alphard/stable-diffusion.cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: flexible model architecture for dit models (Flux & SD3) (#490)

Browse Source 
* Refactor: wtype per tensor

* Fix default args

* refactor: fix flux

* Refactor photmaker v2 support

* unet: refactor the refactoring

* Refactor: fix controlnet and tae

* refactor: upscaler

* Refactor: fix runtime type override

* upscaler: use fp16 again

* Refactor: Flexible sd3 arch

* Refactor: Flexible Flux arch

* format code

---------

Co-authored-by: leejet <leejet714@gmail.com>
This commit is contained in:
stduhpf 2024-11-30 07:18:53 +01:00 committed by GitHub
parent 4570715727
commit 7ce63e740c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
21 changed files with 317 additions and 271 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
clip.hpp
View File

@ -545,9 +545,12 @@ protected:
int64_t vocab_size;
int64_t num_positions;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
params["token_embedding.weight"] = ggml_new_tensor_2d(ctx, wtype, embed_dim, vocab_size);
params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, embed_dim, 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;
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);
}
public:
@ -591,11 +594,14 @@ protected:
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, ggml_type wtype) {
params["patch_embedding.weight"] = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, patch_size, patch_size, num_channels, embed_dim);
params["class_embedding"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, embed_dim);
params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, embed_dim, num_positions);
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);
params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
}
public:
@ -651,9 +657,10 @@ enum CLIPVersion {
class CLIPTextModel : public GGMLBlock {
protected:
void init_params(struct ggml_context* ctx, ggml_type wtype) {
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
if (version == OPEN_CLIP_VIT_BIGG_14) {
params["text_projection"] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, projection_dim, hidden_size);
enum ggml_type wtype = GGML_TYPE_F32; // tensor_types.find(prefix + "text_projection") != tensor_types.end() ? tensor_types[prefix + "text_projection"] : GGML_TYPE_F32;
params["text_projection"] = ggml_new_tensor_2d(ctx, wtype, projection_dim, hidden_size);
}
}
@ -798,9 +805,9 @@ protected:
int64_t out_features;
bool transpose_weight;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
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;
if (transpose_weight) {
LOG_ERROR("transpose_weight");
params["weight"] = ggml_new_tensor_2d(ctx, wtype, out_features, in_features);
} else {
params["weight"] = ggml_new_tensor_2d(ctx, wtype, in_features, out_features);
@ -861,12 +868,13 @@ struct CLIPTextModelRunner : public GGMLRunner {
CLIPTextModel model;
CLIPTextModelRunner(ggml_backend_t backend,
ggml_type wtype,
std::map<std::string, enum ggml_type>& 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, wtype), model(version, clip_skip_value, with_final_ln) {
model.init(params_ctx, wtype);
: GGMLRunner(backend), model(version, clip_skip_value, with_final_ln) {
model.init(params_ctx, tensor_types, prefix);
}
std::string get_desc() {
@ -908,13 +916,13 @@ struct CLIPTextModelRunner : public GGMLRunner {
struct ggml_tensor* embeddings = NULL;
if (num_custom_embeddings > 0 && custom_embeddings_data != NULL) {
auto token_embed_weight = model.get_token_embed_weight();
auto custom_embeddings = ggml_new_tensor_2d(compute_ctx,
wtype,
token_embed_weight->type,
model.hidden_size,
num_custom_embeddings);
set_backend_tensor_data(custom_embeddings, custom_embeddings_data);
auto token_embed_weight = model.get_token_embed_weight();
// concatenate custom embeddings
embeddings = ggml_concat(compute_ctx, token_embed_weight, custom_embeddings, 1);
}

12
common.hpp
View File

@ -182,9 +182,11 @@ protected:
int64_t dim_in;
int64_t dim_out;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
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;
params["proj.weight"] = ggml_new_tensor_2d(ctx, wtype, dim_in, dim_out * 2);
params["proj.bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, dim_out * 2);
params["proj.bias"] = ggml_new_tensor_1d(ctx, bias_wtype, dim_out * 2);
}
public:
@ -438,8 +440,10 @@ public:
class AlphaBlender : public GGMLBlock {
protected:
void init_params(struct ggml_context* ctx, ggml_type wtype) {
params["mix_factor"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& 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;
params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
}
float get_alpha() {

