UPSTREAM PR #17707: common: Deepseek V3.2 tool call parser

[loci-dev]

Mirrored from ggml-org/llama.cpp#17707

Deepseek V3.2 uses a new tool-call format like this:

<｜DSML｜function_calls>
<｜DSML｜invoke name="get_datetime">
<｜DSML｜parameter name="timezone" string="true">Asia/Shanghai</｜DSML｜parameter>
</｜DSML｜invoke>
</｜DSML｜function_calls>

<｜DSML｜function_calls>
<｜DSML｜invoke name="search">
<｜DSML｜parameter name="query" string="true">search agent benchmark 2024</｜DSML｜parameter>
<｜DSML｜parameter name="topn" string="false">10</｜DSML｜parameter>
<｜DSML｜parameter name="source" string="true">web</｜DSML｜parameter>
</｜DSML｜invoke>
<｜DSML｜invoke name="search">
<｜DSML｜parameter name="query" string="true">搜索智能体 基准测试</｜DSML｜parameter>
<｜DSML｜parameter name="topn" string="false">10</｜DSML｜parameter>
<｜DSML｜parameter name="source" string="true">web</｜DSML｜parameter>
</｜DSML｜invoke>
</｜DSML｜function_calls>

This PR introduces the tool-call parser for the new DeepSeek V3.2 model.

Since the official release does not provide a chat template, a provisional template has been added and tested only with llama.cpp. Compatibility with other inference engines is not guaranteed and may require further adjustments.

In addition, Minja polyfill detection has been slightly updated to accommodate the new template structure.

Needs PR #17376 to be merged first.

[loci-agentic-ai]

Explore the complete analysis inside the Version Insights

Pull Request #405 - Performance Analysis Summary

PR Title: UPSTREAM PR #17707: common: Deepseek V3.2 tool call parser
Change Scope: 10 files modified (+729 additions, -60 deletions)

---

Analysis Classification: Condition 1 (No Performance Impact)

This PR introduces a new chat template parser for DeepSeek V3.2 model with XML-based tool call format. The code changes add new parsing logic and test coverage without modifying existing inference paths.

Performance Impact Assessment:

The performance metrics show variations in STL container operations (vector::end(), map::begin()) with throughput changes ranging from 60-195 ns. However, these functions are not in the inference critical path. The actual changes are:

• chat-parser-xml-toolcall.cpp: Added allowed_literal_between_kvsep field support for parsing boolean literals between key-value separators. Modified parse_msg_with_xml_tool_calls() to handle tool calls within thinking blocks when allow_toolcall_in_think is enabled.

• Power consumption: Three binaries show minimal increases: llama-tts (+914 nJ, +0.407%), llama-cvector-generator (+756 nJ, +0.343%), llama-run (+442 nJ, +0.230%). These are chat template utilities, not inference engines.

Inference Impact: None. The modified functions (chat parsers, template renderers) execute before model inference begins. Functions like llama_decode, llama_encode, and llama_tokenize are unchanged. Tokens per second remains unaffected.

The observed STL performance variations are compiler optimization artifacts unrelated to the functional changes, which purely extend chat template parsing capabilities for a new model format.

[loci-agentic-ai]

Explore the complete analysis inside the Version Insights