Commit 5e558492 authored by jameskrw's avatar jameskrw
Browse files

reformatted debug scripts

parent 7be06cfb

vagen/examples/sokoban/debug_qwen0_5_2_gpu_ppo_single_turn_ppo.sh→vagen/examples/sokoban/debug_multi_action_1_turn_ppo_kl.sh

+17 −10
Original line number Diff line number Diff line
@@ -12,10 +12,16 @@ python -m vagen.env.sokoban.create_dataset \ 
    --search_depth 30 \

    --start_seed 0 \

    --train_ratio 0.8 \

    --max_action_per_step 1 \

    --max_action_per_step 5 \

    --max_action_penalty -0.1 \

    --format_reward 0.5 \

    --n_candidate 20000

    --n_candidate 20000 \

    --force-gen



if [ $? -ne 0 ]; then

    echo "Failed to generate dataset"

    exit 1

fi



# max_trajectory_length = max_prompt_length + max_response_length
@@ -24,7 +30,7 @@ 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=16 \

    data.max_prompt_length=768 \

    data.max_response_length=128 \

    data.max_trajectory_length=1024 \
@@ -34,9 +40,9 @@ python3 -m vagen.trainer.main_ppo \ 
    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.use_kl_loss=True \

    actor_rollout_ref.actor.kl_loss_coef=0.001 \

    actor_rollout_ref.actor.kl_loss_type=low_var_kl \

    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 \
@@ -48,6 +54,7 @@ python3 -m vagen.trainer.main_ppo \ 
    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.use_ref=False \

    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 \
@@ -61,15 +68,15 @@ python3 -m vagen.trainer.main_ppo \ 
    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.experiment_name='debug_multi_action_1_turn_ppo_kl' \

    trainer.n_gpus_per_node=2 \

    trainer.nnodes=1 \

    trainer.save_freq=100 \

    trainer.test_freq=10 \

    trainer.test_freq=5 \

    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
 
    trainer.val_generations_to_log_to_wandb=8 \

    rollout_manager.n_trajectory=8 \

    2>&1 | tee debug_multi_action_1_turn_ppo_kl.log

vagen/examples/sokoban/debug_qwen0_5_1_gpu_ppo.sh

deleted100644 → 0
+0 −74
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-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 \

    --max_action_per_step 2 \

    --max_action_penalty -0.1 \

    --format_reward 0.5 \

    --n_candidate 20000



# max_trajectory_length = max_prompt_length + max_response_length



python3 -m vagen.trainer.main_ppo \

    algorithm.adv_estimator=gae \

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

    data.val_files=data/sokoban-text-2-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/Qwen2.5-0.5B-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.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.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.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.enable_gradient_checkpointing=True \

    critic.ppo_micro_batch_size_per_gpu=2 \

    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' \

    trainer.n_gpus_per_node=1 \

    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_qwen0_5_1_gpu_ppo.log

vagen/examples/sokoban/debug_qwen0_5_4_gpu_grpo.sh

deleted100644 → 0
+0 −67
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-1-step \

    --max_action_length 1 \

    --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=grpo \

    data.train_files=data/sokoban-text-1-step/train.parquet \

    data.val_files=data/sokoban-text-1-step/test.parquet \

    data.train_batch_size=16 \

    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/Qwen2.5-0.5B-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=128 \

    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=8 \

    actor_rollout_ref.actor.use_kl_loss=True \

    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=8 \

    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=8 \

    actor_rollout_ref.ref.fsdp_config.param_offload=True \

    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_4_gpu_grpo' \

    trainer.n_gpus_per_node=4 \

    trainer.nnodes=1 \

    trainer.save_freq=50 \

    trainer.test_freq=2 \

    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=2 \

    rollout_manager.n_trajectory=8 \

    2>&1 | tee debug_qwen0_5_4_gpu_grpo.log



# NOTE change gpu_memory_utilization to a smaller value (0.4) to avoid oom error
 
 No newline at end of file

vagen/examples/sokoban/debug_qwen0_5_4_gpu_ppo.sh

deleted100644 → 0
+0 −53
Original line number Diff line number Diff line 
# set -x



# export VLLM_ATTENTION_BACKEND=XFORMERS



# python -m vagen.env.sokoban.create_dataset --data_dir data/sokoban-text



# # max_trajectory_length = max_prompt_length + max_response_length



# python3 -m vagen.trainer.main_ppo \

#     algorithm.adv_estimator=grpo \

#     data.train_files=data/sokoban-text/train.parquet \

#     data.val_files=data/sokoban-text/test.parquet \

#     data.train_batch_size=32 \

#     data.max_prompt_length=512 \

#     data.max_response_length=1536 \

#     +data.max_trajectory_length=2048 \

#     +data.max_response_per_turn=256 \

#     +actor_rollout_ref.rollout.max_response_per_turn=256 \

#     +actor_rollout_ref.rollout.max_trajectory_length=2048 \

#     data.image_key=images \

#     actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B-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=4 \

#     actor_rollout_ref.actor.use_kl_loss=True \

#     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=4 \

#     actor_rollout_ref.rollout.tensor_model_parallel_size=1 \

#     actor_rollout_ref.rollout.name=vllm \

#     actor_rollout_ref.rollout.gpu_memory_utilization=0.7 \

#     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=4 \

#     actor_rollout_ref.ref.fsdp_config.param_offload=True \

#     algorithm.kl_ctrl.kl_coef=0.001 \

#     trainer.critic_warmup=0 \

#     trainer.logger=['console','wandb'] \

#     trainer.project_name='vagen' \

#     trainer.experiment_name='qwen2_5_05b_function_rm' \

#     trainer.n_gpus_per_node=1 \

#     trainer.nnodes=1 \

#     trainer.save_freq=-1 \

#     trainer.test_freq=-1 \

#     trainer.total_epochs=15 \

#     +max_turns=5 \

#     2>&1 | tee debug_qwen0_5_4_gpu_ppo.log

vagen/examples/sokoban/debug_qwen2_5_vl_4gpu_grpo.sh

deleted100644 → 0
+0 −63
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 \

    --visual_env \

    --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



python3 -m vagen.trainer.main_ppo \

    algorithm.adv_estimator=grpo \

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

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

    data.train_batch_size=16 \

    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=4 \

    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \

    actor_rollout_ref.actor.use_kl_loss=True \

    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=4 \

    actor_rollout_ref.rollout.name=vllm \

    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \

    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 \

    algorithm.kl_ctrl.kl_coef=0.001 \

    trainer.critic_warmup=0 \

    trainer.logger=['console','wandb'] \

    trainer.project_name='vagen' \

    trainer.experiment_name='debug_qwen2_5_vl_4gpu_grpo' \

    trainer.n_gpus_per_node=4 \

    trainer.nnodes=1 \

    trainer.save_freq=50 \

    trainer.test_freq=2 \

    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=2 \

    rollout_manager.n_trajectory=8 \

    2>&1 | tee debug_qwen2_5_vl_4gpu_grpo.log
 
 No newline at end of file

CRA Git | Maintained and supported by SUSTech CRA and CCSE