ok so, I would recommend trimming down the configs even further to understand the cause of the leak. There could be two possibilities here:
- There is a fundamental bug within rllib that was not discovered so far
- It could be observation space related which either could be wrong usage of something or again a bug for certain class of spaces.
If it’s a bug. you can share the trimmed down version of repro code on a github issue and we will triage it and respond as soon as possible. There is a chance that in the process you may find the solution and in that case your contribution is ofc. welcomed.
We recently also shared a high level debugging guide. I hope that helps in narrowing down the issue.