44
conditioner.hpp
View File

@ -46,7 +46,6 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
SDVersion version = VERSION_SD1;
PMVersion pm_version = PM_VERSION_1;
CLIPTokenizer tokenizer;
ggml_type wtype;
std::shared_ptr<CLIPTextModelRunner> text_model;
std::shared_ptr<CLIPTextModelRunner> text_model2;
@ -57,12 +56,12 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
std::vector<std::string> readed_embeddings;
FrozenCLIPEmbedderWithCustomWords(ggml_backend_t backend,
ggml_type wtype,
std::map<std::string, enum ggml_type>& 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(version == VERSION_SD2 ? 0 : 49407), embd_dir(embd_dir), wtype(wtype) {
: version(version), pm_version(pv), tokenizer(version == VERSION_SD2 ? 0 : 49407), embd_dir(embd_dir) {
if (clip_skip <= 0) {
clip_skip = 1;
if (version == VERSION_SD2 || version == VERSION_SDXL) {
@ -70,12 +69,12 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
}
}
if (version == VERSION_SD1) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, wtype, 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, clip_skip);
} else if (version == VERSION_SD2) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, wtype, 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, clip_skip);
} else if (version == VERSION_SDXL) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, wtype, OPENAI_CLIP_VIT_L_14, clip_skip, false);
text_model2 = std::make_shared<CLIPTextModelRunner>(backend, wtype, 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, 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);
}
}
@ -138,14 +137,14 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
LOG_DEBUG("embedding wrong hidden size, got %i, expected %i", tensor_storage.ne[0], hidden_size);
return false;
}
embd = ggml_new_tensor_2d(embd_ctx, wtype, hidden_size, tensor_storage.n_dims > 1 ? tensor_storage.ne[1] : 1);
embd = ggml_new_tensor_2d(embd_ctx, tensor_storage.type, hidden_size, tensor_storage.n_dims > 1 ? tensor_storage.ne[1] : 1);
*dst_tensor = embd;
return true;
};
model_loader.load_tensors(on_load, NULL);
readed_embeddings.push_back(embd_name);
token_embed_custom.resize(token_embed_custom.size() + ggml_nbytes(embd));
memcpy((void*)(token_embed_custom.data() + num_custom_embeddings * hidden_size * ggml_type_size(wtype)),
memcpy((void*)(token_embed_custom.data() + num_custom_embeddings * hidden_size * ggml_type_size(embd->type)),
embd->data,
ggml_nbytes(embd));
for (int i = 0; i < embd->ne[1]; i++) {
@ -590,9 +589,9 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
struct FrozenCLIPVisionEmbedder : public GGMLRunner {
CLIPVisionModelProjection vision_model;
FrozenCLIPVisionEmbedder(ggml_backend_t backend, ggml_type wtype)
: vision_model(OPEN_CLIP_VIT_H_14, true), GGMLRunner(backend, wtype) {
vision_model.init(params_ctx, wtype);
FrozenCLIPVisionEmbedder(ggml_backend_t backend, std::map<std::string, enum ggml_type>& tensor_types)
: vision_model(OPEN_CLIP_VIT_H_14, true), GGMLRunner(backend) {
vision_model.init(params_ctx, tensor_types, "cond_stage_model.transformer");
}
std::string get_desc() {
@ -627,7 +626,6 @@ struct FrozenCLIPVisionEmbedder : public GGMLRunner {
};
struct SD3CLIPEmbedder : public Conditioner {
ggml_type wtype;
CLIPTokenizer clip_l_tokenizer;
CLIPTokenizer clip_g_tokenizer;
T5UniGramTokenizer t5_tokenizer;
@ -636,15 +634,15 @@ struct SD3CLIPEmbedder : public Conditioner {
std::shared_ptr<T5Runner> t5;
SD3CLIPEmbedder(ggml_backend_t backend,
ggml_type wtype,
std::map<std::string, enum ggml_type>& tensor_types,
int clip_skip = -1)
: wtype(wtype), clip_g_tokenizer(0) {
: clip_g_tokenizer(0) {
if (clip_skip <= 0) {
clip_skip = 2;
}
clip_l = std::make_shared<CLIPTextModelRunner>(backend, wtype, OPENAI_CLIP_VIT_L_14, clip_skip, false);
clip_g = std::make_shared<CLIPTextModelRunner>(backend, wtype, OPEN_CLIP_VIT_BIGG_14, clip_skip, false);
t5 = std::make_shared<T5Runner>(backend, wtype);
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);
t5 = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
}
void set_clip_skip(int clip_skip) {
@ -974,21 +972,19 @@ struct SD3CLIPEmbedder : public Conditioner {
};
struct FluxCLIPEmbedder : public Conditioner {
ggml_type wtype;
CLIPTokenizer clip_l_tokenizer;
T5UniGramTokenizer t5_tokenizer;
std::shared_ptr<CLIPTextModelRunner> clip_l;
std::shared_ptr<T5Runner> t5;
FluxCLIPEmbedder(ggml_backend_t backend,
ggml_type wtype,
int clip_skip = -1)
: wtype(wtype) {
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, wtype, OPENAI_CLIP_VIT_L_14, clip_skip, true);
t5 = std::make_shared<T5Runner>(backend, wtype);
clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, clip_skip, true);
t5 = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
}
void set_clip_skip(int clip_skip) {

6
control.hpp
View File

@ -317,10 +317,10 @@ struct ControlNet : public GGMLRunner {
bool guided_hint_cached = false;
ControlNet(ggml_backend_t backend,
ggml_type wtype,
std::map<std::string, enum ggml_type>& tensor_types,
SDVersion version = VERSION_SD1)
: GGMLRunner(backend, wtype), control_net(version) {
control_net.init(params_ctx, wtype);
: GGMLRunner(backend), control_net(version) {
control_net.init(params_ctx, tensor_types, "");
}
~ControlNet() {

14
diffusion_model.hpp
View File

@ -31,10 +31,10 @@ struct UNetModel : public DiffusionModel {
UNetModelRunner unet;
UNetModel(ggml_backend_t backend,
ggml_type wtype,
std::map<std::string, enum ggml_type>& tensor_types,
SDVersion version = VERSION_SD1,
bool flash_attn = false)
: unet(backend, wtype, version, flash_attn) {
: unet(backend, tensor_types, "model.diffusion_model", version, flash_attn) {
}
void alloc_params_buffer() {
@ -83,9 +83,8 @@ struct MMDiTModel : public DiffusionModel {
MMDiTRunner mmdit;
MMDiTModel(ggml_backend_t backend,
ggml_type wtype,
SDVersion version = VERSION_SD3_2B)
: mmdit(backend, wtype, version) {
std::map<std::string, enum ggml_type>& tensor_types)
: mmdit(backend, tensor_types, "model.diffusion_model") {
}
void alloc_params_buffer() {
@ -133,10 +132,9 @@ struct FluxModel : public DiffusionModel {
Flux::FluxRunner flux;
FluxModel(ggml_backend_t backend,
ggml_type wtype,
SDVersion version = VERSION_FLUX_DEV,
std::map<std::string, enum ggml_type>& tensor_types,
bool flash_attn = false)
: flux(backend, wtype, version, flash_attn) {
: flux(backend, tensor_types, "model.diffusion_model", flash_attn) {
}
void alloc_params_buffer() {

7
esrgan.hpp
View File

@ -142,10 +142,9 @@ struct ESRGAN : public GGMLRunner {
int scale = 4;
int tile_size = 128; // avoid cuda OOM for 4gb VRAM
ESRGAN(ggml_backend_t backend,
ggml_type wtype)
: GGMLRunner(backend, wtype) {
rrdb_net.init(params_ctx, wtype);
ESRGAN(ggml_backend_t backend, std::map<std::string, enum ggml_type>& tensor_types)
: GGMLRunner(backend) {
rrdb_net.init(params_ctx, tensor_types, "");
}
std::string get_desc() {

