TimeSearch-R-raw / debug.sh
CSDDSFSFSAFSAF's picture
Add files using upload-large-folder tool
bbaf249 verified
# 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=2,3,4,5
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=4 \
--master_port=24137 \
time_r1/reflect_no_tool_three.py \
--input_path /data/shuimu.chen/LongVideoBench/LongVideoHaystack_all.json \
--save_path /data/shuimu.chen/TimeSearch-R/tool_output_reflect_V3 \
--data_root /data/shuimu.chen/LongVideoBench \
--model_base /data/shuimu.chen/TimeSearch-R/Qwen2.5-VL-GRPO \
--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