This repository contains the code for training K2-65B, a 65 billion parameter large language model from LLM360.
Note
This repository is under active development. If you have suggestions or find bugs, please open a GitHub issue or reach out.
To launch training, run:
bash scripts/pretrain_65b.sh
To convert model checkpoints from Megatron to HuggingFace format, run:
python convert_ckpt_to_hf.py --load_path <megatron_ckpt_dir> --save_path <huggingface_ckpt_dir>