3
examples/cli/main.cpp
View File

@ -1010,8 +1010,7 @@ int main(int argc, const char* argv[]) {
int upscale_factor = 4; // unused for RealESRGAN_x4plus_anime_6B.pth
if (params.esrgan_path.size() > 0 && params.upscale_repeats > 0) {
upscaler_ctx_t* upscaler_ctx = new_upscaler_ctx(params.esrgan_path.c_str(),
params.n_threads,
params.wtype);
params.n_threads);
if (upscaler_ctx == NULL) {
printf("new_upscaler_ctx failed\n");

51
flux.hpp
View File

@ -35,8 +35,9 @@ namespace Flux {
int64_t hidden_size;
float eps;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
params["scale"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
ggml_type wtype = GGML_TYPE_F32; //(tensor_types.find(prefix + "scale") != tensor_types.end()) ? tensor_types[prefix + "scale"] : GGML_TYPE_F32;
params["scale"] = ggml_new_tensor_1d(ctx, wtype, hidden_size);
}
public:
@ -823,25 +824,55 @@ namespace Flux {
};
struct FluxRunner : public GGMLRunner {
static std::map<std::string, enum ggml_type> empty_tensor_types;
public:
FluxParams flux_params;
Flux flux;
std::vector<float> pe_vec; // for cache
FluxRunner(ggml_backend_t backend,
ggml_type wtype,
SDVersion version = VERSION_FLUX_DEV,
std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types,
const std::string prefix = "",
bool flash_attn = false)
: GGMLRunner(backend, wtype) {
: GGMLRunner(backend) {
flux_params.flash_attn = flash_attn;
if (version == VERSION_FLUX_SCHNELL) {
flux_params.guidance_embed = false;
flux_params.depth = 0;
flux_params.depth_single_blocks = 0;
for (auto pair : tensor_types) {
std::string tensor_name = pair.first;
if (tensor_name.find("model.diffusion_model.") == std::string::npos)
continue;
if (tensor_name.find("guidance_in.in_layer.weight") != std::string::npos) {
// not schnell
flux_params.guidance_embed = true;
}
if (version == VERSION_FLUX_LITE) {
flux_params.depth = 8;
size_t db = tensor_name.find("double_blocks.");
if (db != std::string::npos) {
tensor_name = tensor_name.substr(db); // remove prefix
int block_depth = atoi(tensor_name.substr(14, tensor_name.find(".", 14)).c_str());
if (block_depth + 1 > flux_params.depth) {
flux_params.depth = block_depth + 1;
}
}
size_t sb = tensor_name.find("single_blocks.");
if (sb != std::string::npos) {
tensor_name = tensor_name.substr(sb); // remove prefix
int block_depth = atoi(tensor_name.substr(14, tensor_name.find(".", 14)).c_str());
if (block_depth + 1 > flux_params.depth_single_blocks) {
flux_params.depth_single_blocks = block_depth + 1;
}
}
}
LOG_INFO("Flux blocks: %d double, %d single", flux_params.depth, flux_params.depth_single_blocks);
if (!flux_params.guidance_embed) {
LOG_INFO("Flux guidance is disabled (Schnell mode)");
}
flux = Flux(flux_params);
flux.init(params_ctx, wtype);
flux.init(params_ctx, tensor_types, prefix);
}
std::string get_desc() {
@ -959,7 +990,7 @@ namespace Flux {
// ggml_backend_t backend = ggml_backend_cuda_init(0);
ggml_backend_t backend = ggml_backend_cpu_init();
ggml_type model_data_type = GGML_TYPE_Q8_0;
std::shared_ptr<FluxRunner> flux = std::shared_ptr<FluxRunner>(new FluxRunner(backend, model_data_type));
std::shared_ptr<FluxRunner> flux = std::shared_ptr<FluxRunner>(new FluxRunner(backend));
{
LOG_INFO("loading from '%s'", file_path.c_str());

