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
tests/test-chat-auto-parser.cpp
-1
View File
@@ -1887,7 +1887,6 @@ static void test_role_markers_all_templates(testing & t) {
{ "Qwen-Qwen3-0.6B.jinja", "<|im_start|>user", "<|im_start|>assistant" },
{ "Qwen-QwQ-32B.jinja", "<|im_start|>user", "<|im_start|>assistant" },
{ "StepFun3.5-Flash.jinja", "<|im_start|>user", "<|im_start|>assistant" },
{ "stepfun-ai-Step-3.5-Flash.jinja", "<|im_start|>user", "<|im_start|>assistant" },
// DeepSeek family
{ "deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja", "<｜User｜>", "<｜Assistant｜>" },
tests/test-chat.cpp
+53
View File
@@ -3155,6 +3155,59 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
}
}
}
{
// StepFun trimming regression test (see https://github.com/ggml-org/llama.cpp/pull/25238)
auto tmpls = read_templates("models/templates/StepFun3.5-Flash.jinja");
common_chat_msg message_chatbot = simple_assist_msg("Let me check.\n\n", "I am thinking.\n\n");
{
common_chat_templates_inputs inputs;
inputs.messages = { message_chatbot };
inputs.add_generation_prompt = true;
auto params = common_chat_templates_apply(tmpls.get(), inputs);
if (params.prompt.find("Let me check.\n\n") != std::string::npos) {
throw std::runtime_error("StepFun 3.5: content not trimmed");
}
if (params.prompt.find("I am thinking.\n\n") != std::string::npos) {
throw std::runtime_error("StepFun 3.5: reasoning_content not trimmed");
}
}
{
// Trimming must also reach typed (text) content parts, not just string content
// (see https://github.com/ggml-org/llama.cpp/pull/25238)
common_chat_msg message_parts;
message_parts.role = "user";
message_parts.content_parts = {
{ /* .type = */ "text", /* .text = */ "First part.\n\n" },
{ /* .type = */ "media_marker", /* .text = */ "<__media__>" },
{ /* .type = */ "text", /* .text = */ "Second part.\n\n" },
};
common_chat_templates_inputs inputs;
inputs.messages = { message_parts };
inputs.add_generation_prompt = true;
auto params = common_chat_templates_apply(tmpls.get(), inputs);
if (params.prompt.find("First part.\n\n") != std::string::npos ||
params.prompt.find("Second part.\n\n") != std::string::npos) {
throw std::runtime_error("StepFun 3.5: text content parts not trimmed");
}
// the trimmed text itself must still be present
if (params.prompt.find("First part.") == std::string::npos ||
params.prompt.find("Second part.") == std::string::npos) {
throw std::runtime_error("StepFun 3.5: text content parts missing after trim");
}
}
}
}
{