ci: skip docker image build job on pull requests (#1439)

.github/workflows/build.yml

@@ -21,6 +21,7 @@ on:
         "**/*.c",
         "**/*.cpp",
         "**/*.cu",
+        "examples/server/frontend",
         "examples/server/frontend/**",
       ]
   pull_request:
@@ -35,6 +36,7 @@ on:
         "**/*.c",
         "**/*.cpp",
         "**/*.cu",
+        "examples/server/frontend",
         "examples/server/frontend/**",
       ]
 
@@ -174,6 +176,7 @@ jobs:
 
   build-and-push-docker-images:
     name: Build and push container images
+    if: ${{ github.event_name != 'pull_request' }}
     runs-on: ubuntu-latest
 
     permissions:

.gitmodules

@@ -3,7 +3,7 @@
 	url = https://github.com/ggml-org/ggml.git
 [submodule "examples/server/frontend"]
 	path = examples/server/frontend
-	url = https://github.com/leejet/stable-ui.git
+	url = https://github.com/leejet/sdcpp-webui.git
 [submodule "thirdparty/libwebp"]
 	path = thirdparty/libwebp
 	url = https://github.com/webmproject/libwebp.git

CMakeLists.txt

@@ -156,8 +156,10 @@ file(GLOB SD_LIB_SOURCES
     "src/*.h"
     "src/*.cpp"
     "src/*.hpp"
-    "src/vocab/*.h"
-    "src/vocab/*.cpp"
+    "src/tokenizers/*.h"
+    "src/tokenizers/*.cpp"
+    "src/tokenizers/vocab/*.h"
+    "src/tokenizers/vocab/*.cpp"
 )
 
 find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
@@ -250,7 +252,7 @@ endif()
 add_subdirectory(thirdparty)
 
 target_link_libraries(${SD_LIB} PUBLIC ggml zip)
-target_include_directories(${SD_LIB} PUBLIC . include)
+target_include_directories(${SD_LIB} PUBLIC . src include)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
 target_compile_features(${SD_LIB} PUBLIC c_std_11 cxx_std_17)
 

README.md

@@ -15,6 +15,9 @@ API and command-line option may change frequently.***
 
 ## 🔥Important News
 
+* **2026/04/11** 🚀 stable-diffusion.cpp now uses a brand-new embedded web UI.  
+  👉 Details: [PR #1408](https://github.com/leejet/stable-diffusion.cpp/pull/1408)
+
 * **2026/01/18** 🚀 stable-diffusion.cpp now supports **FLUX.2-klein**  
   👉 Details: [PR #1193](https://github.com/leejet/stable-diffusion.cpp/pull/1193)
 
@@ -54,6 +57,7 @@ API and command-line option may change frequently.***
     - [Z-Image](./docs/z_image.md)
     - [Ovis-Image](./docs/ovis_image.md)
     - [Anima](./docs/anima.md)
+    - [ERNIE-Image](./docs/ernie_image.md)
   - Image Edit Models
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)
@@ -93,6 +97,7 @@ API and command-line option may change frequently.***
     - `DPM++ 2M`
     - [`DPM++ 2M v2`](https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8457)
     - `DPM++ 2S a`
+    - `ER-SDE`
     - [`LCM`](https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/13952)
 - Cross-platform reproducibility
     - `--rng cuda`, default, consistent with the `stable-diffusion-webui GPU RNG`
@@ -141,6 +146,7 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 - [🔥Z-Image](./docs/z_image.md)
 - [Ovis-Image](./docs/ovis_image.md)
 - [Anima](./docs/anima.md)
+- [ERNIE-Image](./docs/ernie_image.md)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
 - [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)

docs/distilled_sd.md

@@ -87,51 +87,32 @@ pipe.save_pretrained("segmindtiny-sd", safe_serialization=True)
 ```bash
 python convert_diffusers_to_original_stable_diffusion.py \
       --model_path  ./segmindtiny-sd \
-      --checkpoint_path ./segmind_tiny-sd.ckpt --half
+      --checkpoint_path ./segmind_tiny-sd.safetensors  --half --use_safetensors
 ```
 
-The file segmind_tiny-sd.ckpt will be generated and is now ready for use with sd.cpp. You can follow a similar process for the other models mentioned above.
+The file segmind_tiny-sd.safetensors will be generated and is now ready for use with sd.cpp. You can follow a similar process for the other models mentioned above.
 
 
-##### Another available .ckpt file:
-
- * https://huggingface.co/ClashSAN/small-sd/resolve/main/tinySDdistilled.ckpt
-
-To use this file, you must first adjust its non-contiguous tensors:
-
-```python
-import torch
-ckpt = torch.load("tinySDdistilled.ckpt", map_location=torch.device('cpu'))
-for key, value in ckpt['state_dict'].items():
-    if isinstance(value, torch.Tensor):
-        ckpt['state_dict'][key] = value.contiguous()
-torch.save(ckpt, "tinySDdistilled_fixed.ckpt")
-```
-
-
-### SDXS-512
+### SDXS-512-DreamShaper
 
 Another very tiny and **incredibly fast**  model is SDXS by IDKiro et al.  The authors refer to it as *"Real-Time One-Step Latent Diffusion Models with Image Conditions"*. For details read the paper: https://arxiv.org/pdf/2403.16627 . Once again the authors removed some more blocks of U-Net part and unlike other SD1 models they use an adjusted _AutoEncoderTiny_ instead of default _AutoEncoderKL_ for the VAE part.
+##### Some ready-to-run SDXS-512 model files are available online, such as:
 
-##### 1. Download the diffusers model from  Hugging Face using Python:
-
-```python
-from diffusers import StableDiffusionPipeline
-pipe = StableDiffusionPipeline.from_pretrained("IDKiro/sdxs-512-dreamshaper")
-pipe.save_pretrained(save_directory="sdxs")
-```
-##### 2. Create a safetensors file
-
-```bash
-python convert_diffusers_to_original_stable_diffusion.py \
-    --model_path  sdxs  --checkpoint_path sdxs.safetensors --half --use_safetensors
-```
-
-##### 3. Run the model as follows:
+* https://huggingface.co/akleine/sdxs-512
+* https://huggingface.co/concedo/sdxs-512-tinySDdistilled-GGUF
 
+##### Run the model as follows:
 ```bash
 ~/stable-diffusion.cpp/build/bin/sd-cli -m sdxs.safetensors -p "portrait of a lovely cat" \
   --cfg-scale 1 --steps 1
 ```
+Both options: ``` --cfg-scale 1 ``` and  ``` --steps 1 ``` are mandatory here.
 
-Both options: ``` --cfg-scale 1 ``` and  ``` --steps 1 ``` are mandatory here.                                                 
+### SDXS-512-0.9
+
+Even though the name "SDXS-512-0.9" is similar to "SDXS-512-DreamShaper", it is *completely different* but also **incredibly fast**. Sometimes it is preferred, so try it yourself.
+##### Download a ready-to-run file from here:
+
+* https://huggingface.co/akleine/sdxs-09
+
+For the use of this model, both options ``` --cfg-scale 1 ``` and ``` --steps 1 ``` are again absolutely necessary.

docs/ernie_image.md

@@ -0,0 +1,35 @@
+# How to Use
+
+You can run ERNIE-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or even less.
+
+## Download weights
+
+- Download ERNIE-Image-Turbo
+    - safetensors: https://huggingface.co/Comfy-Org/ERNIE-Image/tree/main/diffusion_models
+    - gguf: https://huggingface.co/unsloth/ERNIE-Image-Turbo-GGUF/tree/main
+- Download ERNIE-Image
+    - safetensors: https://huggingface.co/Comfy-Org/ERNIE-Image/tree/main/diffusion_models
+    - gguf: https://huggingface.co/unsloth/ERNIE-Image-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/Comfy-Org/ERNIE-Image/tree/main/vae
+- Download ministral 3b
+    - safetensors: https://huggingface.co/Comfy-Org/ERNIE-Image/tree/main/text_encoders
+    - gguf: https://huggingface.co/unsloth/Ministral-3-3B-Instruct-2512-GGUF/tree/main
+
+## Examples
+
+### ERNIE-Image-Turbo
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ernie-image-turbo.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 8 -v --offload-to-cpu --diffusion-fa
+```
+
+<img width="256" alt="ERNIE-Image Turbo example" src="../assets/ernie_image/turbo_example.png" />
+
+### ERNIE-Image
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ernie-image-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\ministral-3-3b.safetensors -p "a lovely cat" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa
+```
+
+<img width="256" alt="ERNIE-Image example" src="../assets/ernie_image/example.png" />

examples/cli/README.md

@@ -114,7 +114,7 @@ Generation Options:
                                            medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a and dpm++2s_a)
+  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
   --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
@@ -122,7 +122,7 @@ Generation Options:
   --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a and dpm++2s_a)
+  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
   --pm-style-strength <float>
   --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
@@ -133,10 +133,10 @@ Generation Options:
   --disable-image-metadata                 do not embed generation metadata on image files
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
   --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           tcd, res_multistep, res_2s] (default: euler for Flux/SD3/Wan, euler_a
+                                           tcd, res_multistep, res_2s, er_sde] (default: euler for Flux/SD3/Wan, euler_a
                                            otherwise)
   --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
-                                           ddim_trailing, tcd, res_multistep, res_2s] default: euler for Flux/SD3/Wan,
+                                           ddim_trailing, tcd, res_multistep, res_2s, er_sde] default: euler for Flux/SD3/Wan,
                                            euler_a otherwise
   --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
                                            kl_optimal, lcm, bong_tangent], default: discrete

examples/common/common.cpp

@@ -855,7 +855,7 @@ ArgOptions SDGenerationParams::get_options() {
          &sample_params.guidance.slg.layer_end},
         {"",
          "--eta",
-         "noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a and dpm++2s_a)",
+         "noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)",
          &sample_params.eta},
         {"",
          "--flow-shift",
@@ -887,7 +887,7 @@ ArgOptions SDGenerationParams::get_options() {
          &high_noise_sample_params.guidance.slg.layer_end},
         {"",
          "--high-noise-eta",
-         "(high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a and dpm++2s_a)",
+         "(high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)",
          &high_noise_sample_params.eta},
         {"",
          "--strength",
@@ -1185,12 +1185,12 @@ ArgOptions SDGenerationParams::get_options() {
          on_seed_arg},
         {"",
          "--sampling-method",
-         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s] "
+         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde] "
          "(default: euler for Flux/SD3/Wan, euler_a otherwise)",
          on_sample_method_arg},
         {"",
          "--high-noise-sampling-method",
-         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s]"
+         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde]"
          " default: euler for Flux/SD3/Wan, euler_a otherwise",
          on_high_noise_sample_method_arg},
         {"",
@@ -1589,10 +1589,18 @@ bool SDGenerationParams::from_json_str(
         LOG_ERROR("invalid init_image");
         return false;
     }
+    if (!parse_image_json_field(j, "end_image", 3, width, height, end_image)) {
+        LOG_ERROR("invalid end_image");
+        return false;
+    }
     if (!parse_image_array_json_field(j, "ref_images", 3, width, height, ref_images)) {
         LOG_ERROR("invalid ref_images");
         return false;
     }
+    if (!parse_image_array_json_field(j, "control_frames", 3, width, height, control_frames)) {
+        LOG_ERROR("invalid control_frames");
+        return false;
+    }
     if (!parse_image_json_field(j, "mask_image", 1, width, height, mask_image)) {
         LOG_ERROR("invalid mask_image");
         return false;

examples/common/media_io.cpp

@@ -95,6 +95,57 @@ using WebPMuxPtr         = std::unique_ptr<WebPMux, WebPMuxDeleter>;
 using WebPAnimEncoderPtr = std::unique_ptr<WebPAnimEncoder, WebPAnimEncoderDeleter>;
 #endif
 
+#ifdef SD_USE_WEBM
+class MemoryMkvWriter : public mkvmuxer::IMkvWriter {
+public:
+    mkvmuxer::int32 Write(const void* buf, mkvmuxer::uint32 len) override {
+        if (buf == nullptr && len > 0) {
+            return -1;
+        }
+        const size_t end_pos = position_ + static_cast<size_t>(len);
+        if (end_pos > data_.size()) {
+            data_.resize(end_pos);
+        }
+        if (len > 0) {
+            memcpy(data_.data() + position_, buf, len);
+        }
+        position_ = end_pos;
+        return 0;
+    }
+
+    mkvmuxer::int64 Position() const override {
+        return static_cast<mkvmuxer::int64>(position_);
+    }
+
+    mkvmuxer::int32 Position(mkvmuxer::int64 position) override {
+        if (position < 0) {
+            return -1;
+        }
+        const size_t target = static_cast<size_t>(position);
+        if (target > data_.size()) {
+            data_.resize(target);
+        }
+        position_ = target;
+        return 0;
+    }
+
+    bool Seekable() const override {
+        return true;
+    }
+
+    void ElementStartNotify(mkvmuxer::uint64, mkvmuxer::int64) override {
+    }
+
+    const std::vector<uint8_t>& data() const {
+        return data_;
+    }
+
+private:
+    std::vector<uint8_t> data_;
+    size_t position_ = 0;
+};
+#endif
+
 bool read_binary_file_bytes(const char* path, std::vector<uint8_t>& data) {
     std::ifstream fin(fs::path(path), std::ios::binary);
     if (!fin) {
@@ -570,6 +621,32 @@ void write_u16_le(FILE* f, uint16_t val) {
     fwrite(&val, 2, 1, f);
 }
 
+void write_u32_le(std::vector<uint8_t>& data, uint32_t val) {
+    data.push_back(static_cast<uint8_t>(val & 0xFF));
+    data.push_back(static_cast<uint8_t>((val >> 8) & 0xFF));
+    data.push_back(static_cast<uint8_t>((val >> 16) & 0xFF));
+    data.push_back(static_cast<uint8_t>((val >> 24) & 0xFF));
+}
+
+void write_u16_le(std::vector<uint8_t>& data, uint16_t val) {
+    data.push_back(static_cast<uint8_t>(val & 0xFF));
+    data.push_back(static_cast<uint8_t>((val >> 8) & 0xFF));
+}
+
+void patch_u32_le(std::vector<uint8_t>& data, size_t offset, uint32_t val) {
+    if (offset + 4 > data.size()) {
+        return;
+    }
+    data[offset + 0] = static_cast<uint8_t>(val & 0xFF);
+    data[offset + 1] = static_cast<uint8_t>((val >> 8) & 0xFF);
+    data[offset + 2] = static_cast<uint8_t>((val >> 16) & 0xFF);
+    data[offset + 3] = static_cast<uint8_t>((val >> 24) & 0xFF);
+}
+
+void write_fourcc(std::vector<uint8_t>& data, const char* fourcc) {
+    data.insert(data.end(), fourcc, fourcc + 4);
+}
+
 EncodedImageFormat encoded_image_format_from_path(const std::string& path) {
     std::string ext = fs::path(path).extension().string();
     std::transform(ext.begin(), ext.end(), ext.begin(), ::tolower);
@@ -699,95 +776,96 @@ uint8_t* load_image_from_memory(const char* image_bytes,
     return load_image_common(true, image_bytes, len, width, height, expected_width, expected_height, expected_channel);
 }
 
-int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
+std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images, int num_images, int fps, int quality) {
     if (num_images == 0) {
         fprintf(stderr, "Error: Image array is empty.\n");
-        return -1;
+        return {};
     }
 
-    FilePtr file(fopen(filename, "wb"));
-    if (!file) {
-        perror("Error opening file for writing");
-        return -1;
-    }
-    FILE* f = file.get();
-
     uint32_t width    = images[0].width;
     uint32_t height   = images[0].height;
     uint32_t channels = images[0].channel;
     if (channels != 3 && channels != 4) {
         fprintf(stderr, "Error: Unsupported channel count: %u\n", channels);
-        return -1;
+        return {};
     }
 
-    fwrite("RIFF", 4, 1, f);
-    long riff_size_pos = ftell(f);
-    write_u32_le(f, 0);
-    fwrite("AVI ", 4, 1, f);
+    // stb_image_write changes JPEG sampling behavior above quality 90.
+    // MJPG AVI playback is more compatible when we keep the encoder on the
+    // <= 90 path.
+    const int mjpg_quality = std::clamp(quality, 1, 90);
 
-    fwrite("LIST", 4, 1, f);
-    write_u32_le(f, 4 + 8 + 56 + 8 + 4 + 8 + 56 + 8 + 40);
-    fwrite("hdrl", 4, 1, f);
+    std::vector<uint8_t> avi_data;
+    avi_data.reserve(static_cast<size_t>(num_images) * 1024);
 
-    fwrite("avih", 4, 1, f);
-    write_u32_le(f, 56);
-    write_u32_le(f, 1000000 / fps);
-    write_u32_le(f, 0);
-    write_u32_le(f, 0);
-    write_u32_le(f, 0x110);
-    write_u32_le(f, num_images);
-    write_u32_le(f, 0);
-    write_u32_le(f, 1);
-    write_u32_le(f, width * height * 3);
-    write_u32_le(f, width);
-    write_u32_le(f, height);
-    write_u32_le(f, 0);
-    write_u32_le(f, 0);
-    write_u32_le(f, 0);
-    write_u32_le(f, 0);
+    write_fourcc(avi_data, "RIFF");
+    const size_t riff_size_pos = avi_data.size();
+    write_u32_le(avi_data, 0);
+    write_fourcc(avi_data, "AVI ");
 
-    fwrite("LIST", 4, 1, f);
-    write_u32_le(f, 4 + 8 + 56 + 8 + 40);
-    fwrite("strl", 4, 1, f);
+    write_fourcc(avi_data, "LIST");
+    write_u32_le(avi_data, 4 + 8 + 56 + 8 + 4 + 8 + 56 + 8 + 40);
+    write_fourcc(avi_data, "hdrl");
 
-    fwrite("strh", 4, 1, f);
-    write_u32_le(f, 56);
-    fwrite("vids", 4, 1, f);
-    fwrite("MJPG", 4, 1, f);
-    write_u32_le(f, 0);
-    write_u16_le(f, 0);
-    write_u16_le(f, 0);
-    write_u32_le(f, 0);
-    write_u32_le(f, 1);
-    write_u32_le(f, fps);
-    write_u32_le(f, 0);
-    write_u32_le(f, num_images);
-    write_u32_le(f, width * height * 3);
-    write_u32_le(f, (uint32_t)-1);
-    write_u32_le(f, 0);
-    write_u16_le(f, 0);
-    write_u16_le(f, 0);
-    write_u16_le(f, 0);
-    write_u16_le(f, 0);
+    write_fourcc(avi_data, "avih");
+    write_u32_le(avi_data, 56);
+    write_u32_le(avi_data, 1000000 / fps);
+    write_u32_le(avi_data, 0);
+    write_u32_le(avi_data, 0);
+    write_u32_le(avi_data, 0x110);
+    write_u32_le(avi_data, num_images);
+    write_u32_le(avi_data, 0);
+    write_u32_le(avi_data, 1);
+    write_u32_le(avi_data, width * height * 3);
+    write_u32_le(avi_data, width);
+    write_u32_le(avi_data, height);
+    write_u32_le(avi_data, 0);
+    write_u32_le(avi_data, 0);
+    write_u32_le(avi_data, 0);
+    write_u32_le(avi_data, 0);
 
-    fwrite("strf", 4, 1, f);
-    write_u32_le(f, 40);
-    write_u32_le(f, 40);
-    write_u32_le(f, width);
-    write_u32_le(f, height);
-    write_u16_le(f, 1);
-    write_u16_le(f, 24);
-    fwrite("MJPG", 4, 1, f);
-    write_u32_le(f, width * height * 3);
-    write_u32_le(f, 0);
-    write_u32_le(f, 0);
-    write_u32_le(f, 0);
-    write_u32_le(f, 0);
+    write_fourcc(avi_data, "LIST");
+    write_u32_le(avi_data, 4 + 8 + 56 + 8 + 40);
+    write_fourcc(avi_data, "strl");
 
-    fwrite("LIST", 4, 1, f);
-    long movi_size_pos = ftell(f);
-    write_u32_le(f, 0);
-    fwrite("movi", 4, 1, f);
+    write_fourcc(avi_data, "strh");
+    write_u32_le(avi_data, 56);
+    write_fourcc(avi_data, "vids");
+    write_fourcc(avi_data, "MJPG");
+    write_u32_le(avi_data, 0);
+    write_u16_le(avi_data, 0);
+    write_u16_le(avi_data, 0);
+    write_u32_le(avi_data, 0);
+    write_u32_le(avi_data, 1);
+    write_u32_le(avi_data, fps);
+    write_u32_le(avi_data, 0);
+    write_u32_le(avi_data, num_images);
+    write_u32_le(avi_data, width * height * 3);
+    write_u32_le(avi_data, static_cast<uint32_t>(-1));
+    write_u32_le(avi_data, 0);
+    write_u16_le(avi_data, 0);
+    write_u16_le(avi_data, 0);
+    write_u16_le(avi_data, 0);
+    write_u16_le(avi_data, 0);
+
+    write_fourcc(avi_data, "strf");
+    write_u32_le(avi_data, 40);
+    write_u32_le(avi_data, 40);
+    write_u32_le(avi_data, width);
+    write_u32_le(avi_data, height);
+    write_u16_le(avi_data, 1);
+    write_u16_le(avi_data, 24);
+    write_fourcc(avi_data, "MJPG");
+    write_u32_le(avi_data, width * height * 3);
+    write_u32_le(avi_data, 0);
+    write_u32_le(avi_data, 0);
+    write_u32_le(avi_data, 0);
+    write_u32_le(avi_data, 0);
+
+    write_fourcc(avi_data, "LIST");
+    const size_t movi_size_pos = avi_data.size();
+    write_u32_le(avi_data, 0);
+    write_fourcc(avi_data, "movi");
 
     std::vector<avi_index_entry> index(static_cast<size_t>(num_images));
     std::vector<uint8_t> jpeg_data;
@@ -801,55 +879,61 @@ int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int
             buffer->insert(buffer->end(), src, src + size);
         };
 
-        if (!stbi_write_jpg_to_func(write_to_buf, &jpeg_data, images[i].width, images[i].height, channels, images[i].data, quality)) {
+        if (!stbi_write_jpg_to_func(write_to_buf, &jpeg_data, images[i].width, images[i].height, channels, images[i].data, mjpg_quality)) {
             fprintf(stderr, "Error: Failed to encode JPEG frame.\n");
-            return -1;
+            return {};
         }
 
-        fwrite("00dc", 4, 1, f);
-        write_u32_le(f, (uint32_t)jpeg_data.size());
-        index[i].offset = ftell(f) - 8;
-        index[i].size   = (uint32_t)jpeg_data.size();
-        fwrite(jpeg_data.data(), 1, jpeg_data.size(), f);
+        index[i].offset = static_cast<uint32_t>(avi_data.size());
+        write_fourcc(avi_data, "00dc");
+        write_u32_le(avi_data, static_cast<uint32_t>(jpeg_data.size()));
+        index[i].size = (uint32_t)jpeg_data.size();
+        avi_data.insert(avi_data.end(), jpeg_data.begin(), jpeg_data.end());
 
         if (jpeg_data.size() % 2) {
-            fputc(0, f);
+            avi_data.push_back(0);
         }
     }
 
-    long cur_pos   = ftell(f);
-    long movi_size = cur_pos - movi_size_pos - 4;
-    fseek(f, movi_size_pos, SEEK_SET);
-    write_u32_le(f, movi_size);
-    fseek(f, cur_pos, SEEK_SET);
+    const size_t movi_size = avi_data.size() - movi_size_pos - 4;
+    patch_u32_le(avi_data, movi_size_pos, static_cast<uint32_t>(movi_size));
 
-    fwrite("idx1", 4, 1, f);
-    write_u32_le(f, num_images * 16);
+    write_fourcc(avi_data, "idx1");
+    write_u32_le(avi_data, num_images * 16);
     for (int i = 0; i < num_images; i++) {
-        fwrite("00dc", 4, 1, f);
-        write_u32_le(f, 0x10);
-        write_u32_le(f, index[i].offset);
-        write_u32_le(f, index[i].size);
+        write_fourcc(avi_data, "00dc");
+        write_u32_le(avi_data, 0x10);
+        write_u32_le(avi_data, index[i].offset);
+        write_u32_le(avi_data, index[i].size);
     }
 
-    cur_pos        = ftell(f);
-    long file_size = cur_pos - riff_size_pos - 4;
-    fseek(f, riff_size_pos, SEEK_SET);
-    write_u32_le(f, file_size);
-    fseek(f, cur_pos, SEEK_SET);
+    const size_t file_size = avi_data.size() - riff_size_pos - 4;
+    patch_u32_le(avi_data, riff_size_pos, static_cast<uint32_t>(file_size));
 
+    return avi_data;
+}
+
+int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
+    std::vector<uint8_t> avi_data = create_mjpg_avi_from_sd_images_to_vector(images, num_images, fps, quality);
+    if (avi_data.empty()) {
+        return -1;
+    }
+    if (!write_binary_file_bytes(filename, avi_data)) {
+        perror("Error opening file for writing");
+        return -1;
+    }
     return 0;
 }
 
 #ifdef SD_USE_WEBP
-int create_animated_webp_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
+std::vector<uint8_t> create_animated_webp_from_sd_images_to_vector(sd_image_t* images, int num_images, int fps, int quality) {
     if (num_images == 0) {
         fprintf(stderr, "Error: Image array is empty.\n");
-        return -1;
+        return {};
     }
     if (fps <= 0) {
         fprintf(stderr, "Error: FPS must be positive.\n");
-        return -1;
+        return {};
     }
 
     const int width    = static_cast<int>(images[0].width);
@@ -857,14 +941,14 @@ int create_animated_webp_from_sd_images(const char* filename, sd_image_t* images
     const int channels = static_cast<int>(images[0].channel);
     if (channels != 1 && channels != 3 && channels != 4) {
         fprintf(stderr, "Error: Unsupported channel count: %d\n", channels);
-        return -1;
+        return {};
     }
 
     WebPAnimEncoderOptions anim_options;
     WebPConfig config;
     if (!WebPAnimEncoderOptionsInit(&anim_options) || !WebPConfigInit(&config)) {
         fprintf(stderr, "Error: Failed to initialize WebP animation encoder.\n");
-        return -1;
+        return {};
     }
 
     config.quality      = static_cast<float>(quality);
@@ -875,13 +959,13 @@ int create_animated_webp_from_sd_images(const char* filename, sd_image_t* images
     }
     if (!WebPValidateConfig(&config)) {
         fprintf(stderr, "Error: Invalid WebP encoder configuration.\n");
-        return -1;
+        return {};
     }
 
