EBFT: Effective and Block-Wise Fine-Tuning for Sparse LLMs

fine-tuning

1. LlamaV1-7B

python main.py --model decapoda-research/llama-7b-hf \
    --prune_method wanda \
    --density 0.5 \
    --sparsity_type unstructured \
    --learning_rate 0.0002 \
    --eval_zero_shot \

Acknowledgments

Our implementation partially reuses Wanda's code.