File size: 1,532 Bytes
bb7f76d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 |
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"
# For resume training: --resume_from_checkpoint Model_Path \
# Set temporal to choose between T-GRPO and GRPO, and len_control to enable or disable the length control reward.
# Qwen/Qwen2.5-VL-3B-Instruct
CUDA_VISIBLE_DEVICES=0,1,2,3,5,6,7 torchrun --nproc_per_node="8" \
--nnodes="1" \
--node_rank="0" \
--master_addr="127.0.0.1" \
--master_port="12365" \
src/open_r1/grpo.py \
--output_dir "./log/Qwen2.5-VL-3B-GRPO" \
--model_name_or_path 'Qwen/Qwen2.5-VL-3B-Instruct' \
--dataset_name "./Video-R1-data/Video-R1-260k.json" \
--deepspeed local_scripts/zero3.json \
--max_prompt_length 16384 \
--max_completion_length 768 \
--per_device_train_batch_size 1 \
--gradient_accumulation_steps 1 \
--learning_rate 1e-6 \
--lr_scheduler_type "cosine" \
--weight_decay 0.01 \
--bf16 \
--logging_steps 1 \
--gradient_checkpointing true \
--temporal true \
--len_control true \
--attn_implementation flash_attention_2 \
--max_pixels 401408 \
--num_train_epochs 1 \
--run_name Video-R1 \
--save_steps 100 \
--beta 0.04 \
--max_grad_norm 5 \
--save_only_model false \
--num_generations 8 # number of outputs G in grpo, reduce it would lead to faster training and smaller memory cost but higher variance
python /apdcephfs_sh2/share_300000800/user/zongxia/Video-R1/gpu_burn.py |