The only difference between original repo (https://github.com/facebookresearch/codellama) is modifying example_completion.py -> example_completion_colab.py
Because torchrun doens't work well in basic colab environment (in my case), I used torch.distributed.launch instead of torchrun.
https://colab.research.google.com/drive/1X0AcfX0m2o2kaN9x8_DtiEUoa9f23w_D?usp=sharing
This is my implementation on colab notebook. (instructions are written in Korean. But almost same as original repo.)
Requirements
- Google colab pro plan for High system RAM. (If you do hyperparameter tuning, you might be able to do it in the free version, but I'm not sure.)
- at least 12.5GB of free space on your Google Drive. (7b models = 12.5 GB)
- 7b models can be run with a T4 GPU (Free).