| #SBATCH -p Gveval # 队列名称 | |
| #SBATCH --quotatype=spot # 竞价类型 | |
| #SBATCH --nodes=1 # 节点数量 | |
| #SBATCH --ntasks=8 # 总进程数(与GPU数量一致) | |
| #SBATCH --gres=gpu:8 # 每张卡分配的GPU数量 | |
| #SBATCH --cpus-per-task=16 # 每个进程分配的CPU核心数 | |
| #SBATCH --job-name=qwen_train # 任务名称(自定义) | |
| #SBATCH --requeue # 任务重排,当任务被别人抢断后,可以重新排队,但需要程序自动处理resume | |
| #SBATCH --open-mode append # 日志写入方式 | |
| export http_proxy=http://hanjiaming:DXtIkuMPmgy3M3UnCrRhGIxSMMaZn8iit2Br6UdG32fscs2l1bKKQ690WYTC@10.1.20.50:23128/ | |
| export https_proxy=http://hanjiaming:DXtIkuMPmgy3M3UnCrRhGIxSMMaZn8iit2Br6UdG32fscs2l1bKKQ690WYTC@10.1.20.50:23128/ | |
| export HTTP_PROXY=http://hanjiaming:DXtIkuMPmgy3M3UnCrRhGIxSMMaZn8iit2Br6UdG32fscs2l1bKKQ690WYTC@10.1.20.50:23128/ | |
| export HTTPS_PROXY=http://hanjiaming:DXtIkuMPmgy3M3UnCrRhGIxSMMaZn8iit2Br6UdG32fscs2l1bKKQ690WYTC@10.1.20.50:23128/ ; | |
| export LMMS_EVAL_LAUNCHER="accelerate" | |
| benchmark=scannet_6frames # choices: [scan2cap, scanrefer, scannet_4frames, scannet_6frames] | |
| output_path=logs/$(TZ="Asia/Shanghai" date "+%Y%m%d") | |
| model_path=ckpts/Qwen2.5-Omni-3B-sftv2-full | |
| model_args_str="pretrained=$model_path,attn_implementation=flash_attention_2,max_num_frames=32" | |
| export PYTHONPATH=$(pwd)/src:$PYTHONPATH | |
| export TOKENIZERS_PARALLELISM=true | |
| apptainer exec -f -w --nv --bind /mnt:/mnt /mnt/petrelfs/hanjiaming/llama_factory/ \ | |
| accelerate launch --num_processes=8 --main_process_port 29503 -m lmms_eval \ | |
| --model qwen2_5_omni \ | |
| --model_args "$model_args_str" \ | |
| --tasks ${benchmark} \ | |
| --batch_size 1 \ | |
| --log_samples_suffix original \ | |
| --log_samples \ | |
| --output_path $output_path |