furyhawk/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-07-21 10:15:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

chat : fix reasoning leak with force-opened bare <think> templates (#24674)

Browse Source 
* chat : fix reasoning leak with force-opened bare <think> templates

The reasoning start tag inferred from prior turns can carry trailing
whitespace (e.g. <think>\n) while a force-open template prefills a bare
<think>. Trim the tag used for the prefix split so the bare prefill is
matched instead of being swallowed into content.

* chat : fix Nemotron Nano v2 regression

---------

Co-authored-by: Alde Rojas <hello@alde.dev>
This commit is contained in:
Frosty40
2026-07-13 09:45:10 +02:00
committed by GitHub
co-authored by Alde Rojas
parent efb3036c18
commit 91c631b21d
3 changed files with 13 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/chat-auto-parser-generator.cpp
+2 -1
View File
@@ -147,7 +147,8 @@ common_peg_arena autoparser::build_parser(const generation_params & inputs, cons
} else {
parser = content.build_parser(ctx);
}
return pure_content ? p.prefix(generation_prompt, reasoning.start) + parser : p.prefix(generation_prompt, reasoning.start) << parser;
const std::string reasoning_start = trim_whitespace(reasoning.start);
return pure_content ? p.prefix(generation_prompt, reasoning_start) + parser : p.prefix(generation_prompt, reasoning_start) << parser;
});
}
common/chat-diff-analyzer.cpp
+2 -2
View File
@@ -124,16 +124,16 @@ static std::vector<std::function<void(const common_chat_template & tmpl, autopar
analysis.tools.format.section_end = "";
analysis.tools.format.per_call_start = "<TOOLCALL>";
analysis.tools.format.per_call_end = "</TOOLCALL>";
analysis.tools.format.tools_array_wrapped = true;
analysis.content.mode = content_mode::PLAIN;
analysis.content.start = "";
analysis.content.end = "";
analysis.reasoning.mode = reasoning_mode::TAG_BASED;
analysis.reasoning.start = "<think>\n\n";
analysis.reasoning.start = "<think>\n";
analysis.reasoning.end = "</think>";
analysis.assistant_start = "<SPECIAL_11>Assistant";
analysis.user_start = "<SPECIAL_11>User";
analysis.preserved_tokens.clear();
analysis.preserved_tokens.push_back("<SPECIAL_12>");
analysis.preserved_tokens.push_back("<SPECIAL_11>");
analysis.preserved_tokens.push_back("</think>");
analysis.preserved_tokens.push_back("<TOOLCALL>");