A new LLM, LLaMA was written in PyTorch. So to train LLaMA with TPU, we habe to convert to Tensorflow or run it on PyTorch/XLA.
This repository trains LLaMA with huggingface on PyTorch/XLA.
This repository is intended as a minimal, hackable and readable example to load LLaMA (arXiv) models and run inference. In order to download the checkpoints and tokenizer, fill this google form
In a conda env with pytorch / cuda available, run:
pip install -r requirements.txt
Then in this repository:
pip install -e .
Once your request is approved, you will receive links to download the tokenizer and model files.
Edit the download.sh script with the signed url provided in the email to download the model weights and tokenizer.
The provided example.py can be run on a single or multi-gpu node with torchrun and will output completions for two pre-defined prompts. Using TARGET_FOLDER as defined in download.sh:
torchrun --nproc_per_node MP example.py --ckpt_dir $TARGET_FOLDER/model_size --tokenizer_path $TARGET_FOLDER/tokenizer.model
Different models require different MP values:
| Model | MP |
|---|---|
| 7B | 1 |
| 13B | 2 |
| 33B | 4 |
| 65B | 8 |
- 1. The download.sh script doesn't work on default bash in MacOS X
- 2. Generations are bad!
- 3. CUDA Out of memory errors
- 4. Other languages
LLaMA: Open and Efficient Foundation Language Models -- https://arxiv.org/abs/2302.13971
@article{touvron2023llama,
title={LLaMA: Open and Efficient Foundation Language Models},
author={Touvron, Hugo and Lavril, Thibaut and Izacard, Gautier and Martinet, Xavier and Lachaux, Marie-Anne and Lacroix, Timoth{\'e}e and Rozi{\`e}re, Baptiste and Goyal, Naman and Hambro, Eric and Azhar, Faisal and Rodriguez, Aurelien and Joulin, Armand and Grave, Edouard and Lample, Guillaume},
journal={arXiv preprint arXiv:2302.13971},
year={2023}
}
See MODEL_CARD.md
See the LICENSE file.