vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-04-28 11:40:43 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

Henk's Gemma4 31B Magic (#2096)

Browse source
This commit is contained in:
henk717 2026-04-06 12:49:19 +02:00 • committed by GitHub
parent 59f863746d
commit 4e30294cb1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 20 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

20
gpttype_adapter.cpp
View file

@ -3847,6 +3847,26 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
}
}
// Round two, gemma4 boogalo
// If it breaks your stuff you can blame me again (Or thank me because you can actually use gemma 31B stable now). - Henk
// For the record, the GLM4 one didn't break anyone and everyone forgot GLM4 needed this :D
if (file_format == FileFormat::GGUF_GENERIC && (file_format_meta.model_architecture == llm_arch::LLM_ARCH_GEMMA4)) {
std::string temp = gpttype_get_chat_template();
if (temp.find("<|channel>thought") != std::string::npos) {
const std::string channel_open = "<|channel>";
const std::string channel_close = "<channel|>";
const std::string channel_prefix = channel_open + channel_close;
const bool has_open = kcpp_data->prompt.find(channel_open) != std::string::npos;
const bool has_close = kcpp_data->prompt.find(channel_close) != std::string::npos;
// If neither opening nor closing tag is present anywhere, prepend both
if (!has_open && !has_close) {
kcpp_data->prompt = channel_prefix + kcpp_data->prompt;
}
}
}
bool stream_sse = inputs.stream_sse;
bool allow_regular_prints = (!is_quiet && debugmode!=-1);