Troubleshooting

If you encounter any issue when using PEFT, please check the following list of common issues and their solutions.

Examples don't work

Examples often rely on the most recent package versions, so please ensure they’re up-to-date. In particular, check the version of the following packages:

• peft
• transformers
• accelerate
• torch

In general, you can update the package version by running this command inside your Python environment:

python -m pip install -U <package_name>

Installing PEFT from source is useful for keeping up with the latest developments:

python -m pip install git+https://github.com/huggingface/peft

Bad results from a loaded PEFT model

There can be several reasons for getting a poor result from a loaded PEFT model, which are listed below. If you’re still unable to troubleshoot the problem, see if anyone else had a similar issue on GitHub, and if you can’t find any, open a new issue.

When opening an issue, it helps a lot if you provide a minimal code example that reproduces the issue. Also, please report if the loaded model performs at the same level as the model did before fine-tuning, if it performs at a random level, or if it is only slightly worse than expected. This information helps us identify the problem more quickly.

Random deviations

If your model outputs are not exactly the same as previous runs, there could be an issue with random elements. For example:

1. please ensure it is in .eval() mode, which is important, for instance, if the model uses dropout
2. if you use generate on a language model, there could be random sampling, so obtaining the same result requires setting a random seed
3. if you used quantization and merged the weights, small deviations are expected due to rounding errors

Incorrectly loaded model

Please ensure that you load the model correctly. A common error is trying to load a trained model with get_peft_model, which is incorrect. Instead, the loading code should look like this:

from peft import PeftModel, PeftConfig

base_model = ...  # to load the base model, use the same code as when you trained it
config = PeftConfig.from_pretrained(peft_model_id)
peft_model = PeftModel.from_pretrained(base_model, peft_model_id)

Randomly initialized layers

For some tasks, it is important to correctly configure modules_to_save in the config to account for randomly initialized layers.

As an example, this is necessary if you use LoRA to fine-tune a language model for sequence classification because 🤗 Transformers adds a randomly initialized classification head on top of the model. If you do not add this layer to modules_to_save, the classification head won’t be saved. The next time you load the model, you’ll get a different randomly initialized classification head, resulting in completely different results.

In PEFT, we try to correctly guess the modules_to_save if you provide the task_type argument in the config. This should work for transformers models that follow the standard naming scheme. It is always a good idea to double check though because we can’t guarantee all models follow the naming scheme.

When you load a transformers model that has randomly initialized layers, you should see a warning along the lines of:

Some weights of <MODEL> were not initialized from the model checkpoint at <ID> and are newly initialized: [<LAYER_NAMES>].
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.

The mentioned layers should be added to modules_to_save in the config to avoid the described problem.