server : hint preserve_thinking when supported by chat template

[ggerganov]

Overview

ref #24093 (comment)

Print a hint to enable preserve_thinking kwarg when the chat template supports it.

# llama serve -hf unsloth/Qwen3.6-27B-MTP-GGUF:Q4_K_M
0.00.571.359 I cmn  common_param: common_params_print_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
0.00.573.173 I srv    load_model: loading model 'unsloth/Qwen3.6-27B-MTP-GGUF:Q4_K_M'
0.08.345.590 I srv    load_model: initializing, n_slots = 4, n_ctx_slot = 131072, kv_unified = 'true'
0.08.378.594 W srv          init: chat template supports 'preserve_thinking' - consider using --chat-template-kwargs "{\"preserve_thinking\": true}" (ref: https://docs.z.ai/guides/capabilities/thinking-mode#preserved-thinking)
0.08.378.603 I srv  llama_server: model loaded
0.08.378.607 I srv  llama_server: listening on http://0.0.0.0:8013

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. pi:llama.cpp/Qwen3.6-27B

[ngxson]

IMO it's could better be one of the jinja caps (to make it more generic), although I'm not sure how other templates support this function (i.e. do they use the same preserve_thinking, or another mechanism?), cc @pwilkin @aldehir if you have any insights on this

[aldehir]

A cap seems more applicable for consistency but I think either way works for something this simple.

As far as other templates go, I believe only Qwen 3.6 supports this for now. There's no competing variables to enable this.

[ngxson]

@aldehir I'm thinking about a more broader case where we can somehow make this compatible with other templates. It seems like most templates only preserve reasoning_content for last assistant message, not the whole history. We may need a hack to make it work (I'm investigating that)

I think such feature would still be quite useful. From time to time I've seen issues asking for such feature

Update: seems like GLM-4.7 has clear_thinking that is the opposite of preserve_thinking

[aldehir]

I see. We can generalize this feature with capabilities, i.e. storing the field name and normalizing the value (clear_thinking requires false while preserve_thinking needs true).

That said, not sure how easy it would be to generalize implementation for templates that don't natively support this. I also don't know the impact since they are likely not trained to retain thinking.

[ngxson]

I scanned through templates inside models/templates and realize that most models check for last_user_index or last_user_idx to remove the reasoning content, so I'm quite confident that the solution could be quite simple: always force last_user_index=0 and ignore value from set statement. (So, we need to add a notion of "read-only" key)

That being said, it's still technically a hack, but just not a (too) messy one. Will try to push a PoC tomorrow to see how it goes.

[ngxson]

Hmm ok so hacking it is more complicated than I though, so I ended up abandon it. My idea was to record any if_statement and try to flip them to see which one control the reasoning output, then force them to true later.

In anyway, I added #25105 that simply translate a generic --reasoning-preserve flag into model-specific flag, I've found 3 of them:

• preserve_thinking
• clear_thinking (GLM-4.7)
• truncate_history_thinking (NVIDIA-Nemotron-3-Nano-30B-A3B-BF16)