Too many No parser definition detected, assuming pure content parser. warning messages when I run llama-server with --no-jinja on purpose.

[beebopkim]

Same as the title, when I intentionally run llama-server with —no-jinia option for a specific purpose, too many No parser definition detected, assuming pure content parser. warning messages are printed.

Using the —no-jinja option intentionally avoids the model's built-in prompt format written in Jinja, then is this error message still necessary?

I wish this warning message would not be output when using the —no-jinja option.

[kuon]

I am starting it like this:

/home/kuon/gamilab/llama.cpp/build-vulkan/bin/llama-server \
  --model /srv/models/coding/Qwen3.5/Qwen3.5-35B-A3B-UD-Q4_K_XL.gguf \
  --device Vulkan2       \
  --alias coding       \
  --ctx-size 200000     \
  --batch-size 1024    \
  --ubatch-size 512    \
  --fit off \
  --seed 1234 \
  --temp 0.7 \
  --top-p 0.9 \
  --min-p 0.05 \
  --top-k 20 \
  --repeat-last-n 64 \
  --repeat-penalty 1.1 \
  --flash-attn  on     \
  --no-mmap            \
  -dio                 \
  --threads 8          \
  --host 0.0.0.0       \
  --port 8002

And here is my test request (on /completion)

{"stream":true,"seed":42,"prompt":"<|im_start|>system\nYou are a data extraction assistant specialized in English text.\nReturn data in the specified JSON format.\nThe provided text is a memo, analyse it and extract tasks.\n\n# CONTEXT\n- Today's date: 2026-03-08T18:13:36.434070Z\n- Current user: Anonymous\n- Response language: English\n\n# INSTRUCTIONS\n- When the output contain an array limit to a few items and do not insist\n- Resolve relative dates\n- Use ISO Dates\n- If you have no data for field, set it to null\n- Use ONLY the response language in output: English\n\n# DATA STRUCTURE/FIELD FORMAT\n.memo_datetime : try to guess the datetime of when the user recorded the memo, use RFC 3339 format\n.memo_location : try to guess the place where the memo was recorded\n.memo_weather : try to guess the weather of the place where the memo was recorded\n.tasks[i].description : task description with all details and context\n.tasks[i].location : location where the task take place\n.tasks[i].people : names of the people involved in the task\n.tasks[i].time : when is the task due, format as RFC 3339, relative to memo_datetime\n\nYou are a data extraction assistant specialized in English text.\nReturn data in the specified JSON format.\nThe provided text is a memo, analyse it and extract tasks.\n\n# CONTEXT\n- Today's date: 2026-03-08T18:13:36.434070Z\n- Current user: Anonymous\n- Response language: English\n\n# INSTRUCTIONS\n- When the output contain an array limit to a few items and do not insist\n- Resolve relative dates\n- Use ISO Dates\n- If you have no data for field, set it to null\n- Use ONLY the response language in output: English\n\n# DATA STRUCTURE/FIELD FORMAT\n.memo_datetime : try to guess the datetime of when the user recorded the memo, use RFC 3339 format\n.memo_location : try to guess the place where the memo was recorded\n.memo_weather : try to guess the weather of the place where the memo was recorded\n.tasks[i].description : task description with all details and context\n.tasks[i].location : location where the task take place\n.tasks[i].people : names of the people involved in the task\n.tasks[i].time : when is the task due, format as RFC 3339, relative to memo_datetime\n<|im_end|>\n<|im_start|>user\n\"Okay… uh, Thursday, 12 March 2026, I think it's… 7:14 a.m.? The dashboard clock\nkeeps blinking at me but let's call it 7-ish. Right, so I'm still parked in\nfront of the house because Milo jammed Pokémon cards into the seat-belt latch\nagain—good times. Deep breath—weather's crisp, little patch of blue sky, maybe 9\ndegrees. Energy's like a 2 ½—only one espresso so far—but mood's pretty decent,\nmaybe a 4; stress… eh, let's say 3.\n\nWork brain: yesterday was a small win—AcelaBio finally returned the signed MSA.