Gemma 4 31B-it trips vLLM bug

The interesting part here is that the model is probably fine; the serving stack is what’s misbehaving. vLLM’s Gemma 4 reasoning parser appears to be gating structured decoding in a way that can leave the request in a broken reasoning state.

• –The reported symptom matches a known class of vLLM parser bugs: reasoning/parsing state not cleanly transitioning when thinking is disabled
• –If structured output or tool use is involved, this can silently degrade correctness instead of failing loudly
• –The issue is operationally relevant for anyone self-hosting Gemma 4 behind OpenAI-compatible APIs
• –The repeated `nvarchar(max)` text strongly suggests template contamination or token leakage from the serving layer, not a normal model completion
• –For now, the safest workaround is likely to remove the reasoning parser or adjust the chat template/thinking settings until the vLLM side is fixed