65
ggml_extend.hpp
View File

@ -25,6 +25,8 @@
#include "ggml-cpu.h"
#include "ggml.h"
#include "model.h"
#ifdef SD_USE_CUBLAS
#include "ggml-cuda.h"
#endif
@ -964,7 +966,6 @@ protected:
std::map<struct ggml_tensor*, const void*> backend_tensor_data_map;
ggml_type wtype = GGML_TYPE_F32;
ggml_backend_t backend = NULL;
void alloc_params_ctx() {
@ -1040,8 +1041,8 @@ protected:
public:
virtual std::string get_desc() = 0;
GGMLRunner(ggml_backend_t backend, ggml_type wtype = GGML_TYPE_F32)
: backend(backend), wtype(wtype) {
GGMLRunner(ggml_backend_t backend)
: backend(backend) {
alloc_params_ctx();
}
@ -1170,20 +1171,22 @@ protected:
GGMLBlockMap blocks;
ParameterMap params;
void init_blocks(struct ggml_context* ctx, ggml_type wtype) {
void init_blocks(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
for (auto& pair : blocks) {
auto& block = pair.second;
block->init(ctx, wtype);
block->init(ctx, tensor_types, prefix + pair.first);
}
}
virtual void init_params(struct ggml_context* ctx, ggml_type wtype) {}
virtual void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {}
public:
void init(struct ggml_context* ctx, ggml_type wtype) {
init_blocks(ctx, wtype);
init_params(ctx, wtype);
void init(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
if (prefix.size() > 0) {
prefix = prefix + ".";
}
init_blocks(ctx, tensor_types, prefix);
init_params(ctx, tensor_types, prefix);
}
size_t get_params_num() {
@ -1239,13 +1242,15 @@ protected:
bool bias;
bool force_f32;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
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;
if (in_features % ggml_blck_size(wtype) != 0 || force_f32) {
wtype = GGML_TYPE_F32;
}
params["weight"] = ggml_new_tensor_2d(ctx, wtype, in_features, out_features);
if (bias) {
params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_features);
enum ggml_type wtype = GGML_TYPE_F32; //(tensor_types.ypes.find(prefix + "bias") != tensor_types.end()) ? tensor_types[prefix + "bias"] : GGML_TYPE_F32;
params["bias"] = ggml_new_tensor_1d(ctx, wtype, out_features);
}
}
@ -1273,8 +1278,8 @@ class Embedding : public UnaryBlock {
protected:
int64_t embedding_dim;
int64_t num_embeddings;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
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;
params["weight"] = ggml_new_tensor_2d(ctx, wtype, embedding_dim, num_embeddings);
}
@ -1313,10 +1318,12 @@ protected:
std::pair<int, int> dilation;
bool bias;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
params["weight"] = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kernel_size.second, kernel_size.first, in_channels, out_channels);
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
enum ggml_type wtype = GGML_TYPE_F16; //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F16;
params["weight"] = ggml_new_tensor_4d(ctx, wtype, kernel_size.second, kernel_size.first, in_channels, out_channels);
if (bias) {
params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
enum ggml_type wtype = GGML_TYPE_F32; // (tensor_types.find(prefix + "bias") != tensor_types.end()) ? tensor_types[prefix + "bias"] : GGML_TYPE_F32;
params["bias"] = ggml_new_tensor_1d(ctx, wtype, out_channels);
}
}
@ -1356,10 +1363,12 @@ protected:
int64_t dilation;
bool bias;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
params["weight"] = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 1, kernel_size, in_channels, out_channels); // 5d => 4d
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
enum ggml_type wtype = GGML_TYPE_F16; //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F16;
params["weight"] = ggml_new_tensor_4d(ctx, wtype, 1, kernel_size, in_channels, out_channels); // 5d => 4d
if (bias) {
params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
enum ggml_type wtype = GGML_TYPE_F32; //(tensor_types.find(prefix + "bias") != tensor_types.end()) ? tensor_types[prefix + "bias"] : GGML_TYPE_F32;
params["bias"] = ggml_new_tensor_1d(ctx, wtype, out_channels);
}
}
@ -1398,11 +1407,13 @@ protected:
bool elementwise_affine;
bool bias;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
if (elementwise_affine) {
params["weight"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, normalized_shape);
enum ggml_type wtype = GGML_TYPE_F32; //(tensor_types.ypes.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
params["weight"] = ggml_new_tensor_1d(ctx, wtype, normalized_shape);
if (bias) {
params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, normalized_shape);
enum ggml_type wtype = GGML_TYPE_F32; //(tensor_types.ypes.find(prefix + "bias") != tensor_types.end()) ? tensor_types[prefix + "bias"] : GGML_TYPE_F32;
params["bias"] = ggml_new_tensor_1d(ctx, wtype, normalized_shape);
}
}
}
@ -1438,10 +1449,12 @@ protected:
float eps;
bool affine;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
if (affine) {
params["weight"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_channels);
params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_channels);
enum ggml_type wtype = GGML_TYPE_F32; //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
enum ggml_type bias_wtype = GGML_TYPE_F32; //(tensor_types.find(prefix + "bias") != tensor_types.end()) ? tensor_types[prefix + "bias"] : GGML_TYPE_F32;
params["weight"] = ggml_new_tensor_1d(ctx, wtype, num_channels);
params["bias"] = ggml_new_tensor_1d(ctx, bias_wtype, num_channels);
}
}

5
lora.hpp
View File

@ -16,10 +16,9 @@ struct LoraModel : public GGMLRunner {
ggml_tensor* zero_index = NULL;
LoraModel(ggml_backend_t backend,
ggml_type wtype,
const std::string& file_path = "",
const std::string& prefix = "")
: file_path(file_path), GGMLRunner(backend, wtype) {
const std::string prefix = "")
: file_path(file_path), GGMLRunner(backend) {
if (!model_loader.init_from_file(file_path, prefix)) {
load_failed = true;
}

99
mmdit.hpp
View File

@ -147,8 +147,9 @@ protected:
int64_t hidden_size;
float eps;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
params["weight"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
enum ggml_type wtype = GGML_TYPE_F32; //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
params["weight"] = ggml_new_tensor_1d(ctx, wtype, hidden_size);
}
public:
@ -636,7 +637,6 @@ public:
struct MMDiT : public GGMLBlock {
// Diffusion model with a Transformer backbone.
protected:
SDVersion version = VERSION_SD3_2B;
int64_t input_size = -1;
int64_t patch_size = 2;
int64_t in_channels = 16;
@ -652,13 +652,13 @@ protected:
int64_t hidden_size;
std::string qk_norm;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
params["pos_embed"] = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hidden_size, num_patchs, 1);
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
enum ggml_type wtype = GGML_TYPE_F32; //(tensor_types.find(prefix + "pos_embed") != tensor_types.end()) ? tensor_types[prefix + "pos_embed"] : GGML_TYPE_F32;
params["pos_embed"] = ggml_new_tensor_3d(ctx, wtype, hidden_size, num_patchs, 1);
}
public:
MMDiT(SDVersion version = VERSION_SD3_2B)
: version(version) {
MMDiT(std::map<std::string, enum ggml_type>& tensor_types) {
// input_size is always None
// learn_sigma is always False
// register_length is alwalys 0
@ -670,48 +670,44 @@ public:
// pos_embed_scaling_factor is not used
// pos_embed_offset is not used
// context_embedder_config is always {'target': 'torch.nn.Linear', 'params': {'in_features': 4096, 'out_features': 1536}}
if (version == VERSION_SD3_2B) {
input_size = -1;
patch_size = 2;
in_channels = 16;
depth = 24;
mlp_ratio = 4.0f;
adm_in_channels = 2048;
out_channels = 16;
pos_embed_max_size = 192;
num_patchs = 36864; // 192 * 192
context_size = 4096;
context_embedder_out_dim = 1536;
} else if (version == VERSION_SD3_5_8B) {
input_size = -1;
patch_size = 2;
in_channels = 16;
depth = 38;
mlp_ratio = 4.0f;
adm_in_channels = 2048;
out_channels = 16;
pos_embed_max_size = 192;
num_patchs = 36864; // 192 * 192
context_size = 4096;
context_embedder_out_dim = 2432;
qk_norm = "rms";
} else if (version == VERSION_SD3_5_2B) {
input_size = -1;
patch_size = 2;
in_channels = 16;
depth = 24;
d_self = 12;
mlp_ratio = 4.0f;
adm_in_channels = 2048;
out_channels = 16;
pos_embed_max_size = 384;
num_patchs = 147456;
context_size = 4096;
context_embedder_out_dim = 1536;
// read tensors from tensor_types
for (auto pair : tensor_types) {
std::string tensor_name = pair.first;
if (tensor_name.find("model.diffusion_model.") == std::string::npos)
continue;
size_t jb = tensor_name.find("joint_blocks.");
if (jb != std::string::npos) {
tensor_name = tensor_name.substr(jb); // remove prefix
int block_depth = atoi(tensor_name.substr(13, tensor_name.find(".", 13)).c_str());
if (block_depth + 1 > depth) {
depth = block_depth + 1;
}
if (tensor_name.find("attn.ln") != std::string::npos) {
if (tensor_name.find(".bias") != std::string::npos) {
qk_norm = "ln";
} else {
qk_norm = "rms";
}
}
if (tensor_name.find("attn2") != std::string::npos) {
if (block_depth > d_self) {
d_self = block_depth;
}
}
}
}
if (d_self >= 0) {
pos_embed_max_size *= 2;
num_patchs *= 4;
}
LOG_INFO("MMDiT layers: %d (including %d MMDiT-x layers)", depth, d_self + 1);
int64_t default_out_channels = in_channels;
hidden_size = 64 * depth;
context_embedder_out_dim = 64 * depth;
int64_t num_heads = depth;
blocks["x_embedder"] = std::shared_ptr<GGMLBlock>(new PatchEmbed(input_size, patch_size, in_channels, hidden_size, true));
@ -870,15 +866,16 @@ public:
return x;
}
};
struct MMDiTRunner : public GGMLRunner {
MMDiT mmdit;
static std::map<std::string, enum ggml_type> empty_tensor_types;
MMDiTRunner(ggml_backend_t backend,
ggml_type wtype,
SDVersion version = VERSION_SD3_2B)
: GGMLRunner(backend, wtype), mmdit(version) {
mmdit.init(params_ctx, wtype);
std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types,
const std::string prefix = "")
: GGMLRunner(backend), mmdit(tensor_types) {
mmdit.init(params_ctx, tensor_types, prefix);
}
std::string get_desc() {
@ -975,7 +972,7 @@ struct MMDiTRunner : public GGMLRunner {
// ggml_backend_t backend = ggml_backend_cuda_init(0);
ggml_backend_t backend = ggml_backend_cpu_init();
ggml_type model_data_type = GGML_TYPE_F16;
std::shared_ptr<MMDiTRunner> mmdit = std::shared_ptr<MMDiTRunner>(new MMDiTRunner(backend, model_data_type));
std::shared_ptr<MMDiTRunner> mmdit = std::shared_ptr<MMDiTRunner>(new MMDiTRunner(backend));
{
LOG_INFO("loading from '%s'", file_path.c_str());

