Skip to content

WIP: update stable-diffusion.cpp to 5900ef6605c6 (new API) #1669

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 9 commits into from
Aug 12, 2025
Merged

WIP: update stable-diffusion.cpp to 5900ef6605c6 (new API) #1669

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
784b1ec
Update stable-diffusion.cpp to 5900ef6605c6 (new API)
wbruna Aug 4, 2025
b728882
Clean up pending LoRA code and simplify LoRA changes to upstream
wbruna Aug 5, 2025
36e23b2
Move VAE tiling disabling for TAESD to sdtype_adapter.cpp
wbruna Aug 5, 2025
594fd5c
Move auxiliary ctx functions to sdtype_adapter.cpp
wbruna Aug 5, 2025
f5eac10
Use ref_images parameter for Kontext images
wbruna Aug 8, 2025
5fec0cc
Drop clip skip workaround (fixed upstream)
wbruna Aug 10, 2025
2c65a96
Workaround for flash attention with img2img
wbruna Aug 10, 2025
6cc7f3e
Workaround for Chroma with flash attention, debug prints
wbruna Aug 10, 2025
c433877
Disable forcing CLIP weights to F32 for reduced memory usage
wbruna Aug 11, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
44 changes: 25 additions & 19 deletions otherarch/sdcpp/clip.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -545,9 +545,15 @@ class CLIPEmbeddings : public GGMLBlock {
int64_t vocab_size;
int64_t num_positions;

void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
enum ggml_type token_wtype = (tensor_types.find(prefix + "token_embedding.weight") != tensor_types.end()) ? tensor_types[prefix + "token_embedding.weight"] : GGML_TYPE_F32;
enum ggml_type position_wtype = GGML_TYPE_F32; //(tensor_types.find(prefix + "position_embedding.weight") != tensor_types.end()) ? tensor_types[prefix + "position_embedding.weight"] : GGML_TYPE_F32;
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
enum ggml_type token_wtype = GGML_TYPE_F32;
#if 1
// kcpp reduce memory usage (reverts https://github.com/leejet/stable-diffusion.cpp/pull/601)
auto tensor_type = tensor_types.find(prefix + "token_embedding.weight");
if (tensor_type != tensor_types.end())
token_wtype = tensor_type->second;
#endif
enum ggml_type position_wtype = GGML_TYPE_F32;

params["token_embedding.weight"] = ggml_new_tensor_2d(ctx, token_wtype, embed_dim, vocab_size);
params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
Expand Down Expand Up @@ -594,10 +600,10 @@ class CLIPVisionEmbeddings : public GGMLBlock {
int64_t image_size;
int64_t num_patches;
int64_t num_positions;
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
enum ggml_type patch_wtype = GGML_TYPE_F16; // tensor_types.find(prefix + "patch_embedding.weight") != tensor_types.end() ? tensor_types[prefix + "patch_embedding.weight"] : GGML_TYPE_F16;
enum ggml_type class_wtype = GGML_TYPE_F32; // tensor_types.find(prefix + "class_embedding") != tensor_types.end() ? tensor_types[prefix + "class_embedding"] : GGML_TYPE_F32;
enum ggml_type position_wtype = GGML_TYPE_F32; // tensor_types.find(prefix + "position_embedding.weight") != tensor_types.end() ? tensor_types[prefix + "position_embedding.weight"] : GGML_TYPE_F32;
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
enum ggml_type patch_wtype = GGML_TYPE_F16;
enum ggml_type class_wtype = GGML_TYPE_F32;
enum ggml_type position_wtype = GGML_TYPE_F32;

params["patch_embedding.weight"] = ggml_new_tensor_4d(ctx, patch_wtype, patch_size, patch_size, num_channels, embed_dim);
params["class_embedding"] = ggml_new_tensor_1d(ctx, class_wtype, embed_dim);
Expand Down Expand Up @@ -657,9 +663,9 @@ enum CLIPVersion {

class CLIPTextModel : public GGMLBlock {
protected:
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
if (version == OPEN_CLIP_VIT_BIGG_14) {
enum ggml_type wtype = GGML_TYPE_F32; // tensor_types.find(prefix + "text_projection") != tensor_types.end() ? tensor_types[prefix + "text_projection"] : GGML_TYPE_F32;
enum ggml_type wtype = GGML_TYPE_F32;
params["text_projection"] = ggml_new_tensor_2d(ctx, wtype, projection_dim, hidden_size);
}
}
Expand All @@ -678,8 +684,8 @@ class CLIPTextModel : public GGMLBlock {
bool with_final_ln = true;

CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
int clip_skip_value = -1,
bool with_final_ln = true)
bool with_final_ln = true,
int clip_skip_value = -1)
: version(version), with_final_ln(with_final_ln) {
if (version == OPEN_CLIP_VIT_H_14) {
hidden_size = 1024;
Expand All @@ -701,7 +707,7 @@ class CLIPTextModel : public GGMLBlock {

void set_clip_skip(int skip) {
if (skip <= 0) {
return;
skip = -1;
}
clip_skip = skip;
}
Expand Down Expand Up @@ -805,8 +811,8 @@ class CLIPProjection : public UnaryBlock {
int64_t out_features;
bool transpose_weight;

void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
enum ggml_type wtype = tensor_types.find(prefix + "weight") != tensor_types.end() ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
if (transpose_weight) {
params["weight"] = ggml_new_tensor_2d(ctx, wtype, out_features, in_features);
} else {
Expand Down Expand Up @@ -868,12 +874,12 @@ struct CLIPTextModelRunner : public GGMLRunner {
CLIPTextModel model;

CLIPTextModelRunner(ggml_backend_t backend,
std::map<std::string, enum ggml_type>& tensor_types,
const String2GGMLType& tensor_types,
const std::string prefix,
CLIPVersion version = OPENAI_CLIP_VIT_L_14,
int clip_skip_value = 1,
bool with_final_ln = true)
: GGMLRunner(backend), model(version, clip_skip_value, with_final_ln) {
bool with_final_ln = true,
int clip_skip_value = -1)
: GGMLRunner(backend), model(version, with_final_ln, clip_skip_value) {
model.init(params_ctx, tensor_types, prefix);
}

Expand Down Expand Up @@ -949,4 +955,4 @@ struct CLIPTextModelRunner : public GGMLRunner {
}
};

#endif // __CLIP_HPP__
#endif // __CLIP_HPP__
14 changes: 7 additions & 7 deletions otherarch/sdcpp/common.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,8 +56,8 @@ class UpSampleBlock : public GGMLBlock {
// x: [N, channels, h, w]
auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);

x = ggml_upscale(ctx, x, 2, ggml_scale_mode::GGML_SCALE_MODE_NEAREST); // [N, channels, h*2, w*2]
x = conv->forward(ctx, x); // [N, out_channels, h*2, w*2]
x = ggml_upscale(ctx, x, 2, GGML_SCALE_MODE_NEAREST); // [N, channels, h*2, w*2]
x = conv->forward(ctx, x); // [N, out_channels, h*2, w*2]
return x;
}
};
Expand Down Expand Up @@ -182,9 +182,9 @@ class GEGLU : public GGMLBlock {
int64_t dim_in;
int64_t dim_out;

void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
enum ggml_type wtype = (tensor_types.find(prefix + "proj.weight") != tensor_types.end()) ? tensor_types[prefix + "proj.weight"] : GGML_TYPE_F32;
enum ggml_type bias_wtype = GGML_TYPE_F32; //(tensor_types.find(prefix + "proj.bias") != tensor_types.end()) ? tensor_types[prefix + "proj.bias"] : GGML_TYPE_F32;
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
enum ggml_type wtype = get_type(prefix + "proj.weight", tensor_types, GGML_TYPE_F32);
enum ggml_type bias_wtype = GGML_TYPE_F32;
params["proj.weight"] = ggml_new_tensor_2d(ctx, wtype, dim_in, dim_out * 2);
params["proj.bias"] = ggml_new_tensor_1d(ctx, bias_wtype, dim_out * 2);
}
Expand Down Expand Up @@ -440,9 +440,9 @@ class SpatialTransformer : public GGMLBlock {

class AlphaBlender : public GGMLBlock {
protected:
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
// Get the type of the "mix_factor" tensor from the input tensors map with the specified prefix
enum ggml_type wtype = GGML_TYPE_F32; //(tensor_types.ypes.find(prefix + "mix_factor") != tensor_types.end()) ? tensor_types[prefix + "mix_factor"] : GGML_TYPE_F32;
enum ggml_type wtype = GGML_TYPE_F32;
params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
}

Expand Down
71 changes: 37 additions & 34 deletions otherarch/sdcpp/conditioner.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,29 +57,30 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
std::vector<std::string> readed_embeddings;

FrozenCLIPEmbedderWithCustomWords(ggml_backend_t backend,
std::map<std::string, enum ggml_type>& tensor_types,
const String2GGMLType& tensor_types,
const std::string& embd_dir,
SDVersion version = VERSION_SD1,
PMVersion pv = PM_VERSION_1,
int clip_skip = -1)
: version(version), pm_version(pv), tokenizer(sd_version_is_sd2(version) ? 0 : 49407), embd_dir(embd_dir) {
if (clip_skip <= 0) {
clip_skip = 1;
if (sd_version_is_sd2(version) || sd_version_is_sdxl(version)) {
clip_skip = 2;
}
}
if (sd_version_is_sd1(version)) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, clip_skip);
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14);
} else if (sd_version_is_sd2(version)) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14, clip_skip);
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14);
} else if (sd_version_is_sdxl(version)) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, clip_skip, false);
text_model2 = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, clip_skip, false);
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
text_model2 = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
}
set_clip_skip(clip_skip);
}

