Misc. bug: llama-server webui with --jinja flag does not show thinking when using reasoning models

[littlett77]

Name and Version

./build/bin/llama-cli --version version: 5586 (3ac6753) built with clang version 20.1.4 for aarch64-unknown-linux-android24

tested on Chrome browser for Android

Operating systems

Linux

Which llama.cpp modules do you know to be affected?

llama-server

Command line

./build/bin/llama-server     --model ~/_memorycard/_y/Qwen3-1.7B-Q8_K_XL.gguf     --temp 0.6     --top_p 0.95  --top_k 20    --min_p 0.01      --seed 666     -fa  --host 0.0.0.0 --no-warmup --jinja

Problem description & steps to reproduce

as by title, when using the --jinja flag with models like Qwen3 or Deepseek-R1-0528-Qwen3 during the thinking proces the animated 3 dots keep spinning and after a while the token counter gets updated but no "thinking" or "thought process" icon/tags appear and no elucubration is written as it never happened; when the reasoning ends the actual answer start showing as usual. When not using the --jinja flag the reasoning stuff happens regularly. When instead using the llama-cli command the reasoning appear inside the <think>...</think> tags as expected

First Bad Commit

No response

Relevant log output

[cgruver]

I am seeing similar behavior when using llama-server with a Qwen3 model. I am using the Continue.dev extension for VS Code.

Previously, there was a "thinking" icon that could be expanded to see the streaming thought process. That has Gone away in the last few days. So, I suspect a recent change.

I am using the OpenAI API compatibility.

[BVEsun]

same bug here,
in order to show the reasoning, I had to add --reasoning-format none
but it still didn't work as good as the old one,

First Bad Commit

b5576
#13933

[littlett77]

I've also tried using the flag --jinja along with --reasoning-format deepseek (that I used before with the Qwen3 models) but no thinking is showed as without it. I've also tried --reasoning-format none as suggested that effectively makes the thinking show again but it is roughly equivalent to not using the --jinja flag in the first place

[segmond]

I have confirmed this as well, I don't see the thinking tokens when I use --jinja option.
Running deepseek-r1-0528

[jxy]

Shouldn't the webui be updated to show reasoning_content?

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.