[Bug] After storing the MoE's checkpoint, the OPD will become slower in training.

Bug Description

When I was using OPD to train the MoE, the model training speed would become extremely slow after storing the checkpoint. However, the Dense model does not behave like this.

MoE model（save at 200step)：

Dense model（save each 20 step)：

I don't know how to identify the problem. Could someone offer some suggestions?

Steps to Reproduce

Expected Behavior

Actual Behavior

Environment

• slime version: latest
• Python version: 3.11
• PyTorch version: -
• CUDA/ROCm version: 12.8
• GPU type and count: 32*H100
• OS:
• SGLang version (if relevant):
• Megatron-LM version (if relevant):

Logs

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and my issue is not addressed there.
• I have searched for existing issues and this is not a duplicate.
• I have provided a minimal, reproducible example.

No source file, test, reproduction steps, logs, or PyTorch version are provided. Start by reproducing the slowdown with OPD and an MoE checkpoint saved at 200 steps, comparing it with the Dense case on the stated environment; done means the checkpoint-related performance regression is isolated and addressed with evidence that training speed remains stable.