the LoRA like guides into a single guide while focusing on usage of different LoRA variants. Thank you for the HF Space collection of notebook examples.
Left a few suggestions and comments.
| </hfoption> | ||
| <hfoption id="AdaLoRA"> | ||
|
|
||
| [AdaLoRA](../conceptual_guides/adapter#adaptive-low-rank-adaptation-adalora) efficiently manages the LoRA parameter budget by assigning important weight matrices more parameters and pruning less important ones. In contrast, LoRA evenly distributes parameters across all modules. You can control the average desired *rank* or `r` of the matrices, and which modules to apply AdaLoRA to with `target_modules`. Other important parameters to set are `lora_alpha` (scaling factor), and `modules_to_save` (the modules apart from the AdaLoRA layers to be trained and saved). All of these parameters - and more - are found in the [`AdaLoraConfig`]. |
There was a problem hiding this comment.
There is a bit of an issue with the AdaLora example. For AdaLora, we need to call the method update_and_allocate during each training step, otherwise, the adaptation step is not performed. This requires to either write a custom training loop or to subclass the Trainer to add this step. An example of a custom training loop that calls this method is here:
Given this guide, as is, the reader may not be aware that this is required and thus use AdaLora incorrectly. These solutions come to mind:
- Don't provide the AdaLora example. Instead, refer to the existing examples with custom training loops.
- Don't use
Trainer, instead use a custom training loop that calls this method when AdaLora is being used. - Subclass
Trainerto perform this step if AdaLora is being used. I don't know enough aboutTrainerto be really sure how to do that correctly.
There was a problem hiding this comment.
Great point! I mention update_and_allocate + link to a custom training loop in a Warning in the training section for more visibility. I think we should still keep the example of showing the AdaLoraConfig though so users still know how to set it up.
* loras * review * fix * feedback * feedback
As discussed in #1332, this PR condenses the LoRA guides by showcasing one specific task type (image classification) and redirecting to other use cases/tasks stored on the PEFT Hub org. Instead, this guide expands on the other LoRA methods such as LoHa, LoKr, and AdaLoRA. The configs of these methods aren't optimized for the best performance, so any feedback you have on that would be super useful!