Misc. bug: error parsing grammar: number of repetitions exceeds sane defaults, please reduce the number of repetitions

[matiashegoburu]

Name and Version

llama-server full cuda docker v b7139

Looks like this might be related to #17381

Operating systems

Linux

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

llama-server

Command line

Problem description & steps to reproduce

I am using latest docker version of llama-server (b7139)
And i am getting this error when trying to use openweebui with native tool calls:

parse: error parsing grammar: number of repetitions exceeds sane defaults, please reduce the number of repetitions

I didnt have this issue with previous versions (with the exact same setup), not sure exactly when it broke, as i update llama server package once a week.

First Bad Commit

#17381

Relevant log output

[matiashegoburu]

going back to docker container before #17381 fixes the issue

[o7si]

Hi @matiashegoburu, #17381 imposed a limit on the maximum number of repetitions to prevent DoS attacks. If you need a higher repetition count, you can adjust MAX_REPETITION_THRESHOLD.

llama.cpp/src/llama-grammar.cpp

Line 12 in b61de2b

#define MAX_REPETITION_THRESHOLD 2000

[matiashegoburu]

@o7si, could this be adjusted via env variables? I do not want to have to recompile from source every time there is an update

[o7si]

@matiashegoburu Currently, it's not supported via environment variables. If you don't need this limitation, you can set it to a very large value as a workaround. However, unless there's a deliberate attack, the limit of 2000 should be sufficient for normal use cases.

Could you describe your specific scenario? If it's a common use case, supporting configuration via parameters or environment variables might be a good option to consider.

[matiashegoburu]

@o7si , i understand i could recompile from source, but i do not want to have to be constantly compiling from source. I have setup my servers with the docker image for this reason, so hardcoding this limit is a problem for me. As of the use case, I am running qwen3-30b-a3b with native tool calls enabled, and setting a couple of tools (specifically i enabled basic memory https://github.com/basicmachines-co/basic-memory). Before your update, runs with no problem, after your update it breaks. Not sure what the limit needs to be for this to work, but i dont think hardcoding it is a good idea, a parameter or env variable would be greatly appreciated, until then i am unable to update the docker image to the latest version anymore.

[pwilkin]

@matiashegoburu I really didn't think this would be an actual issue; I'll add a configuration parameter for it.