训练Qwen3_32B_GRPO的时候rank0会重命名dataset_infos.json这个文件,训练报错

[Zikhjm]

Describe the bug
What the bug is, and how to reproduce, better with screenshots(描述bug以及复现过程，最好有截图)
训练Qwen3_32B_GRPO的时候rank0会重命名dataset_infos.json这个文件，导致其他rank找不到该文件

Your hardware and system info
Write your system info like CUDA version/system/GPU/torch version here(在这里给出硬件信息和系统信息，如CUDA版本，系统，GPU型号和torch版本等)

Additional context
Add any other context about the problem here(在这里补充其他信息)

[Zikhjm]

补充训练脚本如下：
export NPROC_PER_NODE=8
export NNODES=2
export MASTER_ADDR=172.16.210.230 # master ip
export BKCL_SOCKET_IFNAME=eth0
export BKCL_RDMA_NICS=mlx5_0,mlx5_0,mlx5_1,mlx5_1,mlx5_2,mlx5_2,mlx5_3,mlx5_3
NODE_RANK=0 swift rlhf --rlhf_type grpo --model model_zoo/llm-pytorch/Qwen3-32B/ --external_plugins /workspace/ms-swift/examples/train/grpo/plugin/plugin.py --reward_funcs external_countdown format --train_type full --torch_dtype float16 --attn_impl flash_attn --dataset '/home/bml/storage/Qwen3_32B_GRPO_case/Countdown-Tasks-3to4' --max_length 1024 --max_completion_length 512 --num_train_epochs 1 --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --learning_rate 1e-6 --gradient_accumulation_steps 8 --eval_steps 500 --save_steps 100 --save_total_limit 20 --logging_steps 1 --output_dir output/GRPO_COUNTDOWN --warmup_ratio 0.01 --dataloader_num_workers 4 --num_generations 4 --temperature 1.0 --system 'You are a helpful assistant. You first thinks about the reasoning process in the mind and then provides the user with the answer.' --deepspeed zero3_offload --num_iterations 1 --beta 0.001 --log_completions true --use_vllm true --vllm_mode colocate --vllm_gpu_memory_utilization 0.6 --vllm_max_model_len 8192 --vllm_enforce_eager true --vllm_tensor_parallel_size 8