Ray OOM when saving checkpoint unless both --no-save-optim and --disable-weights-backuper are enabled

Your Question

Hi, thanks for the great work.

I encountered a Ray OOM issue when saving checkpoints during training. The training can run normally if I add both of the following arguments:

--no-save-optim --disable-weights-backuper

However, both flags are required. If either one is missing, Ray OOM still happens during checkpoint saving.

The problem is that with --no-save-optim, the optimizer state is not saved, so I cannot properly resume training from checkpoints. I would like to know whether this is a known issue and what the recommended solution is.

Error log
(raylet, ip=10.22.72.139) [2026-04-17 15:06:03,782 E 622 622] (raylet) node_manager.cc:3250:
1 Workers (tasks / actors) killed due to memory pressure (OOM), 0 Workers crashed due to other reasons at node

Refer to the documentation on how to address the out of memory issue:
https://docs.ray.io/en/latest/ray-core/scheduling/ray-oom-prevention.html

Have you encountered this issue before?

What is the recommended way to avoid Ray OOM during checkpoint saving while still keeping the checkpoint resumable, including optimizer states?

Is there a suggested configuration for saving optimizer states or disabling the weights backuper without losing resume capability?

What I've Tried

--no-save-optim --disable-weights-backuper

Environment (if relevant)

• slime version: 0.2.4
• Python version: python3.12

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and FAQ and my question is not answered there.
• I have searched for existing issues and my question has not been asked before.

The issue names no source file or test; start by reproducing checkpoint saving in slime 0.2.4 with optimizer-state saving enabled and compare the four combinations of --no-save-optim and --disable-weights-backuper. Read the linked Ray OOM-prevention guidance, and consider the issue resolved only when a verified configuration preserves resumable checkpoints with optimizer states.