| cd src/r1-v | |
| export DEBUG_MODE="true" # Enable Debug if you want to see the rollout of model during RL | |
| export LOG_PATH="./debug_log_2b.txt" | |
| CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node="4" \ | |
| --nnodes="1" \ | |
| --node_rank="0" \ | |
| --master_addr="127.0.0.1" \ | |
| --master_port="12351" \ | |
| src/open_r1/grpo.py \ | |
| --output_dir "YOUR_PATH/log_dvd" \ | |
| --model_name_or_path "Qwen/Qwen2-VL-7B-Instruct" \ | |
| --dataset_name "YOUR_PATH/data/train_dvd.jsonl" \ | |
| --deepspeed local_scripts/zero3.json \ | |
| --max_prompt_length 4096 \ | |
| --max_completion_length 512 \ | |
| --per_device_train_batch_size 1 \ | |
| --gradient_accumulation_steps 1 \ | |
| --learning_rate 1e-6 \ | |
| --logging_steps 1 \ | |
| --bf16 \ | |
| --report_to wandb \ | |
| --gradient_checkpointing true \ | |
| --attn_implementation flash_attention_2 \ | |
| --max_pixels 401408 \ | |
| --num_train_epochs 2 \ | |
| --run_name Qwen2-VL-7B-Video-dvd \ | |
| --save_steps 100 \ | |
| --max_grad_norm 20 \ | |
| --save_only_model true \ | |
| --num_generations 8 # number of outputs G in grpo, reduce it would lead to faster training and smaller memory cost but higher variance | |