server : guard chat-template thinking probe against apply-time jinja errors

[palios-taey]

Hit a SIGABRT in llama-cli --jinja loading a model whose chat template parses at init but fails parser generation at apply time (one using {% call %}). The throw — std::invalid_argument from common_chat_templates_support_enable_thinking() — escapes because that call sits just outside the try/catch guarding common_chat_templates_init(), so it is uncaught and the process aborts instead of failing to load.

Moved the thinking probe inside that try/catch. Apply-time errors now fail load the same way init parse errors do — chat template parsing error, clean return. llama-server already degraded gracefully here; this fixes the llama-cli path and keeps both calls behind one guard.

Verified before/after on a model with such a template: SIGABRT → clean rc 1. Normal model still loads and runs under --jinja (CLI + server).

AI usage disclosure: AI-assisted diagnosing the abort and drafting the fix. Reviewed it, own it, verified the before/after myself.

[ggml-gh-bot]

Hi @palios-taey, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[pwilkin]

Yeah, change looks fine, thanks.

[palios-taey]

Rebased on master and resolved the conflict in tools/server/server-context.cpp — the change itself is unchanged (the thinking probe + log stay inside the init try/catch).

[pwilkin]

@ggml-org/maintainers could use an approval please.

[ggerganov]

@pwilkin Btw, slightly related - can we detect and print a hint for preserve_thinking when it is supported by the template? I think that this chat option (i.e. --chat-template-kwargs "{\"preserve_thinking\": true}") is a must for Qwen3.6 models, yet likely many users don't use it because it's quite obscure. Printing a hint to enable it on startup would be helpful.

[palios-taey]

Rebased on master to clear the conflict from the server-context split. The fix itself is unchanged — enable_thinking is declared before the try and the template_supports_thinking probe moved inside it, same as before, just re-applied on top of the refactored code (kept the SRV_TRC line). Still +7/-7 in the one file. llama-server builds clean. Ready for a re-approval whenever you have a moment, @pwilkin.

[palios-taey]

Friendly bump — this is green and mergeable, and I think it's just stuck on a technicality: both approvals (@pwilkin on 5f0afe7, @forforever73 on d36fc59) predate the 6/28 rebase, so GitHub counts them as stale against the current head dc30a3a and still shows "review required". The rebase was only to clear the conflict from the server-context split — the change itself is unchanged, +7/−7 in one file. A re-approval of the current head would let it land. Thanks for the reviews.

[Moore2877]

I've been working on a custom chat template for Qwen and this morning has been a rollercoaster with this and other pushes while testing after updating llama.cpp. At first I thought I broke something but now things are working better than before yesterday. https://huggingface.co/Moore2877/Qwen-Fixed-Chat-Templates-llamacpp