Description
The --reasoning-format deepseek flag doesn't extract reasoning_content when using external chat templates via --chat-template-file. The feature/turboquant-kv-cache branch (AtomicBot-ai fork) has this issue — the reasoning format parsing only works with built-in templates.
Repro
llama-server -m model.gguf \
--reasoning on --reasoning-format deepseek \
--chat-template-file /path/to/qwen3.jinja
The template generates thinking\n markers. With --reasoning-format deepseek, the response should include a separate reasoning_content field, but the markers stay as raw text in content.
Expected
reasoning_content field with extracted thinking text. content field with just the answer.
Actual
reasoning_content absent. thinking\n[text] in content.
Cause
The turboquant fork is based on an older llama.cpp version where --reasoning-format support was incomplete for external chat templates. Upstream mainline handles this correctly. The fix would be syncing the server-side reasoning parsing from upstream.
Description
The
--reasoning-format deepseekflag doesn't extract reasoning_content when using external chat templates via--chat-template-file. Thefeature/turboquant-kv-cachebranch (AtomicBot-ai fork) has this issue — the reasoning format parsing only works with built-in templates.Repro
The template generates
thinking\nmarkers. With--reasoning-format deepseek, the response should include a separatereasoning_contentfield, but the markers stay as raw text incontent.Expected
reasoning_contentfield with extracted thinking text.contentfield with just the answer.Actual
reasoning_contentabsent.thinking\n[text]incontent.Cause
The turboquant fork is based on an older llama.cpp version where
--reasoning-formatsupport was incomplete for external chat templates. Upstream mainline handles this correctly. The fix would be syncing the server-side reasoning parsing from upstream.