DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-09-25 12:40:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: scale Qwen Image 2.1 VAE convolutions (#2054)

Browse Source
This commit is contained in:
leejet 2026-09-25 15:52:35 +08:00 committed by GitHub
parent b167b942f7
commit 0a9340c599
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 30 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

BIN
assets/qwen/qwen-image-2.1-alpha-out2.png
View File

Binary file not shown.
Side by Side Swipe Overlay

Before

Width:  |  Height:  |  Size: 1.7 MiB

After

Width:  |  Height:  |  Size: 1.1 MiB

9
src/core/ggml_extend.cpp
View File

@ -465,7 +465,11 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
int d1,
int d2,
bool force_prec_f32,
bool direct) {
bool direct,
float scale) {
if (scale != 1.f) {
x = ggml_ext_scale(ctx, x, scale);
}
if (direct) {
int64_t OC = w->ne[3] / IC;
int64_t N = x->ne[3] / IC;
@ -502,6 +506,9 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
}
}
if (scale != 1.f) {
x = ggml_ext_scale(ctx, x, 1.f / scale);
}
if (b != nullptr) {
b = ggml_reshape_4d(ctx, b, 1, 1, 1, b->ne[0]); // [OC, 1, 1, 1]
x = ggml_add_inplace(ctx, x, b);

3
src/core/ggml_extend.h
View File

@ -154,7 +154,8 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
int d1 = 1,
int d2 = 1,
bool force_prec_f32 = false,
bool direct = false);
bool direct = false,
float scale = 1.f);
// w: [OC，IC, KD, 1 * 1]
// x: [N, IC, ID, IH*IW]

22
src/model/vae/wan_vae.hpp
View File

@ -24,6 +24,7 @@ namespace WAN {
std::tuple<int, int, int> padding;
std::tuple<int, int, int> dilation;
bool bias;
float scale = 1.f;
void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
auto weight = tensor_storage_map.find(prefix + "weight");
@ -60,6 +61,10 @@ namespace WAN {
dilation(std::move(dilation)),
bias(bias) {}
void set_scale(float scale_value) {
scale = scale_value;
}
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* cache_x = nullptr) {
// x: [N*IC, ID, IH, IW]
// result: x: [N*OC, ID, IH, IW]
@ -93,14 +98,14 @@ namespace WAN {
x2 = ggml_ext_conv_2d(ctx->ggml_ctx, x2, w2, b,
std::get<2>(stride), std::get<1>(stride), 0, 0,
std::get<2>(dilation), std::get<1>(dilation),
ctx->conv2d_direct_enabled);
ctx->conv2d_direct_enabled, false, false, scale);
return ggml_reshape_4d(ctx->ggml_ctx, x2, x2->ne[0], x2->ne[1], 1, out_channels);
}
return ggml_ext_conv_3d(ctx->ggml_ctx, ctx->backend, x, w, b, in_channels,
std::get<2>(stride), std::get<1>(stride), std::get<0>(stride),
0, 0, 0,
std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation),
false, ctx->conv3d_direct_enabled);
false, ctx->conv3d_direct_enabled, scale);
}
};
@ -1117,6 +1122,19 @@ namespace WAN {
} else {
blocks["conv2"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(z_dim, z_dim, {1, 1, 1}));
}
if (version == VERSION_QWEN_IMAGE_2_1) {
// Keep large VAE activations within the FP16 convolution range.
const float conv_scale = 1.f / 128.f;
std::vector<GGMLBlock*> all_blocks;
get_all_blocks(all_blocks);
for (auto block : all_blocks) {
if (auto conv = dynamic_cast<Conv2d*>(block)) {
conv->set_scale(conv_scale);
} else if (auto conv = dynamic_cast<CausalConv3d*>(block)) {
conv->set_scale(conv_scale);
}
}
}
}
static ggml_tensor* patchify(ggml_context* ctx,