| set -e | |
| export LMMS_EVAL_LAUNCHER="accelerate" | |
| export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 | |
| export NCCL_NVLS_ENABLE=0 | |
| output_path=logs/$(TZ="Asia/Shanghai" date "+%Y%m%d") | |
| model_path=GeoThinker-8B | |
| benchmark=vsibench | |
| accelerate launch --num_processes=8 --main_process_port 29505 -m lmms_eval \ | |
| --model geothinker \ | |
| --model_args pretrained=$model_path,use_flash_attention_2=true,max_num_frames=32,max_length=12800,max_pixels=451584,min_pixels=12544 \ | |
| --tasks ${benchmark} \ | |
| --batch_size 1 \ | |
| --output_path $output_path |