Merge branch 'main' of github.com:RAGEN-AI/vagen into prompt_dev (bcee16d0) · Commits · 张泽凯 / VAGEN

README.md

+2 −1

Original line number	Diff line number	Diff line
		@@ -159,7 +159,8 @@ If you find our repo useful, we appreciate it if you could cite our work at:
		```bibtex
		@misc{VAGEN,
		title={VAGEN: Training VLM Agents with Multi-Turn Reinforcement Learning},
		author={Kangrui Wang* and Pingyue Zhang* and Zihan Wang* and Qineng Wang* and Linjie Li* and Zhengyuan Yang and Chi Wan and Yiping Lu and Manling Li},
		author={Kangrui Wang* and Pingyue Zhang* and Zihan Wang* and Qineng Wang* and Yaning Gao* and Linjie Li* and Zhengyuan Yang and Chi Wan and Hanyang Chen and Yiping Lu and Manling Li},
		url={https://github.com/RAGEN-AI/VAGEN},
		year={2025},
		}
		```

scripts/examples/crossview_grpo/env_config.yaml

0 → 100644

+8 −0

Original line number	Diff line number	Diff line
		env1:
		env_name: crossview
		env_config:
		render_mode: vision
		train_size: 10000
		test_size: 512

scripts/examples/crossview_grpo/run.sh

0 → 100755

+74 −0

Original line number	Diff line number	Diff line
		set -x


		export VLLM_ATTENTION_BACKEND=XFORMERS
		export PYTHONHASHSEED=0

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

		python -m vagen.env.create_dataset \
		--yaml_path "$SCRIPT_DIR/env_config.yaml" \
		--train_path "data/crossview/train.parquet" \
		--test_path "data/crossview/test.parquet" \

		# max_trajectory_length = max_prompt_length + max_response_length

		python3 -m vagen.trainer.main_ppo \
		algorithm.adv_estimator=grpo \
		algorithm.high_level_gamma=0.95 \
		data.train_files=data/crossview/train.parquet \
		data.val_files=data/crossview/test.parquet \
		data.train_batch_size=64 \
		data.max_prompt_length=1024 \
		data.max_response_length=648 \
		data.max_trajectory_length=3600 \
		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=8 \
		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=1 \
		actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
		actor_rollout_ref.rollout.name=vllm \
		actor_rollout_ref.rollout.gpu_memory_utilization=0.3 \
		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=1 \
		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','wandb'] \
		trainer.project_name='vagen_crossview' \
		trainer.experiment_name='grpo_crossview' \
		trainer.n_gpus_per_node=4 \
		trainer.nnodes=1 \
		trainer.save_freq=100 \
		trainer.test_freq=20 \
		trainer.total_training_steps=300 \
		rollout_manager.max_turns=1 \
		rollout_manager.window_size=5 \
		rollout_manager.use_multi_turn_reward=False \
		rollout_manager.use_loss_mask=True \
		trainer.val_before_train=True \
		trainer.val_generations_to_log_to_wandb=8 \
		rollout_manager.n_trajectory=8 \
		2>&1 \| tee grpo_crossview.log

scripts/examples/navigation_aico/run.sh

+1 −1

Original line number	Diff line number	Diff line
		@@ -78,5 +78,5 @@ python3 -m vagen.trainer.main_ppo \
		rollout_manager.n_trajectory=1 \
		rollout_manager.use_service=True \
		rollout_manager.timeout=240 \
		rollout_manager.base_url="http://localhost:5001" \
		rollout_manager.base_url="http://localhost:5000" \
		2>&1 \| tee aico_navigation_vision.log
		No newline at end of file

scripts/exps/format/maniskill/free_think/env_config.yaml

0 → 100644

+36 −0

Original line number	Diff line number	Diff line
		env1:
		env_name: primitive_skill
		env_config:
		render_mode: vision
		prompt_format: free_think
		use_accuracy_reward: false
		env_id: "AlignTwoCube" # AlignTwoCube,PlaceTwoCube,PutAppleInDrawer,StackThreeCube
		train_size: 10000
		test_size: 8
		env2:
		env_name: primitive_skill
		env_config:
		render_mode: vision
		prompt_format: free_think
		use_accuracy_reward: false
		env_id: "PlaceTwoCube" # AlignTwoCube,PlaceTwoCube,PutAppleInDrawer,StackThreeCube
		train_size: 10000
		test_size: 8
		env3:
		env_name: primitive_skill
		env_config:
		render_mode: vision
		prompt_format: free_think
		use_accuracy_reward: false
		env_id: "PutAppleInDrawer" # AlignTwoCube,PlaceTwoCube,PutAppleInDrawer,StackThreeCube
		train_size: 10000
		test_size: 8
		env4:
		env_name: primitive_skill
		env_config:
		render_mode: vision
		prompt_format: free_think
		use_accuracy_reward: false
		env_id: "StackThreeCube" # AlignTwoCube,PlaceTwoCube,PutAppleInDrawer,StackThreeCube
		train_size: 10000
		test_size: 8

Admin message