| # The IP address from the above step | |
| export SIGLIP_URL=grpc://127.0.0.1:51000 | |
| # 1. 确保这里列出了所有你想用的卡,比如 4 张卡 | |
| export CUDA_VISIBLE_DEVICES=0,2,3,4 | |
| export PYTHONPATH=$PYTHONPATH:/data/shuimu.chen/TimeSearch-R | |
| # 2. 将 nproc_per_node 改为 4 (与上面的卡数一致) | |
| torchrun \ | |
| --nproc_per_node=1 \ | |
| --master_port=24137 \ | |
| time_r1/test.py \ | |
| --input_path /data/shuimu.chen/LongVideoBench/LongVideoHaystack_all_id.json \ | |
| --save_path /data/shuimu.chen/TimeSearch-R/COT_DATA \ | |
| --data_root /data/shuimu.chen/LongVideoBench \ | |
| --model_base /data/shuimu.chen/qwen2.5vl-Instruct-72b \ | |
| --prompt_template v4 \ | |
| --use_env True \ | |
| --use_vllm True \ | |
| --batch_size 1 \ | |
| --num_data_workers 0 \ | |
| --total_video_tokens 24000 \ | |
| --max_frames 768 \ | |
| --max_tokens 256 |