[Bug] slime-v0.3.0 版本在跑 qwen3.6 35B A3B 模型的时候，在第二次 rollout 会有乱码。怀疑 镜像&sglang 版本导致

[gtarcoder]

Bug Description

slime-v0.3.0 版本在跑 qwen3.6 35B A3B 模型的时候，在第二次 rollout 会有乱码

1. 使用slime-v0.3.0 配套的 镜像 slimerl/sglang:v0.5.12.post1-cu129 会出现乱码
2. 使用 slime-v0.2.4 配套的 镜像 slimerl/sglang:v0.5.9 则结果正常

Steps to Reproduce

分别使用：

1. slime-v0.3.0 配套的 镜像 slimerl/sglang:v0.5.12.post1-cu129 去跑 Qwen-3.6 MOE 35B A3B RL 任务。
2. slime-v0.2.4 配套的 镜像 slimerl/sglang:v0.5.9 去跑 去跑 Qwen-3.6 MOE 35B A3B RL 任务。

Expected Behavior

两个任务在第二次 rollout的时候都没有乱码

Actual Behavior

能看到 前者会在 第二次 rollout的时候有乱码，而后者正常

Environment

• slime version: v0.3.0
• Python version: 3.12.3
• PyTorch version: 2.9.1+cu129
• CUDA/ROCm version: 13.0
• GPU type and count: H800 x 8
• OS: ubuntu-24.04
• SGLang version (if relevant): 镜像自带
• 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.

[zhuzilin]

貌似 0.5.12 的 sglang 在某些配置下确实会有一些问题。可以和这个 e2e 测试对比一下配置，这个配置是通过了 ci 的：https://github.com/THUDM/slime/blob/main/tests/test_qwen3.6_35B_A3B_pd_mooncake.py

[EazyReal]

Ran into the same thing with Qwen3.6 35B-A3B (MoE + GDN) on slime v0.3.0. First rollout is fine, then it produces garbage on the first generation after the weight sync. Same split for us too: sglang 0.5.12.post1 is broken, <=0.5.10 is fine.

It's the CUDA graphs getting replayed against stale memory after the release/resume memory-occupation cycle that the online weight update does. The post-update replay ends up reading buffers that were freed or moved, so MoE+GDN serving comes out garbled.

The fix is sgl-project/sglang#27140 ("recapture CUDA graphs for RL online weight updates"), still open. So either cherry-pick that, or pin sglang to <=0.5.10 until it lands.

One thing worth flagging: turning off piecewise cuda graph on its own did not fix it for us, you need the actual recapture. We backported sgl-project/sglang#27140 onto 0.5.12.post1 and post-sync generation was coherent again with graphs left on (about 19s per sync to recapture), tested on 35B-A3B.

[Gao016]

是的，我在两周前跑也遇到了同样的情况，因此给sglang提了PR：sgl-project/sglang#27140
修改之后，qwen35b-RL训练正常收敛