DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-10 19:06:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat: add max-vram based segmented param offload (#1476)

Browse Source
This commit is contained in:
leejet authored and GitHub committed 2026-05-06 21:56:02 +08:00
1 parent 586b6f1481
commit 90e87bc846
26 files changed
+1576 -66
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/qwen_image.hpp
+5
View File
@@ -412,6 +412,9 @@ namespace Qwen {
auto img = img_in->forward(ctx, x);
auto txt = txt_norm->forward(ctx, context);
txt = txt_in->forward(ctx, txt);
sd::ggml_graph_cut::mark_graph_cut(img, "qwen_image.prelude", "img");
sd::ggml_graph_cut::mark_graph_cut(txt, "qwen_image.prelude", "txt");
// sd::ggml_graph_cut::mark_graph_cut(t_emb, "qwen_image.prelude", "t_emb");
for (int i = 0; i < params.num_layers; i++) {
auto block = std::dynamic_pointer_cast<QwenImageTransformerBlock>(blocks["transformer_blocks." + std::to_string(i)]);
@@ -419,6 +422,8 @@ namespace Qwen {
auto result = block->forward(ctx, img, txt, t_emb, pe, modulate_index);
img = result.first;
txt = result.second;
sd::ggml_graph_cut::mark_graph_cut(img, "qwen_image.transformer_blocks." + std::to_string(i), "img");
sd::ggml_graph_cut::mark_graph_cut(txt, "qwen_image.transformer_blocks." + std::to_string(i), "txt");
}
if (params.zero_cond_t) {