void set_clip_skip(int clip_skip) {
if (clip_skip <= 0) {
clip_skip = 1;
if (sd_version_is_sd2(version) || sd_version_is_sdxl(version)) {
clip_skip = 2;
}
}
text_model->set_clip_skip(clip_skip);
if (sd_version_is_sdxl(version)) {
text_model2->set_clip_skip(clip_skip);
Expand Down Expand Up @@ -458,8 +459,8 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
if (sd_version_is_sdxl(version)) {
text_model2->compute(n_threads,
input_ids2,
0,
NULL,
num_custom_embeddings,
token_embed_custom.data(),
max_token_idx,
false,
&chunk_hidden_states2, work_ctx);
Expand All @@ -469,8 +470,8 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
if (chunk_idx == 0) {
text_model2->compute(n_threads,
input_ids2,
0,
NULL,
num_custom_embeddings,
token_embed_custom.data(),
max_token_idx,
true,
&pooled,
Expand Down Expand Up @@ -617,7 +618,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
struct FrozenCLIPVisionEmbedder : public GGMLRunner {
CLIPVisionModelProjection vision_model;

FrozenCLIPVisionEmbedder(ggml_backend_t backend, std::map<std::string, enum ggml_type>& tensor_types)
FrozenCLIPVisionEmbedder(ggml_backend_t backend, const String2GGMLType& tensor_types = {})
: vision_model(OPEN_CLIP_VIT_H_14, true), GGMLRunner(backend) {
vision_model.init(params_ctx, tensor_types, "cond_stage_model.transformer");
}
Expand Down Expand Up @@ -662,18 +663,19 @@ struct SD3CLIPEmbedder : public Conditioner {
std::shared_ptr<T5Runner> t5;

SD3CLIPEmbedder(ggml_backend_t backend,
std::map<std::string, enum ggml_type>& tensor_types,
int clip_skip = -1)
const String2GGMLType& tensor_types = {},
int clip_skip = -1)
: clip_g_tokenizer(0) {
if (clip_skip <= 0) {
clip_skip = 2;
}
clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, clip_skip, false);
clip_g = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, clip_skip, false);
clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
clip_g = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
t5 = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
set_clip_skip(clip_skip);
}

void set_clip_skip(int clip_skip) {
if (clip_skip <= 0) {
clip_skip = 2;
}
clip_l->set_clip_skip(clip_skip);
clip_g->set_clip_skip(clip_skip);
}
Expand Down Expand Up @@ -1008,16 +1010,17 @@ struct FluxCLIPEmbedder : public Conditioner {
size_t chunk_len = 256;

FluxCLIPEmbedder(ggml_backend_t backend,
std::map<std::string, enum ggml_type>& tensor_types,
int clip_skip = -1) {
if (clip_skip <= 0) {
clip_skip = 2;
}
clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, clip_skip, true);
const String2GGMLType& tensor_types = {},
int clip_skip = -1) {
clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, true);
t5 = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
set_clip_skip(clip_skip);
}

void set_clip_skip(int clip_skip) {
if (clip_skip <= 0) {
clip_skip = 2;
}
clip_l->set_clip_skip(clip_skip);
}

Expand Down Expand Up @@ -1228,10 +1231,10 @@ struct PixArtCLIPEmbedder : public Conditioner {
int mask_pad = 1;

PixArtCLIPEmbedder(ggml_backend_t backend,
std::map<std::string, enum ggml_type>& tensor_types,
int clip_skip = -1,
bool use_mask = false,
int mask_pad = 1)
const String2GGMLType& tensor_types = {},
int clip_skip = -1,
bool use_mask = false,
int mask_pad = 1)
: use_mask(use_mask), mask_pad(mask_pad) {
t5 = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
}
Expand Down Expand Up @@ -1422,4 +1425,4 @@ struct PixArtCLIPEmbedder : public Conditioner {
}
};

#endif
#endif
15 changes: 13 additions & 2 deletions otherarch/sdcpp/control.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -317,12 +317,23 @@ struct ControlNet : public GGMLRunner {
bool guided_hint_cached = false;

ControlNet(ggml_backend_t backend,
std::map<std::string, enum ggml_type>& tensor_types,
SDVersion version = VERSION_SD1)
const String2GGMLType& tensor_types = {},
SDVersion version = VERSION_SD1)
: GGMLRunner(backend), control_net(version) {
control_net.init(params_ctx, tensor_types, "");
}

void enable_conv2d_direct() {
std::vector<GGMLBlock*> blocks;
control_net.get_all_blocks(blocks);
for (auto block : blocks) {
if (block->get_desc() == "Conv2d") {
auto conv_block = (Conv2d*)block;
conv_block->enable_direct();
}
}
}

~ControlNet() {
free_control_ctx();
}
Expand Down
Loading