57
model.cpp
View File

@ -927,6 +927,7 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
GGML_ASSERT(ggml_nbytes(dummy) == tensor_storage.nbytes());
tensor_storages.push_back(tensor_storage);
tensor_storages_types[tensor_storage.name] = tensor_storage.type;
}
gguf_free(ctx_gguf_);
@ -1071,6 +1072,7 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
}
tensor_storages.push_back(tensor_storage);
tensor_storages_types[tensor_storage.name] = tensor_storage.type;
// LOG_DEBUG("%s %s", tensor_storage.to_string().c_str(), dtype.c_str());
}
@ -1296,7 +1298,7 @@ bool ModelLoader::parse_data_pkl(uint8_t* buffer,
zip_t* zip,
std::string dir,
size_t file_index,
const std::string& prefix) {
const std::string prefix) {
uint8_t* buffer_end = buffer + buffer_size;
if (buffer[0] == 0x80) { // proto
if (buffer[1] != 2) {
@ -1401,6 +1403,8 @@ bool ModelLoader::parse_data_pkl(uint8_t* buffer,
// printf(" ZIP got tensor %s \n ", reader.tensor_storage.name.c_str());
reader.tensor_storage.name = prefix + reader.tensor_storage.name;
tensor_storages.push_back(reader.tensor_storage);
tensor_storages_types[reader.tensor_storage.name] = reader.tensor_storage.type;
// LOG_DEBUG("%s", reader.tensor_storage.name.c_str());
// reset
reader = PickleTensorReader();
@ -1455,28 +1459,12 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
SDVersion ModelLoader::get_sd_version() {
TensorStorage token_embedding_weight;
bool is_flux = false;
bool is_schnell = true;
bool is_lite = true;
bool is_sd3 = false;
for (auto& tensor_storage : tensor_storages) {
if (tensor_storage.name.find("model.diffusion_model.guidance_in.in_layer.weight") != std::string::npos) {
is_schnell = false;
}
if (tensor_storage.name.find("model.diffusion_model.double_blocks.") != std::string::npos) {
is_flux = true;
return VERSION_FLUX;
}
if (tensor_storage.name.find("model.diffusion_model.double_blocks.8") != std::string::npos) {
is_lite = false;
}
if (tensor_storage.name.find("joint_blocks.0.x_block.attn2.ln_q.weight") != std::string::npos) {
return VERSION_SD3_5_2B;
}
if (tensor_storage.name.find("joint_blocks.37.x_block.attn.ln_q.weight") != std::string::npos) {
return VERSION_SD3_5_8B;
}
if (tensor_storage.name.find("model.diffusion_model.joint_blocks.23.") != std::string::npos) {
is_sd3 = true;
if (tensor_storage.name.find("model.diffusion_model.joint_blocks.") != std::string::npos) {
return VERSION_SD3;
}
if (tensor_storage.name.find("conditioner.embedders.1") != std::string::npos) {
return VERSION_SDXL;
@ -1498,19 +1486,7 @@ SDVersion ModelLoader::get_sd_version() {
// break;
}
}
if (is_flux) {
if (is_schnell) {
GGML_ASSERT(!is_lite);
return VERSION_FLUX_SCHNELL;
} else if (is_lite) {
return VERSION_FLUX_LITE;
} else {
return VERSION_FLUX_DEV;
}
}
if (is_sd3) {
return VERSION_SD3_2B;
}
if (token_embedding_weight.ne[0] == 768) {
return VERSION_SD1;
} else if (token_embedding_weight.ne[0] == 1024) {
@ -1603,6 +1579,21 @@ ggml_type ModelLoader::get_vae_wtype() {
return GGML_TYPE_COUNT;
}
void ModelLoader::set_wtype_override(ggml_type wtype, std::string prefix) {
for (auto& pair : tensor_storages_types) {
if (prefix.size() < 1 || pair.first.substr(0, prefix.size()) == prefix) {
for (auto& tensor_storage : tensor_storages) {
if (tensor_storage.name == pair.first) {
if (tensor_should_be_converted(tensor_storage, wtype)) {
pair.second = wtype;
}
break;
}
}
}
}
}
std::string ModelLoader::load_merges() {
std::string merges_utf8_str(reinterpret_cast<const char*>(merges_utf8_c_str), sizeof(merges_utf8_c_str));
return merges_utf8_str;