     WebPAnimEncoderPtr enc(WebPAnimEncoderNew(width, height, &anim_options));
     if (enc == nullptr) {
         fprintf(stderr, "Error: Could not create WebPAnimEncoder object.\n");
-        return -1;
+        return {};
     }
 
     const int frame_duration_ms = std::max(1, static_cast<int>(std::lround(1000.0 / static_cast<double>(fps))));
@@ -891,13 +975,13 @@ int create_animated_webp_from_sd_images(const char* filename, sd_image_t* images
         const sd_image_t& image = images[i];
         if (static_cast<int>(image.width) != width || static_cast<int>(image.height) != height) {
             fprintf(stderr, "Error: Frame dimensions do not match.\n");
-            return -1;
+            return {};
         }
 
         WebPPictureGuard picture;
         if (!picture.initialized) {
             fprintf(stderr, "Error: Failed to initialize WebPPicture.\n");
-            return -1;
+            return {};
         }
         picture.picture.use_argb = 1;
         picture.picture.width    = width;
@@ -921,12 +1005,12 @@ int create_animated_webp_from_sd_images(const char* filename, sd_image_t* images
 
         if (!picture_ok) {
             fprintf(stderr, "Error: Failed to import frame into WebPPicture.\n");
-            return -1;
+            return {};
         }
 
         if (!WebPAnimEncoderAdd(enc.get(), &picture.picture, timestamp_ms, &config)) {
             fprintf(stderr, "Error: Failed to add frame to animated WebP: %s\n", WebPAnimEncoderGetError(enc.get()));
-            return -1;
+            return {};
         }
 
         timestamp_ms += frame_duration_ms;
@@ -934,52 +1018,50 @@ int create_animated_webp_from_sd_images(const char* filename, sd_image_t* images
 
     if (!WebPAnimEncoderAdd(enc.get(), nullptr, timestamp_ms, nullptr)) {
         fprintf(stderr, "Error: Failed to finalize animated WebP frames: %s\n", WebPAnimEncoderGetError(enc.get()));
-        return -1;
+        return {};
     }
 
     WebPDataGuard webp_data;
     if (!WebPAnimEncoderAssemble(enc.get(), &webp_data.data)) {
         fprintf(stderr, "Error: Failed to assemble animated WebP: %s\n", WebPAnimEncoderGetError(enc.get()));
-        return -1;
+        return {};
     }
 
-    FilePtr f(fopen(filename, "wb"));
-    if (!f) {
+    return std::vector<uint8_t>(webp_data.data.bytes, webp_data.data.bytes + webp_data.data.size);
+}
+
+int create_animated_webp_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
+    std::vector<uint8_t> webp_data = create_animated_webp_from_sd_images_to_vector(images, num_images, fps, quality);
+    if (webp_data.empty()) {
+        return -1;
+    }
+    if (!write_binary_file_bytes(filename, webp_data)) {
         perror("Error opening file for writing");
         return -1;
     }
-    if (webp_data.data.size > 0 && fwrite(webp_data.data.bytes, 1, webp_data.data.size, f.get()) != webp_data.data.size) {
-        fprintf(stderr, "Error: Failed to write animated WebP file.\n");
-        return -1;
-    }
-
     return 0;
 }
 #endif
 
 #ifdef SD_USE_WEBM
-int create_webm_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
+std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images, int num_images, int fps, int quality) {
     if (num_images == 0) {
         fprintf(stderr, "Error: Image array is empty.\n");
-        return -1;
+        return {};
     }
     if (fps <= 0) {
         fprintf(stderr, "Error: FPS must be positive.\n");
-        return -1;
+        return {};
     }
 
     const int width  = static_cast<int>(images[0].width);
     const int height = static_cast<int>(images[0].height);
     if (width <= 0 || height <= 0) {
         fprintf(stderr, "Error: Invalid frame dimensions.\n");
-        return -1;
+        return {};
     }
 
-    mkvmuxer::MkvWriter writer;
-    if (!writer.Open(filename)) {
-        fprintf(stderr, "Error: Could not open WebM file for writing.\n");
-        return -1;
-    }
+    MemoryMkvWriter writer;
 
     const int ret = [&]() -> int {
         mkvmuxer::Segment segment;
@@ -1045,30 +1127,63 @@ int create_webm_from_sd_images(const char* filename, sd_image_t* images, int num
         }
         return 0;
     }();
-    writer.Close();
-    return ret;
+    if (ret != 0) {
+        return {};
+    }
+    return writer.data();
+}
+
+int create_webm_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
+    std::vector<uint8_t> webm_data = create_webm_from_sd_images_to_vector(images, num_images, fps, quality);
+    if (webm_data.empty()) {
+        return -1;
+    }
+    if (!write_binary_file_bytes(filename, webm_data)) {
+        perror("Error opening file for writing");
+        return -1;
+    }
+    return 0;
 }
 #endif
 
-int create_video_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
-    std::string path = filename ? filename : "";
-    auto pos         = path.find_last_of('.');
-    std::string ext  = pos == std::string::npos ? "" : path.substr(pos);
-    for (char& ch : ext) {
-        ch = static_cast<char>(tolower(static_cast<unsigned char>(ch)));
+std::vector<uint8_t> create_video_from_sd_images_to_vector(const std::string& output_format,
+                                                           sd_image_t* images,
+                                                           int num_images,
+                                                           int fps,
+                                                           int quality) {
+    std::string format = output_format;
+    std::transform(format.begin(), format.end(), format.begin(),
+                   [](unsigned char c) { return static_cast<char>(tolower(c)); });
+    if (!format.empty() && format[0] == '.') {
+        format.erase(format.begin());
     }
 
 #ifdef SD_USE_WEBM
-    if (ext == ".webm") {
-        return create_webm_from_sd_images(filename, images, num_images, fps, quality);
+    if (format == "webm") {
+        return create_webm_from_sd_images_to_vector(images, num_images, fps, quality);
     }
 #endif
 
 #ifdef SD_USE_WEBP
-    if (ext == ".webp") {
-        return create_animated_webp_from_sd_images(filename, images, num_images, fps, quality);
+    if (format == "webp") {
+        return create_animated_webp_from_sd_images_to_vector(images, num_images, fps, quality);
     }
 #endif
 
-    return create_mjpg_avi_from_sd_images(filename, images, num_images, fps, quality);
+    return create_mjpg_avi_from_sd_images_to_vector(images, num_images, fps, quality);
+}
+
+int create_video_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality) {
+    std::string path                = filename ? filename : "";
+    auto pos                        = path.find_last_of('.');
+    std::string ext                 = pos == std::string::npos ? "" : path.substr(pos);
+    std::vector<uint8_t> video_data = create_video_from_sd_images_to_vector(ext, images, num_images, fps, quality);
+    if (video_data.empty()) {
+        return -1;
+    }
+    if (!write_binary_file_bytes(filename, video_data)) {
+        perror("Error opening file for writing");
+        return -1;
+    }
+    return 0;
 }

examples/common/media_io.h

@@ -58,6 +58,10 @@ int create_mjpg_avi_from_sd_images(const char* filename,
                                    int num_images,
                                    int fps,
                                    int quality = 90);
+std::vector<uint8_t> create_mjpg_avi_from_sd_images_to_vector(sd_image_t* images,
+                                                              int num_images,
+                                                              int fps,
+                                                              int quality = 90);
 
 #ifdef SD_USE_WEBP
 int create_animated_webp_from_sd_images(const char* filename,
@@ -65,6 +69,10 @@ int create_animated_webp_from_sd_images(const char* filename,
                                         int num_images,
                                         int fps,
                                         int quality = 90);
+std::vector<uint8_t> create_animated_webp_from_sd_images_to_vector(sd_image_t* images,
+                                                                   int num_images,
+                                                                   int fps,
+                                                                   int quality = 90);
 #endif
 
 #ifdef SD_USE_WEBM
@@ -73,6 +81,10 @@ int create_webm_from_sd_images(const char* filename,
                                int num_images,
                                int fps,
                                int quality = 90);
+std::vector<uint8_t> create_webm_from_sd_images_to_vector(sd_image_t* images,
+                                                          int num_images,
+                                                          int fps,
+                                                          int quality = 90);
 #endif
 
 int create_video_from_sd_images(const char* filename,
@@ -80,5 +92,10 @@ int create_video_from_sd_images(const char* filename,
                                 int num_images,
                                 int fps,
                                 int quality = 90);
+std::vector<uint8_t> create_video_from_sd_images_to_vector(const std::string& output_format,
+                                                           sd_image_t* images,
+                                                           int num_images,
+                                                           int fps,
+                                                           int quality = 90);
 
 #endif  // __MEDIA_IO_H__

examples/server/CMakeLists.txt

@@ -50,7 +50,13 @@ if(SD_SERVER_BUILD_FRONTEND AND EXISTS "${FRONTEND_DIR}")
 
         set_source_files_properties("${GENERATED_HTML_HEADER}" PROPERTIES GENERATED TRUE)
     else()
-        message(WARNING "pnpm not found, frontend build disabled")
+        if(EXISTS "${GENERATED_HTML_HEADER}")
+            message(STATUS "pnpm not found; using pre-built frontend header detected at ${GENERATED_HTML_HEADER}")
+            set(HAVE_FRONTEND_BUILD ON)
+            add_custom_target(${TARGET}_frontend)
+        else()
+            message(WARNING "pnpm not found; frontend build disabled.")
+        endif()
     endif()
 else()
     message(STATUS "Frontend disabled or directory not found: ${FRONTEND_DIR}")

examples/server/README.md

@@ -1,3 +1,33 @@
+# Example
+
+The following example starts `sd-server` with a standalone diffusion model, VAE, and LLM text encoder:
+
+```
+.\bin\Release\sd-server.exe --diffusion-model  ..\models\diffusion_models\z_image_turbo_bf16.safetensors --vae ..\models\vae\ae.sft  --llm ..\models\text_encoders\qwen_3_4b.safetensors --diffusion-fa --offload-to-cpu -v --cfg-scale 1.0
+```
+
+What this example does:
+
+* `--diffusion-model` selects the standalone diffusion model
+* `--vae` selects the VAE decoder
+* `--llm` selects the text encoder / language model used by this pipeline
+* `--diffusion-fa` enables flash attention in the diffusion model
+* `--offload-to-cpu` reduces VRAM pressure by keeping weights in RAM when possible
+* `-v` enables verbose logging
+* `--cfg-scale 1.0` sets the default CFG scale for generation
+
+After the server starts successfully:
+
+* the web UI is available at `http://127.0.0.1:1234/`
+* the native async API is available under `/sdcpp/v1/...`
+* the compatibility APIs are available under `/v1/...` and `/sdapi/v1/...`
+
+If you want to use a different host or port, pass:
+
+```bash
+--listen-ip <ip> --listen-port <port>
+```
+
 # Frontend
 
 ## Build with Frontend
@@ -8,7 +38,7 @@ The server can optionally build the web frontend and embed it into the binary as
 
 Install the following tools:
 
-* **Node.js** ≥ 22.18
+* **Node.js** ≥ 20
   https://nodejs.org/
 
 * **pnpm** ≥ 10
@@ -54,7 +84,7 @@ and embed the generated frontend into the server binary.
 
 ## Frontend Repository
 
-The web frontend is maintained in a **separate repository**, https://github.com/leejet/stable-ui.
+The web frontend is maintained in a **separate repository**, https://github.com/leejet/sdcpp-webui.
 
 If you want to modify the UI or frontend logic, please submit pull requests to the **frontend repository**.
 
@@ -189,7 +219,7 @@ Default Generation Options:
                                            medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a and dpm++2s_a)
+  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
   --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
@@ -197,7 +227,7 @@ Default Generation Options:
   --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a and dpm++2s_a)
+  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
   --pm-style-strength <float>
   --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
@@ -208,10 +238,10 @@ Default Generation Options:
   --disable-image-metadata                 do not embed generation metadata on image files
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
   --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           tcd, res_multistep, res_2s] (default: euler for Flux/SD3/Wan, euler_a
+                                           tcd, res_multistep, res_2s, er_sde] (default: euler for Flux/SD3/Wan, euler_a
                                            otherwise)
   --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
-                                           ddim_trailing, tcd, res_multistep, res_2s] default: euler for Flux/SD3/Wan,
+                                           ddim_trailing, tcd, res_multistep, res_2s, er_sde] default: euler for Flux/SD3/Wan,
                                            euler_a otherwise
   --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
                                            kl_optimal, lcm, bong_tangent], default: discrete

examples/server/api.md

@@ -9,7 +9,7 @@ The server currently exposes three API families:
 - `sdcpp API` under `/sdcpp/v1/...`
 
 The `sdcpp API` is the native API surface.
-Its request schema is also the canonical schema for `sd_cpp_extra_args`.
+Its request schema is the same schema used by `sd_cpp_extra_args`.
 
 Global LoRA rule:
 
@@ -55,8 +55,6 @@ Current endpoints include:
 - `POST /sdcpp/v1/jobs/{id}/cancel`
 - `POST /sdcpp/v1/vid_gen`
 
-`POST /sdcpp/v1/vid_gen` is currently exposed but returns `501 Not Implemented`.
-
 ## `sd_cpp_extra_args`
 
 `sd_cpp_extra_args` is an extension mechanism for the compatibility APIs.
@@ -79,12 +77,12 @@ Behavior:
 - The JSON block is parsed using the same field rules as the `sdcpp API`.
 - The block is removed from the final prompt before generation.
 
-Intended use:
+Supported use:
 
 - extend `OpenAI API` requests with native `stable-diffusion.cpp` controls
 - extend `sdapi` requests with native `stable-diffusion.cpp` controls
 
-Not intended use:
+Unsupported use:
 
 - do not use `sd_cpp_extra_args` with `/sdcpp/v1/*`
 
@@ -372,20 +370,25 @@ Field types:
 
 Returns frontend-friendly capability metadata.
 
-Typical contents:
+The mode-aware fields are the primary interface. The top-level compatibility fields are deprecated mirrors kept for older clients.
 
-| Field | Type |
-| --- | --- |
-| `model` | `object` |
-| `defaults` | `object` |
-| `loras` | `array<object>` |
-| `samplers` | `array<string>` |
-| `schedulers` | `array<string>` |
-| `output_formats` | `array<string>` |
-| `limits` | `object` |
-| `features` | `object` |
+Top-level fields:
 
-Nested fields currently returned:
+| Field | Type | Notes |
+| --- | --- | --- |
+| `model` | `object` | Loaded model metadata |
+| `current_mode` | `string` | The native generation mode mirrored by top-level compatibility fields |
+| `supported_modes` | `array<string>` | Supported native modes such as `img_gen` or `vid_gen` |
+| `defaults` | `object` | Deprecated compatibility mirror of `defaults_by_mode[current_mode]` |
+| `output_formats` | `array<string>` | Deprecated compatibility mirror of `output_formats_by_mode[current_mode]` |
+| `features` | `object` | Deprecated compatibility mirror of `features_by_mode[current_mode]` |
+| `defaults_by_mode` | `object` | Explicit defaults for each supported mode |
+| `output_formats_by_mode` | `object` | Explicit output formats for each supported mode |
+| `features_by_mode` | `object` | Explicit feature flags for each supported mode |
+| `samplers` | `array<string>` | Available sampling methods |
+| `schedulers` | `array<string>` | Available schedulers |
+| `loras` | `array<object>` | Available LoRA entries |
+| `limits` | `object` | Shared queue and size limits |
 
 `model`
 
@@ -395,50 +398,24 @@ Nested fields currently returned:
 | `model.stem` | `string` |
 | `model.path` | `string` |
 
-`defaults`
+Compatibility rules:
+
+- `defaults`, `output_formats`, and `features` are deprecated compatibility mirrors
+- those three top-level fields always mirror `current_mode`
+- `supported_modes`, `defaults_by_mode`, `output_formats_by_mode`, and `features_by_mode` are the mode-aware fields
+
+Mode-aware objects:
 
 | Field | Type |
 | --- | --- |
-| `defaults.prompt` | `string` |
-| `defaults.negative_prompt` | `string` |
-| `defaults.clip_skip` | `integer` |
-| `defaults.width` | `integer` |
-| `defaults.height` | `integer` |
-| `defaults.strength` | `number` |
-| `defaults.seed` | `integer` |
-| `defaults.batch_count` | `integer` |
-| `defaults.auto_resize_ref_image` | `boolean` |
-| `defaults.increase_ref_index` | `boolean` |
-| `defaults.control_strength` | `number` |
-| `defaults.sample_params` | `object` |
-| `defaults.sample_params.scheduler` | `string` |
-| `defaults.sample_params.sample_method` | `string` |
-| `defaults.sample_params.sample_steps` | `integer` |
-| `defaults.sample_params.eta` | `number \| null` |
-| `defaults.sample_params.shifted_timestep` | `integer` |
-| `defaults.sample_params.flow_shift` | `number \| null` |
-| `defaults.sample_params.guidance` | `object` |
-| `defaults.sample_params.guidance.txt_cfg` | `number` |
-| `defaults.sample_params.guidance.img_cfg` | `number \| null` |
-| `defaults.sample_params.guidance.distilled_guidance` | `number` |
-| `defaults.sample_params.guidance.slg` | `object` |
-| `defaults.sample_params.guidance.slg.layers` | `array<integer>` |
-| `defaults.sample_params.guidance.slg.layer_start` | `number` |
-| `defaults.sample_params.guidance.slg.layer_end` | `number` |
-| `defaults.sample_params.guidance.slg.scale` | `number` |
-| `defaults.vae_tiling_params` | `object` |
-| `defaults.vae_tiling_params.enabled` | `boolean` |
-| `defaults.vae_tiling_params.tile_size_x` | `integer` |
-| `defaults.vae_tiling_params.tile_size_y` | `integer` |
-| `defaults.vae_tiling_params.target_overlap` | `number` |
-| `defaults.vae_tiling_params.rel_size_x` | `number` |
-| `defaults.vae_tiling_params.rel_size_y` | `number` |
-| `defaults.cache_mode` | `string` |
-| `defaults.cache_option` | `string` |
-| `defaults.scm_mask` | `string` |
-| `defaults.scm_policy_dynamic` | `boolean` |
-| `defaults.output_format` | `string` |
-| `defaults.output_compression` | `integer` |
+| `defaults_by_mode.img_gen` | `object` |
+| `defaults_by_mode.vid_gen` | `object` |
+| `output_formats_by_mode.img_gen` | `array<string>` |
+| `output_formats_by_mode.vid_gen` | `array<string>` |
+| `features_by_mode.img_gen` | `object` |
+| `features_by_mode.vid_gen` | `object` |
+
+Shared nested fields:
 
 `loras`
 
@@ -458,19 +435,100 @@ Nested fields currently returned:
 | `limits.max_batch_count` | `integer` |
 | `limits.max_queue_size` | `integer` |
 
-`features`
+Shared default fields used by both `img_gen` and `vid_gen`:
 
 | Field | Type |
 | --- | --- |
-| `features.init_image` | `boolean` |
-| `features.mask_image` | `boolean` |
-| `features.control_image` | `boolean` |
-| `features.ref_images` | `boolean` |
-| `features.lora` | `boolean` |
-| `features.vae_tiling` | `boolean` |
-| `features.cache` | `boolean` |
-| `features.cancel_queued` | `boolean` |
-| `features.cancel_generating` | `boolean` |
+| `prompt` | `string` |
+| `negative_prompt` | `string` |
+| `clip_skip` | `integer` |
+| `width` | `integer` |
+| `height` | `integer` |
+| `strength` | `number` |
+| `seed` | `integer` |
+| `sample_params` | `object` |
+| `sample_params.scheduler` | `string` |
+| `sample_params.sample_method` | `string` |
+| `sample_params.sample_steps` | `integer` |
+| `sample_params.eta` | `number \| null` |
+| `sample_params.shifted_timestep` | `integer` |
+| `sample_params.flow_shift` | `number \| null` |
+| `sample_params.guidance.txt_cfg` | `number` |
+| `sample_params.guidance.img_cfg` | `number \| null` |
+| `sample_params.guidance.distilled_guidance` | `number` |
+| `sample_params.guidance.slg.layers` | `array<integer>` |
+| `sample_params.guidance.slg.layer_start` | `number` |
+| `sample_params.guidance.slg.layer_end` | `number` |
+| `sample_params.guidance.slg.scale` | `number` |
+| `vae_tiling_params` | `object` |
+| `vae_tiling_params.enabled` | `boolean` |
+| `vae_tiling_params.tile_size_x` | `integer` |
+| `vae_tiling_params.tile_size_y` | `integer` |
+| `vae_tiling_params.target_overlap` | `number` |
+| `vae_tiling_params.rel_size_x` | `number` |
+| `vae_tiling_params.rel_size_y` | `number` |
+| `cache_mode` | `string` |
+| `cache_option` | `string` |
+| `scm_mask` | `string` |
+| `scm_policy_dynamic` | `boolean` |
+| `output_format` | `string` |
+| `output_compression` | `integer` |
+
+`img_gen`-specific default fields:
+
+| Field | Type |
+| --- | --- |
+| `batch_count` | `integer` |
+| `auto_resize_ref_image` | `boolean` |
+| `increase_ref_index` | `boolean` |
+| `control_strength` | `number` |
+
+`vid_gen`-specific default fields:
+
+| Field | Type |
+| --- | --- |
+| `video_frames` | `integer` |
+| `fps` | `integer` |
+| `moe_boundary` | `number` |
+| `vace_strength` | `number` |
+| `high_noise_sample_params` | `object` |
+| `high_noise_sample_params.scheduler` | `string` |
+| `high_noise_sample_params.sample_method` | `string` |
+| `high_noise_sample_params.sample_steps` | `integer` |
+| `high_noise_sample_params.eta` | `number \| null` |
+| `high_noise_sample_params.shifted_timestep` | `integer` |
+| `high_noise_sample_params.flow_shift` | `number \| null` |
+| `high_noise_sample_params.guidance.txt_cfg` | `number` |
+| `high_noise_sample_params.guidance.img_cfg` | `number \| null` |
+| `high_noise_sample_params.guidance.distilled_guidance` | `number` |
+| `high_noise_sample_params.guidance.slg.layers` | `array<integer>` |
+| `high_noise_sample_params.guidance.slg.layer_start` | `number` |
+| `high_noise_sample_params.guidance.slg.layer_end` | `number` |
+| `high_noise_sample_params.guidance.slg.scale` | `number` |
+
+Fields returned in `features_by_mode.img_gen`:
+
+- `init_image`
+- `mask_image`
+- `control_image`
+- `ref_images`
+- `lora`
+- `vae_tiling`
+- `cache`
+- `cancel_queued`
+- `cancel_generating`
+
+Fields returned in `features_by_mode.vid_gen`:
+
+- `init_image`
+- `end_image`
+- `control_frames`
+- `high_noise_sample_params`
+- `lora`
+- `vae_tiling`
+- `cache`
+- `cancel_queued`
+- `cancel_generating`
 
 #### `POST /sdcpp/v1/img_gen`
 
@@ -521,9 +579,7 @@ Typical status codes:
 - `409 Conflict`
 - `410 Gone`
 
-### Canonical Request Schema
-
-The `sdcpp API` request body is the canonical native schema.
+### Request Body
 
 Example:
 
@@ -612,7 +668,7 @@ Channel expectations:
 If omitted or null:
 
 - single-image fields map to an empty `sd_image_t`
-- array fields map to `nullptr + count = 0`
+- array fields map to an empty C-style array, represented as `pointer = nullptr` and `count = 0`
 
 ### Field Mapping Summary
 
@@ -686,11 +742,11 @@ HTTP-only output fields:
 | `output_format` | `string` |
 | `output_compression` | `integer` |
 
-### Optional Field Semantics
+### Optional Field Handling
 
-Clients should preserve unset semantics for optional sampling fields.
+Optional sampling fields may be omitted.
 
