[Bug] Models fail to produce non-empty answer while evaluated for reasoning on long context.

### Checklist

- [x] I searched related issues but found no solution.
- [x] The bug persists in the latest version.
- [x] Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
- [x] 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.
- [x] Please use English. Otherwise, it will be closed.

### Describe the bug

While ran on SGLang, the reasoning models (e.g. GLM-5.1, Kimi-K2.5) most of the time fail to produce output both in the chain-of-thought and/or conclusive stages. That is measured with testing on the LongBench V2 questions.

Both on CUDA and ROCm platforms, the model doesn't always produce an answer, the rate of successful thought-conclusion chains on a given question is not zero though. I.e. if the same question is asked repeatedly, eventually the answer is non-empty.

If the collected non-empty responses are evaluated for correctness, the resulting level of accuracy matches SOTA.

### Reproduction

sglang serve --model-path zai-org/GLM-5.1 --tp-size 8 --host 127.0.0.1 --port {port of choice} --trust-remote-code --reasoning-parser glm45

Client:
python3 ./pred.py --model GLM-5.1 --cot --n_proc 503

### Environment

Docker:
lmsysorg/sglang:latest (as of July, 15, 2026)

Start by reproducing the issue with the provided sglang serve command and pred.py on LongBench V2, noting whether empty chain-of-thought or conclusion responses occur. Compare repeated requests and both CUDA and ROCm runs. Done means reasoning evaluations consistently produce non-empty thought and/or conclusive output without reducing the reported accuracy.

Written by the indexing model from the issue text.