Commit 97f9140d authored by williamzhangNU's avatar williamzhangNU
Browse files

debug

parent 3ec8a5cf

vagen/env/sokoban/env.py

+6 −6
Original line number Diff line number Diff line
@@ -209,6 +209,7 @@ class SokobanInterface(BaseInterface): 
    FORMAT_REWARD = 0.5

    FORMAT_PENALTY = -0.5

    VALID_ACTION_REWARD = 0.5

    MAX_ACTION_PER_STEP = 1 # NOTE hard coded here

    ACTION_LOOKUP = {

        0: "None",

        1: "Up",
@@ -299,7 +300,7 @@ class SokobanInterface(BaseInterface): 
                preprocess_result.action_list.append(action)

            else:

                break

        

        preprocess_result.action_list = preprocess_result.action_list[:cls.MAX_ACTION_PER_STEP]

        return preprocess_result

        

    @classmethod
@@ -381,13 +382,12 @@ class SokobanInterface(BaseInterface): 
        answer = preprocess_result.answer

        final_info['llm_raw_response'] = preprocess_result.llm_raw_response



        # deal with format

        if think and answer: # format is correct

        if think and answer: # format reward

            reward += self.FORMAT_REWARD

            if action_list:

                reward += self.VALID_ACTION_REWARD

        else:

        else: # format penalty

            reward += self.FORMAT_PENALTY

        if action_list: # valid action reward

            reward += self.VALID_ACTION_REWARD



        info = {}

        for action in action_list:

vagen/examples/sokoban/debug_qwen0_5_1_gpu_ppo_multi_turn_ppo.sh→vagen/examples/sokoban/debug_qwen0_5_2_gpu_ppo_multi_turn_ppo.sh

+13 −13
Original line number Diff line number Diff line
@@ -21,52 +21,52 @@ python3 -m vagen.trainer.main_ppo \ 
    algorithm.high_level_gamma=0.95 \

    data.train_files=data/sokoban-text-3-step/train.parquet \

    data.val_files=data/sokoban-text-3-step/test.parquet \

    data.train_batch_size=64 \

    data.train_batch_size=8 \

    data.max_prompt_length=1408 \

    data.max_response_length=128 \

    data.max_trajectory_length=1536 \

    data.image_key=images \

    actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B-Instruct \

    actor_rollout_ref.model.path=Qwen/Qwen2.5-1.5B-Instruct \

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

    actor_rollout_ref.model.use_remove_padding=False \

    actor_rollout_ref.actor.ppo_mini_batch_size=32 \

    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2 \

    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=low_var_kl \

    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=1 \

    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=0.4 \

    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.log_prob_micro_batch_size_per_gpu=1 \

    actor_rollout_ref.ref.fsdp_config.param_offload=True \

    critic.optim.lr=1e-5 \

    critic.model.use_remove_padding=False \

    critic.model.path=Qwen/Qwen2.5-0.5B-Instruct \

    critic.model.path=Qwen/Qwen2.5-1.5B-Instruct \

    critic.model.enable_gradient_checkpointing=True \

    critic.ppo_micro_batch_size_per_gpu=2 \

    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' \

    trainer.experiment_name='debug_qwen0_5_1_gpu_ppo_multi_turn_ppo' \

    trainer.n_gpus_per_node=1 \

    trainer.experiment_name='debug_qwen1_5_1_gpu_ppo_multi_turn_ppo' \

    trainer.n_gpus_per_node=2 \

    trainer.nnodes=1 \

    trainer.save_freq=100 \

    trainer.test_freq=5 \

    trainer.test_freq=10 \

    trainer.total_epochs=15 \

    rollout_manager.max_turns=3 \

    rollout_manager.window_size=5 \

    trainer.val_before_train=True \

    trainer.val_generations_to_log_to_wandb=4 \

    rollout_manager.n_trajectory=2 \

    2>&1 | tee debug_qwen0_5_1_gpu_ppo_multi_turn_ppo.log
 
    rollout_manager.n_trajectory=16 \

    2>&1 | tee debug_qwen1_5_1_gpu_ppo_multi_turn_ppo.log

vagen/examples/sokoban/debug_qwen0_5_2_gpu_ppo_single_turn_ppo.sh

0 → 100644
+72 −0
Original line number Diff line number Diff line 
set -x



export VLLM_ATTENTION_BACKEND=XFORMERS

export PYTHONHASHSEED=0



python -m vagen.env.sokoban.create_dataset \

    --data_dir data/sokoban-text-3-step \

    --max_action_length 3 \

    --dim_room 6 6 \

    --num_boxes 1 \

    --max_steps 100 \

    --search_depth 30 \

    --start_seed 0 \

    --train_ratio 0.8 \

    --n_candidate 20000



# max_trajectory_length = max_prompt_length + max_response_length



python3 -m vagen.trainer.main_ppo \

    algorithm.adv_estimator=gae \

    algorithm.high_level_gamma=0.95 \

    data.train_files=data/sokoban-text-3-step/train.parquet \

    data.val_files=data/sokoban-text-3-step/test.parquet \

    data.train_batch_size=64 \

    data.max_prompt_length=768 \

    data.max_response_length=128 \

    data.max_trajectory_length=1024 \

    data.image_key=images \

    actor_rollout_ref.model.path=Qwen/Qwen2.5-1.5B-Instruct \

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

    actor_rollout_ref.model.use_remove_padding=False \

    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=low_var_kl \

    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=0.4 \

    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 \

    critic.optim.lr=1e-5 \

    critic.model.use_remove_padding=False \

    critic.model.path=Qwen/Qwen2.5-1.5B-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' \

