vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 09:04:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

GLM4 batch clamp

Browse source
This commit is contained in:
Concedo 2025-04-26 09:42:17 +08:00
parent 3f545eadbe
commit 4decd6bea1
3 changed files with 9 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

4
gpttype_adapter.cpp
View file

@ -1907,6 +1907,10 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
printf("Warning: Only GGUF models can use max context above 16k. Max context lowered to 16k.\n");
clamped_max_context_length = 16384;
}
if (isGguf && file_format_meta.model_architecture == GGUFArch::ARCH_GLM4 && kcpp_data->n_batch > 16) {
printf("GLM-4 is broken on larger batch sizes. Clamping batch size to 16.\n");
kcpp_data->n_batch = kcpp_data->n_ubatch = 16;
}
kcpp_data->n_ctx = clamped_max_context_length;
max_context_limit_at_load = clamped_max_context_length;

4
model_adapter.cpp
View file

@ -329,6 +329,10 @@ void print_tok_vec(std::vector<float> &embd)
{
fileformatmeta->model_architecture = GGUFArch::ARCH_RWKV;
}
else if(modelarch=="glm4")
{
fileformatmeta->model_architecture = GGUFArch::ARCH_GLM4;
}
printf("Arch Category: %d\n",fileformatmeta->model_architecture);
}

1
model_adapter.h
View file

@ -61,6 +61,7 @@ enum GGUFArch
ARCH_RWKV = 6,
ARCH_QWEN2VL = 7,
ARCH_GEMMA3 = 8,
ARCH_GLM4 = 9,
};
struct FileFormatExtraMeta