This repo provides easy means of inference, training, and evaluation of LLMs It uses Modal Labs compute to peform its operations.
IN PROGRESS: Currently working on implementing open source version of this previously private repo.
Currently only these tasks have been confirmed functional:
- text completion
Todo:
- add example data processing for LORA training
- fix testing loop to evaluate LORA training
- fix analyze outputs task
- fix lora inference script
pip install modal-client
modal token new
FIRST:
set the MODEL to the model name you'd like. see CACHED_MODELS_MAP for available entries or to add new ones
Run:
modal run modal_training.py --task task
e.g. training the llm
modal run modal_training.py --task train_llm