Why is memory being allocated on GPU?

reinforcement-learning
1

pytorch 1.0

Screenshot%20from%202019-01-23%2021-53-51
Screenshot%20from%202019-01-23%2021-53-51727×733 124 KB

relevant source code: https://github.com/heronsystems/adeptRL/blob/master/adept/environments/managers/subproc_env_manager.py#L148-L193

even if I set os.environ[“CUDA_VISIBLE_DEVICES”] = “”, torch.cuda.is_available() returns False, but memory is still being allocated on GPU? maybe something to do with shared memory?

4

hmm, maybe sc2 is not actually headless