torch_dtype Can not passed in Qwen2VLGRPOTrainer。

[robinjoe93]

torch_dtype = model_init_kwargs.get("torch_dtype")

torchrun --nproc_per_node="2" \ --nnodes="1" \ --node_rank="0" \ --master_addr="127.0.0.1" \ --master_port="12345" \ src/open_r1/grpo.py \ --output_dir ./output/Qwen2-VL-2B-GRPO-CLEVR-70k \ --model_name_or_path /root/data/Qwen2.5-VL-3B-Instruct \ --dataset_name leonardPKU/GEOQA_R1V_Train_8K \ --max_prompt_length 512 \ --max_completion_length 512 \ --deepspeed local_scripts/zero3_offload.json \ --per_device_train_batch_size 1 \ --gradient_accumulation_steps 8 \ --logging_steps 1 \ --bf16 \ --beta 1e-3 \ --report_to wandb \ --gradient_checkpointing true \ --attn_implementation flash_attention_2 \ --max_pixels 401408 \ --num_train_epochs 2 \ --run_name Qwen2-VL-2B-GRPO \ --save_steps 200 \ --save_only_model true \ --use_vllm false \ --num_generations 8 \ --torch_dtype bfloat16

torch_dtype bfloat16 not useful in Qwen2VLGRPOTrainer