grpo训练qwen2.5-vl报错

[h66840]

Describe the bug
脚本执行中会卡住
MAX_PIXELS=200704
CUDA_VISIBLE_DEVICES=1,2,3,4,5,6
NPROC_PER_NODE=6
swift rlhf
--rlhf_type grpo
--model Qwen/Qwen2.5-VL-3B-Instruct
--external_plugins /rest/lhy/ms-swift-3.4.1/examples/train/grpo/plugin/plugin.py
--reward_funcs external_r1v_acc format
--use_vllm false
--train_type full
--torch_dtype bfloat16
--dataset /rest/lhy/.cache/modelscope/hub/datasets/lmms-lab___multimodal-open-r1-8k-verified
--max_completion_length 1536
--num_train_epochs 1
--per_device_train_batch_size 4
--per_device_eval_batch_size 4
--learning_rate 1e-7
--eval_steps 1000
--save_steps 1000
--save_total_limit 2
--logging_steps 5
--output_dir output
--warmup_ratio 0.05
--dataloader_num_workers 4
--dataset_num_proc 4
--num_generations 24
--temperature 1.0
--top_p 0.9
--top_k 50
--async_generate false
--system '/rest/lhy/ms-swift-3.4.1/examples/train/grpo/prompt.txt'
--deepspeed zero2
--log_completions true
--num_iterations 1
--num_infer_workers 2
--report_to tensorboard wandb

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)

[hjh0119]

CUDA_VISIBLE_DEVICES=1,2,3,4,5,6
NPROC_PER_NODE=6
--num_infer_workers 2

does not meet the requirements for async or colocate mode. If you are using the source code, please refer to the usage in the documentation.