-If a user has not explicitly provided one of these fields, the client should omit it instead of injecting a guessed fallback:
+When omitted, backend defaults apply to these fields:
 
 - `sample_params.scheduler`
 - `sample_params.sample_method`
@@ -766,29 +822,394 @@ Example cancelled job:
 }
 ```
 
-### Validation and Retention
+### Submission Errors
 
-Recommended behavior:
+`POST /sdcpp/v1/img_gen` may return:
 
-- malformed JSON returns `400`
-- invalid image payloads return `400`
-- invalid parameter structure returns `400`
-- queue full returns `429` or `503`
-- accepted runtime failures transition the job to `failed`
-- unsupported in-progress cancellation may return `409`
+- `202 Accepted` when the job is created
+- `400 Bad Request` for an empty body, unsupported model mode, invalid JSON, or invalid generation parameters
+- `429 Too Many Requests` when the job queue is full
+- `500 Internal Server Error` for unexpected server exceptions during submission
 
-Recommended retention controls:
+### `vid_gen`
 
-- pending job limit
-- completed job TTL
-- failed job TTL
+The following section documents the native async contract for video generation.
 
-### Future `vid_gen`
+#### `POST /sdcpp/v1/vid_gen`
 
-Future `vid_gen` should reuse the same async job model:
+Submits an async video generation job.
 
-- `POST /sdcpp/v1/vid_gen`
-- `GET /sdcpp/v1/jobs/{id}`
-- `POST /sdcpp/v1/jobs/{id}/cancel`
+Successful submission returns `202 Accepted`.
 
-Its request body should mirror `sd_vid_gen_params_t` in the same way that `img_gen` mirrors `sd_img_gen_params_t`.
+Example response:
+
+```json
+{
+  "id": "job_01HTXYZVID",
+  "kind": "vid_gen",
+  "status": "queued",
+  "created": 1775401200,
+  "poll_url": "/sdcpp/v1/jobs/job_01HTXYZVID"
+}
+```
+
+Response fields:
+
+| Field | Type |
+| --- | --- |
+| `id` | `string` |
+| `kind` | `string` |
+| `status` | `string` |
+| `created` | `integer` |
+| `poll_url` | `string` |
+
+### Request Body
+
+Compared with `img_gen`, the `vid_gen` request body:
+
+- `vid_gen` is a single video sequence job, so `batch_count` is not part of the request schema
+- `ref_images`, `mask_image`, `control_image`, `control_strength`, and `embed_image_metadata` are not part of the request schema
+- `vid_gen` adds `end_image`, `control_frames`, `high_noise_sample_params`, `video_frames`, `fps`, `moe_boundary`, and `vace_strength`
+
+Example:
+
+```json
+{
+  "prompt": "a cat walking through a rainy alley",
+  "negative_prompt": "",
+  "clip_skip": -1,
+  "width": 832,
+  "height": 480,
+  "strength": 0.75,
+  "seed": -1,
+  "video_frames": 33,
+  "fps": 16,
+  "moe_boundary": 0.875,
+  "vace_strength": 1.0,
+
+  "init_image": null,
+  "end_image": null,
+  "control_frames": [],
+
+  "sample_params": {
+    "scheduler": "discrete",
+    "sample_method": "euler",
+    "sample_steps": 28,
+    "eta": 1.0,
+    "shifted_timestep": 0,
+    "custom_sigmas": [],
+    "flow_shift": 0.0,
+    "guidance": {
+      "txt_cfg": 7.0,
+      "img_cfg": 7.0,
+      "distilled_guidance": 3.5,
+      "slg": {
+        "layers": [7, 8, 9],
+        "layer_start": 0.01,
+        "layer_end": 0.2,
+        "scale": 0.0
+      }
+    }
+  },
+
+  "high_noise_sample_params": {
+    "scheduler": "discrete",
+    "sample_method": "euler",
+    "sample_steps": -1,
+    "eta": 1.0,
+    "shifted_timestep": 0,
+    "flow_shift": 0.0,
+    "guidance": {
+      "txt_cfg": 7.0,
+      "img_cfg": 7.0,
+      "distilled_guidance": 3.5,
+      "slg": {
+        "layers": [7, 8, 9],
+        "layer_start": 0.01,
+        "layer_end": 0.2,
+        "scale": 0.0
+      }
+    }
+  },
+
+  "lora": [],
+
+  "vae_tiling_params": {
+    "enabled": false,
+    "tile_size_x": 0,
+    "tile_size_y": 0,
+    "target_overlap": 0.5,
+    "rel_size_x": 0.0,
+    "rel_size_y": 0.0
+  },
+
+  "cache_mode": "disabled",
+  "cache_option": "",
+  "scm_mask": "",
+  "scm_policy_dynamic": true,
+
+  "output_format": "webm",
+  "output_compression": 100
+}
+```
+
+### LoRA Rules
+
+- The server only accepts explicit LoRA entries from the `lora` field.
+- Prompt-embedded `<lora:...>` tags are intentionally unsupported.
+- `lora[].is_high_noise` controls whether a LoRA applies only to the high-noise stage.
+
+### Image and Frame Encoding Rules
+
+Any image field accepts:
+
+- a raw base64 string, or
+- a data URL such as `data:image/png;base64,...`
+
+Channel expectations:
+
+- `init_image`: 3 channels
+- `end_image`: 3 channels
+- `control_frames[]`: 3 channels
+
+Frame ordering rules:
+
+- `control_frames[]` order is the conditioning frame order
+- `control_frames[]` is preserved in request order
+
+If omitted or null:
+
+- single-image fields map to an empty `sd_image_t`
+- array fields map to an empty C-style array, represented as `pointer = nullptr` and `count = 0`
+
+### Field Mapping Summary
+
+Top-level scalar fields:
+
+| Field | Type |
+| --- | --- |
+| `prompt` | `string` |
+| `negative_prompt` | `string` |
+| `clip_skip` | `integer` |
+| `width` | `integer` |
+| `height` | `integer` |
+| `strength` | `number` |
+| `seed` | `integer` |
+| `video_frames` | `integer` |
+| `fps` | `integer` |
+| `moe_boundary` | `number` |
+| `vace_strength` | `number` |
+
+Image and frame fields:
+
+| Field | Type |
+| --- | --- |
+| `init_image` | `string \| null` |
+| `end_image` | `string \| null` |
+| `control_frames` | `array<string>` |
+
+LoRA fields:
+
+| Field | Type |
+| --- | --- |
+| `lora[].path` | `string` |
+| `lora[].multiplier` | `number` |
+| `lora[].is_high_noise` | `boolean` |
+
+Sampling fields:
+
+| Field | Type |
+| --- | --- |
+| `sample_params.scheduler` | `string` |
+| `sample_params.sample_method` | `string` |
+| `sample_params.sample_steps` | `integer` |
+| `sample_params.eta` | `number` |
+| `sample_params.shifted_timestep` | `integer` |
+| `sample_params.custom_sigmas` | `array<number>` |
+| `sample_params.flow_shift` | `number` |
+| `sample_params.guidance.txt_cfg` | `number` |
+| `sample_params.guidance.img_cfg` | `number` |
+| `sample_params.guidance.distilled_guidance` | `number` |
+| `sample_params.guidance.slg.layers` | `array<integer>` |
+| `sample_params.guidance.slg.layer_start` | `number` |
+| `sample_params.guidance.slg.layer_end` | `number` |
+| `sample_params.guidance.slg.scale` | `number` |
+
+High-noise sampling fields:
+
+| Field | Type |
+| --- | --- |
+| `high_noise_sample_params.scheduler` | `string` |
+| `high_noise_sample_params.sample_method` | `string` |
+| `high_noise_sample_params.sample_steps` | `integer` |
+| `high_noise_sample_params.eta` | `number` |
+| `high_noise_sample_params.shifted_timestep` | `integer` |
+| `high_noise_sample_params.flow_shift` | `number` |
+| `high_noise_sample_params.guidance.txt_cfg` | `number` |
+| `high_noise_sample_params.guidance.img_cfg` | `number` |
+| `high_noise_sample_params.guidance.distilled_guidance` | `number` |
+| `high_noise_sample_params.guidance.slg.layers` | `array<integer>` |
+| `high_noise_sample_params.guidance.slg.layer_start` | `number` |
+| `high_noise_sample_params.guidance.slg.layer_end` | `number` |
+| `high_noise_sample_params.guidance.slg.scale` | `number` |
+
+Other native fields:
+
+| Field | Type |
+| --- | --- |
+| `vae_tiling_params` | `object` |
+| `cache_mode` | `string` |
+| `cache_option` | `string` |
+| `scm_mask` | `string` |
+| `scm_policy_dynamic` | `boolean` |
+
+HTTP-only output fields:
+
+| Field | Type |
+| --- | --- |
+| `output_format` | `string` |
+| `output_compression` | `integer` |
+
+For `vid_gen`, `output_format` and `output_compression` control container encoding.
+`fps` is request metadata for the generated sequence and is echoed in the completed job result.
+
+Allowed `output_format` values:
+
+- `webm`
+- `webp`
+- `avi`
+
+Output format behavior:
+
+- `output_format` defaults to `webm`
+- `webp` means animated WebP
+- `avi` means MJPG AVI
+- `webm` requires the server to be built with WebM support; otherwise the request returns `400`
+
+### Result Payload
+
+Completed jobs return one encoded container payload, not a list of per-frame images.
+
+Result fields:
+
+- `result.b64_json` contains the whole encoded container file as base64
+- `result.mime_type` identifies the media type
+- `result.output_format` echoes the selected container format
+- `result.fps` echoes the effective playback FPS
+- `result.frame_count` reports the actual decoded frame count used to build the container
+
+Expected MIME types:
+
+| `output_format` | `mime_type` |
+| --- | --- |
+| `webm` | `video/webm` |
+| `webp` | `image/webp` |
+| `avi` | `video/x-msvideo` |
+
+### Optional Field Handling
+
+Optional sampling fields may be omitted.
+
+When omitted, backend defaults apply to these fields:
+
+- `sample_params.scheduler`
+- `sample_params.sample_method`
+- `sample_params.eta`
+- `sample_params.flow_shift`
+- `sample_params.guidance.img_cfg`
+- `high_noise_sample_params.scheduler`
+- `high_noise_sample_params.sample_method`
+- `high_noise_sample_params.eta`
+- `high_noise_sample_params.flow_shift`
+- `high_noise_sample_params.guidance.img_cfg`
+
+`high_noise_sample_params` may also be omitted entirely.
+
+### Frame Count Semantics
+
+`video_frames` is the requested target length, but the current core video path internally normalizes the effective frame count to the largest `4n + 1` value that does not exceed the requested count.
+
+Examples:
+
+- `video_frames = 33` stays `33`
+- `video_frames = 34` becomes `33`
+- `video_frames = 32` becomes `29`
+
+The completed job payload includes the actual decoded `frame_count`.
+
+### Completion Result
+
+Example completed job:
+
+```json
+{
+  "id": "job_01HTXYZVID",
+  "kind": "vid_gen",
+  "status": "completed",
+  "created": 1775401200,
+  "started": 1775401203,
+  "completed": 1775401215,
+  "queue_position": 0,
+  "result": {
+    "output_format": "webm",
+    "mime_type": "video/webm",
+    "fps": 16,
+    "frame_count": 33,
+    "b64_json": "GkXfo59ChoEBQveBAULygQRC84EIQo..."
+  },
+  "error": null
+}
+```
+
+The response returns the encoded `.webm`, animated `.webp`, or `.avi` container payload directly.
+
+### Failure Result
+
+Example failed job:
+
+```json
+{
+  "id": "job_01HTXYZVID",
+  "kind": "vid_gen",
+  "status": "failed",
+  "created": 1775401200,
+  "started": 1775401203,
+  "completed": 1775401204,
+  "queue_position": 0,
+  "result": null,
+  "error": {
+    "code": "generation_failed",
+    "message": "generate_video returned no results"
+  }
+}
+```
+
+### Cancelled Result
+
+Example cancelled job:
+
+```json
+{
+  "id": "job_01HTXYZVID",
+  "kind": "vid_gen",
+  "status": "cancelled",
+  "created": 1775401200,
+  "started": null,
+  "completed": 1775401202,
+  "queue_position": 0,
+  "result": null,
+  "error": {
+    "code": "cancelled",
+    "message": "job cancelled by client"
+  }
+}
+```
+
+### Submission Errors
+
+`POST /sdcpp/v1/vid_gen` may return:
+
+- `202 Accepted` when the job is created
+- `400 Bad Request` for an empty body, unsupported model mode, invalid JSON, invalid generation parameters, or an unsupported output format
+- `429 Too Many Requests` when the job queue is full
+- `500 Internal Server Error` for unexpected server exceptions during submission

examples/server/async_jobs.cpp

@@ -95,8 +95,12 @@ bool cancel_queued_job(AsyncJobManager& manager, AsyncGenerationJob& job) {
     job.status       = AsyncJobStatus::Cancelled;
     job.completed_at = unix_timestamp_now();
     job.result_images_b64.clear();
-    job.error_code    = "cancelled";
-    job.error_message = "job cancelled by client";
+    job.result_media_b64.clear();
+    job.result_media_mime_type.clear();
+    job.result_frame_count = 0;
+    job.result_fps         = 0;
+    job.error_code         = "cancelled";
+    job.error_message      = "job cancelled by client";
     return true;
 }
 
@@ -122,14 +126,24 @@ json make_async_job_json(const AsyncJobManager& manager, const AsyncGenerationJo
     }
 
     if (job.status == AsyncJobStatus::Completed) {
-        json images = json::array();
-        for (size_t i = 0; i < job.result_images_b64.size(); ++i) {
-            images.push_back({{"index", i}, {"b64_json", job.result_images_b64[i]}});
+        if (job.kind == AsyncJobKind::VidGen) {
+            result["result"] = {
+                {"output_format", job.vid_gen.output_format},
+                {"mime_type", job.result_media_mime_type},
+                {"fps", job.result_fps},
+                {"frame_count", job.result_frame_count},
+                {"b64_json", job.result_media_b64},
+            };
+        } else {
+            json images = json::array();
+            for (size_t i = 0; i < job.result_images_b64.size(); ++i) {
+                images.push_back({{"index", i}, {"b64_json", job.result_images_b64[i]}});
+            }
+            result["result"] = {
+                {"output_format", job.img_gen.output_format},
+                {"images", images},
+            };
         }
-        result["result"] = {
-            {"output_format", job.img_gen.output_format},
-            {"images", images},
-        };
         result["error"] = nullptr;
     } else if (job.status == AsyncJobStatus::Failed ||
                job.status == AsyncJobStatus::Cancelled) {
@@ -156,16 +170,15 @@ bool execute_img_gen_job(ServerRuntime& runtime,
     sd_img_gen_params_t params = job.img_gen.to_sd_img_gen_params_t();
 
     SDImageVec results;
-    int num_results = 0;
 
     {
         std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
         sd_image_t* raw_results = generate_image(runtime.sd_ctx, &params);
-        num_results             = params.batch_count;
-        results.adopt(raw_results, num_results);
+        results.adopt(raw_results, params.batch_count);
     }
 
-    if (results.empty() || num_results <= 0) {
+    const int num_results = results.count();
+    if (num_results <= 0) {
         error_message = "generate_image returned no results";
         return false;
     }
@@ -208,6 +221,47 @@ bool execute_img_gen_job(ServerRuntime& runtime,
     return true;
 }
 
+bool execute_vid_gen_job(ServerRuntime& runtime,
+                         AsyncGenerationJob& job,
+                         std::string& output_media_b64,
+                         std::string& output_media_mime_type,
+                         int& output_frame_count,
+                         int& output_fps,
+                         std::string& error_message) {
+    sd_vid_gen_params_t params = job.vid_gen.to_sd_vid_gen_params_t();
+
+    SDImageVec results;
+    int num_results = 0;
+
+    {
+        std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
+        sd_image_t* raw_results = generate_video(runtime.sd_ctx, &params, &num_results);
+        results.adopt(raw_results, num_results);
+    }
+
+    num_results = results.count();
+    if (num_results <= 0) {
+        error_message = "generate_video returned no results";
+        return false;
+    }
+
+    std::vector<uint8_t> video_bytes = create_video_from_sd_images_to_vector(job.vid_gen.output_format,
+                                                                             results.data(),
+                                                                             num_results,
+                                                                             job.vid_gen.gen_params.fps,
+                                                                             job.vid_gen.output_compression);
+    if (video_bytes.empty()) {
+        error_message = "failed to encode generated video container";
+        return false;
+    }
+
+    output_media_b64       = base64_encode(video_bytes);
+    output_media_mime_type = video_mime_type(job.vid_gen.output_format);
+    output_frame_count     = num_results;
+    output_fps             = job.vid_gen.gen_params.fps;
+    return true;
+}
+
 void async_job_worker(ServerRuntime& runtime) {
     AsyncJobManager& manager = *runtime.async_job_manager;
 
@@ -240,11 +294,23 @@ void async_job_worker(ServerRuntime& runtime) {
         }
 
         std::vector<std::string> output_images;
+        std::string output_media_b64;
+        std::string output_media_mime_type;
+        int output_frame_count = 0;
+        int output_fps         = 0;
         std::string error_message;
         bool ok = false;
 
         if (job->kind == AsyncJobKind::ImgGen) {
             ok = execute_img_gen_job(runtime, *job, output_images, error_message);
+        } else if (job->kind == AsyncJobKind::VidGen) {
+            ok = execute_vid_gen_job(runtime,
+                                     *job,
+                                     output_media_b64,
+                                     output_media_mime_type,
+                                     output_frame_count,
+                                     output_fps,
+                                     error_message);
         } else {
             error_message = "unsupported job kind";
         }
@@ -258,8 +324,12 @@ void async_job_worker(ServerRuntime& runtime) {
 
             job->completed_at = unix_timestamp_now();
             if (ok) {
-                job->status            = AsyncJobStatus::Completed;
-                job->result_images_b64 = std::move(output_images);
+                job->status                 = AsyncJobStatus::Completed;
+                job->result_images_b64      = std::move(output_images);
+                job->result_media_b64       = std::move(output_media_b64);
+                job->result_media_mime_type = std::move(output_media_mime_type);
+                job->result_frame_count     = output_frame_count;
+                job->result_fps             = output_fps;
                 job->error_code.clear();
                 job->error_message.clear();
             } else {
@@ -267,6 +337,10 @@ void async_job_worker(ServerRuntime& runtime) {
                 job->error_code    = "generation_failed";
                 job->error_message = error_message.empty() ? "unknown generation error" : error_message;
                 job->result_images_b64.clear();
+                job->result_media_b64.clear();
+                job->result_media_mime_type.clear();
+                job->result_frame_count = 0;
+                job->result_fps         = 0;
             }
 
             purge_expired_jobs(manager);

examples/server/async_jobs.h

@@ -36,7 +36,12 @@ struct AsyncGenerationJob {
     int64_t started_at    = 0;
     int64_t completed_at  = 0;
     ImgGenJobRequest img_gen;
+    VidGenJobRequest vid_gen;
     std::vector<std::string> result_images_b64;
+    std::string result_media_b64;
+    std::string result_media_mime_type;
+    int result_frame_count = 0;
+    int result_fps         = 0;
     std::string error_code;
     std::string error_message;
 };
@@ -63,4 +68,11 @@ bool execute_img_gen_job(ServerRuntime& runtime,
                          AsyncGenerationJob& job,
                          std::vector<std::string>& output_images,
                          std::string& error_message);
+bool execute_vid_gen_job(ServerRuntime& runtime,
+                         AsyncGenerationJob& job,
+                         std::string& output_media_b64,
+                         std::string& output_media_mime_type,
+                         int& output_frame_count,
+                         int& output_fps,
+                         std::string& error_message);
 void async_job_worker(ServerRuntime& runtime);

examples/server/routes_openai.cpp

@@ -253,6 +253,12 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
 
     svr.Post("/v1/images/generations", [runtime](const httplib::Request& req, httplib::Response& res) {
         try {
+            if (!runtime_supports_generation_mode(*runtime, IMG_GEN)) {
+                res.status = 400;
+                res.set_content(json({{"error", unsupported_generation_mode_error(IMG_GEN)}}).dump(), "application/json");
+                return;
+            }
+
             ImgGenJobRequest request;
             std::string error_message;
             if (!build_openai_generation_request(req, *runtime, request, error_message)) {
@@ -319,6 +325,12 @@ void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
 
     svr.Post("/v1/images/edits", [runtime](const httplib::Request& req, httplib::Response& res) {
         try {
+            if (!runtime_supports_generation_mode(*runtime, IMG_GEN)) {
+                res.status = 400;
+                res.set_content(json({{"error", unsupported_generation_mode_error(IMG_GEN)}}).dump(), "application/json");
+                return;
+            }
+
             ImgGenJobRequest request;
             std::string error_message;
             if (!build_openai_edit_request(req, *runtime, request, error_message)) {

examples/server/routes_sdapi.cpp

@@ -246,6 +246,11 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 res.set_content(R"({"error":"empty body"})", "application/json");
                 return;
             }
+            if (!runtime_supports_generation_mode(*runtime, IMG_GEN)) {
+                res.status = 400;
+                res.set_content(json({{"error", unsupported_generation_mode_error(IMG_GEN)}}).dump(), "application/json");
+                return;
+            }
 
             json j = json::parse(req.body);
             ImgGenJobRequest request;

examples/server/routes_sdcpp.cpp

@@ -75,48 +75,33 @@ static fs::path resolve_display_model_path(const ServerRuntime& runtime) {
     return {};
 }
 
-static json make_capabilities_json(ServerRuntime& runtime) {
-    refresh_lora_cache(runtime);
-
-    AsyncJobManager& manager  = *runtime.async_job_manager;
-    const auto& defaults      = *runtime.default_gen_params;
-    const auto& sample_params = defaults.sample_params;
-    const auto& guidance      = sample_params.guidance;
-    const fs::path model_path = resolve_display_model_path(runtime);
-    json samplers             = json::array();
-    json schedulers           = json::array();
-    json output_formats       = json::array({"png", "jpeg"});
-    json available_loras      = json::array();
-
-    for (int i = 0; i < SAMPLE_METHOD_COUNT; ++i) {
-        samplers.push_back(sd_sample_method_name((sample_method_t)i));
-    }
-
-    for (int i = 0; i < SCHEDULER_COUNT; ++i) {
-        schedulers.push_back(sd_scheduler_name((scheduler_t)i));
-    }
-
-#ifdef SD_USE_WEBP
-    output_formats.push_back("webp");
-#endif
-
-    {
-        std::lock_guard<std::mutex> lock(*runtime.lora_mutex);
-        for (const auto& entry : *runtime.lora_cache) {
-            available_loras.push_back({
-                {"name", entry.name},
-                {"path", entry.path},
-            });
-        }
-    }
-
-    json result;
-    result["model"] = {
-        {"name", model_path.filename().u8string()},
-        {"stem", model_path.stem().u8string()},
-        {"path", model_path.u8string()},
+static json make_sample_params_json(const sd_sample_params_t& sample_params, const std::vector<int>& skip_layers) {
+    const auto& guidance = sample_params.guidance;
+    return {
+        {"scheduler", capability_scheduler_name(sample_params.scheduler)},
+        {"sample_method", capability_sample_method_name(sample_params.sample_method)},
+        {"sample_steps", sample_params.sample_steps},
+        {"eta", finite_number_or_null(sample_params.eta)},
+        {"shifted_timestep", sample_params.shifted_timestep},
+        {"flow_shift", finite_number_or_null(sample_params.flow_shift)},
+        {"guidance",
+         {
+             {"txt_cfg", guidance.txt_cfg},
+             {"img_cfg", finite_number_or_null(guidance.img_cfg)},
+             {"distilled_guidance", guidance.distilled_guidance},
+             {"slg",
+              {
+                  {"layers", skip_layers},
+                  {"layer_start", guidance.slg.layer_start},
+                  {"layer_end", guidance.slg.layer_end},
+                  {"scale", guidance.slg.scale},
+              }},
+         }},
     };
-    result["defaults"] = {
+}
+
+static json make_img_gen_defaults_json(const SDGenerationParams& defaults, const std::string& output_format) {
+    return {
         {"prompt", defaults.prompt},
         {"negative_prompt", defaults.negative_prompt},
         {"clip_skip", defaults.clip_skip},
@@ -128,59 +113,177 @@ static json make_capabilities_json(ServerRuntime& runtime) {
         {"auto_resize_ref_image", defaults.auto_resize_ref_image},
         {"increase_ref_index", defaults.increase_ref_index},
         {"control_strength", defaults.control_strength},
-        {"sample_params",
-         {
-             {"scheduler", capability_scheduler_name(sample_params.scheduler)},
-             {"sample_method", capability_sample_method_name(sample_params.sample_method)},
-             {"sample_steps", sample_params.sample_steps},
-             {"eta", finite_number_or_null(sample_params.eta)},
-             {"shifted_timestep", sample_params.shifted_timestep},
-             {"flow_shift", finite_number_or_null(sample_params.flow_shift)},
-             {"guidance",
-              {
-                  {"txt_cfg", guidance.txt_cfg},
-                  {"img_cfg", finite_number_or_null(guidance.img_cfg)},
-                  {"distilled_guidance", guidance.distilled_guidance},
-                  {"slg",
-                   {
-                       {"layers", defaults.skip_layers},
-                       {"layer_start", guidance.slg.layer_start},
-                       {"layer_end", guidance.slg.layer_end},
-                       {"scale", guidance.slg.scale},
-                   }},
-              }},
-         }},
+        {"sample_params", make_sample_params_json(defaults.sample_params, defaults.skip_layers)},
         {"vae_tiling_params", make_vae_tiling_json(defaults.vae_tiling_params)},
         {"cache_mode", defaults.cache_mode},
         {"cache_option", defaults.cache_option},
         {"scm_mask", defaults.scm_mask},
         {"scm_policy_dynamic", defaults.scm_policy_dynamic},
-        {"output_format", "png"},
+        {"output_format", output_format},
         {"output_compression", 100},
     };
-    result["limits"] = {
-        {"min_width", 64},
-        {"max_width", 4096},
-        {"min_height", 64},
-        {"max_height", 4096},
-        {"max_batch_count", 8},
-        {"max_queue_size", manager.max_pending_jobs},
+}
+
+static json make_vid_gen_defaults_json(const SDGenerationParams& defaults, const std::string& output_format) {
+    return {
+        {"prompt", defaults.prompt},
+        {"negative_prompt", defaults.negative_prompt},
+        {"clip_skip", defaults.clip_skip},
+        {"width", defaults.width > 0 ? defaults.width : 512},
+        {"height", defaults.height > 0 ? defaults.height : 512},
+        {"strength", defaults.strength},
+        {"seed", defaults.seed},
+        {"video_frames", defaults.video_frames},
+        {"fps", defaults.fps},
+        {"moe_boundary", defaults.moe_boundary},
+        {"vace_strength", defaults.vace_strength},
+        {"sample_params", make_sample_params_json(defaults.sample_params, defaults.skip_layers)},
+        {"high_noise_sample_params", make_sample_params_json(defaults.high_noise_sample_params, defaults.high_noise_skip_layers)},
+        {"vae_tiling_params", make_vae_tiling_json(defaults.vae_tiling_params)},
+        {"cache_mode", defaults.cache_mode},
+        {"cache_option", defaults.cache_option},
+        {"scm_mask", defaults.scm_mask},
+        {"scm_policy_dynamic", defaults.scm_policy_dynamic},
+        {"output_format", output_format},
+        {"output_compression", 100},
     };
-    result["samplers"]       = samplers;
-    result["schedulers"]     = schedulers;
-    result["output_formats"] = output_formats;
-    result["features"]       = {
-              {"init_image", true},
-              {"mask_image", true},
-              {"control_image", true},
-              {"ref_images", true},
-              {"lora", true},
-              {"vae_tiling", true},
-              {"cache", true},
+}
+
+static json make_img_gen_features_json() {
+    return {
+        {"init_image", true},
+        {"mask_image", true},
+        {"control_image", true},
+        {"ref_images", true},
+        {"lora", true},
+        {"vae_tiling", true},
+        {"cache", true},
+        {"cancel_queued", true},
+        {"cancel_generating", false},
+    };
+}
+
+static json make_vid_gen_features_json() {
+    return {
+        {"init_image", true},
+        {"end_image", true},
+        {"control_frames", true},
+        {"high_noise_sample_params", true},
+        {"lora", true},
+        {"vae_tiling", true},
+        {"cache", true},
+        {"cancel_queued", true},
+        {"cancel_generating", false},
+    };
+}
+
+static json make_capabilities_json(ServerRuntime& runtime) {
+    refresh_lora_cache(runtime);
+
+    AsyncJobManager& manager  = *runtime.async_job_manager;
+    const auto& defaults      = *runtime.default_gen_params;
+    const fs::path model_path = resolve_display_model_path(runtime);
+    const bool supports_img   = runtime_supports_generation_mode(runtime, IMG_GEN);