17
model.h
View File

@ -22,24 +22,20 @@ enum SDVersion {
VERSION_SD2,
VERSION_SDXL,
VERSION_SVD,
VERSION_SD3_2B,
VERSION_FLUX_DEV,
VERSION_FLUX_SCHNELL,
VERSION_SD3_5_8B,
VERSION_SD3_5_2B,
VERSION_FLUX_LITE,
VERSION_SD3,
VERSION_FLUX,
VERSION_COUNT,
};
static inline bool sd_version_is_flux(SDVersion version) {
if (version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL || version == VERSION_FLUX_LITE) {
if (version == VERSION_FLUX) {
return true;
}
return false;
}
static inline bool sd_version_is_sd3(SDVersion version) {
if (version == VERSION_SD3_2B || version == VERSION_SD3_5_8B || version == VERSION_SD3_5_2B) {
if (version == VERSION_SD3) {
return true;
}
return false;
@ -170,7 +166,7 @@ protected:
zip_t* zip,
std::string dir,
size_t file_index,
const std::string& prefix);
const std::string prefix);
bool init_from_gguf_file(const std::string& file_path, const std::string& prefix = "");
bool init_from_safetensors_file(const std::string& file_path, const std::string& prefix = "");
@ -178,12 +174,15 @@ protected:
bool init_from_diffusers_file(const std::string& file_path, const std::string& prefix = "");
public:
std::map<std::string, enum ggml_type> tensor_storages_types;
bool init_from_file(const std::string& file_path, const std::string& prefix = "");
SDVersion get_sd_version();
ggml_type get_sd_wtype();
ggml_type get_conditioner_wtype();
ggml_type get_diffusion_model_wtype();
ggml_type get_vae_wtype();
void set_wtype_override(ggml_type wtype, std::string prefix = "");
bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend_t backend);
bool load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
ggml_backend_t backend,

11
pmid.hpp
View File

@ -623,15 +623,15 @@ public:
std::vector<float> zeros_right;
public:
PhotoMakerIDEncoder(ggml_backend_t backend, ggml_type wtype, SDVersion version = VERSION_SDXL, PMVersion pm_v = PM_VERSION_1, float sty = 20.f)
: GGMLRunner(backend, wtype),
PhotoMakerIDEncoder(ggml_backend_t backend, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix, SDVersion version = VERSION_SDXL, PMVersion pm_v = PM_VERSION_1, float sty = 20.f)
: GGMLRunner(backend),
version(version),
pm_version(pm_v),
style_strength(sty) {
if (pm_version == PM_VERSION_1) {
id_encoder.init(params_ctx, wtype);
id_encoder.init(params_ctx, tensor_types, prefix);
} else if (pm_version == PM_VERSION_2) {
id_encoder2.init(params_ctx, wtype);
id_encoder2.init(params_ctx, tensor_types, prefix);
}
}
@ -780,11 +780,10 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
bool applied = false;
PhotoMakerIDEmbed(ggml_backend_t backend,
ggml_type wtype,
ModelLoader* ml,
const std::string& file_path = "",
const std::string& prefix = "")
: file_path(file_path), GGMLRunner(backend, wtype), model_loader(ml) {
: file_path(file_path), GGMLRunner(backend), model_loader(ml) {
if (!model_loader->init_from_file(file_path, prefix)) {
load_failed = true;
}

