An updated version can be found in a new repo
https://github.com/gmongaras/Wizard_QLoRA_Finetuning
A working example of a 4bit QLoRA Falcon/Llama2 model using huggingface
To start finetuning, edit and run main.py
Once finetuning is complete, you should have checkpoints in ./outputs. Before running inference, we can combine the LoRA weights with the original weights for faster inference and smaller GPU requirements during inference. To do this, run the merge_weights.py script with your paths.
Finally, you can run generate.py for example generation given the merged model.
The python requirements to run the script are located in requirements.txt
You should also download the Falcon weights of the 7B model here https://huggingface.co/tiiuae/falcon-7b and put the files in a directory ./tiiuae/falcon-7b or download the Llama-2 weights here https://huggingface.co/meta-llama/Llama-2-7b-hf and put them in a directory named ./llama-2
This script does not support multi-gpus on 4-bit finetuning. If I find a way to do this, I will update the script.
- The base model takes about 6 GB of memory.
- Finetuning depends on the adapter size, batch size, max length, etc. In the current configuration, the memory usage is about 8GB.
- If there is a shape error upon training, then bitsandbytes and/or peft are having issues. The best way to get around this issue is to completely uninstall them and reinstall them from the source:
python -m pip uninstall bitsandbytes transformers accelerate peft -y
python -m pip install git+https://github.com/huggingface/transformers.git git+https://github.com/huggingface/peft.git git+https://github.com/huggingface/accelerate.git git+https://github.com/timdettmers/bitsandbytes.git -U
- If you get the error
CUDA Setup failed despite GPU being available. Please run the following command to get more information, then you need to build bitsandbytes from the source and put it in your bits and bytes site-package by followinghttps://github.com/oobabooga/text-generation-webui/issues/147