+    const bool supports_vid   = runtime_supports_generation_mode(runtime, VID_GEN);
+    json samplers             = json::array();
+    json schedulers           = json::array();
+    json image_output_formats = supported_img_output_formats();
+    json video_output_formats = supported_vid_output_formats();
+    json available_loras      = json::array();
+    json supported_modes      = json::array();
+
+    for (int i = 0; i < SAMPLE_METHOD_COUNT; ++i) {
+        samplers.push_back(sd_sample_method_name((sample_method_t)i));
+    }
+
+    for (int i = 0; i < SCHEDULER_COUNT; ++i) {
+        schedulers.push_back(sd_scheduler_name((scheduler_t)i));
+    }
+
+    {
+        std::lock_guard<std::mutex> lock(*runtime.lora_mutex);
+        for (const auto& entry : *runtime.lora_cache) {
+            available_loras.push_back({
+                {"name", entry.name},
+                {"path", entry.path},
+            });
+        }
+    }
+
+    if (supports_img) {
+        supported_modes.push_back("img_gen");
+    }
+    if (supports_vid) {
+        supported_modes.push_back("vid_gen");
+    }
+
+    std::string default_img_output_format = "png";
+    std::string default_vid_output_format = "avi";
+    if (!image_output_formats.empty()) {
+        default_img_output_format = image_output_formats[0].get<std::string>();
+    }
+    if (!video_output_formats.empty()) {
+        default_vid_output_format = video_output_formats[0].get<std::string>();
+    }
+
+    json defaults_by_mode       = json::object();
+    json output_formats_by_mode = json::object();
+    json features_by_mode       = json::object();
+    if (supports_img) {
+        defaults_by_mode["img_gen"]       = make_img_gen_defaults_json(defaults, default_img_output_format);
+        output_formats_by_mode["img_gen"] = image_output_formats;
+        features_by_mode["img_gen"]       = make_img_gen_features_json();
+    }
+    if (supports_vid) {
+        defaults_by_mode["vid_gen"]       = make_vid_gen_defaults_json(defaults, default_vid_output_format);
+        output_formats_by_mode["vid_gen"] = video_output_formats;
+        features_by_mode["vid_gen"]       = make_vid_gen_features_json();
+    }
+
+    json top_level_defaults       = json::object();
+    json top_level_output_formats = json::array();
+    json top_level_features       = {
               {"cancel_queued", true},
               {"cancel_generating", false},
     };
-    result["loras"] = available_loras;
+    std::string current_mode = "";
+    if (supports_img) {
+        current_mode             = "img_gen";
+        top_level_defaults       = defaults_by_mode["img_gen"];
+        top_level_output_formats = output_formats_by_mode["img_gen"];
+        top_level_features       = features_by_mode["img_gen"];
+    } else if (supports_vid) {
+        current_mode             = "vid_gen";
+        top_level_defaults       = defaults_by_mode["vid_gen"];
+        top_level_output_formats = output_formats_by_mode["vid_gen"];
+        top_level_features       = features_by_mode["vid_gen"];
+    }
+
+    json result;
+    result["model"] = {
+        {"name", model_path.filename().u8string()},
+        {"stem", model_path.stem().u8string()},
+        {"path", model_path.u8string()},
+    };
+    result["current_mode"]     = current_mode;
+    result["supported_modes"]  = supported_modes;
+    result["defaults"]         = top_level_defaults;
+    result["defaults_by_mode"] = defaults_by_mode;
+    result["limits"]           = {
+                  {"min_width", 64},
+                  {"max_width", 4096},
+                  {"min_height", 64},
+                  {"max_height", 4096},
+                  {"max_batch_count", 8},
+                  {"max_queue_size", manager.max_pending_jobs},
+    };
+    result["samplers"]               = samplers;
+    result["schedulers"]             = schedulers;
+    result["output_formats"]         = top_level_output_formats;
+    result["output_formats_by_mode"] = output_formats_by_mode;
+    result["features"]               = top_level_features;
+    result["features_by_mode"]       = features_by_mode;
+    result["loras"]                  = available_loras;
     return result;
 }
 
@@ -211,6 +314,33 @@ static bool parse_img_gen_request(const json& body,
     return true;
 }
 
+static bool parse_vid_gen_request(const json& body,
+                                  ServerRuntime& runtime,
+                                  VidGenJobRequest& request,
+                                  std::string& error_message) {
+    request.gen_params = *runtime.default_gen_params;
+
+    refresh_lora_cache(runtime);
+    if (!request.gen_params.from_json_str(body.dump(), [&](const std::string& path) {
+            return get_lora_full_path(runtime, path);
+        })) {
+        error_message = "invalid generation parameters";
+        return false;
+    }
+
+    std::string output_format = body.value("output_format", "webm");
+    int output_compression    = body.value("output_compression", 100);
+    if (!assign_output_options(request, output_format, output_compression, error_message)) {
+        return false;
+    }
+    // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
+    if (!request.gen_params.resolve_and_validate(VID_GEN, "", true)) {
+        error_message = "invalid generation parameters";
+        return false;
+    }
+    return true;
+}
+
 void register_sdcpp_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
     ServerRuntime* runtime = &rt;
 
@@ -226,6 +356,11 @@ void register_sdcpp_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
                 res.set_content(R"({"error":"empty body"})", "application/json");
                 return;
             }
+            if (!runtime_supports_generation_mode(*runtime, IMG_GEN)) {
+                res.status = 400;
+                res.set_content(json({{"error", unsupported_generation_mode_error(IMG_GEN)}}).dump(), "application/json");
+                return;
+            }
 
             json body = json::parse(req.body);
             ImgGenJobRequest request;
@@ -276,9 +411,66 @@ void register_sdcpp_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
         }
     });
 
-    svr.Post("/sdcpp/v1/vid_gen", [](const httplib::Request&, httplib::Response& res) {
-        res.status = 501;
-        res.set_content(R"({"error":"vid_gen is reserved and not implemented yet"})", "application/json");
+    svr.Post("/sdcpp/v1/vid_gen", [runtime](const httplib::Request& req, httplib::Response& res) {
+        try {
+            if (req.body.empty()) {
+                res.status = 400;
+                res.set_content(R"({"error":"empty body"})", "application/json");
+                return;
+            }
+            if (!runtime_supports_generation_mode(*runtime, VID_GEN)) {
+                res.status = 400;
+                res.set_content(json({{"error", unsupported_generation_mode_error(VID_GEN)}}).dump(), "application/json");
+                return;
+            }
+
+            json body = json::parse(req.body);
+            VidGenJobRequest request;
+            std::string error_message;
+            if (!parse_vid_gen_request(body, *runtime, request, error_message)) {
+                res.status = 400;
+                res.set_content(json({{"error", error_message}}).dump(), "application/json");
+                return;
+            }
+
+            AsyncJobManager& manager                = *runtime->async_job_manager;
+            std::shared_ptr<AsyncGenerationJob> job = std::make_shared<AsyncGenerationJob>();
+            job->kind                               = AsyncJobKind::VidGen;
+            job->status                             = AsyncJobStatus::Queued;
+            job->created_at                         = unix_timestamp_now();
+            job->vid_gen                            = std::move(request);
+
+            {
+                std::lock_guard<std::mutex> lock(manager.mutex);
+                purge_expired_jobs(manager);
+                if (count_pending_jobs(manager) >= manager.max_pending_jobs) {
+                    res.status = 429;
+                    res.set_content(R"({"error":"job queue is full"})", "application/json");
+                    return;
+                }
+                job->id               = make_async_job_id(manager);
+                manager.jobs[job->id] = job;
+                manager.queue.push_back(job->id);
+            }
+
+            manager.cv.notify_one();
+
+            json out;
+            out["id"]       = job->id;
+            out["kind"]     = async_job_kind_name(job->kind);
+            out["status"]   = async_job_status_name(job->status);
+            out["created"]  = job->created_at;
+            out["poll_url"] = "/sdcpp/v1/jobs/" + job->id;
+
+            res.status = 202;
+            res.set_content(out.dump(), "application/json");
+        } catch (const json::parse_error& e) {
+            res.status = 400;
+            res.set_content(json({{"error", "invalid json"}, {"message", e.what()}}).dump(), "application/json");
+        } catch (const std::exception& e) {
+            res.status = 500;
+            res.set_content(json({{"error", "server_error"}, {"message", e.what()}}).dump(), "application/json");
+        }
     });
 
     svr.Get(R"(/sdcpp/v1/jobs/([A-Za-z0-9_\-]+))", [runtime](const httplib::Request& req, httplib::Response& res) {

examples/server/runtime.cpp

@@ -45,6 +45,44 @@ std::string normalize_output_format(std::string output_format) {
     return output_format;
 }
 
+std::vector<std::string> supported_img_output_formats(bool allow_webp) {
+    std::vector<std::string> formats = {"png", "jpeg"};
+#ifdef SD_USE_WEBP
+    if (allow_webp) {
+        formats.push_back("webp");
+    }
+#else
+    (void)allow_webp;
+#endif
+    return formats;
+}
+
+std::vector<std::string> supported_vid_output_formats() {
+    std::vector<std::string> formats;
+#ifdef SD_USE_WEBM
+    formats.push_back("webm");
+#endif
+#ifdef SD_USE_WEBP
+    formats.push_back("webp");
+#endif
+    formats.push_back("avi");
+    return formats;
+}
+
+static std::string valid_vid_output_formats_message() {
+    const std::vector<std::string> formats = supported_vid_output_formats();
+
+    std::string message = "invalid output_format, must be one of [";
+    for (size_t i = 0; i < formats.size(); ++i) {
+        if (i > 0) {
+            message += ", ";
+        }
+        message += formats[i];
+    }
+    message += "]";
+    return message;
+}
+
 bool assign_output_options(ImgGenJobRequest& request,
                            std::string output_format,
                            int output_compression,
@@ -53,19 +91,88 @@ bool assign_output_options(ImgGenJobRequest& request,
     request.output_format      = normalize_output_format(std::move(output_format));
     request.output_compression = std::clamp(output_compression, 0, 100);
 
-    const bool valid_format = request.output_format == "png" ||
-                              request.output_format == "jpeg" ||
-                              (allow_webp && request.output_format == "webp");
+    const std::vector<std::string> valid_formats = supported_img_output_formats(allow_webp);
+    const bool valid_format                      = std::find(valid_formats.begin(),
+                                                             valid_formats.end(),
+                                                             request.output_format) != valid_formats.end();
     if (!valid_format) {
-        error_message = allow_webp
-                            ? "invalid output_format, must be one of [png, jpeg, webp]"
-                            : "invalid output_format, must be one of [png, jpeg]";
+        error_message = "invalid output_format, must be one of [";
+        for (size_t i = 0; i < valid_formats.size(); ++i) {
+            if (i > 0) {
+                error_message += ", ";
+            }
+            error_message += valid_formats[i];
+        }
+        error_message += "]";
         return false;
     }
 
     return true;
 }
 
+bool assign_output_options(VidGenJobRequest& request,
+                           std::string output_format,
+                           int output_compression,
+                           std::string& error_message) {
+    request.output_format      = normalize_output_format(std::move(output_format));
+    request.output_compression = std::clamp(output_compression, 0, 100);
+
+    if (request.output_format == "avi") {
+        return true;
+    }
+
+    if (request.output_format == "webm") {
+#ifdef SD_USE_WEBM
+        return true;
+#else
+        error_message = valid_vid_output_formats_message();
+        return false;
+#endif
+    }
+
+    if (request.output_format == "webp") {
+#ifdef SD_USE_WEBP
+        return true;
+#else
+        error_message = valid_vid_output_formats_message();
+        return false;
+#endif
+    }
+
+    error_message = valid_vid_output_formats_message();
+    return false;
+}
+
+std::string video_mime_type(const std::string& output_format) {
+    if (output_format == "webm") {
+        return "video/webm";
+    }
+    if (output_format == "webp") {
+        return "image/webp";
+    }
+    return "video/x-msvideo";
+}
+
+bool runtime_supports_generation_mode(const ServerRuntime& runtime, SDMode mode) {
+    if (mode == VID_GEN) {
+        return sd_ctx_supports_video_generation(runtime.sd_ctx);
+    }
+    if (mode == IMG_GEN) {
+        return sd_ctx_supports_image_generation(runtime.sd_ctx);
+    }
+    return true;
+}
+
+std::string unsupported_generation_mode_error(SDMode mode) {
+    if (mode == VID_GEN) {
+        return "loaded model does not support vid_gen";
+    }
+    if (mode == IMG_GEN) {
+        return "loaded model does not support img_gen";
+    }
+    return "loaded model does not support requested mode";
+}
+
 ArgOptions SDSvrParams::get_options() {
     ArgOptions options;
 

examples/server/runtime.h

@@ -58,13 +58,32 @@ struct ImgGenJobRequest {
     }
 };
 
+struct VidGenJobRequest {
+    SDGenerationParams gen_params;
+    std::string output_format = "webm";
+    int output_compression    = 100;
+
+    sd_vid_gen_params_t to_sd_vid_gen_params_t() {
+        return gen_params.to_sd_vid_gen_params_t();
+    }
+};
+
 std::string base64_encode(const std::vector<uint8_t>& bytes);
 std::string normalize_output_format(std::string output_format);
+std::vector<std::string> supported_img_output_formats(bool allow_webp = true);
+std::vector<std::string> supported_vid_output_formats();
 bool assign_output_options(ImgGenJobRequest& request,
                            std::string output_format,
                            int output_compression,
                            bool allow_webp,
                            std::string& error_message);
+bool assign_output_options(VidGenJobRequest& request,
+                           std::string output_format,
+                           int output_compression,
+                           std::string& error_message);
+std::string video_mime_type(const std::string& output_format);
+bool runtime_supports_generation_mode(const ServerRuntime& runtime, SDMode mode);
+std::string unsupported_generation_mode_error(SDMode mode);
 void refresh_lora_cache(ServerRuntime& rt);
 std::string get_lora_full_path(ServerRuntime& rt, const std::string& path);
 int64_t unix_timestamp_now();

format-code.sh

@@ -1,4 +1,4 @@
-for f in src/*.cpp src/*.h src/*.hpp src/vocab/*.h src/vocab/*.cpp \
+for f in src/*.cpp src/*.h src/*.hpp src/tokenizers/*.h src/tokenizers/*.cpp src/tokenizers/vocab/*.h src/tokenizers/vocab/*.cpp \
          examples/cli/*.cpp examples/cli/*.h examples/server/*.cpp \
          examples/common/*.hpp examples/common/*.h examples/common/*.cpp; do
   [[ "$f" == vocab* ]] && continue

include/stable-diffusion.h

@@ -50,6 +50,7 @@ enum sample_method_t {
     TCD_SAMPLE_METHOD,
     RES_MULTISTEP_SAMPLE_METHOD,
     RES_2S_SAMPLE_METHOD,
+    ER_SDE_SAMPLE_METHOD,
     SAMPLE_METHOD_COUNT
 };
 
@@ -347,6 +348,8 @@ SD_API void sd_set_progress_callback(sd_progress_cb_t cb, void* data);
 SD_API void sd_set_preview_callback(sd_preview_cb_t cb, enum preview_t mode, int interval, bool denoised, bool noisy, void* data);
 SD_API int32_t sd_get_num_physical_cores();
 SD_API const char* sd_get_system_info();
+SD_API bool sd_ctx_supports_image_generation(const sd_ctx_t* sd_ctx);
+SD_API bool sd_ctx_supports_video_generation(const sd_ctx_t* sd_ctx);
 
 SD_API const char* sd_type_name(enum sd_type_t type);
 SD_API enum sd_type_t str_to_sd_type(const char* str);

src/auto_encoder_kl.hpp

@@ -533,7 +533,7 @@ public:
                        const std::string& prefix                      = "")
         : version(version), decode_only(decode_only), use_video_decoder(use_video_decoder) {
         if (sd_version_is_dit(version)) {
-            if (sd_version_is_flux2(version)) {
+            if (sd_version_uses_flux2_vae(version)) {
                 dd_config.z_channels = 32;
                 embed_dim            = 32;
             } else {
@@ -578,7 +578,7 @@ public:
 
     ggml_tensor* decode(GGMLRunnerContext* ctx, ggml_tensor* z) {
         // z: [N, z_channels, h, w]
-        if (sd_version_is_flux2(version)) {
+        if (sd_version_uses_flux2_vae(version)) {
             // [N, C*p*p, h, w] -> [N, C, h*p, w*p]
             int64_t p = 2;
 
@@ -617,7 +617,7 @@ public:
             auto quant_conv = std::dynamic_pointer_cast<Conv2d>(blocks["quant_conv"]);
             z               = quant_conv->forward(ctx, z);  // [N, 2*embed_dim, h/8, w/8]
         }
-        if (sd_version_is_flux2(version)) {
+        if (sd_version_uses_flux2_vae(version)) {
             z = ggml_ext_chunk(ctx->ggml_ctx, z, 2, 2)[0];
 
             // [N, C, H, W] -> [N, C*p*p, H/p, W/p]
@@ -640,7 +640,7 @@ public:
 
     int get_encoder_output_channels() {
         int factor = dd_config.double_z ? 2 : 1;
-        if (sd_version_is_flux2(version)) {
+        if (sd_version_uses_flux2_vae(version)) {
             return dd_config.z_channels * 4;
         }
         return dd_config.z_channels * factor;
@@ -673,7 +673,7 @@ struct AutoEncoderKL : public VAE {
         } else if (sd_version_is_flux(version) || sd_version_is_z_image(version)) {
             scale_factor = 0.3611f;
             shift_factor = 0.1159f;
-        } else if (sd_version_is_flux2(version)) {
+        } else if (sd_version_uses_flux2_vae(version)) {
             scale_factor = 1.0f;
             shift_factor = 0.f;
         }
@@ -747,7 +747,7 @@ struct AutoEncoderKL : public VAE {
     }
 
     sd::Tensor<float> vae_output_to_latents(const sd::Tensor<float>& vae_output, std::shared_ptr<RNG> rng) override {
-        if (sd_version_is_flux2(version)) {
+        if (sd_version_uses_flux2_vae(version)) {
             return vae_output;
         } else if (version == VERSION_SD1_PIX2PIX) {
             return sd::ops::chunk(vae_output, 2, 2)[0];
@@ -758,7 +758,7 @@ struct AutoEncoderKL : public VAE {
 
     std::pair<sd::Tensor<float>, sd::Tensor<float>> get_latents_mean_std(const sd::Tensor<float>& latents, int channel_dim) {
         GGML_ASSERT(channel_dim >= 0 && static_cast<size_t>(channel_dim) < static_cast<size_t>(latents.dim()));
-        if (sd_version_is_flux2(version)) {
+        if (sd_version_uses_flux2_vae(version)) {
             GGML_ASSERT(latents.shape()[channel_dim] == 128);
             std::vector<int64_t> stats_shape(static_cast<size_t>(latents.dim()), 1);
             stats_shape[static_cast<size_t>(channel_dim)] = latents.shape()[channel_dim];
@@ -804,7 +804,7 @@ struct AutoEncoderKL : public VAE {
     }
 
     sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
-        if (sd_version_is_flux2(version)) {
+        if (sd_version_uses_flux2_vae(version)) {
             int channel_dim                = 2;
             auto [mean_tensor, std_tensor] = get_latents_mean_std(latents, channel_dim);
             return (latents * std_tensor) / scale_factor + mean_tensor;
@@ -813,7 +813,7 @@ struct AutoEncoderKL : public VAE {
     }
 
     sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents) override {
-        if (sd_version_is_flux2(version)) {
+        if (sd_version_uses_flux2_vae(version)) {
             int channel_dim                = 2;
             auto [mean_tensor, std_tensor] = get_latents_mean_std(latents, channel_dim);
             return ((latents - mean_tensor) * scale_factor) / std_tensor;

src/clip.hpp

@@ -3,455 +3,7 @@
 
 #include "ggml_extend.hpp"
 #include "model.h"
-#include "tokenize_util.h"
-#include "vocab/vocab.h"
-
-/*================================================== CLIPTokenizer ===================================================*/
-
-__STATIC_INLINE__ std::vector<std::pair<int, std::u32string>> bytes_to_unicode() {
-    std::vector<std::pair<int, std::u32string>> byte_unicode_pairs;
-    std::set<int> byte_set;
-    for (int b = static_cast<int>('!'); b <= static_cast<int>('~'); ++b) {
-        byte_set.insert(b);
-        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
-    }
-    for (int b = 161; b <= 172; ++b) {
-        byte_set.insert(b);
-        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
-    }
-    for (int b = 174; b <= 255; ++b) {
-        byte_set.insert(b);
-        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
-    }
-    int n = 0;
-    for (int b = 0; b < 256; ++b) {
-        if (byte_set.find(b) == byte_set.end()) {
-            byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(n + 256)));
-            ++n;
-        }
-    }
-    // LOG_DEBUG("byte_unicode_pairs %d", byte_unicode_pairs.size());
-    return byte_unicode_pairs;
-}
-
-// Ref: https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py
-
-typedef std::function<bool(std::string&, std::vector<int32_t>&)> on_new_token_cb_t;
-
-class CLIPTokenizer {
-private:
-    std::map<int, std::u32string> byte_encoder;
-    std::map<std::u32string, int> byte_decoder;
-    std::map<std::u32string, int> encoder;
-    std::map<int, std::u32string> decoder;
-    std::map<std::pair<std::u32string, std::u32string>, int> bpe_ranks;
-    std::regex pat;
-    int encoder_len;
-    int bpe_len;
-
-    std::vector<std::string> special_tokens;
-
-public:
-    const std::string UNK_TOKEN = "<|endoftext|>";
-    const std::string BOS_TOKEN = "<|startoftext|>";
-    const std::string EOS_TOKEN = "<|endoftext|>";
-    const std::string PAD_TOKEN = "<|endoftext|>";
-
-    const int UNK_TOKEN_ID = 49407;
-    const int BOS_TOKEN_ID = 49406;
-    const int EOS_TOKEN_ID = 49407;
-    const int PAD_TOKEN_ID = 49407;
-
-private:
-    static std::string strip(const std::string& str) {
-        std::string::size_type start = str.find_first_not_of(" \t\n\r\v\f");
-        std::string::size_type end   = str.find_last_not_of(" \t\n\r\v\f");
-
-        if (start == std::string::npos) {
-            // String contains only whitespace characters
-            return "";
-        }
-
-        return str.substr(start, end - start + 1);
-    }
-
-    static std::string whitespace_clean(std::string text) {
-        text = std::regex_replace(text, std::regex(R"(\s+)"), " ");
-        text = strip(text);
-        return text;
-    }
-
-    static std::set<std::pair<std::u32string, std::u32string>> get_pairs(const std::vector<std::u32string>& subwords) {
-        std::set<std::pair<std::u32string, std::u32string>> pairs;
-        if (subwords.size() == 0) {
-            return pairs;
-        }
-        std::u32string prev_subword = subwords[0];
-        for (int i = 1; i < subwords.size(); i++) {
-            std::u32string subword = subwords[i];
-            std::pair<std::u32string, std::u32string> pair(prev_subword, subword);
-            pairs.insert(pair);
-            prev_subword = subword;
-        }
-        return pairs;
-    }
-
-    bool is_special_token(const std::string& token) {
-        for (auto& special_token : special_tokens) {
-            if (special_token == token) {
-                return true;
-            }
-        }
-        return false;
-    }
-
-public:
-    CLIPTokenizer(int pad_token_id = 49407, const std::string& merges_utf8_str = "")
-        : PAD_TOKEN_ID(pad_token_id) {
-        if (merges_utf8_str.size() > 0) {
-            load_from_merges(merges_utf8_str);
-        } else {
-            load_from_merges(load_clip_merges());
-        }
-        add_special_token("<|startoftext|>");
-        add_special_token("<|endoftext|>");
-    }
-
-    void load_from_merges(const std::string& merges_utf8_str) {
-        auto byte_unicode_pairs = bytes_to_unicode();
-        // printf("byte_unicode_pairs have %lu pairs \n", byte_unicode_pairs.size());
-        byte_encoder = std::map<int, std::u32string>(byte_unicode_pairs.begin(), byte_unicode_pairs.end());
-        for (auto& pair : byte_unicode_pairs) {
-            byte_decoder[pair.second] = pair.first;
-        }
-        // for (auto & pair: byte_unicode_pairs) {
-        //     std::cout << pair.first << ": " << pair.second << std::endl;
-        // }
-        std::vector<std::u32string> merges;
-        size_t start = 0;
-        size_t pos;
-        std::u32string merges_utf32_str = utf8_to_utf32(merges_utf8_str);
-        while ((pos = merges_utf32_str.find('\n', start)) != std::string::npos) {
-            merges.push_back(merges_utf32_str.substr(start, pos - start));
-            start = pos + 1;
-        }
-        // LOG_DEBUG("merges size %llu", merges.size());
-        GGML_ASSERT(merges.size() == 48895);
-        merges = std::vector<std::u32string>(merges.begin() + 1, merges.end());
-        std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
-        for (const auto& merge : merges) {
-            size_t space_pos = merge.find(' ');
-            merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
-            // LOG_DEBUG("%s", utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
-            // printf("%s :: %s | %s \n", utf32_to_utf8(merge).c_str(), utf32_to_utf8(merge.substr(0, space_pos)).c_str(),
-            //                     utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
-        }
-        std::vector<std::u32string> vocab;
-        for (const auto& pair : byte_unicode_pairs) {
-            vocab.push_back(pair.second);
-        }
-        for (const auto& pair : byte_unicode_pairs) {
-            vocab.push_back(pair.second + utf8_to_utf32("</w>"));
-        }
-        for (const auto& merge : merge_pairs) {
-            vocab.push_back(merge.first + merge.second);
-        }
-        vocab.push_back(utf8_to_utf32("<|startoftext|>"));
-        vocab.push_back(utf8_to_utf32("<|endoftext|>"));
-        LOG_DEBUG("vocab size: %llu", vocab.size());
-        int i = 0;
-        for (const auto& token : vocab) {
-            encoder[token] = i;
-            decoder[i]     = token;
-            i++;
-        }
-        encoder_len = i;
-
-        auto it = encoder.find(utf8_to_utf32("img</w>"));
-        if (it != encoder.end()) {
-            LOG_DEBUG("trigger word img already in vocab");
-        } else {
-            LOG_DEBUG("trigger word img not in vocab yet");
-        }
-
-        int rank = 0;
-        for (const auto& merge : merge_pairs) {
-            bpe_ranks[merge] = rank++;
-        }
-        bpe_len = rank;
-    };
-
-    void add_token(const std::string& text) {
-        std::u32string token = utf8_to_utf32(text);
-        auto it              = encoder.find(token);
-        if (it != encoder.end()) {
-            encoder[token]       = encoder_len;
-            decoder[encoder_len] = token;
-            encoder_len++;
-        }
-    }
-
-    void add_special_token(const std::string& token) {
-        special_tokens.push_back(token);
-    }
-
-    std::u32string bpe(const std::u32string& token) {
-        std::vector<std::u32string> word;
-
-        for (int i = 0; i < token.size() - 1; i++) {
-            word.emplace_back(1, token[i]);
-        }
-        word.push_back(token.substr(token.size() - 1) + utf8_to_utf32("</w>"));
-
-        std::set<std::pair<std::u32string, std::u32string>> pairs = get_pairs(word);
-
-        if (pairs.empty()) {
-            return token + utf8_to_utf32("</w>");
-        }
-
-        while (true) {
-            auto min_pair_iter = std::min_element(pairs.begin(),
-                                                  pairs.end(),
-                                                  [&](const std::pair<std::u32string, std::u32string>& a,
-                                                      const std::pair<std::u32string, std::u32string>& b) {
-                                                      if (bpe_ranks.find(a) == bpe_ranks.end()) {
-                                                          return false;
-                                                      } else if (bpe_ranks.find(b) == bpe_ranks.end()) {
-                                                          return true;
-                                                      }