62
stable-diffusion.cpp
View File

@ -29,12 +29,8 @@ const char* model_version_to_str[] = {
"SD 2.x",
"SDXL",
"SVD",
"SD3 2B",
"Flux Dev",
"Flux Schnell",
"SD3.5 8B",
"SD3.5 2B",
"Flux Lite 8B"};
"SD3.x",
"Flux"};
const char* sampling_methods_str[] = {
"Euler A",
@ -264,16 +260,18 @@ public:
conditioner_wtype = wtype;
diffusion_model_wtype = wtype;
vae_wtype = wtype;
model_loader.set_wtype_override(wtype);
}
if (version == VERSION_SDXL) {
vae_wtype = GGML_TYPE_F32;
model_loader.set_wtype_override(GGML_TYPE_F32, "vae.");
}
LOG_INFO("Weight type: %s", ggml_type_name(model_wtype));
LOG_INFO("Conditioner weight type: %s", ggml_type_name(conditioner_wtype));
LOG_INFO("Diffusion model weight type: %s", ggml_type_name(diffusion_model_wtype));
LOG_INFO("VAE weight type: %s", ggml_type_name(vae_wtype));
LOG_INFO("Weight type: %s", model_wtype != SD_TYPE_COUNT ? ggml_type_name(model_wtype) : "??");
LOG_INFO("Conditioner weight type: %s", conditioner_wtype != SD_TYPE_COUNT ? ggml_type_name(conditioner_wtype) : "??");
LOG_INFO("Diffusion model weight type: %s", diffusion_model_wtype != SD_TYPE_COUNT ? ggml_type_name(diffusion_model_wtype) : "??");
LOG_INFO("VAE weight type: %s", vae_wtype != SD_TYPE_COUNT ? ggml_type_name(vae_wtype) : "??");
LOG_DEBUG("ggml tensor size = %d bytes", (int)sizeof(ggml_tensor));
@ -294,15 +292,15 @@ public:
}
if (version == VERSION_SVD) {
clip_vision = std::make_shared<FrozenCLIPVisionEmbedder>(backend, conditioner_wtype);
clip_vision = std::make_shared<FrozenCLIPVisionEmbedder>(backend, model_loader.tensor_storages_types);
clip_vision->alloc_params_buffer();
clip_vision->get_param_tensors(tensors);
diffusion_model = std::make_shared<UNetModel>(backend, diffusion_model_wtype, version);
diffusion_model = std::make_shared<UNetModel>(backend, model_loader.tensor_storages_types, version);
diffusion_model->alloc_params_buffer();
diffusion_model->get_param_tensors(tensors);
first_stage_model = std::make_shared<AutoEncoderKL>(backend, vae_wtype, vae_decode_only, true, version);
first_stage_model = std::make_shared<AutoEncoderKL>(backend, model_loader.tensor_storages_types, "first_stage_model", vae_decode_only, true, version);
LOG_DEBUG("vae_decode_only %d", vae_decode_only);
first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model");
@ -327,19 +325,20 @@ public:
if (diffusion_flash_attn) {
LOG_WARN("flash attention in this diffusion model is currently unsupported!");
}
cond_stage_model = std::make_shared<SD3CLIPEmbedder>(clip_backend, conditioner_wtype);
diffusion_model = std::make_shared<MMDiTModel>(backend, diffusion_model_wtype, version);
cond_stage_model = std::make_shared<SD3CLIPEmbedder>(clip_backend, model_loader.tensor_storages_types);
diffusion_model = std::make_shared<MMDiTModel>(backend, model_loader.tensor_storages_types);
} else if (sd_version_is_flux(version)) {
cond_stage_model = std::make_shared<FluxCLIPEmbedder>(clip_backend, conditioner_wtype);
diffusion_model = std::make_shared<FluxModel>(backend, diffusion_model_wtype, version, diffusion_flash_attn);
cond_stage_model = std::make_shared<FluxCLIPEmbedder>(clip_backend, model_loader.tensor_storages_types);
diffusion_model = std::make_shared<FluxModel>(backend, model_loader.tensor_storages_types, diffusion_flash_attn);
} else {
if (id_embeddings_path.find("v2") != std::string::npos) {
cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend, conditioner_wtype, embeddings_path, version, PM_VERSION_2);
cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend, model_loader.tensor_storages_types, embeddings_path, version, PM_VERSION_2);
} else {
cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend, conditioner_wtype, embeddings_path, version);
cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend, model_loader.tensor_storages_types, embeddings_path, version);
}
diffusion_model = std::make_shared<UNetModel>(backend, diffusion_model_wtype, version, diffusion_flash_attn);
diffusion_model = std::make_shared<UNetModel>(backend, model_loader.tensor_storages_types, version, diffusion_flash_attn);
}
cond_stage_model->alloc_params_buffer();
cond_stage_model->get_param_tensors(tensors);
@ -353,11 +352,11 @@ public:
} else {
vae_backend = backend;
}
first_stage_model = std::make_shared<AutoEncoderKL>(vae_backend, vae_wtype, vae_decode_only, false, version);
first_stage_model = std::make_shared<AutoEncoderKL>(vae_backend, model_loader.tensor_storages_types, "first_stage_model", vae_decode_only, false, version);
first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model");
} else {
tae_first_stage = std::make_shared<TinyAutoEncoder>(backend, vae_wtype, vae_decode_only);
tae_first_stage = std::make_shared<TinyAutoEncoder>(backend, model_loader.tensor_storages_types, "decoder.layers", vae_decode_only);
}
// first_stage_model->get_param_tensors(tensors, "first_stage_model.");
@ -369,17 +368,17 @@ public:
} else {
controlnet_backend = backend;
}
control_net = std::make_shared<ControlNet>(controlnet_backend, diffusion_model_wtype, version);
control_net = std::make_shared<ControlNet>(controlnet_backend, model_loader.tensor_storages_types, version);
}
if (id_embeddings_path.find("v2") != std::string::npos) {
pmid_model = std::make_shared<PhotoMakerIDEncoder>(backend, model_wtype, version, PM_VERSION_2);
pmid_model = std::make_shared<PhotoMakerIDEncoder>(backend, model_loader.tensor_storages_types, "pmid", version, PM_VERSION_2);
LOG_INFO("using PhotoMaker Version 2");
} else {
pmid_model = std::make_shared<PhotoMakerIDEncoder>(backend, model_wtype, version);
pmid_model = std::make_shared<PhotoMakerIDEncoder>(backend, model_loader.tensor_storages_types, "pmid", version);
}
if (id_embeddings_path.size() > 0) {
pmid_lora = std::make_shared<LoraModel>(backend, model_wtype, id_embeddings_path, "");
pmid_lora = std::make_shared<LoraModel>(backend, id_embeddings_path, "");
if (!pmid_lora->load_from_file(true)) {
LOG_WARN("load photomaker lora tensors from %s failed", id_embeddings_path.c_str());
return false;
@ -532,9 +531,12 @@ public:
denoiser = std::make_shared<DiscreteFlowDenoiser>();
} else if (sd_version_is_flux(version)) {
LOG_INFO("running in Flux FLOW mode");
float shift = 1.15f;
if (version == VERSION_FLUX_SCHNELL) {
shift = 1.0f; // TODO: validate
float shift = 1.0f; // TODO: validate
for (auto pair : model_loader.tensor_storages_types) {
if (pair.first.find("model.diffusion_model.guidance_in.in_layer.weight") != std::string::npos) {
shift = 1.15f;
break;
}
}
denoiser = std::make_shared<FluxFlowDenoiser>(shift);
} else if (is_using_v_parameterization) {
@ -633,7 +635,7 @@ public:
LOG_WARN("can not find %s or %s for lora %s", st_file_path.c_str(), ckpt_file_path.c_str(), lora_name.c_str());
return;
}
LoraModel lora(backend, model_wtype, file_path);
LoraModel lora(backend, file_path);
if (!lora.load_from_file()) {
LOG_WARN("load lora tensors from %s failed", file_path.c_str());
return;

3
stable-diffusion.h
View File

@ -215,8 +215,7 @@ SD_API sd_image_t* img2vid(sd_ctx_t* sd_ctx,
typedef struct upscaler_ctx_t upscaler_ctx_t;
SD_API upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path,
int n_threads,
enum sd_type_t wtype);
int n_threads);
SD_API void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx);
SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx, sd_image_t input_image, uint32_t upscale_factor);

