run_behavior_qwendual.sh

# export NCCL_SOCKET_IFNAME=bond0
# export NCCL_IB_HCA=mlx5_2,mlx5_3

# # used for check save when communication
# export NCCL_BLOCKING_WAIT=1
# export NCCL_ASYNC_ERROR_HANDLING=1
# export NCCL_TIMEOUT=1000  # timeout set to 1 hour (unit: seconds)

###########################################################################################
# === Please modify the following paths according to your environment ===
# Framework_name=QwenDual
Framework_name=QwenDual
base_vlm=./playground/Pretrained_models/nora # must be a local path, due to simpler will run in other where
base_vlm=./playground/Pretrained_models/Qwen2.5-VL-3B-Instruct
base_vlm=./playground/Pretrained_models/Qwen3-VL-4B-Instruct
freeze_module_list="qwen_vl_interface" # just for fast debug, sota is under fully FT, i.g., freeze_module_list=""
freeze_module_list=''
connect_layer_index=37

config_yaml=./starVLA/config/training/starvla_cotrain_behavior.yaml
oxe_data_root=playground/Datasets
data_mix=BEHAVIOR_challenge
# data_mix=BEHAVIORDataConfig
run_root_dir=./results/Checkpoints
export task_id=0

export run_id=1117_${data_mix}_${Framework_name}_task${task_id}

# BEHAVIOR_rgp_seg
# === End of environment variable configuration ===
###########################################################################################

export future_action_window_size=49
export action_horizon=50

# export WANDB_MODE=disabled

output_dir=${run_root_dir}/${run_id}
mkdir -p ${output_dir}
# mv this script to the output dir
cp $0 ${output_dir}/
  # --framework.action_model.connect_layer_index ${connect_layer_index} \

accelerate launch \
  --config_file starVLA/config/deepseeds/deepspeed_zero2.yaml \
  --num_processes 8 \
  starVLA/training/train_starvla.py \
  --config_yaml ${config_yaml} \
  --framework.name ${Framework_name} \
  --framework.qwenvl.base_vlm ${base_vlm} \
  --framework.action_model.future_action_window_size ${future_action_window_size} \
  --framework.action_model.action_horizon ${action_horizon} \
  --framework.action_model.state_dim 44 \
  --datasets.vla_data.data_root_dir ${oxe_data_root} \
  --datasets.vla_data.data_mix ${data_mix} \
  --datasets.vla_data.task_id ${task_id} \
  --datasets.vla_data.include_state True \
  --datasets.vla_data.per_device_batch_size 32 \
  --trainer.freeze_modules ${freeze_module_list} \
  --trainer.max_train_steps 100000 \
  --trainer.save_interval 2000 \
  --trainer.logging_frequency 100 \
  --trainer.eval_interval 100 \
  --trainer.learning_rate.base 4e-5 \
  --run_root_dir ${run_root_dir} \
  --run_id ${run_id} \
  --wandb_project Behavior_17 \
  --wandb_entity zwanggk \
  # --is_debug True