-                                                      return bpe_ranks.at(a) < bpe_ranks.at(b);
-                                                  });
-
-            const std::pair<std::u32string, std::u32string>& bigram = *min_pair_iter;
-
-            if (bpe_ranks.find(bigram) == bpe_ranks.end()) {
-                break;
-            }
-
-            std::u32string first  = bigram.first;
-            std::u32string second = bigram.second;
-            std::vector<std::u32string> new_word;
-            int32_t i = 0;
-
-            while (i < word.size()) {
-                auto it = std::find(word.begin() + i, word.end(), first);
-                if (it == word.end()) {
-                    new_word.insert(new_word.end(), word.begin() + i, word.end());
-                    break;
-                }
-                new_word.insert(new_word.end(), word.begin() + i, it);
-                i = static_cast<int32_t>(std::distance(word.begin(), it));
-
-                if (word[i] == first && i < static_cast<int32_t>(word.size()) - 1 && word[i + 1] == second) {
-                    new_word.push_back(first + second);
-                    i += 2;
-                } else {
-                    new_word.push_back(word[i]);
-                    i += 1;
-                }
-            }
-
-            word = new_word;
-
-            if (word.size() == 1) {
-                break;
-            }
-            pairs = get_pairs(word);
-        }
-
-        std::u32string result;
-        for (int i = 0; i < word.size(); i++) {
-            result += word[i];
-            if (i != word.size() - 1) {
-                result += utf8_to_utf32(" ");
-            }
-        }
-
-        return result;
-    }
-
-    std::vector<int> tokenize(std::string text,
-                              on_new_token_cb_t on_new_token_cb,
-                              size_t max_length = 0,
-                              bool padding      = false) {
-        std::vector<int32_t> tokens = encode(text, on_new_token_cb);
-
-        tokens.insert(tokens.begin(), BOS_TOKEN_ID);
-        if (max_length > 0) {
-            if (tokens.size() > max_length - 1) {
-                tokens.resize(max_length - 1);
-                tokens.push_back(EOS_TOKEN_ID);
-            } else {
-                tokens.push_back(EOS_TOKEN_ID);
-                if (padding) {
-                    tokens.insert(tokens.end(), max_length - tokens.size(), PAD_TOKEN_ID);
-                }
-            }
-        }
-
-        return tokens;
-    }
-
-    void pad_tokens(std::vector<int>& tokens,
-                    std::vector<float>& weights,
-                    size_t max_length = 0,
-                    bool padding      = false) {
-        if (max_length > 0 && padding) {
-            size_t n = static_cast<size_t>(std::ceil(tokens.size() * 1.0 / (max_length - 2)));
-            if (n == 0) {
-                n = 1;
-            }
-            size_t length = max_length * n;
-            LOG_DEBUG("token length: %llu", length);
-            std::vector<int> new_tokens;
-            std::vector<float> new_weights;
-            new_tokens.push_back(BOS_TOKEN_ID);
-            new_weights.push_back(1.0);
-            int token_idx = 0;
-            for (int i = 1; i < length; i++) {
-                if (token_idx >= tokens.size()) {
-                    break;
-                }
-                if (i % max_length == 0) {
-                    new_tokens.push_back(BOS_TOKEN_ID);
-                    new_weights.push_back(1.0);
-                } else if (i % max_length == max_length - 1) {
-                    new_tokens.push_back(EOS_TOKEN_ID);
-                    new_weights.push_back(1.0);
-                } else {
-                    new_tokens.push_back(tokens[token_idx]);
-                    new_weights.push_back(weights[token_idx]);
-                    token_idx++;
-                }
-            }
-
-            new_tokens.push_back(EOS_TOKEN_ID);
-            new_weights.push_back(1.0);
-            tokens  = new_tokens;
-            weights = new_weights;
-
-            if (padding) {
-                tokens.insert(tokens.end(), length - tokens.size(), PAD_TOKEN_ID);
-                weights.insert(weights.end(), length - weights.size(), 1.0);
-            }
-        }
-    }
-
-    std::string clean_up_tokenization(std::string& text) {
-        std::regex pattern(R"( ,)");
-        // Replace " ," with ","
-        std::string result = std::regex_replace(text, pattern, ",");
-        return result;
-    }
-
-    std::string decode(const std::vector<int>& tokens) {
-        std::string text = "";
-        for (int t : tokens) {
-            if (t == 49406 || t == 49407)
-                continue;
-            std::u32string ts = decoder[t];
-            // printf("%d, %s \n", t,  utf32_to_utf8(ts).c_str());
-            std::string s = utf32_to_utf8(ts);
-            if (s.length() >= 4) {
-                if (ends_with(s, "</w>")) {
-                    text += s.replace(s.length() - 4, s.length() - 1, "") + " ";
-                } else {
-                    text += s;
-                }
-            } else {
-                text += " " + s;
-            }
-        }
-        // std::vector<unsigned char> bytes;
-        // for (auto c : text){
-        //     bytes.push_back(byte_decoder[c]);
-        // }
-
-        // std::string s((char *)bytes.data());
-        // std::string s = "";
-        text = clean_up_tokenization(text);
-        return trim(text);
-    }
-
-    std::vector<std::string> token_split(const std::string& text) {
-        std::regex pat(R"('s|'t|'re|'ve|'m|'ll|'d|[[:alpha:]]+|[[:digit:]]|[^[:space:][:alpha:][:digit:]]+)",
-                       std::regex::icase);
-        std::sregex_iterator iter(text.begin(), text.end(), pat);
-        std::sregex_iterator end;
-
-        std::vector<std::string> result;
-        for (; iter != end; ++iter) {
-            result.emplace_back(iter->str());
-        }
-
-        return result;
-    }
-
-    std::vector<int> encode(std::string text, on_new_token_cb_t on_new_token_cb) {
-        std::string original_text = text;
-        std::vector<int32_t> bpe_tokens;
-        text = whitespace_clean(text);
-        std::transform(text.begin(), text.end(), text.begin(), [](unsigned char c) { return std::tolower(c); });
-
-        std::string str = text;
-        std::vector<std::string> token_strs;
-
-        auto splited_texts = split_with_special_tokens(text, special_tokens);
-
-        for (auto& splited_text : splited_texts) {
-            LOG_DEBUG("token %s", splited_text.c_str());
-            if (is_special_token(splited_text)) {
-                LOG_DEBUG("special %s", splited_text.c_str());
-                bool skip = on_new_token_cb(splited_text, bpe_tokens);
-                if (skip) {
-                    token_strs.push_back(splited_text);
-                    continue;
-                }
-                continue;
-            }
-
-            auto tokens = token_split(splited_text);
-            for (auto& token : tokens) {
-                if (on_new_token_cb != nullptr) {
-                    bool skip = on_new_token_cb(token, bpe_tokens);
-                    if (skip) {
-                        token_strs.push_back(token);
-                        continue;
-                    }
-                }
-
-                std::string token_str = token;
-                std::u32string utf32_token;
-                for (int i = 0; i < token_str.length(); i++) {
-                    unsigned char b = token_str[i];
-                    utf32_token += byte_encoder[b];
-                }
-                auto bpe_strs = bpe(utf32_token);
-                size_t start  = 0;
-                size_t pos;
-                while ((pos = bpe_strs.find(' ', start)) != std::u32string::npos) {
-                    auto bpe_str = bpe_strs.substr(start, pos - start);
-                    bpe_tokens.push_back(encoder[bpe_str]);
-                    token_strs.push_back(utf32_to_utf8(bpe_str));
-
-                    start = pos + 1;
-                }
-                auto bpe_str = bpe_strs.substr(start, bpe_strs.size() - start);
-                bpe_tokens.push_back(encoder[bpe_str]);
-                token_strs.push_back(utf32_to_utf8(bpe_str));
-            }
-        }
-        // std::stringstream ss;
-        // ss << "[";
-        // for (auto token : token_strs) {
-        //     ss << "\"" << token << "\", ";
-        // }
-        // ss << "]";
-        // LOG_DEBUG("split prompt \"%s\" to tokens %s", original_text.c_str(), ss.str().c_str());
-        // printf("split prompt \"%s\" to tokens %s \n", original_text.c_str(), ss.str().c_str());
-        return bpe_tokens;
-    }
-};
+#include "tokenizers/clip_tokenizer.h"
 
 /*================================================ FrozenCLIPEmbedder ================================================*/
 

src/common_block.hpp

@@ -277,6 +277,7 @@ protected:
     int64_t context_dim;
     int64_t n_head;
     int64_t d_head;
+    bool xtra_dim = false;
 
 public:
     CrossAttention(int64_t query_dim,
@@ -288,7 +289,11 @@ public:
           query_dim(query_dim),
           context_dim(context_dim) {
         int64_t inner_dim = d_head * n_head;
-
+        if (context_dim == 320 && d_head == 320) {
+            // LOG_DEBUG("CrossAttention: temp set dim to 1024 for sdxs_09");
+            xtra_dim    = true;
+            context_dim = 1024;
+        }
         blocks["to_q"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, false));
         blocks["to_k"] = std::shared_ptr<GGMLBlock>(new Linear(context_dim, inner_dim, false));
         blocks["to_v"] = std::shared_ptr<GGMLBlock>(new Linear(context_dim, inner_dim, false));
@@ -313,10 +318,16 @@ public:
         int64_t n_context = context->ne[1];
         int64_t inner_dim = d_head * n_head;
 
-        auto q = to_q->forward(ctx, x);        // [N, n_token, inner_dim]
+        auto q = to_q->forward(ctx, x);  // [N, n_token, inner_dim]
+        if (xtra_dim) {
+            // LOG_DEBUG("CrossAttention: temp set dim to 1024 for sdxs_09");
+            context->ne[0] = 1024;  // patch dim
+        }
         auto k = to_k->forward(ctx, context);  // [N, n_context, inner_dim]
         auto v = to_v->forward(ctx, context);  // [N, n_context, inner_dim]
-
+        if (xtra_dim) {
+            context->ne[0] = 320;  // reset dim to orig
+        }
         x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, inner_dim]
 
         x = to_out_0->forward(ctx, x);  // [N, n_token, query_dim]

src/conditioner.hpp

@@ -256,15 +256,6 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         return true;
     }
 
-    std::tuple<std::vector<int>, std::vector<float>, std::vector<bool>>
-    tokenize_with_trigger_token(std::string text,
-                                int num_input_imgs,
-                                int32_t image_token,
-                                bool padding = false) {
-        return tokenize_with_trigger_token(text, num_input_imgs, image_token,
-                                           text_model->model.n_token, padding);
-    }
-
     std::vector<int> convert_token_to_id(std::string text) {
         auto on_new_token_cb = [&](std::string& str, std::vector<int32_t>& bpe_tokens) -> bool {
             auto iter = embedding_map.find(str);
@@ -288,9 +279,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
     std::tuple<std::vector<int>, std::vector<float>, std::vector<bool>>
     tokenize_with_trigger_token(std::string text,
                                 int num_input_imgs,
-                                int32_t image_token,
-                                size_t max_length = 0,
-                                bool padding      = false) {
+                                int32_t image_token) {
         auto parsed_attention = parse_prompt_attention(text);
 
         {
@@ -377,7 +366,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         // tokens.insert(tokens.begin(), tokenizer.BOS_TOKEN_ID);
         // weights.insert(weights.begin(), 1.0);
 
-        tokenizer.pad_tokens(tokens, weights, max_length, padding);
+        tokenizer.pad_tokens(tokens, &weights, nullptr, text_model->model.n_token, text_model->model.n_token, true);
         int offset = pm_version == PM_VERSION_2 ? 2 * num_input_imgs : num_input_imgs;
         for (int i = 0; i < tokens.size(); i++) {
             // if (class_idx + 1 <= i && i < class_idx + 1 + 2*num_input_imgs) // photomaker V2 has num_tokens(=2)*num_input_imgs
@@ -403,13 +392,9 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
     }
 
     std::pair<std::vector<int>, std::vector<float>> tokenize(std::string text,
-                                                             bool padding = false) {
-        return tokenize(text, text_model->model.n_token, padding);
-    }
-
-    std::pair<std::vector<int>, std::vector<float>> tokenize(std::string text,
-                                                             size_t max_length = 0,
-                                                             bool padding      = false) {
+                                                             size_t min_length          = 0,
+                                                             size_t max_length          = 0,
+                                                             bool allow_overflow_expand = true) {
         auto parsed_attention = parse_prompt_attention(text);
 
         {
@@ -460,7 +445,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             weights.insert(weights.end(), curr_tokens.size(), curr_weight);
         }
 
-        tokenizer.pad_tokens(tokens, weights, max_length, padding);
+        tokenizer.pad_tokens(tokens, &weights, nullptr, min_length, max_length, allow_overflow_expand);
 
         // for (int i = 0; i < tokens.size(); i++) {
         //     std::cout << tokens[i] << ":" << weights[i] << ", ";
@@ -603,8 +588,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         GGML_ASSERT(image_tokens.size() == 1);
         auto tokens_and_weights     = tokenize_with_trigger_token(conditioner_params.text,
                                                                   conditioner_params.num_input_imgs,
-                                                                  image_tokens[0],
-                                                                  true);
+                                                                  image_tokens[0]);
         std::vector<int>& tokens    = std::get<0>(tokens_and_weights);
         std::vector<float>& weights = std::get<1>(tokens_and_weights);
         std::vector<bool>& clsm     = std::get<2>(tokens_and_weights);
@@ -630,7 +614,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
     std::string remove_trigger_from_prompt(const std::string& prompt) override {
         auto image_tokens = convert_token_to_id(trigger_word);
         GGML_ASSERT(image_tokens.size() == 1);
-        auto tokens_and_weights  = tokenize(prompt, false);
+        auto tokens_and_weights  = tokenize(prompt);
         std::vector<int>& tokens = tokens_and_weights.first;
         auto it                  = std::find(tokens.begin(), tokens.end(), image_tokens[0]);
         GGML_ASSERT(it != tokens.end());  // prompt must have trigger word
@@ -640,7 +624,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
 
     SDCondition get_learned_condition(int n_threads,
                                       const ConditionerParams& conditioner_params) override {
-        auto tokens_and_weights     = tokenize(conditioner_params.text, true);
+        auto tokens_and_weights     = tokenize(conditioner_params.text, text_model->model.n_token, text_model->model.n_token, true);
         std::vector<int>& tokens    = tokens_and_weights.first;
         std::vector<float>& weights = tokens_and_weights.second;
         return get_learned_condition_common(n_threads,
@@ -822,8 +806,9 @@ struct SD3CLIPEmbedder : public Conditioner {
     }
 
     std::vector<std::pair<std::vector<int>, std::vector<float>>> tokenize(std::string text,
-                                                                          size_t max_length = 0,
-                                                                          bool padding      = false) {
+                                                                          size_t min_length          = 0,
+                                                                          size_t max_length          = 0,
+                                                                          bool allow_overflow_expand = true) {
         auto parsed_attention = parse_prompt_attention(text);
 
         {
@@ -860,20 +845,20 @@ struct SD3CLIPEmbedder : public Conditioner {
                 clip_g_weights.insert(clip_g_weights.end(), curr_tokens.size(), curr_weight);
             }
             if (t5) {
-                std::vector<int> curr_tokens = t5_tokenizer.Encode(curr_text, true);
+                std::vector<int> curr_tokens = t5_tokenizer.encode(curr_text);
                 t5_tokens.insert(t5_tokens.end(), curr_tokens.begin(), curr_tokens.end());
                 t5_weights.insert(t5_weights.end(), curr_tokens.size(), curr_weight);
             }
         }
 
         if (clip_l) {
-            clip_l_tokenizer.pad_tokens(clip_l_tokens, clip_l_weights, max_length, padding);
+            clip_l_tokenizer.pad_tokens(clip_l_tokens, &clip_l_weights, nullptr, min_length, max_length, allow_overflow_expand);
         }
         if (clip_g) {
-            clip_g_tokenizer.pad_tokens(clip_g_tokens, clip_g_weights, max_length, padding);
+            clip_g_tokenizer.pad_tokens(clip_g_tokens, &clip_g_weights, nullptr, min_length, max_length, allow_overflow_expand);
         }
         if (t5) {
-            t5_tokenizer.pad_tokens(t5_tokens, t5_weights, nullptr, max_length, padding);
+            t5_tokenizer.pad_tokens(t5_tokens, &t5_weights, nullptr, min_length, max_length, true);
         }
 
         // for (int i = 0; i < clip_l_tokens.size(); i++) {
@@ -1056,7 +1041,7 @@ struct SD3CLIPEmbedder : public Conditioner {
 
     SDCondition get_learned_condition(int n_threads,
                                       const ConditionerParams& conditioner_params) override {
-        auto tokens_and_weights = tokenize(conditioner_params.text, 77, true);
+        auto tokens_and_weights = tokenize(conditioner_params.text, 77, 77, true);
         return get_learned_condition_common(n_threads,
                                             tokens_and_weights,
                                             conditioner_params.clip_skip,
@@ -1158,8 +1143,8 @@ struct FluxCLIPEmbedder : public Conditioner {
     }
 
     std::vector<std::pair<std::vector<int>, std::vector<float>>> tokenize(std::string text,
-                                                                          size_t max_length = 0,
-                                                                          bool padding      = false) {
+                                                                          size_t min_length = 0,
+                                                                          size_t max_length = 0) {
         auto parsed_attention = parse_prompt_attention(text);
 
         {
@@ -1189,17 +1174,17 @@ struct FluxCLIPEmbedder : public Conditioner {
                 clip_l_weights.insert(clip_l_weights.end(), curr_tokens.size(), curr_weight);
             }
             if (t5) {
-                std::vector<int> curr_tokens = t5_tokenizer.Encode(curr_text, true);
+                std::vector<int> curr_tokens = t5_tokenizer.encode(curr_text);
                 t5_tokens.insert(t5_tokens.end(), curr_tokens.begin(), curr_tokens.end());
                 t5_weights.insert(t5_weights.end(), curr_tokens.size(), curr_weight);
             }
         }
 
         if (clip_l) {
-            clip_l_tokenizer.pad_tokens(clip_l_tokens, clip_l_weights, 77, padding);
+            clip_l_tokenizer.pad_tokens(clip_l_tokens, &clip_l_weights, nullptr, 77, 77, true);
         }
         if (t5) {
-            t5_tokenizer.pad_tokens(t5_tokens, t5_weights, nullptr, max_length, padding);
+            t5_tokenizer.pad_tokens(t5_tokens, &t5_weights, nullptr, min_length, max_length, true);
         }
 
         // for (int i = 0; i < clip_l_tokens.size(); i++) {
@@ -1300,7 +1285,7 @@ struct FluxCLIPEmbedder : public Conditioner {
 
     SDCondition get_learned_condition(int n_threads,
                                       const ConditionerParams& conditioner_params) override {
-        auto tokens_and_weights = tokenize(conditioner_params.text, chunk_len, true);
+        auto tokens_and_weights = tokenize(conditioner_params.text, chunk_len, chunk_len);
         return get_learned_condition_common(n_threads,
                                             tokens_and_weights,
                                             conditioner_params.clip_skip,
@@ -1377,8 +1362,8 @@ struct T5CLIPEmbedder : public Conditioner {
     }
 
     std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> tokenize(std::string text,
-                                                                                  size_t max_length = 0,
-                                                                                  bool padding      = false) {
+                                                                                  size_t min_length = 0,
+                                                                                  size_t max_length = 0) {
         auto parsed_attention = parse_prompt_attention(text);
 
         {
@@ -1403,12 +1388,15 @@ struct T5CLIPEmbedder : public Conditioner {
                 const std::string& curr_text = item.first;
                 float curr_weight            = item.second;
 
-                std::vector<int> curr_tokens = t5_tokenizer.Encode(curr_text, true);
+                std::vector<int> curr_tokens = t5_tokenizer.encode(curr_text);
                 t5_tokens.insert(t5_tokens.end(), curr_tokens.begin(), curr_tokens.end());
                 t5_weights.insert(t5_weights.end(), curr_tokens.size(), curr_weight);
             }
 
-            t5_tokenizer.pad_tokens(t5_tokens, t5_weights, &t5_mask, max_length, padding);
+            t5_tokenizer.pad_tokens(t5_tokens, &t5_weights, &t5_mask, min_length, max_length, true);
+            for (auto& mask_value : t5_mask) {
+                mask_value = mask_value > 0.0f ? 0.0f : -HUGE_VALF;
+            }
         }
         return {t5_tokens, t5_weights, t5_mask};
     }
@@ -1496,7 +1484,7 @@ struct T5CLIPEmbedder : public Conditioner {
 
     SDCondition get_learned_condition(int n_threads,
                                       const ConditionerParams& conditioner_params) override {
-        auto tokens_and_weights = tokenize(conditioner_params.text, chunk_len, true);
+        auto tokens_and_weights = tokenize(conditioner_params.text, chunk_len, chunk_len);
         return get_learned_condition_common(n_threads,
                                             tokens_and_weights,
                                             conditioner_params.clip_skip,
@@ -1505,14 +1493,14 @@ struct T5CLIPEmbedder : public Conditioner {
 };
 
 struct AnimaConditioner : public Conditioner {
-    std::shared_ptr<LLM::BPETokenizer> qwen_tokenizer;
+    std::shared_ptr<BPETokenizer> qwen_tokenizer;
     T5UniGramTokenizer t5_tokenizer;
     std::shared_ptr<LLM::LLMRunner> llm;
 
     AnimaConditioner(ggml_backend_t backend,
                      bool offload_params_to_cpu,
                      const String2TensorStorage& tensor_storage_map = {}) {
-        qwen_tokenizer = std::make_shared<LLM::Qwen2Tokenizer>();
+        qwen_tokenizer = std::make_shared<Qwen2Tokenizer>();
         llm            = std::make_shared<LLM::LLMRunner>(LLM::LLMArch::QWEN3,
                                                backend,
                                                offload_params_to_cpu,
@@ -1578,7 +1566,7 @@ struct AnimaConditioner : public Conditioner {
         for (const auto& item : parsed_attention) {
             const std::string& curr_text = item.first;
             float curr_weight            = item.second;
-            std::vector<int> curr_tokens = t5_tokenizer.Encode(curr_text, true);
+            std::vector<int> curr_tokens = t5_tokenizer.tokenize(curr_text, nullptr, true);
             t5_tokens.insert(t5_tokens.end(), curr_tokens.begin(), curr_tokens.end());
             t5_weights.insert(t5_weights.end(), curr_tokens.size(), curr_weight);
         }
@@ -1620,7 +1608,7 @@ struct AnimaConditioner : public Conditioner {
 
 struct LLMEmbedder : public Conditioner {
     SDVersion version;
-    std::shared_ptr<LLM::BPETokenizer> tokenizer;
+    std::shared_ptr<BPETokenizer> tokenizer;
     std::shared_ptr<LLM::LLMRunner> llm;
 
     LLMEmbedder(ggml_backend_t backend,
@@ -1633,13 +1621,15 @@ struct LLMEmbedder : public Conditioner {
         LLM::LLMArch arch = LLM::LLMArch::QWEN2_5_VL;
         if (version == VERSION_FLUX2) {
             arch = LLM::LLMArch::MISTRAL_SMALL_3_2;
+        } else if (sd_version_is_ernie_image(version)) {
+            arch = LLM::LLMArch::MINISTRAL_3_3B;
         } else if (sd_version_is_z_image(version) || version == VERSION_OVIS_IMAGE || version == VERSION_FLUX2_KLEIN) {
             arch = LLM::LLMArch::QWEN3;
         }
-        if (arch == LLM::LLMArch::MISTRAL_SMALL_3_2) {
-            tokenizer = std::make_shared<LLM::MistralTokenizer>();
+        if (arch == LLM::LLMArch::MISTRAL_SMALL_3_2 || arch == LLM::LLMArch::MINISTRAL_3_3B) {
+            tokenizer = std::make_shared<MistralTokenizer>();
         } else {
-            tokenizer = std::make_shared<LLM::Qwen2Tokenizer>();
+            tokenizer = std::make_shared<Qwen2Tokenizer>();
         }
         llm = std::make_shared<LLM::LLMRunner>(arch,
                                                backend,
@@ -1677,20 +1667,24 @@ struct LLMEmbedder : public Conditioner {
         }
     }
 
-    std::tuple<std::vector<int>, std::vector<float>> tokenize(std::string text,
-                                                              const std::pair<int, int>& attn_range,
-                                                              size_t max_length = 0,
-                                                              bool padding      = false) {
+    std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> tokenize(std::string text,
+                                                                                  const std::pair<int, int>& attn_range,
+                                                                                  size_t min_length = 0,
+                                                                                  size_t max_length = 100000000) {
         std::vector<std::pair<std::string, float>> parsed_attention;
         if (attn_range.first >= 0 && attn_range.second > 0) {
-            parsed_attention.emplace_back(text.substr(0, attn_range.first), 1.f);
+            if (attn_range.first > 0) {
+                parsed_attention.emplace_back(text.substr(0, attn_range.first), 1.f);
+            }
             if (attn_range.second - attn_range.first > 0) {
                 auto new_parsed_attention = parse_prompt_attention(text.substr(attn_range.first, attn_range.second - attn_range.first));
                 parsed_attention.insert(parsed_attention.end(),
                                         new_parsed_attention.begin(),
                                         new_parsed_attention.end());
             }
-            parsed_attention.emplace_back(text.substr(attn_range.second), 1.f);
+            if (attn_range.second < text.size()) {
+                parsed_attention.emplace_back(text.substr(attn_range.second), 1.f);
+            }
         } else {
             parsed_attention.emplace_back(text, 1.f);
         }
@@ -1710,39 +1704,34 @@ struct LLMEmbedder : public Conditioner {
         for (const auto& item : parsed_attention) {
             const std::string& curr_text = item.first;
             float curr_weight            = item.second;
-            std::vector<int> curr_tokens = tokenizer->tokenize(curr_text, nullptr);
+            std::vector<int> curr_tokens = tokenizer->encode(curr_text, nullptr);
             tokens.insert(tokens.end(), curr_tokens.begin(), curr_tokens.end());
             weights.insert(weights.end(), curr_tokens.size(), curr_weight);
         }
 
-        tokenizer->pad_tokens(tokens, weights, max_length, padding);
+        std::vector<float> mask;
+        tokenizer->pad_tokens(tokens, &weights, &mask, min_length, max_length);
 
         // for (int i = 0; i < tokens.size(); i++) {
         //     std::cout << tokens[i] << ":" << weights[i] << ", " << i << std::endl;
         // }
         // std::cout << std::endl;
 
-        return {tokens, weights};
+        return {tokens, weights, mask};
     }
 
     sd::Tensor<float> encode_prompt(int n_threads,
                                     const std::string prompt,
