DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2025-12-13 05:48:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: resolve precision issues in SDXL VAE under fp16

Browse Source
This commit is contained in:
leejet 2025-10-14 23:12:39 +08:00
parent 2e9242e37f
commit 1d13041aa2
6 changed files with 55 additions and 44 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
README.md
View File

@ -17,7 +17,6 @@ API and command-line option may change frequently.***
- Image Models
- SD1.x, SD2.x, [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo)
- SDXL, [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo)
- !!!The VAE in SDXL encounters NaN issues under FP16, but unfortunately, the ggml_conv_2d only operates under FP16. Hence, a parameter is needed to specify the VAE that has fixed the FP16 NaN issue. You can find it here: [SDXL VAE FP16 Fix](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors).
- [SD3/SD3.5](./docs/sd3.md)
- [Flux-dev/Flux-schnell](./docs/flux.md)
- [Chroma](./docs/chroma.md)

2
conditioner.hpp
View File

@ -1457,7 +1457,7 @@ struct Qwen2_5_VLCLIPEmbedder : public Conditioner {
const ConditionerParams& conditioner_params) {
std::string prompt;
std::vector<std::pair<int, ggml_tensor*>> image_embeds;
size_t system_prompt_length = 0;
size_t system_prompt_length = 0;
int prompt_template_encode_start_idx = 34;
if (qwenvl->enable_vision && conditioner_params.ref_images.size() > 0) {
LOG_INFO("QwenImageEditPlusPipeline");

70
ggml_extend.hpp
View File

@ -975,38 +975,28 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_conv_2d(struct ggml_context* ctx,
struct ggml_tensor* x,
struct ggml_tensor* w,
struct ggml_tensor* b,
int s0 = 1,
int s1 = 1,
int p0 = 0,
int p1 = 0,
int d0 = 1,
int d1 = 1) {
x = ggml_conv_2d(ctx, w, x, s0, s1, p0, p1, d0, d1);
if (b != NULL) {
b = ggml_reshape_4d(ctx, b, 1, 1, b->ne[0], 1);
// b = ggml_repeat(ctx, b, x);
x = ggml_add_inplace(ctx, x, b);
int s0 = 1,
int s1 = 1,
int p0 = 0,
int p1 = 0,
int d0 = 1,
int d1 = 1,
bool direct = false,
float scale = 1.f) {
if (scale != 1.f) {
x = ggml_scale(ctx, x, scale);
}
if (direct) {
x = ggml_conv_2d_direct(ctx, w, x, s0, s1, p0, p1, d0, d1);
} else {
x = ggml_conv_2d(ctx, w, x, s0, s1, p0, p1, d0, d1);
}
if (scale != 1.f) {
x = ggml_scale(ctx, x, 1.f / scale);
}
return x;
}
// w: [OC*IC, KD, KH, KW]
// x: [N*IC, ID, IH, IW]
__STATIC_INLINE__ struct ggml_tensor* ggml_nn_conv_2d_direct(struct ggml_context* ctx,
struct ggml_tensor* x,
struct ggml_tensor* w,
struct ggml_tensor* b,
int s0 = 1,
int s1 = 1,
int p0 = 0,
int p1 = 0,
int d0 = 1,
int d1 = 1) {
x = ggml_conv_2d_direct(ctx, w, x, s0, s1, p0, p1, d0, d1);
if (b != NULL) {
b = ggml_reshape_4d(ctx, b, 1, 1, b->ne[0], 1);
// b = ggml_repeat(ctx, b, x);
x = ggml_add(ctx, x, b);
x = ggml_add_inplace(ctx, x, b);
}
return x;
}
@ -2067,6 +2057,7 @@ protected:
std::pair<int, int> dilation;
bool bias;
bool direct = false;
float scale = 1.f;
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types, const std::string prefix = "") {
enum ggml_type wtype = GGML_TYPE_F16;
@ -2097,6 +2088,10 @@ public:
direct = true;
}
void set_scale(float scale_value) {
scale = scale_value;
}
std::string get_desc() {
return "Conv2d";
}
@ -2107,11 +2102,18 @@ public:
if (bias) {
b = params["bias"];
}
if (direct) {
return ggml_nn_conv_2d_direct(ctx, x, w, b, stride.second, stride.first, padding.second, padding.first, dilation.second, dilation.first);
} else {
return ggml_nn_conv_2d(ctx, x, w, b, stride.second, stride.first, padding.second, padding.first, dilation.second, dilation.first);
}
return ggml_nn_conv_2d(ctx,
x,
w,
b,
stride.second,
stride.first,
padding.second,
padding.first,
dilation.second,
dilation.first,
direct,
scale);
}
};

2
qwen_image.hpp
View File

@ -535,7 +535,7 @@ namespace Qwen {
}
}
LOG_ERROR("qwen_image_params.num_layers: %ld", qwen_image_params.num_layers);
qwen_image = QwenImageModel(qwen_image_params);
qwen_image = QwenImageModel(qwen_image_params);
qwen_image.init(params_ctx, tensor_types, prefix);
}

12
stable-diffusion.cpp
View File

@ -330,13 +330,6 @@ public:
if (sd_version_is_sdxl(version)) {
scale_factor = 0.13025f;
if (strlen(SAFE_STR(sd_ctx_params->vae_path)) == 0 && strlen(SAFE_STR(sd_ctx_params->taesd_path)) == 0) {
LOG_WARN(
"!!!It looks like you are using SDXL model. "
"If you find that the generated images are completely black, "
"try specifying SDXL VAE FP16 Fix with the --vae parameter. "
"You can find it here: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors");
}
} else if (sd_version_is_sd3(version)) {
scale_factor = 1.5305f;
} else if (sd_version_is_flux(version)) {
@ -517,6 +510,11 @@ public:
LOG_INFO("Using Conv2d direct in the vae model");
first_stage_model->enable_conv2d_direct();
}
if (version == VERSION_SDXL && strlen(SAFE_STR(sd_ctx_params->vae_path)) == 0) {
float vae_conv_2d_scale = 1.f / 32.f;
LOG_WARN("No VAE specified with --vae, using Conv2D scale %.3f", vae_conv_2d_scale);
first_stage_model->set_conv2d_scale(vae_conv_2d_scale);
}
first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model");
} else {

12
vae.hpp
View File

@ -530,6 +530,7 @@ struct VAE : public GGMLRunner {
struct ggml_context* output_ctx) = 0;
virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) = 0;
virtual void enable_conv2d_direct(){};
virtual void set_conv2d_scale(float scale) { SD_UNUSED(scale); };
};
struct AutoEncoderKL : public VAE {
@ -558,6 +559,17 @@ struct AutoEncoderKL : public VAE {
}
}
void set_conv2d_scale(float scale) {
std::vector<GGMLBlock*> blocks;
ae.get_all_blocks(blocks);
for (auto block : blocks) {
if (block->get_desc() == "Conv2d") {
auto conv_block = (Conv2d*)block;
conv_block->set_scale(scale);
}
}
}
std::string get_desc() {
return "vae";
}