export MODELSCOPE_CACHE='/slow_disk2/ccl/datasets/多模态数据'
export CUDA_VISIBLE_DEVICES=2,3
export MAX_PIXELS=200704
swift rollout \
    --model Qwen/Qwen2.5-VL-3B-Instruct \
    --vllm_data_parallel_size 2

# sss
# export NO_PROXY=localhost,127.0.0.1,192.168.2.158
export MODELSCOPE_CACHE='/slow_disk2/ccl/datasets/多模态数据'
export WANDB_API_KEY=************************************
export WANDB_MODE=offline
export CUDA_VISIBLE_DEVICES=0,1
export MAX_PIXELS=401408
export NPROC_PER_NODE=2
swift rlhf \
    --rlhf_type grpo \
    --model Qwen/Qwen2.5-VL-3B-Instruct \
    --external_plugins ./ms-swift/examples/train/grpo/plugin/plugin.py \
    --reward_funcs external_r1v_acc format \
    --use_vllm true \
    --vllm_mode server \
    --vllm_server_host 127.0.0.1 \
    --vllm_server_port 8000 \
    --train_type lora \
    --freeze_vit true \
    --torch_dtype bfloat16 \
    --dataset 'AI-ModelScope/GEOQA_R1V_Train_8K' \
    --max_completion_length 1024 \
    --num_train_epochs 1 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 8 \
    --learning_rate 1e-6 \
    --gradient_accumulation_steps 2 \
    --save_strategy 'steps' \
    --eval_strategy 'steps' \
    --eval_steps 400 \
    --save_steps 400 \
    --save_total_limit 10 \
    --logging_steps 1 \
    --output_dir output/GRPO_GEOQA \
    --warmup_ratio 0.05 \
    --dataloader_num_workers 4 \
    --num_generations 8 \
    --temperature 1.0 \
    --repetition_penalty 1.1 \
    --system './ms-swift/examples/train/grpo/prompt.txt' \
    --deepspeed zero3 \
    --log_completions true \
    --report_to wandb \
    --num_iterations 2 \
    --async_generate false \
    --beta 0.001 \
    --max_grad_norm 0.5