\nHuge relief. Means I can nudge Procurement to raise the PO this afternoon. Oh,\nand HelixNova pinged at 6 a.m. wanting \"that updated cost model\"—I told them I'd\nget it to them before tomorrow's close of business, so… note to self: block an\nhour after lunch.\nI've also got that quick check-in with Priya at 10:30 on Teams—nothing formal,\nshe just wants to vent about the buffer tanks. Then at 2 p.m. the FDA\nwalk-through; really need to reread the cleaning log.\n\nWait, back up: blocker—Diego still hasn't run a fresh capacity Gantt. Without\nthat I can't price OmegaGen's proposal. Need to poke him, politely, but\nfirmly—due today, Moira, today.\n\nFamily stuff: call Mum about her physio, she texted at 5 a.m. Twins have algebra\nworksheets—help them tonight after dinner, not during. Dentist reminder—Monday\nthe 16th at 8 a.m.—don't bail again. Health—slept maybe five hours? If the FDA\naudit ends on time I'll sneak a 30-minute treadmill run.\nErrands: pick up dry-cleaning, charger cable's still sitting in my Amazon cart,\nand oh—pay the AmEx bill; it hits on Monday. Social—Lina's birthday drinks\nSaturday. I still need a gift… think she mentioned a ceramic mug set? Add to\nlist.\n\nLearned something neat on the Bioprocess Buzz podcast—episode about continuous\nperfusion; I want to pull the journal article they cited—remind me. Also\ngrateful to Tom for covering the Madison call yesterday so I could make school\npick-up. Idea-flash: what if we built a \"quick-quote calculator\" inside the CRM\nso I'm not juggling spreadsheets in traffic? Might prototype Sunday morning.\n\nBike's still at the shop—collect it tonight unless the FDA folks overrun. Okay,\nseat-belt latch is free, let's roll.\"\n\n<|im_end|>\n<|im_start|>assistant\n","grammar":"boolean ::= (\"true\" | \"false\") space\nchar ::= [^\"\\\\\\x7F\\x00-\\x1F] | [\\\\] ([\"\\\\bfnrt] | \"u\" [0-9a-fA-F]{4})\ndecimal-part ::= [0-9]{1,16}\nf-memo-datetime ::= \"\\\"memo_datetime\\\"\" space \":\" space ( string | null )\nf-memo-location ::= \"\\\"memo_location\\\"\" space \":\" space ( string | null )\nf-memo-weather ::= \"\\\"memo_weather\\\"\" space \":\" space ( string | null )\nf-tasks ::= \"\\\"tasks\\\"\" space \":\" space \"[\" space ( o-tasks (\",\" space o-tasks)* )? \"]\" space\nf-tasks-description ::= \"\\\"description\\\"\" space \":\" space string\nf-tasks-location ::= \"\\\"location\\\"\" space \":\" space ( string | null )\nf-tasks-people ::= \"\\\"people\\\"\" space \":\" space ( \"[\" space ( string (\",\" space string)* )? \"]\" space | null )\nf-tasks-time ::= \"\\\"time\\\"\" space \":\" space ( string | null )\nguard ::=  \"\\\"_\\\": \\\"\\\"\"\ninteger ::= (\"-\"? integral-part) space\nintegral-part ::= [0] | [1-9] [0-9]{0,15}\nnull ::= \"null\" space\nnumber ::= (\"-\"? integral-part) (\".\" decimal-part)? ([eE] [-+]? integral-part)? space\no ::= \"{\" space ( f-memo-datetime \",\" space )? ( f-memo-location \",\" space )? ( f-memo-weather \",\" space )? ( f-tasks \",\" space ) guard \"}\" space\no-tasks ::= \"{\" space ( f-tasks-time \",\" space )? ( f-tasks-description \",\" space ) ( f-tasks-location \",\" space )? ( f-tasks-people \",\" space )? guard \"}\" space\nroot ::= o\nspace ::= | \" \" | \"\\n\"{1,2} [ \\t]{0,20}\nstring ::= \"\\\"\" char* \"\\\"\" space","cache_prompt":false,"repeat_penalty":1.05,"temperature":0.1}

It generates the warning on each token.

[ccebelenski]

Yes - seeing it on every token generation - really annoying and slows things down. This is a recent problem. Also see it with --jinja enabled.

[sinni800]

I'm seeing it with Qwen 3.5 with the Jinja template in the commandline as well. It spams it on every generation, but it doesn't seem to slow things down at least...?

[jan-johansson-mr]

This message is displayed (I am using Qwen 3.5) after a recent update, with our without --no-jinja option. However, it doesn't affect the inference speed. It will fill up logs with this message though...

[beebopkim]

Now I found that 6c770d1 resolved annoying No parser definition detected, assuming pure content parser. error message spamming in the terminal. Yay!