I’m facing a very similar (if not the same) issue;
In my case, I’m trying to trigger a Ray Workflow from Ray Serve. The Workflow itself needs to run on a large node, since it needs to load a heavy model, and run inference with it.
The first couple of Workflow runs work, but at the 3rd/4th try, I start getting OOM, and when looking at the dashboard I see several Ray::IDLE processes, using around 15GB of RAM.