                                     const std::pair<int, int>& prompt_attn_range,
-                                    int max_length,
                                     int min_length,
+                                    int hidden_states_min_length,
                                     const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds,
                                     const std::set<int>& out_layers,
                                     int prompt_template_encode_start_idx) {
-        auto tokens_and_weights = tokenize(prompt, prompt_attn_range);
-        auto& tokens            = std::get<0>(tokens_and_weights);
-        auto& weights           = std::get<1>(tokens_and_weights);
-        std::vector<float> mask;
-
-        if (max_length > 0 && tokens.size() < max_length) {
-            mask.insert(mask.end(), tokens.size(), 1.f);
-            mask.insert(mask.end(), max_length - tokens.size(), 0.f);
-            tokenizer->pad_tokens(tokens, weights, max_length, true);
-        }
+        auto tokens_weights_mask = tokenize(prompt, prompt_attn_range, min_length);
+        auto& tokens             = std::get<0>(tokens_weights_mask);
+        auto& weights            = std::get<1>(tokens_weights_mask);
+        auto& mask               = std::get<2>(tokens_weights_mask);
 
         sd::Tensor<int32_t> input_ids({static_cast<int64_t>(tokens.size())}, tokens);
         sd::Tensor<float> attention_mask;
@@ -1769,9 +1758,9 @@ struct LLMEmbedder : public Conditioner {
         GGML_ASSERT(hidden_states.shape()[1] > prompt_template_encode_start_idx);
 
         int64_t zero_pad_len = 0;
-        if (min_length > 0) {
-            if (hidden_states.shape()[1] - prompt_template_encode_start_idx < min_length) {
-                zero_pad_len = min_length - hidden_states.shape()[1] + prompt_template_encode_start_idx;
+        if (hidden_states_min_length > 0) {
+            if (hidden_states.shape()[1] - prompt_template_encode_start_idx < hidden_states_min_length) {
+                zero_pad_len = hidden_states_min_length - hidden_states.shape()[1] + prompt_template_encode_start_idx;
             }
         }
 
@@ -1798,8 +1787,8 @@ struct LLMEmbedder : public Conditioner {
         std::vector<std::pair<int, int>> extra_prompts_attn_range;
         std::vector<std::pair<int, sd::Tensor<float>>> image_embeds;
         int prompt_template_encode_start_idx = 34;
-        int max_length                       = 0;  // pad tokens
-        int min_length                       = 0;  // zero pad hidden_states
+        int min_length                       = 0;  // pad tokens
+        int hidden_states_min_length         = 0;  // zero pad hidden_states
         std::set<int> out_layers;
 
         int64_t t0 = ggml_time_ms();
@@ -1874,7 +1863,7 @@ struct LLMEmbedder : public Conditioner {
             }
         } else if (version == VERSION_FLUX2) {
             prompt_template_encode_start_idx = 0;
-            min_length                       = 512;
+            hidden_states_min_length         = 512;
             out_layers                       = {10, 20, 30};
 
             prompt = "[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]";
@@ -1884,6 +1873,13 @@ struct LLMEmbedder : public Conditioner {
             prompt_attn_range.second = static_cast<int>(prompt.size());
 
             prompt += "[/INST]";
+        } else if (sd_version_is_ernie_image(version)) {
+            prompt_template_encode_start_idx = 0;
+            out_layers                       = {25};  // -2
+
+            prompt_attn_range.first = 0;
+            prompt += conditioner_params.text;
+            prompt_attn_range.second = static_cast<int>(prompt.size());
         } else if (sd_version_is_z_image(version)) {
             prompt_template_encode_start_idx = 0;
             out_layers                       = {35};  // -2
@@ -1907,7 +1903,7 @@ struct LLMEmbedder : public Conditioner {
             }
         } else if (version == VERSION_FLUX2_KLEIN) {
             prompt_template_encode_start_idx = 0;
-            max_length                       = 512;
+            min_length                       = 512;
             out_layers                       = {9, 18, 27};
 
             prompt = "<|im_start|>user\n";
@@ -1919,7 +1915,7 @@ struct LLMEmbedder : public Conditioner {
             prompt += "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n";
         } else if (version == VERSION_OVIS_IMAGE) {
             prompt_template_encode_start_idx = 28;
-            max_length                       = prompt_template_encode_start_idx + 256;
+            min_length                       = prompt_template_encode_start_idx + 256;
 
             prompt = "<|im_start|>user\nDescribe the image by detailing the color, quantity, text, shape, size, texture, spatial relationships of the objects and background:";
 
@@ -1935,8 +1931,8 @@ struct LLMEmbedder : public Conditioner {
         auto hidden_states = encode_prompt(n_threads,
                                            prompt,
                                            prompt_attn_range,
-                                           max_length,
                                            min_length,
+                                           hidden_states_min_length,
                                            image_embeds,
                                            out_layers,
                                            prompt_template_encode_start_idx);
@@ -1945,8 +1941,8 @@ struct LLMEmbedder : public Conditioner {
             auto extra_hidden_states = encode_prompt(n_threads,
                                                      extra_prompts[i],
                                                      extra_prompts_attn_range[i],
-                                                     max_length,
                                                      min_length,
+                                                     hidden_states_min_length,
                                                      image_embeds,
                                                      out_layers,
                                                      prompt_template_encode_start_idx);

src/denoiser.hpp

@@ -953,8 +953,9 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
             float t_next         = t_fn(sigma_down);
             float h              = t_next - t;
             float s              = t + 0.5f * h;
-            sd::Tensor<float> x2 = (sigma_fn(s) / sigma_fn(t)) * x - (exp(-h * 0.5f) - 1) * denoised;
-            auto denoised2_opt   = model(x2, sigmas[i + 1], i + 1);
+            float sigma_s        = sigma_fn(s);
+            sd::Tensor<float> x2 = (sigma_s / sigma_fn(t)) * x - (exp(-h * 0.5f) - 1) * denoised;
+            auto denoised2_opt   = model(x2, sigma_s, i + 1);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -969,6 +970,102 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
     return x;
 }
 
+static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
+                                                        sd::Tensor<float> x,
+                                                        const std::vector<float>& sigmas,
+                                                        std::shared_ptr<RNG> rng,
+                                                        float eta = 1.0f) {
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        float sigma = sigmas[i];
+        float sigma_to = sigmas[i + 1];
+
+        bool opt_first_step = (1.0 - sigma < 1e-6);
+
+        auto denoised_opt = model(x, sigma, (opt_first_step ? 1 : -1) * (i + 1));
+        if (denoised_opt.empty()) {
+            return {};
+        }
+        sd::Tensor<float> denoised = std::move(denoised_opt);
+
+        if (sigma_to == 0.0f) {
+            // Euler method (final step, no noise)
+            // sigma_to == 0 --> sigma_down = 0, so:
+            //   x +                       d  * (sigma_down - sigma)
+            // = x + ((x - denoised) / sigma) * (sigma_down - sigma)
+            // = x + ((x - denoised) / sigma) * (         0 - sigma)
+            // = x + ((x - denoised)        ) * -1
+            // = x    -x + denoised
+            x = denoised;
+
+        } else {
+            auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step_flow(sigma, sigma_to, eta);
+            sd::Tensor<float> D_i;
+
+            if (opt_first_step) {
+                // the reformulated exp_s calc already accounts for this, but we can avoid
+                // a redundant model call for the typical sigma 1 at the first step:
+                // exp_s = sqrt((1-sigma)/sigma * (1-sigma_down)/sigma_down)
+                //       = sqrt((1-    1)/    1 * (1-sigma_down)/sigma_down)
+                //       = 0
+                // so sigma_s = 1 = sigma, and sigma_s_i_ratio = sigma_s / sigma = 1
+                // u = (x*sigma_s_i_ratio)+(denoised*(1.0f-sigma_s_i_ratio))
+                //   = (x*1)+(denoised*0) = x
+                // so D_i = model(u, sigma_s, i + 1)
+                //        = model(x, sigma,   i + 1)
+                //        = denoised
+                D_i = denoised;
+
+            } else {
+                float sigma_s;
+
+                // ref implementation would be:
+                // auto lambda_fn = [](float sigma) -> float {
+                //     return std::log((1.0f - sigma) / sigma); };
+                // auto sigma_fn = [](float lbda) -> float {
+                //     return 1.0f / (std::exp(lbda) + 1.0f); };
+                // t_i     = lambda_fn(sigma);
+                // t_down  = lambda_fn(sigma_down);
+                // float r = 0.5f;
+                // h       = t_down - t_i;
+                // s       = t_i + r * h;
+                // sigma_s = sigma_fn(s);
+
+                // assuming r is constant, we sidestep the singularity at sigma -> 1 by:
+                // s   = 0.5 * (lambda_fn(sigma)     + lambda_fn(sigma_down))
+                //     = 0.5 * (log((1-sigma)/sigma) + log((1-sigma_down)/sigma_down))
+                //     = 0.5 * log(((1-sigma)/sigma) *    ((1-sigma_down)/sigma_down))
+                //     = log(sqrt (((1-sigma)/sigma) *    ((1-sigma_down)/sigma_down)))
+                // so exp(s) = sqrt((1-sigma)/sigma  *     (1-sigma_down)/sigma_down)
+                // and sigma_s = sigma_fn(s) = 1.0f / (exp(s) + 1.0f)
+
+                float exp_s = std::sqrt(((1 - sigma) / sigma) * ((1 - sigma_down) / sigma_down));
+                sigma_s = 1.0f / (exp_s + 1.0f);
+
+                float sigma_s_i_ratio = sigma_s / sigma;
+                sd::Tensor<float> u = (x * sigma_s_i_ratio) + (denoised * (1.0f - sigma_s_i_ratio));
+
+                auto denoised2_opt = model(u, sigma_s, i + 1);
+                if (denoised2_opt.empty()) {
+                    return {};
+                }
+                D_i = std::move(denoised2_opt);
+            }
+
+            float sigma_down_i_ratio = sigma_down / sigma;
+            x = (x * sigma_down_i_ratio) + (D_i * (1.0f - sigma_down_i_ratio));
+
+            if (sigma_to > 0.0f && eta > 0.0f) {
+                x = alpha_scale * x + sd::Tensor<float>::randn_like(x, rng) * sigma_up;
+            }
+        }
+    }
+
+    return x;
+}
+
+
+
 static sd::Tensor<float> sample_dpmpp_2m(denoise_cb_t model,
                                          sd::Tensor<float> x,
                                          const std::vector<float>& sigmas) {
@@ -1285,6 +1382,140 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
     return x;
 }
 
+static sd::Tensor<float> sample_er_sde(denoise_cb_t model,
+                                       sd::Tensor<float> x,
+                                       std::vector<float> sigmas,
+                                       std::shared_ptr<RNG> rng,
+                                       bool is_flow_denoiser,
+                                       float eta) {
+    constexpr int max_stage                  = 3;
+    constexpr int num_integration_points     = 200;
+    constexpr float num_integration_points_f = 200.0f;
+    float s_noise                            = eta;
+
+    auto er_sde_flow_sigma = [](float sigma) -> float {
+        sigma = std::max(sigma, 1e-6f);
+        sigma = std::min(sigma, 1.0f - 1e-4f);
+        return sigma;
+    };
+
+    auto sigma_to_er_sde_lambda = [&](float sigma, bool is_flow_denoiser) -> float {
+        if (is_flow_denoiser) {
+            sigma = er_sde_flow_sigma(sigma);
+            return sigma / std::max(1.0f - sigma, 1e-6f);
+        }
+        return std::max(sigma, 1e-6f);
+    };
+
+    auto sigma_to_er_sde_alpha = [&](float sigma, bool is_flow_denoiser) -> float {
+        if (is_flow_denoiser) {
+            sigma = er_sde_flow_sigma(sigma);
+            return 1.0f - sigma;
+        }
+        return 1.0f;
+    };
+
+    auto er_sde_noise_scaler = [](float x) -> float {
+        x = std::max(x, 0.0f);
+        return x * (std::exp(std::pow(x, 0.3f)) + 10.0f);
+    };
+
+    if (is_flow_denoiser) {
+        for (size_t i = 0; i + 1 < sigmas.size(); ++i) {
+            if (sigmas[i] > 1.0f) {
+                sigmas[i] = er_sde_flow_sigma(sigmas[i]);
+            }
+        }
+    }
+
+    std::vector<float> er_lambdas(sigmas.size(), 0.0f);
+    for (size_t i = 0; i < sigmas.size(); ++i) {
+        er_lambdas[i] = sigma_to_er_sde_lambda(sigmas[i], is_flow_denoiser);
+    }
+
+    sd::Tensor<float> old_denoised   = x;
+    sd::Tensor<float> old_denoised_d = x;
+    bool have_old_denoised           = false;
+    bool have_old_denoised_d         = false;
+
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        sd::Tensor<float> denoised = model(x, sigmas[i], i + 1);
+        if (denoised.empty()) {
+            return {};
+        }
+
+        int stage_used = std::min(max_stage, i + 1);
+
+        if (sigmas[i + 1] == 0.0f) {
+            x = denoised;
+        } else {
+            float er_lambda_s = er_lambdas[i];
+            float er_lambda_t = er_lambdas[i + 1];
+            float alpha_s     = sigma_to_er_sde_alpha(sigmas[i], is_flow_denoiser);
+            float alpha_t     = sigma_to_er_sde_alpha(sigmas[i + 1], is_flow_denoiser);
+            float scaled_s    = er_sde_noise_scaler(er_lambda_s);
+            float scaled_t    = er_sde_noise_scaler(er_lambda_t);
+            float r_alpha     = alpha_s > 0.0f ? alpha_t / alpha_s : 0.0f;
+            float r           = scaled_s > 0.0f ? scaled_t / scaled_s : 0.0f;
+
+            x = r_alpha * r * x + alpha_t * (1.0f - r) * denoised;
+
+            if (stage_used >= 2 && have_old_denoised) {
+                float dt               = er_lambda_t - er_lambda_s;
+                float lambda_step_size = -dt / num_integration_points_f;
+                float s                = 0.0f;
+                float s_u              = 0.0f;
+
+                for (int p = 0; p < num_integration_points; ++p) {
+                    float lambda_pos = er_lambda_t + p * lambda_step_size;
+                    float scaled_pos = er_sde_noise_scaler(lambda_pos);
+                    if (scaled_pos <= 0.0f) {
+                        continue;
+                    }
+
+                    s += 1.0f / scaled_pos;
+                    if (stage_used >= 3 && have_old_denoised_d) {
+                        s_u += (lambda_pos - er_lambda_s) / scaled_pos;
+                    }
+                }
+
+                s *= lambda_step_size;
+
+                float denom_d = er_lambda_s - er_lambdas[i - 1];
+                if (std::fabs(denom_d) > 1e-12f) {
+                    float coeff_d                = alpha_t * (dt + s * scaled_t);
+                    sd::Tensor<float> denoised_d = (denoised - old_denoised) / denom_d;
+                    x += coeff_d * denoised_d;
+
+                    if (stage_used >= 3 && have_old_denoised_d) {
+                        float denom_u = (er_lambda_s - er_lambdas[i - 2]) * 0.5f;
+                        if (std::fabs(denom_u) > 1e-12f) {
+                            s_u *= lambda_step_size;
+                            float coeff_u                = alpha_t * (0.5f * dt * dt + s_u * scaled_t);
+                            sd::Tensor<float> denoised_u = (denoised_d - old_denoised_d) / denom_u;
+                            x += coeff_u * denoised_u;
+                        }
+                    }
+
+                    old_denoised_d      = denoised_d;
+                    have_old_denoised_d = true;
+                }
+            }
+
+            float noise_scale_sq = er_lambda_t * er_lambda_t - er_lambda_s * er_lambda_s * r * r;
+            if (s_noise > 0.0f && noise_scale_sq > 0.0f) {
+                float noise_scale = alpha_t * std::sqrt(std::max(noise_scale_sq, 0.0f));
+                x += sd::Tensor<float>::randn_like(x, rng) * noise_scale;
+            }
+        }
+
+        old_denoised      = denoised;
+        have_old_denoised = true;
+    }
+    return x;
+}
+
 static sd::Tensor<float> sample_ddim_trailing(denoise_cb_t model,
                                               sd::Tensor<float> x,
                                               const std::vector<float>& sigmas,
@@ -1431,7 +1662,10 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
         case DPM2_SAMPLE_METHOD:
             return sample_dpm2(model, std::move(x), sigmas);
         case DPMPP2S_A_SAMPLE_METHOD:
-            return sample_dpmpp_2s_ancestral(model, std::move(x), sigmas, rng, eta);
+            if (is_flow_denoiser)
+                return sample_dpmpp_2s_ancestral_flow(model, std::move(x), sigmas, rng, eta);
+            else
+                return sample_dpmpp_2s_ancestral(model, std::move(x), sigmas, rng, eta);
         case DPMPP2M_SAMPLE_METHOD:
             return sample_dpmpp_2m(model, std::move(x), sigmas);
         case DPMPP2Mv2_SAMPLE_METHOD:
@@ -1446,6 +1680,8 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
             return sample_res_multistep(model, std::move(x), sigmas, rng, eta);
         case RES_2S_SAMPLE_METHOD:
             return sample_res_2s(model, std::move(x), sigmas, rng, eta);
+        case ER_SDE_SAMPLE_METHOD:
+            return sample_er_sde(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case DDIM_TRAILING_SAMPLE_METHOD:
             return sample_ddim_trailing(model, std::move(x), sigmas, rng, eta);
         case TCD_SAMPLE_METHOD:

src/diffusion_model.hpp

@@ -3,6 +3,7 @@
 
 #include <optional>
 #include "anima.hpp"
+#include "ernie_image.hpp"
 #include "flux.hpp"
 #include "mmdit.hpp"
 #include "qwen_image.hpp"
@@ -516,4 +517,66 @@ struct ZImageModel : public DiffusionModel {
     }
 };
 
+struct ErnieImageModel : public DiffusionModel {
+    std::string prefix;
+    ErnieImage::ErnieImageRunner ernie_image;
+
+    ErnieImageModel(ggml_backend_t backend,
+                    bool offload_params_to_cpu,
+                    const String2TensorStorage& tensor_storage_map = {},
+                    const std::string prefix                       = "model.diffusion_model")
+        : prefix(prefix), ernie_image(backend, offload_params_to_cpu, tensor_storage_map, prefix) {
+    }
+
+    std::string get_desc() override {
+        return ernie_image.get_desc();
+    }
+
+    void alloc_params_buffer() override {
+        ernie_image.alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        ernie_image.free_params_buffer();
+    }
+
+    void free_compute_buffer() override {
+        ernie_image.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        ernie_image.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() override {
+        return ernie_image.get_params_buffer_size();
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        ernie_image.set_weight_adapter(adapter);
+    }
+
+    int64_t get_adm_in_channels() override {
+        return 768;
+    }
+
+    void set_flash_attention_enabled(bool enabled) {
+        ernie_image.set_flash_attention_enabled(enabled);
+    }
+
+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        ernie_image.set_circular_axes(circular_x, circular_y);
+    }
+
+    sd::Tensor<float> compute(int n_threads,
+                              const DiffusionParams& diffusion_params) override {
+        GGML_ASSERT(diffusion_params.x != nullptr);
+        GGML_ASSERT(diffusion_params.timesteps != nullptr);
+        return ernie_image.compute(n_threads,
+                                   *diffusion_params.x,
+                                   *diffusion_params.timesteps,
+                                   tensor_or_empty(diffusion_params.context));
+    }
+};
+
 #endif

src/ernie_image.hpp

@@ -0,0 +1,438 @@
+#ifndef __SD_ERNIE_IMAGE_HPP__
+#define __SD_ERNIE_IMAGE_HPP__
+
+#include <memory>
+#include <vector>
+
+#include "common_dit.hpp"
+#include "flux.hpp"
+#include "qwen_image.hpp"
+#include "rope.hpp"
+
+namespace ErnieImage {
+    constexpr int ERNIE_IMAGE_GRAPH_SIZE = 40960;
+
+    __STATIC_INLINE__ ggml_tensor* timestep_embedding_sin_cos(ggml_context* ctx,
+                                                              ggml_tensor* timesteps,
+                                                              int dim,
+                                                              int max_period = 10000) {
+        auto emb       = ggml_ext_timestep_embedding(ctx, timesteps, dim, max_period, 1.0f);
+        int64_t half   = dim / 2;
+        auto cos_part  = ggml_view_2d(ctx, emb, half, emb->ne[1], emb->nb[1], 0);
+        auto sin_part  = ggml_view_2d(ctx, emb, half, emb->ne[1], emb->nb[1], half * emb->nb[0]);
+        auto sin_first = ggml_concat(ctx, sin_part, cos_part, 0);
+        return sin_first;
+    }
+
+    __STATIC_INLINE__ ggml_tensor* apply_rotary_emb(ggml_context* ctx, ggml_tensor* x, ggml_tensor* pe) {
+        // x: [N, S, heads, head_dim]
+        // pe: [2, S, 1, head_dim], stored as ggml [head_dim, 1, S, 2].
+        int64_t head_dim = x->ne[0];
+        int64_t heads    = x->ne[1];
+        int64_t S        = x->ne[2];
+        int64_t N        = x->ne[3];
+        int64_t rot_dim  = pe->ne[0];
+        GGML_ASSERT(rot_dim <= head_dim);
+        GGML_ASSERT(rot_dim % 2 == 0);
+        GGML_ASSERT(pe->ne[1] == 1 && pe->ne[2] == S && pe->ne[3] == 2);
+
+        x           = ggml_cont(ctx, x);
+        auto x_rot  = ggml_ext_slice(ctx, x, 0, 0, rot_dim, false);
+        auto x_pass = rot_dim < head_dim ? ggml_ext_slice(ctx, x, 0, rot_dim, head_dim, false) : nullptr;
+
+        int64_t half = rot_dim / 2;
+        auto x1      = ggml_view_4d(ctx, x_rot, half, heads, S, N, x_rot->nb[1], x_rot->nb[2], x_rot->nb[3], 0);
+        auto x2      = ggml_view_4d(ctx, x_rot, half, heads, S, N, x_rot->nb[1], x_rot->nb[2], x_rot->nb[3], half * x_rot->nb[0]);
+        x1           = ggml_cont(ctx, x1);
+        x2           = ggml_cont(ctx, x2);
+        auto rotated = ggml_concat(ctx, ggml_neg(ctx, x2), x1, 0);
+
+        auto cos_emb = ggml_ext_slice(ctx, pe, 3, 0, 1, false);
+        auto sin_emb = ggml_ext_slice(ctx, pe, 3, 1, 2, false);
+
+        auto out = ggml_add(ctx, ggml_mul(ctx, x_rot, cos_emb), ggml_mul(ctx, rotated, sin_emb));
+        if (x_pass != nullptr) {
+            out = ggml_concat(ctx, out, x_pass, 0);
+        }
+        return out;
+    }
+
+    struct ErnieImageAttention : public GGMLBlock {
+        int64_t num_heads;
+        int64_t head_dim;
+
+        ErnieImageAttention(int64_t query_dim,
+                            int64_t heads,
+                            int64_t dim_head,
+                            float eps = 1e-6f)
+            : num_heads(heads), head_dim(dim_head) {
+            int64_t inner_dim  = heads * dim_head;
+            blocks["to_q"]     = std::make_shared<Linear>(query_dim, inner_dim, false);
+            blocks["to_k"]     = std::make_shared<Linear>(query_dim, inner_dim, false);
+            blocks["to_v"]     = std::make_shared<Linear>(query_dim, inner_dim, false);
+            blocks["norm_q"]   = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["norm_k"]   = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["to_out.0"] = std::make_shared<Linear>(inner_dim, query_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* pe,
+                             ggml_tensor* attention_mask = nullptr) {
+            // x: [N, S, hidden_size]
+            // pe: [S, head_dim/2, 2, 2], generated in image-token-first order.
+            auto to_q     = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
+            auto to_k     = std::dynamic_pointer_cast<Linear>(blocks["to_k"]);
+            auto to_v     = std::dynamic_pointer_cast<Linear>(blocks["to_v"]);
+            auto norm_q   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"]);
+            auto norm_k   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"]);
+            auto to_out_0 = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+
+            int64_t S = x->ne[1];
+            int64_t N = x->ne[2];
+
+            auto q = to_q->forward(ctx, x);
+            auto k = to_k->forward(ctx, x);
+            auto v = to_v->forward(ctx, x);
+
+            q = ggml_reshape_4d(ctx->ggml_ctx, q, head_dim, num_heads, S, N);  // [N, S, heads, head_dim]
+            k = ggml_reshape_4d(ctx->ggml_ctx, k, head_dim, num_heads, S, N);  // [N, S, heads, head_dim]
+            v = ggml_reshape_4d(ctx->ggml_ctx, v, head_dim, num_heads, S, N);  // [N, S, heads, head_dim]
+
+            q = norm_q->forward(ctx, q);
+            k = norm_k->forward(ctx, k);
+
+            q = apply_rotary_emb(ctx->ggml_ctx, q, pe);
+            k = apply_rotary_emb(ctx->ggml_ctx, k, pe);
+
+            q = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, q, 0, 2, 1, 3));  // [N, heads, S, head_dim]
+            q = ggml_reshape_3d(ctx->ggml_ctx, q, q->ne[0], q->ne[1], q->ne[2] * q->ne[3]);
+
+            k = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, k, 0, 2, 1, 3));  // [N, heads, S, head_dim]
+            k = ggml_reshape_3d(ctx->ggml_ctx, k, k->ne[0], k->ne[1], k->ne[2] * k->ne[3]);
+
+            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, attention_mask, true, ctx->flash_attn_enabled);  // [N, S, hidden_size]
+            x = to_out_0->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct ErnieImageFeedForward : public GGMLBlock {
+    public:
+        ErnieImageFeedForward(int64_t hidden_size, int64_t ffn_hidden_size) {
+            blocks["gate_proj"]  = std::make_shared<Linear>(hidden_size, ffn_hidden_size, false);
+            blocks["up_proj"]    = std::make_shared<Linear>(hidden_size, ffn_hidden_size, false);
+            blocks["linear_fc2"] = std::make_shared<Linear>(ffn_hidden_size, hidden_size, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto gate_proj  = std::dynamic_pointer_cast<Linear>(blocks["gate_proj"]);
+            auto up_proj    = std::dynamic_pointer_cast<Linear>(blocks["up_proj"]);
+            auto linear_fc2 = std::dynamic_pointer_cast<Linear>(blocks["linear_fc2"]);
+
+            auto gate = gate_proj->forward(ctx, x);
+            gate      = ggml_ext_gelu(ctx->ggml_ctx, gate);
+            x         = up_proj->forward(ctx, x);
+            x         = ggml_mul(ctx->ggml_ctx, x, gate);
+            x         = linear_fc2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct ErnieImageSharedAdaLNBlock : public GGMLBlock {
+    public:
+        ErnieImageSharedAdaLNBlock(int64_t hidden_size,
+                                   int64_t num_heads,
+                                   int64_t ffn_hidden_size,
+                                   float eps = 1e-6f) {
+            blocks["adaLN_sa_ln"]    = std::make_shared<RMSNorm>(hidden_size, eps);
+            blocks["self_attention"] = std::make_shared<ErnieImageAttention>(hidden_size,
+                                                                             num_heads,
+                                                                             hidden_size / num_heads,
+                                                                             eps);
+            blocks["adaLN_mlp_ln"]   = std::make_shared<RMSNorm>(hidden_size, eps);
+            blocks["mlp"]            = std::make_shared<ErnieImageFeedForward>(hidden_size, ffn_hidden_size);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* pe,
+                             const std::vector<ggml_tensor*>& temb,
+                             ggml_tensor* attention_mask = nullptr) {
+            // x: [N, image_tokens + text_tokens, hidden_size]
+            auto adaLN_sa_ln    = std::dynamic_pointer_cast<RMSNorm>(blocks["adaLN_sa_ln"]);
+            auto self_attention = std::dynamic_pointer_cast<ErnieImageAttention>(blocks["self_attention"]);
+            auto adaLN_mlp_ln   = std::dynamic_pointer_cast<RMSNorm>(blocks["adaLN_mlp_ln"]);
+            auto mlp            = std::dynamic_pointer_cast<ErnieImageFeedForward>(blocks["mlp"]);
+
+            auto shift_msa = temb[0];
+            auto scale_msa = temb[1];
+            auto gate_msa  = temb[2];
+            auto shift_mlp = temb[3];
+            auto scale_mlp = temb[4];
