feat: add the vlnce training script

set -x

export VLLM_ATTENTION_BACKEND=XFORMERS

export PYTHONHASHSEED=0

SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"

# run python -m vagen.server.server in a tmux session first

# max_trajectory_length = max_prompt_length + max_response_length

python3 -m vagen.trainer.main_ppo \

    algorithm.adv_estimator=masked_gae \

    algorithm.high_level_gamma=0.95 \

    data.train_files=data/vlnce/train.parquet \

    data.val_files=data/vlnce/train.parquet \

    data.train_batch_size=32 \

    data.val_batch_size=32 \

    data.max_prompt_length=1024 \

    data.max_response_length=128 \

    data.max_trajectory_length=20000 \

    data.image_key=images \

    data.truncation=error \

    actor_rollout_ref.model.path=Qwen/Qwen2.5-VL-3B-Instruct \

    actor_rollout_ref.actor.optim.lr=1e-6 \

    actor_rollout_ref.model.use_remove_padding=True \

    actor_rollout_ref.actor.ppo_mini_batch_size=32 \

    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \

    actor_rollout_ref.actor.use_kl_loss=False \

    actor_rollout_ref.actor.kl_loss_coef=0.001 \

    actor_rollout_ref.actor.kl_loss_type=mse \

    actor_rollout_ref.model.enable_gradient_checkpointing=True \

    actor_rollout_ref.actor.fsdp_config.param_offload=False \

    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \

    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=2 \

    actor_rollout_ref.rollout.tensor_model_parallel_size=2 \

    actor_rollout_ref.rollout.name=vllm \

    actor_rollout_ref.rollout.gpu_memory_utilization=0.2 \

    actor_rollout_ref.rollout.enable_chunked_prefill=False \

    actor_rollout_ref.rollout.enforce_eager=False \

    actor_rollout_ref.rollout.free_cache_engine=False \

    actor_rollout_ref.rollout.n=1 \

    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=2 \

    actor_rollout_ref.ref.fsdp_config.param_offload=True \

    actor_rollout_ref.rollout.top_p=0.95 \

    actor_rollout_ref.rollout.temperature=0.7 \

    critic.optim.lr=1e-5 \

    critic.model.use_remove_padding=True \

    critic.model.path=Qwen/Qwen2.5-VL-3B-Instruct \

    critic.model.enable_gradient_checkpointing=True \

    critic.ppo_micro_batch_size_per_gpu=1 \

    critic.model.fsdp_config.param_offload=False \

    critic.model.fsdp_config.optimizer_offload=False \

    algorithm.kl_ctrl.kl_coef=0.001 \

    trainer.critic_warmup=0 \

    trainer.logger=['console'] \

    trainer.project_name='vlnce' \

    trainer.experiment_name='vlnce_service' \

    trainer.n_gpus_per_node=4 \

    trainer.nnodes=1 \

    trainer.save_freq=70 \

    trainer.test_freq=9999999 \

    trainer.total_training_steps=300 \

    rollout_manager.max_turns=9999999 \

    rollout_manager.window_size=9999999 \

    rollout_manager.use_multi_turn_reward=True \

    rollout_manager.use_loss_mask=True \

    rollout_manager.use_gae_mask=True \

    trainer.val_before_train=False \

    trainer.val_generations_to_log_to_wandb=8 \

    rollout_manager.n_trajectory=8 \

    rollout_manager.use_service=True \

    rollout_manager.timeout=9999999 \

    rollout_manager.base_url="http://localhost:5001" \

    2>&1 | tee vlnce.log

 No newline at end of file