Official Repo for paper Let the Expert Stick to His Last: Expert-Specialized Fine-Tuning for Sparse Architectural Large Language Models by Zihan Wang, Deli Chen, Damai Dai, Runxin Xu, Zhuoshu Li and Y. Wu.
ESFT aims to efficiently customize Large Language Models (LLMs) with Mixture-of-Experts (MoE) architecture by adjusting only task-relevant parts, improving efficiency and performance while using fewer resources and storage.
📅 2024.8.11: We now release the ESFT training code! ✨ You can now try it with your own models and dataset!
git clone https://github.com/deepseek-ai/ESFT.git
cd esftpip install transformers torch safetensors acceleratebash scripts/download_adapters.sh- eval_multigpu.py This script evaluates the performance of the model on various datasets. See scripts/eval.sh for detailed configs and explanations.
Usage:
python eval_multigpu.py \
--eval_dataset=translation \
--base_model_path=deepseek-ai/ESFT-vanilla-lite \
--adapter_dir=all_models/adapters/token/translation \
--output_path=results/completions/token/translation.jsonl \
--openai_api_key=YOUR_OPENAI_API_KEY- get_expert_scores.py This script calculates the scores for each expert based on the evaluation datasets. Usage:
python scripts/expert/get_expert_scores.py \
--eval_dataset=translation \
--base_model_path=deepseek-ai/ESFT-vanilla-lite \
--output_dir=results/expert_scores/translation \
--n_sample_tokens=131072 \
--world_size=4 \
--gpus_per_rank=2- generate_expert_config.py This script generates the configuration to convert a MoE model with only task-relevant tasks trained based on evaluation scores. Usage:
python scripts/expert/generate_expert_config.py \
--eval_datasets=intent,summary,law,translation \
--expert_scores_dir=results/expert_scores \
--output_dir=results/expert_configs \
--score_function=token \
--top_p=0.2 # the scoring function and top_p are hyperparameters- train.py and train_ep.py This script trains the model with the expert configuration generated by the previous script. The train_ep.py file uses expert parallel and has been optimized for multi-GPU training. Usage:
python train.py \
--base_model_path=deepseek-ai/ESFT-vanilla-lite \
--expert_config=results/expert_configs/intent.json \
--train_dataset=intent \
--train_config=configs/base.yaml \
--output_dir=results/checkpoints/intent
torchrun --nproc-per-node=8 train_ep.py \
--base_model_path=deepseek-ai/ESFT-vanilla-lite \
--expert_config=results/expert_configs/translation.json \
--train_dataset=translation \
--train_config=configs/base.yaml \
--output_dir=results/checkpoints/translation
For bug reports, feature requests, and general inquiries, please open an issue on our GitHub Issues page. Make sure to include as much detail as possible to help us address your issue quickly.
- ☑️ 📝 Update models, evaluation scripts, and expert selection scripts
- ☑️ 🔧 Update training scripts
- 🔲 🚀 More...
If you find our code or paper useful, please cite:
@article{wang2024letexpertsticklast,
title={Let the Expert Stick to His Last: Expert-Specialized Fine-Tuning for Sparse Architectural Large Language Models},
author={Zihan Wang and Deli Chen and Damai Dai and Runxin Xu and Zhuoshu Li and Y. Wu},
year={2024},
eprint={2407.01906},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2407.01906},
}