CUDA_ERROR_NOT_INITIALIZED when trying to use a GPU accelerated environment

[Sephirex-X]

Hi, very impressive work, however when I try to use a GPU accelerated env I got error like CUDA_ERROR_NOT_INITIALIZED.

The env I used is accelerated by tensorflow and it works pretty well seperately. After a few troubleshoots I noticed that ur work somehow use multiprocess to accelerate data collection on env, therefore I believe the error roots from the confilct between multiprocess and CUDA initialization, I'm wonder if your guys could provide some info to help me out.

thx a lot
here is my code

import eveything
@ENV_REGISTRY.register('myenv')
class Myenv(BaseEnv):
    def __init__(self, cfg: dict) -> None:
       #doing something related to gpu mainly initialization
    def reset(self):
        #doing reset
        lightzero_obs_dict = {'observation': obs, 'action_mask': self.action_mask, 'to_play': -1}
        return lightzero_obs_dict
    def step(self, action):
        #doing env step
        return BaseEnvTimestep(lightzero_obs_dict, rew, done, info)
        
    def seed(self, seed: int, dynamic_seed: bool = True) -> None:
        self._seed = seed
        self._dynamic_seed = dynamic_seed
        np.random.seed(self._seed)
    
    def close(self) -> None:
        if self._init_flag:
            pass
        self._init_flag = False

    def __repr__(self) -> str:
        return "My Env"

    @property
    def observation_space(self) -> gym.spaces.Space:
        return self._observation_space

    @property
    def action_space(self) -> gym.spaces.Space:
        return self._action_space
    
    @property
    def reward_space(self) -> gym.spaces.Space:
        return self._reward_space