chat : keep DSML markup out of DeepSeek V3.2 string tool arguments

[iamganesha]

Overview

I ran into this on 6 September while testing DeepSeek V4 Flash 0731 (Unsloth UD-IQ3_XXS) locally on an AMD Strix Halo system. llama.cpp handles its tool format through the DeepSeek V3.2 DSML parser.

During one of the tool loops, the model misspelled a closing parameter marker. Instead of treating that as malformed DSML, the parser swallowed it as part of the string argument. The model then kept generating until it eventually hit the token ceiling.

Tracing that back led to the DeepSeek V3.2 tool-call parser in common/parsers/deepseek.cpp.

At the moment, string parameter values are parsed using a bare until(PARAM_END). Because of that, the value can consume anything up to the next valid closing parameter tag — including text that itself looks like malformed or unexpected DSML markup.

The same PEG is also used to build the sampling grammar, so constrained generation can technically produce stray or nested <｜DSML｜...> markup inside a string value and still have it accepted by the grammar.

This change tightens that behaviour.

A string value now stops when it encounters either:

• the actual closing parameter tag, or
• anything that looks like DSML markup: <｜ or </｜.

Normal angle brackets are still perfectly valid inside string values, so content such as <east> & Osaka continues to work as expected.

I added two regression tests in tests/test-chat.cpp:

• A parser test with <east> & Osaka inside a string parameter to make sure ordinary angle-bracket content is preserved literally.
• A grammar test that builds the DeepSeek V3.2 tool grammar, confirms that a clean tool call is accepted, and confirms that a call containing a nested <｜DSML｜parameter ...> inside the argument value is rejected.

The rejection test fails against current master and passes with this change.

The full test-chat suite also passes.

Additional information

The original failure I saw was not a nested parameter tag. DeepSeek V4 Flash misspelled a closing parameter marker, the parser treated that malformed DSML as part of the argument string, and generation continued until the token limit.

The nested-tag case in the regression test is a simpler deterministic way of exercising the same parser bug: DSML-looking markup should never be allowed to disappear into an ordinary string parameter.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. The original fix and the two regression tests were produced with the assistance of Codex Sol. I reviewed the generated changes, verified that they address the issue I reproduced locally, ran the relevant test suite, and take responsibility for the submitted change.

[ggml-gh-bot]

Hi @iamganesha, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[iamganesha]

The Description now follows the template with the requirements section included