minor

[submodule "verl"]

	path = verl

	url = https://github.com/volcengine/verl.git

	url = https://github.com/JamesKrW/verl.git

echo "Starting installation process..."

# Initialize and update git submodules

echo "Initializing git submodules..."

git submodule update --init

# First install verl subdirectory as editable package

echo "Installing verl package..."

cd verl

pip install -e .

cd ../

# Now install the main requirements

echo "Installing vagen dependencies..."

pip install 'transformers>=4.49.0'

pip install 'vllm>=0.7.3'

pip install 'qwen-vl-utils'

pip install 'mathruler'

pip install 'gym'

pip install 'gym-sokoban'

pip install 'matplotlib'

# Install flash-attn with special flags

echo "Installing flash-attn with no build isolation..."

pip install flash-attn --no-build-isolation

# Install the main package

echo "Installing vagen package..."

pip install -e .

    def create_dataset(self, start_seed, train_size, test_size):

    def create_dataset(self, start_seed, train_size, test_size, force_gen=False):

        train_file_path = os.path.join(self.data_dir, 'train.parquet')

        test_file_path = os.path.join(self.data_dir, 'test.parquet')

        # Check if files already exist and force_gen is False

        if not force_gen and os.path.exists(train_file_path) and os.path.exists(test_file_path):

            print(f"Dataset files already exist at {self.data_dir}. Skipping generation.")

            print(f"Use --force-gen to override and regenerate the dataset.")

            return

        # Ensure data directory exists

        os.makedirs(self.data_dir, exist_ok=True)

        seeds = range(start_seed, start_seed + train_size + test_size)

        instructions = []

        for seed in seeds:

        train_dataset = train_dataset.map(function=make_map_fn('train'), with_indices=True)

        test_dataset = test_dataset.map(function=make_map_fn('test'), with_indices=True)

        train_dataset.to_parquet(os.path.join(self.data_dir, 'train.parquet'))

        test_dataset.to_parquet(os.path.join(self.data_dir, 'test.parquet'))

        train_dataset.to_parquet(train_file_path)

        test_dataset.to_parquet(test_file_path)

        print(f"Dataset successfully generated at {self.data_dir}")

    def merge_parquet_files(

    parser.add_argument('--start_seed', type=int, default=10000)

    parser.add_argument('--train_size', type=int, default=10000)

    parser.add_argument('--test_size', type=int, default=1000)

    parser.add_argument('--force-gen', action='store_true', 

                        help='Force dataset generation even if files already exist')

    args = parser.parse_args()

    creator = DatasetCreator(config_path=args.config_path)

    creator.create_dataset(start_seed=args.start_seed, train_size=args.train_size, test_size=args.test_size)

    creator.create_dataset(

        start_seed=args.start_seed, 

        train_size=args.train_size, 

        test_size=args.test_size,

        force_gen=args.force_gen

    )

 No newline at end of file

if __name__ == "__main__":

    parser = argparse.ArgumentParser()

    parser.add_argument('--config_path', type=str, default='vagen/env/sokoban.yaml')

    parser.add_argument('--config_path', type=str, default='vagen/env/config/sokoban.yaml')

    parser.add_argument('--start_seed', type=int, default=0)

    parser.add_argument('--train_size', type=int, default=100)

    parser.add_argument('--test_size', type=int, default=100)

    parser.add_argument('--force-gen', action='store_true', 

                        help='Force dataset generation even if files already exist')

    args = parser.parse_args()

    creator = SokobanDatasetCreator(config_path=args.config_path)

    creator.create_filtered_dataset(start_seed=args.start_seed, train_size=args.train_size, test_size=args.test_size)

    #creator.create_filtered_dataset(start_seed=args.start_seed, train_size=args.train_size, test_size=args.test_size)

    creator.create_dataset(start_seed=args.start_seed, train_size=args.train_size, test_size=args.test_size)

export VLLM_ATTENTION_BACKEND=XFORMERS

python -m vagen.env.sokoban.create_dataset 

python3 -m vagen.trainer.main_ppo \

    algorithm.adv_estimator=grpo \

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