File size: 1,134 Bytes
c96df66 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 |
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
model_name_or_path="lmsys/vicuna-7b-v1.5-16k"
# model_name_or_path="meta-llama/Llama-2-7b-chat-hf"
# model_name_or_path="Yukang/LongAlpaca-7B-16k"
# model_name_or_path="lmsys/longchat-7b-v1.5-32k"
super_tokenizer_name_or_path="/share/ninglu_shao/code/PluginTransformer/data/outputs/90k_0104+8-longalpaca_0106/super_tokenizer"
# super_tokenizer_name_or_path="/home/baaiks/ninglu/code/PluginTransformer/data/outputs/90k_0111+8-longalpaca_0111/super_tokenizer"
output_dir="data/results/longbench/test"
mkdir -p ${output_dir}
torchrun --nproc_per_node 8 --master_port=12345 -m main.eval_longbench \
--model_name_or_path ${model_name_or_path} \
--super_tokenizer_name_or_path ${super_tokenizer_name_or_path} \
--super_tokenizer_num_hidden_layers 8 \
--dataset_list "gov_report,qmsum,multi_news" \
--output_dir ${output_dir} \
--max_length 3500 \
--batch_size 2 \
--compression_ratio 32 \
--down_scale_method "uniform" \
| tee "${output_dir}/eval_longbench.log"
# narrativeqa,qasper,multifieldqa_en,hotpotqa,2wikimqa,musique,gov_report,qmsum,multi_news |