    trainer.experiment_name='debug_qwen1_5_1_gpu_ppo_single_turn_ppo' \

    trainer.n_gpus_per_node=2 \

    trainer.nnodes=1 \

    trainer.save_freq=100 \

    trainer.test_freq=10 \

    trainer.total_epochs=15 \

    rollout_manager.max_turns=1 \

    rollout_manager.window_size=5 \

    trainer.val_before_train=True \

    trainer.val_generations_to_log_to_wandb=4 \

    rollout_manager.n_trajectory=2 \

    2>&1 | tee debug_qwen1_5_1_gpu_ppo_single_turn_ppo.log

vagen/examples/sokoban/debug_qwen_vl_2gpu_ppo_multi_turn_ppo.sh

0 → 100644
+72 −0
Original line number Diff line number Diff line 
set -x



export VLLM_ATTENTION_BACKEND=XFORMERS

export PYTHONHASHSEED=0



python -m vagen.env.sokoban.create_dataset \

    --data_dir data/sokoban-vision-2-step \

    --max_action_length 2 \

    --dim_room 6 6 \

    --num_boxes 1 \

    --max_steps 100 \

    --search_depth 30 \

    --start_seed 0 \

    --train_ratio 0.8 \

    --n_candidate 20000



# max_trajectory_length = max_prompt_length + max_response_length



python3 -m vagen.trainer.main_ppo \

    algorithm.adv_estimator=multi_turn_gae \

    algorithm.high_level_gamma=0.95 \

    data.train_files=data/sokoban-vision-2-step/train.parquet \

    data.val_files=data/sokoban-vision-2-step/test.parquet \

    data.train_batch_size=64 \

    data.max_prompt_length=1920 \

    data.max_response_length=128 \

    data.max_trajectory_length=2048 \

    data.image_key=images \

    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=low_var_kl \

    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=0.4 \

    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 \

    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' \

    trainer.experiment_name='debug_qwen_vl_2gpu_ppo_multi_turn_ppo' \

    trainer.n_gpus_per_node=2 \

    trainer.nnodes=1 \

    trainer.save_freq=100 \

    trainer.test_freq=5 \

    trainer.total_epochs=15 \

    rollout_manager.max_turns=2 \

    rollout_manager.window_size=5 \

    trainer.val_before_train=True \

    trainer.val_generations_to_log_to_wandb=4 \

    rollout_manager.n_trajectory=2 \

    2>&1 | tee debug_qwen_vl_2gpu_ppo_multi_turn_ppo.log

vagen/examples/sokoban/debug_qwen_vl_2gpu_ppo_single_turn_ppo.sh

0 → 100644
+72 −0
Original line number Diff line number Diff line 
set -x



export VLLM_ATTENTION_BACKEND=XFORMERS

export PYTHONHASHSEED=0



python -m vagen.env.sokoban.create_dataset \

    --data_dir data/sokoban-text-3-step \

    --max_action_length 3 \

    --dim_room 6 6 \

    --num_boxes 1 \

    --max_steps 100 \

    --search_depth 30 \

    --start_seed 0 \

    --train_ratio 0.8 \

    --n_candidate 20000



# max_trajectory_length = max_prompt_length + max_response_length



python3 -m vagen.trainer.main_ppo \

    algorithm.adv_estimator=gae \

    algorithm.high_level_gamma=0.95 \

    data.train_files=data/sokoban-text-3-step/train.parquet \

    data.val_files=data/sokoban-text-3-step/test.parquet \

    data.train_batch_size=64 \

    data.max_prompt_length=1920 \

    data.max_response_length=128 \

    data.max_trajectory_length=2048 \

    data.image_key=images \

    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=low_var_kl \

    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=0.4 \

    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 \

    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' \

    trainer.experiment_name='debug_qwen_vl_2gpu_ppo_single_turn_ppo' \

    trainer.n_gpus_per_node=2 \

    trainer.nnodes=1 \

    trainer.save_freq=100 \

    trainer.test_freq=10 \

    trainer.total_epochs=15 \

    rollout_manager.max_turns=1 \

    rollout_manager.window_size=5 \

    trainer.val_before_train=True \

    trainer.val_generations_to_log_to_wandb=4 \

    rollout_manager.n_trajectory=2 \

    2>&1 | tee debug_qwen_vl_2gpu_ppo_single_turn_ppo.log

CRA Git | Maintained and supported by SUSTech CRA and CCSE