lm_eval --model hf \
    --model_args pretrained=/project/models/1_qwen2-0.5B \
    --tasks hellaswag \
    --device cuda:0 \
    --batch_size 8