vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-05-11 04:51:25 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

add support for Qwen Image Pruning (#1779)

Browse source 
From leejet/stable-diffusion.cpp#874 .
This commit is contained in:
Wagner Bruna 2025-10-10 05:22:47 -03:00 • committed by GitHub
parent bece22f996
commit bc762fe9b4
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 21 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

21
otherarch/sdcpp/qwen_image.hpp
View file

@ -492,6 +492,27 @@ namespace Qwen {
bool flash_attn = false)
: GGMLRunner(backend, offload_params_to_cpu) {
qwen_image_params.flash_attn = flash_attn;
int model_layers = qwen_image_params.num_layers;
int num_layers = 1;
std::string layer_prefix = "transformer_blocks.";
if (prefix.size() > 0) {
layer_prefix = prefix + "." + layer_prefix;
}
for (int layer = model_layers; layer > num_layers; layer--) {
std::string layer_name = layer_prefix + std::to_string(layer-1) + ".attn.add_k_proj.bias";
for (auto pair : tensor_types) {
if (pair.first.find(layer_name) != std::string::npos) {
num_layers = layer;
break;
}
}
}
if (num_layers < model_layers) {
LOG_INFO("Qwen Image: some layers missing, assuming pruned model");
}
qwen_image_params.num_layers = num_layers;
qwen_image = QwenImageModel(qwen_image_params);
qwen_image.init(params_ctx, tensor_types, prefix);
}