Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

sd: sync to master-582-7023fc4 (#2150)

Browse Source 
* sd: remove sampler alias handling from the C++ layer

It's already handled at the Python layer.

* sd: sync to master-580-7d33d4b

* sd: sync to master-582-7023fc4
This commit is contained in:
Wagner Bruna
2026-04-21 12:01:33 -03:00
committed by GitHub
parent c818716f57
commit bad9b61064
48 changed files with 3376 additions and 2093 deletions
Download Patch File Download Diff File Expand all files Collapse all files
otherarch/sdcpp/stable-diffusion.cpp
+50 -6
View File
@@ -32,7 +32,8 @@ const char* model_version_to_str[] = {
"SD 2.x",
"SD 2.x Inpaint",
"SD 2.x Tiny UNet",
"SDXS",
"SDXS (512-DS)",
"SDXS (09)",
"SDXL",
"SDXL Inpaint",
"SDXL Instruct-Pix2Pix",
@@ -54,6 +55,7 @@ const char* model_version_to_str[] = {
"Flux.2 klein",
"Z-Image",
"Ovis Image",
"Ernie Image",
};
const char* sampling_methods_str[] = {
@@ -71,6 +73,7 @@ const char* sampling_methods_str[] = {
"TCD",
"Res Multistep",
"Res 2s",
"ER-SDE",
};
/*================================================== Helper Functions ================================================*/
@@ -366,7 +369,8 @@ public:
bool isflux2 = sd_version_is_flux2(tempver);
bool is_ovis = (tempver==VERSION_OVIS_IMAGE);
bool is_anima = sd_version_is_anima(tempver);
bool conditioner_is_llm = (is_qwenimg || iszimg || isflux2 || is_ovis || is_anima);
bool is_ernie = sd_version_is_ernie_image(tempver);
bool conditioner_is_llm = (is_qwenimg || iszimg || isflux2 || is_ovis || is_anima || is_ernie);
//kcpp qol fallback: if a llm was loaded as t5 by mistake
if(conditioner_is_llm && t5_path_fixed!="")
@@ -457,7 +461,7 @@ public:
{
to_replace = "taesd_3.embd";
}
else if(sd_version_is_flux2(tempver))
else if(sd_version_uses_flux2_vae(tempver))
{
to_replace = "taesd_f2.embd";
}
@@ -634,7 +638,7 @@ public:
}
bool tae_preview_only = sd_ctx_params->tae_preview_only;
if (version == VERSION_SDXS) {
if (version == VERSION_SDXS_512_DS || version == VERSION_SDXS_09) {
tae_preview_only = false;
use_tae = true;
}
@@ -772,6 +776,15 @@ public:
tensor_storage_map,
"model.diffusion_model",
version);
} else if (sd_version_is_ernie_image(version)) {
cond_stage_model = std::make_shared<LLMEmbedder>(clip_backend,
offload_params_to_cpu,
tensor_storage_map,
version);
diffusion_model = std::make_shared<ErnieImageModel>(backend,
offload_params_to_cpu,
tensor_storage_map,
"model.diffusion_model");
} else { // SD1.x SD2.x SDXL
std::map<std::string, std::string> embbeding_map;
for (uint32_t i = 0; i < sd_ctx_params->embedding_count; i++) {
@@ -1047,6 +1060,10 @@ public:
if (version == VERSION_SVD) {
ignore_tensors.insert("conditioner.embedders.3");
}
if (sd_version_is_ernie_image(version)) {
ignore_tensors.insert("text_encoders.llm.vision_tower.");
ignore_tensors.insert("text_encoders.llm.multi_modal_projector.");
}
bool success = model_loader.load_tensors(tensors, ignore_tensors, n_threads, sd_ctx_params->enable_mmap);
if (!success) {
LOG_ERROR("load tensors from model loader failed");
@@ -1150,10 +1167,13 @@ public:
sd_version_is_wan(version) ||
sd_version_is_qwen_image(version) ||
sd_version_is_anima(version) ||
sd_version_is_ernie_image(version) ||
sd_version_is_z_image(version)) {
pred_type = FLOW_PRED;
if (sd_version_is_wan(version)) {
default_flow_shift = 5.f;
} else if (sd_version_is_ernie_image(version)) {
default_flow_shift = 4.f;
} else {
default_flow_shift = 3.f;
}
@@ -1644,7 +1664,7 @@ public:
uint32_t dim = is_video ? static_cast<uint32_t>(latents.shape()[3]) : static_cast<uint32_t>(latents.shape()[2]);
if (dim == 128) {
if (sd_version_is_flux2(version)) {
if (sd_version_uses_flux2_vae(version)) {
latent_rgb_proj = flux2_latent_rgb_proj;
latent_rgb_bias = flux2_latent_rgb_bias;
patch_sz = 2;
@@ -2093,7 +2113,7 @@ public:
latent_channel = 48;
} else if (version == VERSION_CHROMA_RADIANCE) {
latent_channel = 3;
} else if (sd_version_is_flux2(version)) {
} else if (sd_version_uses_flux2_vae(version)) {
latent_channel = 128;
} else {
latent_channel = 16;
@@ -2241,6 +2261,7 @@ const char* sample_method_to_str[] = {
"tcd",
"res_multistep",
"res_2s",
"er_sde",
};
const char* sd_sample_method_name(enum sample_method_t sample_method) {
@@ -2636,6 +2657,14 @@ struct sd_ctx_t {
StableDiffusionGGML* sd = nullptr;
};
static bool sd_version_supports_video_generation(SDVersion version) {
return version == VERSION_SVD || sd_version_is_wan(version);
}
static bool sd_version_supports_image_generation(SDVersion version) {
return !sd_version_supports_video_generation(version);
}
sd_ctx_t* new_sd_ctx(const sd_ctx_params_t* sd_ctx_params) {
sd_ctx_t* sd_ctx = (sd_ctx_t*)malloc(sizeof(sd_ctx_t));
if (sd_ctx == nullptr) {
@@ -2665,6 +2694,20 @@ void free_sd_ctx(sd_ctx_t* sd_ctx) {
free(sd_ctx);
}
SD_API bool sd_ctx_supports_image_generation(const sd_ctx_t* sd_ctx) {
if (sd_ctx == nullptr || sd_ctx->sd == nullptr) {
return false;
}
return sd_version_supports_image_generation(sd_ctx->sd->version);
}
SD_API bool sd_ctx_supports_video_generation(const sd_ctx_t* sd_ctx) {
if (sd_ctx == nullptr || sd_ctx->sd == nullptr) {
return false;
}
return sd_version_supports_video_generation(sd_ctx->sd->version);
}
enum sample_method_t sd_get_default_sample_method(const sd_ctx_t* sd_ctx) {
if (sd_ctx != nullptr && sd_ctx->sd != nullptr) {
if (sd_version_is_dit(sd_ctx->sd->version)) {
@@ -2723,6 +2766,7 @@ static float resolve_eta(sd_ctx_t* sd_ctx,
return 0.0f;
case EULER_A_SAMPLE_METHOD:
case DPMPP2S_A_SAMPLE_METHOD:
case ER_SDE_SAMPLE_METHOD:
return 1.0f;
default:;
}