+            auto gate_mlp  = temb[5];
+
+            auto residual = x;
+            x             = adaLN_sa_ln->forward(ctx, x);
+            x             = Flux::modulate(ctx->ggml_ctx, x, shift_msa, scale_msa, true);
+            auto attn_out = self_attention->forward(ctx, x, pe, attention_mask);
+            x             = ggml_add(ctx->ggml_ctx, residual, ggml_mul(ctx->ggml_ctx, attn_out, gate_msa));
+
+            residual = x;
+            x        = adaLN_mlp_ln->forward(ctx, x);
+            x        = Flux::modulate(ctx->ggml_ctx, x, shift_mlp, scale_mlp, true);
+            x        = ggml_add(ctx->ggml_ctx, residual, ggml_mul(ctx->ggml_ctx, mlp->forward(ctx, x), gate_mlp));
+            return x;
+        }
+    };
+
+    struct ErnieImageAdaLNContinuous : public GGMLBlock {
+    public:
+        ErnieImageAdaLNContinuous(int64_t hidden_size, float eps = 1e-6f) {
+            blocks["norm"]   = std::make_shared<LayerNorm>(hidden_size, eps, false);
+            blocks["linear"] = std::make_shared<Linear>(hidden_size, hidden_size * 2, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* conditioning) {
+            auto norm   = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+
+            auto mods  = ggml_ext_chunk(ctx->ggml_ctx, linear->forward(ctx, conditioning), 2, 0);
+            auto scale = mods[0];
+            auto shift = mods[1];
+
+            x = norm->forward(ctx, x);
+            x = Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+            return x;
+        }
+    };
+
+    struct ErnieImageParams {
+        int64_t hidden_size       = 4096;
+        int64_t num_heads         = 32;
+        int64_t num_layers        = 36;
+        int64_t ffn_hidden_size   = 12288;
+        int64_t in_channels       = 128;
+        int64_t out_channels      = 128;
+        int patch_size            = 1;
+        int64_t text_in_dim       = 3072;
+        int theta                 = 256;
+        std::vector<int> axes_dim = {32, 48, 48};
+        int axes_dim_sum          = 128;
+        float eps                 = 1e-6f;
+    };
+
+    class ErnieImageModel : public GGMLBlock {
+    public:
+        ErnieImageParams params;
+
+        ErnieImageModel() = default;
+        ErnieImageModel(ErnieImageParams params)
+            : params(params) {
+            blocks["x_embedder.proj"] = std::make_shared<Conv2d>(params.in_channels,
+                                                                 params.hidden_size,
+                                                                 std::pair<int, int>{params.patch_size, params.patch_size},
+                                                                 std::pair<int, int>{params.patch_size, params.patch_size},
+                                                                 std::pair<int, int>{0, 0},
+                                                                 std::pair<int, int>{1, 1},
+                                                                 true);
+            if (params.text_in_dim != params.hidden_size) {
+                blocks["text_proj"] = std::make_shared<Linear>(params.text_in_dim, params.hidden_size, false);
+            }
+            blocks["time_embedding"]     = std::make_shared<Qwen::TimestepEmbedding>(params.hidden_size, params.hidden_size);
+            blocks["adaLN_modulation.1"] = std::make_shared<Linear>(params.hidden_size, 6 * params.hidden_size, true);
+
+            for (int i = 0; i < params.num_layers; i++) {
+                blocks["layers." + std::to_string(i)] = std::make_shared<ErnieImageSharedAdaLNBlock>(params.hidden_size,
+                                                                                                     params.num_heads,
+                                                                                                     params.ffn_hidden_size,
+                                                                                                     params.eps);
+            }
+
+            blocks["final_norm"]   = std::make_shared<ErnieImageAdaLNContinuous>(params.hidden_size, params.eps);
+            blocks["final_linear"] = std::make_shared<Linear>(params.hidden_size,
+                                                              params.patch_size * params.patch_size * params.out_channels,
+                                                              true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timestep,
+                             ggml_tensor* context,
+                             ggml_tensor* pe) {
+            // x: [N, C, H, W]
+            // context: [N, text_tokens, 3072]
+            // pe: [image_tokens + text_tokens, head_dim/2, 2, 2]
+            GGML_ASSERT(context != nullptr);
+            GGML_ASSERT(x->ne[1] % params.patch_size == 0 && x->ne[0] % params.patch_size == 0);
+
+            int64_t W     = x->ne[0];
+            int64_t H     = x->ne[1];
+            int64_t Hp    = H / params.patch_size;
+            int64_t Wp    = W / params.patch_size;
+            int64_t n_img = Hp * Wp;
+            int64_t N     = x->ne[3];
+
+            auto x_embedder_proj = std::dynamic_pointer_cast<Conv2d>(blocks["x_embedder.proj"]);
+            auto time_embedding  = std::dynamic_pointer_cast<Qwen::TimestepEmbedding>(blocks["time_embedding"]);
+            auto adaLN_mod       = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation.1"]);
+            auto final_norm      = std::dynamic_pointer_cast<ErnieImageAdaLNContinuous>(blocks["final_norm"]);
+            auto final_linear    = std::dynamic_pointer_cast<Linear>(blocks["final_linear"]);
+
+            auto img = x_embedder_proj->forward(ctx, x);                                                  // [N, hidden_size, Hp, Wp]
+            img      = ggml_reshape_3d(ctx->ggml_ctx, img, img->ne[0] * img->ne[1], img->ne[2], N);       // [N, hidden_size, image_tokens]
+            img      = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, img, 1, 0, 2, 3));  // [N, image_tokens, hidden_size]
+
+            auto txt       = context;
+            auto text_proj = std::dynamic_pointer_cast<Linear>(blocks["text_proj"]);
+            if (text_proj) {
+                txt = text_proj->forward(ctx, txt);
+            }
+
+            auto hidden_states = ggml_concat(ctx->ggml_ctx, img, txt, 1);  // [N, image_tokens + text_tokens, hidden_size]
+
+            auto sample = timestep_embedding_sin_cos(ctx->ggml_ctx, timestep, static_cast<int>(params.hidden_size));
+            auto c      = time_embedding->forward(ctx, sample);  // [N, hidden_size]
+
+            auto mod_params = adaLN_mod->forward(ctx, ggml_silu(ctx->ggml_ctx, c));  // [N, 6 * hidden_size]
+            auto chunks     = ggml_ext_chunk(ctx->ggml_ctx, mod_params, 6, 0);
+            std::vector<ggml_tensor*> temb;
+            temb.reserve(6);
+            for (auto chunk : chunks) {
+                temb.push_back(ggml_reshape_3d(ctx->ggml_ctx, chunk, chunk->ne[0], 1, chunk->ne[1]));  // [N, 1, hidden_size]
+            }
+
+            for (int i = 0; i < params.num_layers; i++) {
+                auto layer    = std::dynamic_pointer_cast<ErnieImageSharedAdaLNBlock>(blocks["layers." + std::to_string(i)]);
+                hidden_states = layer->forward(ctx, hidden_states, pe, temb);
+            }
+
+            hidden_states = final_norm->forward(ctx, hidden_states, c);
+            hidden_states = final_linear->forward(ctx, hidden_states);                  // [N, image_tokens, p*p*out_channels]
+            auto patches  = ggml_ext_slice(ctx->ggml_ctx, hidden_states, 1, 0, n_img);  // [N, image_tokens, hidden_size]
+
+            auto out = DiT::unpatchify(ctx->ggml_ctx,
+                                       patches,
+                                       Hp,
+                                       Wp,
+                                       params.patch_size,
+                                       params.patch_size,
+                                       false);  // [N, out_channels, H, W]
+            return out;
+        }
+    };
+
+    struct ErnieImageRunner : public GGMLRunner {
+        ErnieImageParams ernie_params;
+        ErnieImageModel ernie_image;
+        std::vector<float> pe_vec;
+
+        ErnieImageRunner(ggml_backend_t backend,
+                         bool offload_params_to_cpu,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         const std::string prefix                       = "")
+            : GGMLRunner(backend, offload_params_to_cpu) {
+            ernie_params.num_layers = 0;
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (ends_with(name, "x_embedder.proj.weight") && tensor_storage.n_dims == 4) {
+                    ernie_params.patch_size  = static_cast<int>(tensor_storage.ne[0]);
+                    ernie_params.in_channels = tensor_storage.ne[2];
+                    ernie_params.hidden_size = tensor_storage.ne[3];
+                } else if (ends_with(name, "text_proj.weight") && tensor_storage.n_dims == 2) {
+                    ernie_params.text_in_dim = tensor_storage.ne[0];
+                } else if (ends_with(name, "layers.0.self_attention.norm_q.weight")) {
+                    int64_t head_dim       = tensor_storage.ne[0];
+                    ernie_params.num_heads = ernie_params.hidden_size / head_dim;
+                } else if (ends_with(name, "layers.0.mlp.gate_proj.weight") && tensor_storage.n_dims == 2) {
+                    ernie_params.ffn_hidden_size = tensor_storage.ne[1];
+                } else if (ends_with(name, "final_linear.weight") && tensor_storage.n_dims == 2) {
+                    int64_t out_dim           = tensor_storage.ne[1];
+                    ernie_params.out_channels = out_dim / ernie_params.patch_size / ernie_params.patch_size;
+                }
+
+                size_t pos = name.find("layers.");
+                if (pos != std::string::npos) {
+                    std::string layer_name = name.substr(pos);
+                    auto items             = split_string(layer_name, '.');
+                    if (items.size() > 1) {
+                        int block_index = atoi(items[1].c_str());
+                        if (block_index + 1 > ernie_params.num_layers) {
+                            ernie_params.num_layers = block_index + 1;
+                        }
+                    }
+                }
+            }
+            if (ernie_params.num_layers == 0) {
+                ernie_params.num_layers = 36;
+            }
+            ernie_params.axes_dim_sum = 0;
+            for (int axis_dim : ernie_params.axes_dim) {
+                ernie_params.axes_dim_sum += axis_dim;
+            }
+
+            LOG_INFO("ernie_image: layers = %" PRId64 ", hidden_size = %" PRId64 ", heads = %" PRId64
+                     ", ffn_hidden_size = %" PRId64 ", in_channels = %" PRId64 ", out_channels = %" PRId64,
+                     ernie_params.num_layers,
+                     ernie_params.hidden_size,
+                     ernie_params.num_heads,
+                     ernie_params.ffn_hidden_size,
+                     ernie_params.in_channels,
+                     ernie_params.out_channels);
+
+            ernie_image = ErnieImageModel(ernie_params);
+            ernie_image.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "ernie_image";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+            ernie_image.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor) {
+            ggml_cgraph* gf        = new_graph_custom(ERNIE_IMAGE_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timesteps = make_input(timesteps_tensor);
+            GGML_ASSERT(x->ne[3] == 1);
+            GGML_ASSERT(!context_tensor.empty());
+            ggml_tensor* context = make_input(context_tensor);
+
+            pe_vec      = Rope::gen_ernie_image_pe(static_cast<int>(x->ne[1]),
+                                                   static_cast<int>(x->ne[0]),
+                                                   ernie_params.patch_size,
+                                                   static_cast<int>(x->ne[3]),
+                                                   static_cast<int>(context->ne[1]),
+                                                   ernie_params.theta,
+                                                   circular_y_enabled,
+                                                   circular_x_enabled,
+                                                   ernie_params.axes_dim);
+            int pos_len = static_cast<int>(pe_vec.size() / ernie_params.axes_dim_sum / 2);
+            auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, ernie_params.axes_dim_sum, 1, pos_len, 2);
+            set_backend_tensor_data(pe, pe_vec.data());
+
+            auto runner_ctx  = get_context();
+            ggml_tensor* out = ernie_image.forward(&runner_ctx, x, timesteps, context, pe);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+        }
+    };
+}  // namespace ErnieImage
+
+#endif  // __SD_ERNIE_IMAGE_HPP__

src/llm.hpp

@@ -14,469 +14,21 @@
 #include <utility>
 #include <vector>
 
-#include "clip.hpp"
 #include "ggml_extend.hpp"
 #include "json.hpp"
 #include "rope.hpp"
-#include "tokenize_util.h"
-#include "vocab/vocab.h"
+#include "tokenizers/bpe_tokenizer.h"
+#include "tokenizers/mistral_tokenizer.h"
+#include "tokenizers/qwen2_tokenizer.h"
 
 namespace LLM {
     constexpr int LLM_GRAPH_SIZE = 10240;
 
-    class BPETokenizer {
-    protected:
-        std::map<int, std::u32string> byte_encoder;
-        std::map<std::u32string, int> byte_decoder;
-        std::map<std::u32string, int> encoder;
-        std::map<int, std::u32string> decoder;
-        std::map<std::pair<std::u32string, std::u32string>, int> bpe_ranks;
-        std::regex pat;
-        int encoder_len;
-        int bpe_len;
-
-        std::string UNK_TOKEN;
-        std::string BOS_TOKEN;
-        std::string EOS_TOKEN;
-        std::string PAD_TOKEN;
-
-        int UNK_TOKEN_ID;
-        int BOS_TOKEN_ID;
-        int EOS_TOKEN_ID;
-        int PAD_TOKEN_ID;
-
-        std::vector<std::string> special_tokens;
-
-        bool add_bos_token = false;
-
-    protected:
-        static std::string strip(const std::string& str) {
-            std::string::size_type start = str.find_first_not_of(" \t\n\r\v\f");
-            std::string::size_type end   = str.find_last_not_of(" \t\n\r\v\f");
-
-            if (start == std::string::npos) {
-                // String contains only whitespace characters
-                return "";
-            }
-
-            return str.substr(start, end - start + 1);
-        }
-
-        static std::string whitespace_clean(std::string text) {
-            text = std::regex_replace(text, std::regex(R"(\s+)"), " ");
-            text = strip(text);
-            return text;
-        }
-
-        static std::set<std::pair<std::u32string, std::u32string>> get_pairs(const std::vector<std::u32string>& subwords) {
-            std::set<std::pair<std::u32string, std::u32string>> pairs;
-            if (subwords.size() == 0) {
-                return pairs;
-            }
-            std::u32string prev_subword = subwords[0];
-            for (int i = 1; i < subwords.size(); i++) {
-                std::u32string subword = subwords[i];
-                std::pair<std::u32string, std::u32string> pair(prev_subword, subword);
-                pairs.insert(pair);
-                prev_subword = subword;
-            }
-            return pairs;
-        }
-
-        bool is_special_token(const std::string& token) {
-            for (auto& special_token : special_tokens) {
-                if (special_token == token) {
-                    return true;
-                }
-            }
-            return false;
-        }
-
-    public:
-        BPETokenizer() = default;
-
-        std::u32string bpe(const std::u32string& token) {
-            std::vector<std::u32string> word;
-
-            for (int i = 0; i < token.size(); i++) {
-                word.emplace_back(1, token[i]);
-            }
-
-            std::set<std::pair<std::u32string, std::u32string>> pairs = get_pairs(word);
-
-            if (pairs.empty()) {
-                return token;
-            }
-
-            while (true) {
-                auto min_pair_iter = std::min_element(pairs.begin(),
-                                                      pairs.end(),
-                                                      [&](const std::pair<std::u32string, std::u32string>& a,
-                                                          const std::pair<std::u32string, std::u32string>& b) {
-                                                          if (bpe_ranks.find(a) == bpe_ranks.end()) {
-                                                              return false;
-                                                          } else if (bpe_ranks.find(b) == bpe_ranks.end()) {
-                                                              return true;
-                                                          }
-                                                          return bpe_ranks.at(a) < bpe_ranks.at(b);
-                                                      });
-
-                const std::pair<std::u32string, std::u32string>& bigram = *min_pair_iter;
-
-                if (bpe_ranks.find(bigram) == bpe_ranks.end()) {
-                    break;
-                }
-
-                std::u32string first  = bigram.first;
-                std::u32string second = bigram.second;
-                std::vector<std::u32string> new_word;
-                int32_t i = 0;
-
-                while (i < word.size()) {
-                    auto it = std::find(word.begin() + i, word.end(), first);
-                    if (it == word.end()) {
-                        new_word.insert(new_word.end(), word.begin() + i, word.end());
-                        break;
-                    }
-                    new_word.insert(new_word.end(), word.begin() + i, it);
-                    i = static_cast<int32_t>(std::distance(word.begin(), it));
-
-                    if (word[i] == first && i < static_cast<int32_t>(word.size()) - 1 && word[i + 1] == second) {
-                        new_word.push_back(first + second);
-                        i += 2;
-                    } else {
-                        new_word.push_back(word[i]);
-                        i += 1;
-                    }
-                }
-
-                word = new_word;
-
-                if (word.size() == 1) {
-                    break;
-                }
-                pairs = get_pairs(word);
-            }
-
-            std::u32string result;
-            for (int i = 0; i < word.size(); i++) {
-                result += word[i];
-                if (i != word.size() - 1) {
-                    result += utf8_to_utf32(" ");
-                }
-            }
-
-            return result;
-        }
-
-        std::vector<int> tokenize(std::string text,
-                                  on_new_token_cb_t on_new_token_cb = nullptr,
-                                  size_t max_length                 = 0,
-                                  bool padding                      = false) {
-            std::vector<int32_t> tokens = encode(text, on_new_token_cb);
-
-            if (max_length > 0) {
-                if (tokens.size() < max_length) {
-                    tokens.resize(max_length);
-                } else {
-                    if (padding) {
-                        tokens.insert(tokens.end(), max_length - tokens.size(), PAD_TOKEN_ID);
-                    }
-                }
-            }
-
-            return tokens;
-        }
-
-        void pad_tokens(std::vector<int>& tokens,
-                        std::vector<float>& weights,
-                        size_t max_length = 0,
-                        bool padding      = false) {
-            if (add_bos_token) {
-                tokens.insert(tokens.begin(), BOS_TOKEN_ID);
-                weights.insert(weights.begin(), 1.f);
-            }
-            if (max_length > 0 && padding) {
-                size_t n = static_cast<size_t>(std::ceil(tokens.size() * 1.f / max_length));
-                if (n == 0) {
-                    n = 1;
-                }
-                size_t length = max_length * n;
-                LOG_DEBUG("token length: %llu", length);
-                tokens.insert(tokens.end(), length - tokens.size(), PAD_TOKEN_ID);
-                weights.insert(weights.end(), length - weights.size(), 1.f);
-            }
-        }
-
-        std::vector<int> encode(std::string text, on_new_token_cb_t on_new_token_cb = nullptr) {
-            std::string original_text = text;
-            std::vector<int32_t> bpe_tokens;
-            std::vector<std::string> token_strs;
-
-            auto splited_texts = split_with_special_tokens(text, special_tokens);
-
-            for (auto& splited_text : splited_texts) {
-                if (is_special_token(splited_text)) {
-                    bpe_tokens.push_back(encoder[utf8_to_utf32(splited_text)]);
-                    token_strs.push_back(splited_text);
-                    continue;
-                }
-                auto tokens = token_split(splited_text);
-                for (auto& token : tokens) {
-                    if (on_new_token_cb != nullptr) {
-                        bool skip = on_new_token_cb(token, bpe_tokens);
-                        if (skip) {
-                            continue;
-                        }
-                    }
-
-                    std::string token_str = token;
-                    std::u32string utf32_token;
-                    for (int i = 0; i < token_str.length(); i++) {
-                        unsigned char b = token_str[i];
-                        utf32_token += byte_encoder[b];
-                    }
-                    auto bpe_strs = bpe(utf32_token);
-                    size_t start  = 0;
-                    size_t pos;
-                    while ((pos = bpe_strs.find(' ', start)) != std::u32string::npos) {
-                        auto bpe_str = bpe_strs.substr(start, pos - start);
-                        bpe_tokens.push_back(encoder[bpe_str]);
-                        token_strs.push_back(utf32_to_utf8(bpe_str));
-
-                        start = pos + 1;
-                    }
-                    auto bpe_str = bpe_strs.substr(start, bpe_strs.size() - start);
-                    bpe_tokens.push_back(encoder[bpe_str]);
-                    token_strs.push_back(utf32_to_utf8(bpe_str));
-                }
-            }
-
-            std::stringstream ss;
-            ss << "[";
-            for (auto token : token_strs) {
-                ss << "\"" << token << "\", ";
-            }
-            ss << "]";
-            LOG_DEBUG("split prompt \"%s\" to tokens %s", original_text.c_str(), ss.str().c_str());
-            // printf("split prompt \"%s\" to tokens %s \n", original_text.c_str(), ss.str().c_str());
-            return bpe_tokens;
-        }
-    };
-
-    class Qwen2Tokenizer : public BPETokenizer {
-    protected:
-        void load_from_merges(const std::string& merges_utf8_str) {
-            auto byte_unicode_pairs = bytes_to_unicode();
-            // printf("byte_unicode_pairs have %lu pairs \n", byte_unicode_pairs.size());
-            byte_encoder = std::map<int, std::u32string>(byte_unicode_pairs.begin(), byte_unicode_pairs.end());
-            for (auto& pair : byte_unicode_pairs) {
-                byte_decoder[pair.second] = pair.first;
-            }
-            // for (auto & pair: byte_unicode_pairs) {
-            //     std::cout << pair.first << ": " << pair.second << std::endl;
-            // }
-            std::vector<std::u32string> merges;
-            size_t start = 0;
-            size_t pos;
-            std::u32string merges_utf32_str = utf8_to_utf32(merges_utf8_str);
-            while ((pos = merges_utf32_str.find('\n', start)) != std::string::npos) {
-                merges.push_back(merges_utf32_str.substr(start, pos - start));
-                start = pos + 1;
-            }
-            LOG_DEBUG("merges size %llu", merges.size());
-            merges = std::vector<std::u32string>(merges.begin(), merges.end());
-            std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
-            // int print_num = 10;
-            for (const auto& merge : merges) {
-                size_t space_pos = merge.find(' ');
-                merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
-                // if (print_num > 0) {
-                //     print_num--;
-                //     printf("%s :: %s | %s \n", utf32_to_utf8(merge).c_str(), utf32_to_utf8(merge.substr(0, space_pos)).c_str(),
-                //                     utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
-                // }
-            }
-
-            std::vector<std::u32string> tokens;
-            for (const auto& pair : byte_unicode_pairs) {
-                tokens.push_back(pair.second);
-            }
-            for (const auto& merge : merge_pairs) {
-                tokens.push_back(merge.first + merge.second);
-            }
-            for (auto& special_token : special_tokens) {
-                tokens.push_back(utf8_to_utf32(special_token));
-            }
-
-            int i = 0;
-            for (const auto& token : tokens) {
-                encoder[token] = i;
-                decoder[i]     = token;
-                i++;
-            }
-            encoder_len = i;
-            LOG_DEBUG("vocab size: %d", encoder_len);
-
-            int rank = 0;
-            for (const auto& merge : merge_pairs) {
-                bpe_ranks[merge] = rank++;
-            }
-            bpe_len = rank;
-        };
-
-    public:
-        explicit Qwen2Tokenizer(const std::string& merges_utf8_str = "") {
-            UNK_TOKEN = "<|endoftext|>";
-            EOS_TOKEN = "<|endoftext|>";
-            PAD_TOKEN = "<|endoftext|>";
-
-            UNK_TOKEN_ID = 151643;
-            EOS_TOKEN_ID = 151643;
-            PAD_TOKEN_ID = 151643;
-
-            special_tokens = {
-                "<|endoftext|>",
-                "<|im_start|>",
-                "<|im_end|>",
-                "<|object_ref_start|>",
-                "<|object_ref_end|>",
-                "<|box_start|>",
-                "<|box_end|>",
-                "<|quad_start|>",
-                "<|quad_end|>",
-                "<|vision_start|>",
-                "<|vision_end|>",
-                "<|vision_pad|>",
-                "<|image_pad|>",
-                "<|video_pad|>",
-                "<tool_call>",
-                "</tool_call>",
-                "<|fim_prefix|>",
-                "<|fim_middle|>",
-                "<|fim_suffix|>",
-                "<|fim_pad|>",
-                "<|repo_name|>",
-                "<|file_sep|>",
-                "<tool_response>",
-                "</tool_response>",
-                "<think>",
-                "</think>",
-            };
-
-            if (merges_utf8_str.size() > 0) {
-                load_from_merges(merges_utf8_str);
-            } else {
-                load_from_merges(load_qwen2_merges());
-            }
-        }
-    };
-
-    class MistralTokenizer : public BPETokenizer {
-    protected:
-        void load_from_merges(const std::string& merges_utf8_str, const std::string& vocab_utf8_str) {
-            nlohmann::json vocab;
-
-            try {
-                vocab = nlohmann::json::parse(vocab_utf8_str);
-            } catch (const nlohmann::json::parse_error&) {
-                GGML_ABORT("invalid vocab json str");
-            }
-            for (const auto& [key, value] : vocab.items()) {
-                std::u32string token = utf8_to_utf32(key);
-                int i                = value;
-                encoder[token]       = i;
-                decoder[i]           = token;
-            }
-            encoder_len = static_cast<int>(vocab.size());
-            LOG_DEBUG("vocab size: %d", encoder_len);
-
-            auto byte_unicode_pairs = bytes_to_unicode();
-            byte_encoder            = std::map<int, std::u32string>(byte_unicode_pairs.begin(), byte_unicode_pairs.end());
-            for (auto& pair : byte_unicode_pairs) {
-                byte_decoder[pair.second] = pair.first;
-            }
-            std::vector<std::u32string> merges;
-            size_t start = 0;
-            size_t pos;
-            std::u32string merges_utf32_str = utf8_to_utf32(merges_utf8_str);
-            while ((pos = merges_utf32_str.find('\n', start)) != std::string::npos) {
-                merges.push_back(merges_utf32_str.substr(start, pos - start));
-                start = pos + 1;
-            }
-            LOG_DEBUG("merges size %llu", merges.size());
-            merges = std::vector<std::u32string>(merges.begin(), merges.end());
-            std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
-            // int print_num = 10;
-            for (const auto& merge : merges) {