21
t5.hpp
View File

@ -441,8 +441,9 @@ protected:
int64_t hidden_size;
float eps;
void init_params(struct ggml_context* ctx, ggml_type wtype) {
params["weight"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
enum ggml_type wtype = GGML_TYPE_F32; //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
params["weight"] = ggml_new_tensor_1d(ctx, wtype, hidden_size);
}
public:
@ -717,14 +718,15 @@ struct T5Runner : public GGMLRunner {
std::vector<int> relative_position_bucket_vec;
T5Runner(ggml_backend_t backend,
ggml_type wtype,
std::map<std::string, enum ggml_type>& tensor_types,
const std::string prefix,
int64_t num_layers = 24,
int64_t model_dim = 4096,
int64_t ff_dim = 10240,
int64_t num_heads = 64,
int64_t vocab_size = 32128)
: GGMLRunner(backend, wtype), model(num_layers, model_dim, ff_dim, num_heads, vocab_size) {
model.init(params_ctx, wtype);
: GGMLRunner(backend), model(num_layers, model_dim, ff_dim, num_heads, vocab_size) {
model.init(params_ctx, tensor_types, prefix);
}
std::string get_desc() {
@ -854,14 +856,17 @@ struct T5Embedder {
T5UniGramTokenizer tokenizer;
T5Runner model;
static std::map<std::string, enum ggml_type> empty_tensor_types;
T5Embedder(ggml_backend_t backend,
ggml_type wtype,
std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types,
const std::string prefix = "",
int64_t num_layers = 24,
int64_t model_dim = 4096,
int64_t ff_dim = 10240,
int64_t num_heads = 64,
int64_t vocab_size = 32128)
: model(backend, wtype, num_layers, model_dim, ff_dim, num_heads, vocab_size) {
: model(backend, tensor_types, prefix, num_layers, model_dim, ff_dim, num_heads, vocab_size) {
}
void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
@ -951,7 +956,7 @@ struct T5Embedder {
// ggml_backend_t backend = ggml_backend_cuda_init(0);
ggml_backend_t backend = ggml_backend_cpu_init();
ggml_type model_data_type = GGML_TYPE_F32;
std::shared_ptr<T5Embedder> t5 = std::shared_ptr<T5Embedder>(new T5Embedder(backend, model_data_type));
std::shared_ptr<T5Embedder> t5 = std::shared_ptr<T5Embedder>(new T5Embedder(backend));
{
LOG_INFO("loading from '%s'", file_path.c_str());

7
tae.hpp
View File

@ -188,12 +188,13 @@ struct TinyAutoEncoder : public GGMLRunner {
bool decode_only = false;
TinyAutoEncoder(ggml_backend_t backend,
ggml_type wtype,
std::map<std::string, enum ggml_type>& tensor_types,
const std::string prefix,
bool decoder_only = true)
: decode_only(decoder_only),
taesd(decode_only),
GGMLRunner(backend, wtype) {
taesd.init(params_ctx, wtype);
GGMLRunner(backend) {
taesd.init(params_ctx, tensor_types, prefix);
}
std::string get_desc() {

7
unet.hpp
View File

@ -532,11 +532,12 @@ struct UNetModelRunner : public GGMLRunner {
UnetModelBlock unet;
UNetModelRunner(ggml_backend_t backend,
ggml_type wtype,
std::map<std::string, enum ggml_type>& tensor_types,
const std::string prefix,
SDVersion version = VERSION_SD1,
bool flash_attn = false)
: GGMLRunner(backend, wtype), unet(version, flash_attn) {
unet.init(params_ctx, wtype);
: GGMLRunner(backend), unet(version, flash_attn) {
unet.init(params_ctx, tensor_types, prefix);
}
std::string get_desc() {

11
upscaler.cpp
View File

@ -32,13 +32,17 @@ struct UpscalerGGML {
LOG_DEBUG("Using SYCL backend");
backend = ggml_backend_sycl_init(0);
#endif
ModelLoader model_loader;
if (!model_loader.init_from_file(esrgan_path)) {
LOG_ERROR("init model loader from file failed: '%s'", esrgan_path.c_str());
}
model_loader.set_wtype_override(model_data_type);
if (!backend) {
LOG_DEBUG("Using CPU backend");
backend = ggml_backend_cpu_init();
}
LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
esrgan_upscaler = std::make_shared<ESRGAN>(backend, model_data_type);
esrgan_upscaler = std::make_shared<ESRGAN>(backend, model_loader.tensor_storages_types);
if (!esrgan_upscaler->load_from_file(esrgan_path)) {
return false;
}
@ -96,8 +100,7 @@ struct upscaler_ctx_t {
};
upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path_c_str,
int n_threads,
enum sd_type_t wtype) {
int n_threads) {
upscaler_ctx_t* upscaler_ctx = (upscaler_ctx_t*)malloc(sizeof(upscaler_ctx_t));
if (upscaler_ctx == NULL) {
return NULL;

12
vae.hpp
View File

@ -163,8 +163,9 @@ public:
class VideoResnetBlock : public ResnetBlock {
protected:
void init_params(struct ggml_context* ctx, ggml_type wtype) {
params["mix_factor"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
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 + "mix_factor") != tensor_types.end()) ? tensor_types[prefix + "mix_factor"] : GGML_TYPE_F32;
params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
}
float get_alpha() {
@ -524,12 +525,13 @@ struct AutoEncoderKL : public GGMLRunner {
AutoencodingEngine ae;
AutoEncoderKL(ggml_backend_t backend,
ggml_type wtype,
std::map<std::string, enum ggml_type>& tensor_types,
const std::string prefix,
bool decode_only = false,
bool use_video_decoder = false,
SDVersion version = VERSION_SD1)
: decode_only(decode_only), ae(decode_only, use_video_decoder, version), GGMLRunner(backend, wtype) {
ae.init(params_ctx, wtype);
: decode_only(decode_only), ae(decode_only, use_video_decoder, version), GGMLRunner(backend) {
ae.init(params_ctx, tensor_types, prefix);
}
std::string get_desc() {