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: trim messages sent to StepFun parser (fixes long reasoning loops) (#25238)

Browse Source 
* chat: trim messages sent to StepFun parser (fixes long reasoning loops)

* add regression test; remove duplicate template

* chat: trim StepFun content parts before rendering

The StepFun trim workaround ran on the already-rendered messages, where
typed content parts have been concatenated into a single string, so the
per-part whitespace could no longer be reached. Move the trim ahead of
rendering and apply it to content_parts text as well as the string
content and reasoning_content. Adds a content-parts regression test.

Co-Authored-By: Piotr Wilkin <ilintar@gmail.com>
Assisted-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: tarruda <tpadilha84@gmail.com>
This commit is contained in:
Piotr Wilkin (ilintar)
2026-07-03 23:12:11 +02:00
committed by GitHub
co-authored by tarruda
parent d4cff114c0
commit 2d973636e2
4 changed files with 80 additions and 82 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/chat.cpp
+27 -1
View File
@@ -2378,6 +2378,23 @@ static void func_args_not_string(json & messages) {
}
}
// Trim leading/trailing whitespace from message contents before rendering. This
// has to run on the messages (not on the rendered JSON) because templates with
// string-only content caps concatenate typed content parts into a single string
// during rendering, after which the per-part whitespace can no longer be reached.
// Both the plain string content and the text of typed content parts are trimmed.
static void trim_all_content(std::vector<common_chat_msg> & messages) {
for (auto & message : messages) {
message.content = trim_whitespace(message.content);
message.reasoning_content = trim_whitespace(message.reasoning_content);
for (auto & part : message.content_parts) {
if (part.type == "text") {
part.text = trim_whitespace(part.text);
}
}
}
}
}
// MiniCPM5 format:
@@ -2634,7 +2651,16 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
params.tools.is_array() && tmpls->template_tool_use ? *tmpls->template_tool_use : *tmpls->template_default;
const auto & src = tmpl.source();
const auto & caps = tmpl.original_caps();
params.messages = render_message_to_json(inputs.messages, tmpl.original_caps());
std::vector<common_chat_msg> trimmed_messages;
const std::vector<common_chat_msg> * messages_to_render = &inputs.messages;
if (src.find("You have access to the following functions in JSONSchema format") != std::string::npos) {
// StepFun: trim message contents (including typed content parts) before rendering,
// otherwise leftover whitespace drives the model into reasoning loops (issue #24181)
trimmed_messages = inputs.messages;
workaround::trim_all_content(trimmed_messages);
messages_to_render = &trimmed_messages;
}
params.messages = render_message_to_json(*messages_to_render, tmpl.original_caps());
params.tool_choice = inputs.tool_choice;
params.reasoning_format = inputs.reasoning_format;
params.enable_thinking = inputs.enable_thinking;