[Bug] Qwen3-next mtp topk>1 accuracy issue

[hanming-lu]

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

GDN model spec dec topk > 1 path has bug and can generate garbage output. The easiest repro test is this: https://github.com/sgl-project/sglang/blob/main/test/registered/4-gpu-models/test_qwen3_next_models_mtp.py#L30

The threshold for mtp should be the same as non-mtp, at "kl_div": 0.0025, instead of currently 0.008.

Reproduction

1. update kl threshold to 0.0025, same as non-mtp threshold https://github.com/sgl-project/sglang/blob/main/test/registered/4-gpu-models/test_qwen3_next_models_mtp.py#L30
2. run TestQwen3NextMTP

Environment

main default environment

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen it if needed.