Checklist

• I searched related issues but found no solution.
• The bug persists in the latest version.
• Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
• If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
• Please use English. Otherwise, it will be closed.

Describe the bug

For Qwen3-next model, the gsm8k accuracy is much higher than 0.93 threshold. But in disaggregation test, sometimes it's lower than the threshold. There may be some data race issues when PD disaggregation enabled. We need to find the root cause and fix it.

Reproduction

https://github.com/sgl-project/sglang/actions/runs/23760688526/job/69259644818#step:9:2711

Environment

/