-                size_t space_pos = merge.find(' ');
-                merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
-                // if (print_num > 0) {
-                //     print_num--;
-                //     printf("%s :: %s | %s \n", utf32_to_utf8(merge).c_str(), utf32_to_utf8(merge.substr(0, space_pos)).c_str(),
-                //                     utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
-                // }
-            }
-
-            int rank = 0;
-            for (const auto& merge : merge_pairs) {
-                bpe_ranks[merge] = rank++;
-            }
-            bpe_len = rank;
-        };
-
-    public:
-        explicit MistralTokenizer(const std::string& merges_utf8_str = "", const std::string& vocab_utf8_str = "") {
-            add_bos_token = true;
-
-            UNK_TOKEN = "<unk>";
-            BOS_TOKEN = "<s>";
-            EOS_TOKEN = "</s>";
-            PAD_TOKEN = "<pad>";
-
-            UNK_TOKEN_ID = 0;
-            BOS_TOKEN_ID = 1;
-            EOS_TOKEN_ID = 2;
-            PAD_TOKEN_ID = 11;
-
-            special_tokens = {
-                "<unk>",
-                "<s>",
-                "</s>",
-                "[INST]",
-                "[/INST]",
-                "[AVAILABLE_TOOLS]",
-                "[/AVAILABLE_TOOLS]",
-                "[TOOL_RESULTS]",
-                "[/TOOL_RESULTS]",
-                "[TOOL_CALLS]",
-                "[IMG]",
-                "<pad>",
-                "[IMG_BREAK]",
-                "[IMG_END]",
-                "[PREFIX]",
-                "[MIDDLE]",
-                "[SUFFIX]",
-                "[SYSTEM_PROMPT]",
-                "[/SYSTEM_PROMPT]",
-                "[TOOL_CONTENT]",
-            };
-            for (int i = 20; i < 1000; i++) {
-                special_tokens.push_back("<SPECIAL_" + std::to_string(i) + ">");
-            }
-
-            if (merges_utf8_str.size() > 0 && vocab_utf8_str.size() > 0) {
-                load_from_merges(merges_utf8_str, vocab_utf8_str);
-            } else {
-                load_from_merges(load_mistral_merges(), load_mistral_vocab_json());
-            }
-        }
-    };
-
     enum class LLMArch {
         QWEN2_5_VL,
         QWEN3,
         MISTRAL_SMALL_3_2,
+        MINISTRAL_3_3B,
         ARCH_COUNT,
     };
 
@@ -484,6 +36,7 @@ namespace LLM {
         "qwen2.5vl",
         "qwen3",
         "mistral_small3.2",
+        "ministral3.3b",
     };
 
     struct LLMVisionParams {
@@ -868,6 +421,9 @@ namespace LLM {
             if (arch == LLMArch::MISTRAL_SMALL_3_2) {
                 q = ggml_rope_ext(ctx->ggml_ctx, q, input_pos, nullptr, 128, GGML_ROPE_TYPE_NORMAL, 8192, 1000000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
                 k = ggml_rope_ext(ctx->ggml_ctx, k, input_pos, nullptr, 128, GGML_ROPE_TYPE_NORMAL, 8192, 1000000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
+            } else if (arch == LLMArch::MINISTRAL_3_3B) {
+                q = ggml_rope_ext(ctx->ggml_ctx, q, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 262144, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
+                k = ggml_rope_ext(ctx->ggml_ctx, k, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 262144, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
             } else if (arch == LLMArch::QWEN3) {
                 q = ggml_rope_ext(ctx->ggml_ctx, q, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 40960, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
                 k = ggml_rope_ext(ctx->ggml_ctx, k, input_pos, nullptr, 128, GGML_ROPE_TYPE_NEOX, 40960, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
@@ -1083,7 +639,7 @@ namespace LLM {
                   bool enable_vision_ = false)
             : GGMLRunner(backend, offload_params_to_cpu), enable_vision(enable_vision_) {
             params.arch = arch;
-            if (arch == LLMArch::MISTRAL_SMALL_3_2) {
+            if (arch == LLMArch::MISTRAL_SMALL_3_2 || arch == LLMArch::MINISTRAL_3_3B) {
                 params.head_dim     = 128;
                 params.num_heads    = 32;
                 params.num_kv_heads = 8;
@@ -1195,7 +751,7 @@ namespace LLM {
             }
 
             int64_t n_tokens = input_ids->ne[0];
-            if (params.arch == LLMArch::MISTRAL_SMALL_3_2 || params.arch == LLMArch::QWEN3) {
+            if (params.arch == LLMArch::MISTRAL_SMALL_3_2 || params.arch == LLMArch::MINISTRAL_3_3B || params.arch == LLMArch::QWEN3) {
                 input_pos_vec.resize(n_tokens);
                 for (int i = 0; i < n_tokens; ++i) {
                     input_pos_vec[i] = i;
@@ -1431,7 +987,7 @@ namespace LLM {
                     const std::string prefix                       = "",
                     bool enable_vision                             = false)
             : model(arch, backend, offload_params_to_cpu, tensor_storage_map, prefix, enable_vision) {
-            if (arch == LLMArch::MISTRAL_SMALL_3_2) {
+            if (arch == LLMArch::MISTRAL_SMALL_3_2 || arch == LLMArch::MINISTRAL_3_3B) {
                 tokenizer = std::make_shared<MistralTokenizer>();
             } else {
                 tokenizer = std::make_shared<Qwen2Tokenizer>();
@@ -1479,7 +1035,7 @@ namespace LLM {
                 weights.insert(weights.end(), curr_tokens.size(), curr_weight);
             }
 
-            tokenizer->pad_tokens(tokens, weights, max_length, padding);
+            tokenizer->pad_tokens(tokens, &weights, nullptr, padding ? max_length : 0, padding ? max_length : 100000000, padding);
 
             // for (int i = 0; i < tokens.size(); i++) {
             //     std::cout << tokens[i] << ":" << weights[i] << ", ";

src/model.cpp

@@ -1019,64 +1019,66 @@ SDVersion ModelLoader::get_sd_version() {
     bool has_middle_block_1          = false;
     bool has_output_block_311        = false;
     bool has_output_block_71         = false;
+    bool has_attn_1024               = false;
 
     for (auto& [name, tensor_storage] : tensor_storage_map) {
-        if (!(is_xl)) {
-            if (tensor_storage.name.find("model.diffusion_model.double_blocks.") != std::string::npos) {
-                is_flux = true;
+        if (tensor_storage.name.find("model.diffusion_model.double_blocks.") != std::string::npos) {
+            is_flux = true;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.nerf_final_layer_conv.") != std::string::npos) {
+            return VERSION_CHROMA_RADIANCE;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.joint_blocks.") != std::string::npos) {
+            return VERSION_SD3;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
+            return VERSION_QWEN_IMAGE;
+        }
+        if (tensor_storage.name.find("llm_adapter.blocks.0.cross_attn.q_proj.weight") != std::string::npos) {
+            return VERSION_ANIMA;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.double_stream_modulation_img.lin.weight") != std::string::npos) {
+            is_flux2 = true;
+        }
+        if (tensor_storage.name.find("single_blocks.47.linear1.weight") != std::string::npos) {
+            has_single_block_47 = true;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.double_blocks.0.img_mlp.gate_proj.weight") != std::string::npos) {
+            return VERSION_OVIS_IMAGE;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.cap_embedder.0.weight") != std::string::npos) {
+            return VERSION_Z_IMAGE;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.layers.0.adaLN_sa_ln.weight") != std::string::npos) {
+            return VERSION_ERNIE_IMAGE;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.blocks.0.cross_attn.norm_k.weight") != std::string::npos) {
+            is_wan = true;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.patch_embedding.weight") != std::string::npos) {
+            patch_embedding_channels = tensor_storage.ne[3];
+        }
+        if (tensor_storage.name.find("model.diffusion_model.img_emb") != std::string::npos) {
+            has_img_emb = true;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.input_blocks.") != std::string::npos ||
+            tensor_storage.name.find("unet.down_blocks.") != std::string::npos) {
+            is_unet = true;
+            if (has_multiple_encoders) {
+                is_xl = true;
             }
-            if (tensor_storage.name.find("model.diffusion_model.nerf_final_layer_conv.") != std::string::npos) {
-                return VERSION_CHROMA_RADIANCE;
-            }
-            if (tensor_storage.name.find("model.diffusion_model.joint_blocks.") != std::string::npos) {
-                return VERSION_SD3;
-            }
-            if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
-                return VERSION_QWEN_IMAGE;
-            }
-            if (tensor_storage.name.find("llm_adapter.blocks.0.cross_attn.q_proj.weight") != std::string::npos) {
-                return VERSION_ANIMA;
-            }
-            if (tensor_storage.name.find("model.diffusion_model.double_stream_modulation_img.lin.weight") != std::string::npos) {
-                is_flux2 = true;
-            }
-            if (tensor_storage.name.find("single_blocks.47.linear1.weight") != std::string::npos) {
-                has_single_block_47 = true;
-            }
-            if (tensor_storage.name.find("model.diffusion_model.double_blocks.0.img_mlp.gate_proj.weight") != std::string::npos) {
-                return VERSION_OVIS_IMAGE;
-            }
-            if (tensor_storage.name.find("model.diffusion_model.cap_embedder.0.weight") != std::string::npos) {
-                return VERSION_Z_IMAGE;
-            }
-            if (tensor_storage.name.find("model.diffusion_model.blocks.0.cross_attn.norm_k.weight") != std::string::npos) {
-                is_wan = true;
-            }
-            if (tensor_storage.name.find("model.diffusion_model.patch_embedding.weight") != std::string::npos) {
-                patch_embedding_channels = tensor_storage.ne[3];
-            }
-            if (tensor_storage.name.find("model.diffusion_model.img_emb") != std::string::npos) {
-                has_img_emb = true;
-            }
-            if (tensor_storage.name.find("model.diffusion_model.input_blocks.") != std::string::npos ||
-                tensor_storage.name.find("unet.down_blocks.") != std::string::npos) {
-                is_unet = true;
-                if (has_multiple_encoders) {
-                    is_xl = true;
-                }
-            }
-            if (tensor_storage.name.find("conditioner.embedders.1") != std::string::npos ||
-                tensor_storage.name.find("cond_stage_model.1") != std::string::npos ||
-                tensor_storage.name.find("te.1") != std::string::npos) {
-                has_multiple_encoders = true;
-                if (is_unet) {
-                    is_xl = true;
-                }
-            }
-            if (tensor_storage.name.find("model.diffusion_model.input_blocks.8.0.time_mixer.mix_factor") != std::string::npos) {
-                return VERSION_SVD;
+        }
+        if (tensor_storage.name.find("conditioner.embedders.1") != std::string::npos ||
+            tensor_storage.name.find("cond_stage_model.1") != std::string::npos ||
+            tensor_storage.name.find("te.1") != std::string::npos) {
+            has_multiple_encoders = true;
+            if (is_unet) {
+                is_xl = true;
             }
         }
+        if (tensor_storage.name.find("model.diffusion_model.input_blocks.8.0.time_mixer.mix_factor") != std::string::npos) {
+            return VERSION_SVD;
+        }
         if (tensor_storage.name.find("model.diffusion_model.middle_block.1.") != std::string::npos ||
             tensor_storage.name.find("unet.mid_block.resnets.1.") != std::string::npos) {
             has_middle_block_1 = true;
@@ -1088,6 +1090,10 @@ SDVersion ModelLoader::get_sd_version() {
         if (tensor_storage.name.find("model.diffusion_model.output_blocks.7.1") != std::string::npos ||
             tensor_storage.name.find("unet.up_blocks.2.attentions.1") != std::string::npos) {
             has_output_block_71 = true;
+            if (tensor_storage.name.find("model.diffusion_model.output_blocks.7.1.transformer_blocks.0.attn1.to_k.weight") != std::string::npos) {
+                if (tensor_storage.ne[0] == 1024)
+                    has_attn_1024 = true;
+            }
         }
         if (tensor_storage.name == "cond_stage_model.transformer.text_model.embeddings.token_embedding.weight" ||
             tensor_storage.name == "cond_stage_model.model.token_embedding.weight" ||
@@ -1161,7 +1167,7 @@ SDVersion ModelLoader::get_sd_version() {
         }
         if (!has_middle_block_1) {
             if (!has_output_block_71) {
-                return VERSION_SDXS;
+                return VERSION_SDXS_512_DS;
             }
             return VERSION_SD1_TINY_UNET;
         }
@@ -1171,7 +1177,7 @@ SDVersion ModelLoader::get_sd_version() {
             return VERSION_SD2_INPAINT;
         }
         if (!has_middle_block_1) {
-            return VERSION_SD2_TINY_UNET;
+            return has_attn_1024 ? VERSION_SDXS_09 : VERSION_SD2_TINY_UNET;
         }
         return VERSION_SD2;
     }

src/model.h

@@ -28,7 +28,8 @@ enum SDVersion {
     VERSION_SD2,
     VERSION_SD2_INPAINT,
     VERSION_SD2_TINY_UNET,
-    VERSION_SDXS,
+    VERSION_SDXS_512_DS,
+    VERSION_SDXS_09,
     VERSION_SDXL,
     VERSION_SDXL_INPAINT,
     VERSION_SDXL_PIX2PIX,
@@ -50,18 +51,19 @@ enum SDVersion {
     VERSION_FLUX2_KLEIN,
     VERSION_Z_IMAGE,
     VERSION_OVIS_IMAGE,
+    VERSION_ERNIE_IMAGE,
     VERSION_COUNT,
 };
 
 static inline bool sd_version_is_sd1(SDVersion version) {
-    if (version == VERSION_SD1 || version == VERSION_SD1_INPAINT || version == VERSION_SD1_PIX2PIX || version == VERSION_SD1_TINY_UNET || version == VERSION_SDXS) {
+    if (version == VERSION_SD1 || version == VERSION_SD1_INPAINT || version == VERSION_SD1_PIX2PIX || version == VERSION_SD1_TINY_UNET || version == VERSION_SDXS_512_DS) {
         return true;
     }
     return false;
 }
 
 static inline bool sd_version_is_sd2(SDVersion version) {
-    if (version == VERSION_SD2 || version == VERSION_SD2_INPAINT || version == VERSION_SD2_TINY_UNET) {
+    if (version == VERSION_SD2 || version == VERSION_SD2_INPAINT || version == VERSION_SD2_TINY_UNET || version == VERSION_SDXS_09) {
         return true;
     }
     return false;
@@ -137,6 +139,20 @@ static inline bool sd_version_is_z_image(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_ernie_image(SDVersion version) {
+    if (version == VERSION_ERNIE_IMAGE) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_uses_flux2_vae(SDVersion version) {
+    if (sd_version_is_flux2(version) || sd_version_is_ernie_image(version)) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_inpaint(SDVersion version) {
     if (version == VERSION_SD1_INPAINT ||
         version == VERSION_SD2_INPAINT ||
@@ -155,7 +171,8 @@ static inline bool sd_version_is_dit(SDVersion version) {
         sd_version_is_wan(version) ||
         sd_version_is_qwen_image(version) ||
         sd_version_is_anima(version) ||
-        sd_version_is_z_image(version)) {
+        sd_version_is_z_image(version) ||
+        sd_version_is_ernie_image(version)) {
         return true;
     }
     return false;

src/name_conversion.cpp

@@ -1120,7 +1120,7 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
         for (const auto& prefix : first_stage_model_prefix_vec) {
             if (starts_with(name, prefix)) {
                 name = convert_first_stage_model_name(name.substr(prefix.size()), prefix);
-                if (version == VERSION_SDXS) {
+                if (version == VERSION_SDXS_512_DS || version == VERSION_SDXS_09) {
                     name = "tae." + name;
                 } else {
                     name = prefix + name;

src/rope.hpp

@@ -7,6 +7,11 @@
 #include "ggml_extend.hpp"
 
 namespace Rope {
+    enum class EmbedNDLayout {
+        Matrix,
+        ErnieImage,
+    };
+
     template <class T>
     __STATIC_INLINE__ std::vector<T> linspace(T start, T end, int num) {
         std::vector<T> result(num);
@@ -169,7 +174,8 @@ namespace Rope {
                                                   int bs,
                                                   const std::vector<float>& axis_thetas,
                                                   const std::vector<int>& axes_dim,
-                                                  const std::vector<std::vector<int>>& wrap_dims = {}) {
+                                                  const std::vector<std::vector<int>>& wrap_dims = {},
+                                                  EmbedNDLayout layout                           = EmbedNDLayout::Matrix) {
         std::vector<std::vector<float>> trans_ids = transpose(ids);
         size_t pos_len                            = ids.size() / bs;
         size_t num_axes                           = axes_dim.size();
@@ -204,6 +210,24 @@ namespace Rope {
             offset += rope_emb[0].size();
         }
 
+        if (layout == EmbedNDLayout::ErnieImage) {
+            int head_dim = emb_dim * 2;
+            std::vector<float> ernie_emb(bs * pos_len * head_dim * 2, 0.0f);
+            for (size_t pos_idx = 0; pos_idx < bs * pos_len; ++pos_idx) {
+                for (int i = 0; i < emb_dim; ++i) {
+                    float cos_val             = emb[pos_idx][4 * i];
+                    float sin_val             = emb[pos_idx][4 * i + 2];
+                    size_t cos_offset         = pos_idx * head_dim + 2 * i;
+                    size_t sin_offset         = bs * pos_len * head_dim + cos_offset;
+                    ernie_emb[cos_offset]     = cos_val;
+                    ernie_emb[cos_offset + 1] = cos_val;
+                    ernie_emb[sin_offset]     = sin_val;
+                    ernie_emb[sin_offset + 1] = sin_val;
+                }
+            }
+            return ernie_emb;
+        }
+
         return flatten(emb);
     }
 
@@ -211,9 +235,10 @@ namespace Rope {
                                                   int bs,
                                                   float theta,
                                                   const std::vector<int>& axes_dim,
-                                                  const std::vector<std::vector<int>>& wrap_dims = {}) {
+                                                  const std::vector<std::vector<int>>& wrap_dims = {},
+                                                  EmbedNDLayout layout                           = EmbedNDLayout::Matrix) {
         std::vector<float> axis_thetas(axes_dim.size(), theta);
-        return embed_nd(ids, bs, axis_thetas, axes_dim, wrap_dims);
+        return embed_nd(ids, bs, axis_thetas, axes_dim, wrap_dims, layout);
     }
 
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_refs_ids(int patch_size,
@@ -437,6 +462,74 @@ namespace Rope {
         return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_ernie_image_ids(int h,
+                                                                          int w,
+                                                                          int patch_size,
+                                                                          int bs,
+                                                                          int context_len) {
+        int h_len = h / patch_size;
+        int w_len = w / patch_size;
+
+        std::vector<std::vector<float>> img_ids(h_len * w_len, std::vector<float>(3, 0.0f));
+        std::vector<float> h_ids = linspace<float>(0.f, static_cast<float>(h_len - 1), h_len);
+        std::vector<float> w_ids = linspace<float>(0.f, static_cast<float>(w_len - 1), w_len);
+        for (int i = 0; i < h_len; ++i) {
+            for (int j = 0; j < w_len; ++j) {
+                img_ids[i * w_len + j][0] = static_cast<float>(context_len);
+                img_ids[i * w_len + j][1] = h_ids[i];
+                img_ids[i * w_len + j][2] = w_ids[j];
+            }
+        }
+
+        std::vector<std::vector<float>> img_ids_repeated(bs * img_ids.size(), std::vector<float>(3, 0.0f));
+        for (int i = 0; i < bs; ++i) {
+            for (int j = 0; j < static_cast<int>(img_ids.size()); ++j) {
+                img_ids_repeated[i * img_ids.size() + j] = img_ids[j];
+            }
+        }
+
+        std::vector<std::vector<float>> txt_ids(bs * context_len, std::vector<float>(3, 0.0f));
+        for (int i = 0; i < bs; ++i) {
+            for (int j = 0; j < context_len; ++j) {
+                txt_ids[i * context_len + j][0] = static_cast<float>(j);
+            }
+        }
+
+        return concat_ids(img_ids_repeated, txt_ids, bs);
+    }
+
+    __STATIC_INLINE__ std::vector<float> gen_ernie_image_pe(int h,
+                                                            int w,
+                                                            int patch_size,
+                                                            int bs,
+                                                            int context_len,
+                                                            int theta,
+                                                            bool circular_h,
+                                                            bool circular_w,
+                                                            const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids = gen_ernie_image_ids(h, w, patch_size, bs, context_len);
+        std::vector<std::vector<int>> wrap_dims;
+        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            int h_len = h / patch_size;
+            int w_len = w / patch_size;
+            if (h_len > 0 && w_len > 0) {
+                size_t pos_len = ids.size() / bs;
+                wrap_dims.assign(axes_dim.size(), std::vector<int>(pos_len, 0));
+                const size_t img_tokens = static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
+                for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
+                    if (circular_h) {
+                        wrap_dims[1][token_i] = h_len;
+                    }
+                    if (circular_w) {
+                        wrap_dims[2][token_i] = w_len;
+                    }
+                }
+            }
+        }
+
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims, EmbedNDLayout::ErnieImage);
+    }
+
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_vid_ids(int t,
                                                                   int h,
                                                                   int w,

src/stable-diffusion.cpp

@@ -30,7 +30,8 @@ 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",
@@ -52,6 +53,7 @@ const char* model_version_to_str[] = {
     "Flux.2 klein",
     "Z-Image",
     "Ovis Image",
+    "Ernie Image",
 };
 
 const char* sampling_methods_str[] = {
@@ -69,6 +71,7 @@ const char* sampling_methods_str[] = {
     "TCD",
     "Res Multistep",
     "Res 2s",
+    "ER-SDE",
 };
 
 /*================================================== Helper Functions ================================================*/
@@ -413,7 +416,7 @@ 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;
         }
@@ -551,6 +554,15 @@ 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++) {
@@ -819,6 +831,10 @@ 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");
@@ -922,10 +938,13 @@ 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;
                     }
@@ -1395,7 +1414,7 @@ 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;
@@ -1844,7 +1863,7 @@ 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;
@@ -1975,6 +1994,7 @@ 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) {
@@ -2370,6 +2390,14 @@ 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) {
@@ -2399,6 +2427,20 @@ 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)) {
@@ -2457,6 +2499,7 @@ 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:;
         }

src/t5.hpp

@@ -10,452 +10,9 @@
 #include <string>
 #include <unordered_map>
 
-#include "darts.h"
 #include "ggml_extend.hpp"
-#include "json.hpp"
 #include "model.h"
-#include "vocab/vocab.h"
-
-// Port from: https://github.com/google/sentencepiece/blob/master/src/unigram_model.h
-// and https://github.com/google/sentencepiece/blob/master/src/unigram_model.h.
-// Original License: https://github.com/google/sentencepiece/blob/master/LICENSE
-//
-// Since tokenization is not the bottleneck in SD, performance was not a major consideration
-// during the migration.
-class MetaspacePreTokenizer {
-private:
-    std::string replacement;
-    bool add_prefix_space;
-
-public:
-    MetaspacePreTokenizer(const std::string replacement = " ", bool add_prefix_space = true)
-        : replacement(replacement), add_prefix_space(add_prefix_space) {}
-
-    std::string tokenize(const std::string& input) const {
-        std::string tokens;
-        std::stringstream ss(input);
-
-        if (add_prefix_space) {
-            tokens += replacement;
-        }
-
-        std::string token;
-        bool firstToken = true;
-        while (std::getline(ss, token, ' ')) {
-            if (!firstToken)
-                tokens += replacement + token;
-            else
-                tokens += token;
-
-            firstToken = false;
-        }
-
-        return tokens;
-    }
-};
-
-using EncodeResult = std::vector<std::pair<std::string, int>>;
-class T5UniGramTokenizer {
-public:
-    enum Status {
-        OK,
-        NO_PIECES_LOADED,
-        NO_ENTRY_FOUND,
-        BUILD_DOUBLE_ARRAY_FAILED,
-        PIECE_ALREADY_DEFINED,
-        INVLIAD_JSON
-    };
-
-protected:
-    MetaspacePreTokenizer pre_tokenizer;
-
-    // all <piece, score> pairs
-    std::vector<std::pair<std::string, float>> piece_score_pairs;
-
-    float min_score_ = 0.0;
-    float max_score_ = 0.0;
-    std::unique_ptr<Darts::DoubleArray> trie_;
-
-    // Maximum size of the return value of Trie, which corresponds
-    // to the maximum size of shared common prefix in the sentence pieces.
-    int trie_results_size_;
-    // unknown id.
-    int unk_id_            = 2;
-    std::string eos_token_ = "</s>";
-    int eos_id_            = 1;
-    int pad_id_            = 0;
-    // status.
-    Status status_ = OK;
-
-    float kUnkPenalty = 10.0;
-
-    std::string replacement;
-    bool add_prefix_space = true;
-
-    void InitializePieces(const std::string& json_str) {
-        nlohmann::json data;
-
-        try {
-            data = nlohmann::json::parse(json_str);
-        } catch (const nlohmann::json::parse_error&) {
-            status_ = INVLIAD_JSON;
-            return;
-        }
-        if (!data.contains("model")) {
-            status_ = INVLIAD_JSON;
-            return;
-        }
-        nlohmann::json model = data["model"];
-        if (!model.contains("vocab")) {
-            status_ = INVLIAD_JSON;
-            return;
-        }
-        if (model.contains("unk_id")) {
-            unk_id_ = model["unk_id"];
-        }
-
-        replacement      = data["pre_tokenizer"]["replacement"];
-        add_prefix_space = data["pre_tokenizer"]["add_prefix_space"];
-
-        pre_tokenizer = MetaspacePreTokenizer(replacement, add_prefix_space);
-
-        for (const auto& item : model["vocab"]) {
-            if (item.size() != 2 || !item[0].is_string() || !item[1].is_number_float()) {
-                status_ = INVLIAD_JSON;
-                return;
-            }
-            std::string piece = item[0];
-            if (piece.empty()) {
-                piece = "<empty_token>";
-            }
-            float score = item[1];
-            piece_score_pairs.emplace_back(piece, score);
-        }
-    }
-
-    // Builds a Trie index.
-    void BuildTrie(std::vector<std::pair<std::string, int>>* pieces) {
-        if (status_ != OK)
-            return;
-
-        if (pieces->empty()) {
-            status_ = NO_PIECES_LOADED;
-            return;
-        }
-
-        // sort by sentencepiece since DoubleArray::build()
-        // only accepts sorted strings.
-        sort(pieces->begin(), pieces->end());
-
-        // Makes key/value set for DoubleArrayTrie.
-        std::vector<const char*> key(pieces->size());
-        std::vector<int> value(pieces->size());
-        for (size_t i = 0; i < pieces->size(); ++i) {
-            // LOG_DEBUG("%s %d", (*pieces)[i].first.c_str(), (*pieces)[i].second);
-            key[i]   = (*pieces)[i].first.data();  // sorted piece.
-            value[i] = (*pieces)[i].second;        // vocab_id
-        }
-
-        trie_ = std::unique_ptr<Darts::DoubleArray>(new Darts::DoubleArray());
-        if (trie_->build(key.size(), const_cast<char**>(&key[0]), nullptr,
-                         &value[0]) != 0) {
-            status_ = BUILD_DOUBLE_ARRAY_FAILED;
-            return;
-        }
-
-        // Computes the maximum number of shared prefixes in the trie.
-        const int kMaxTrieResultsSize = 1024;
-        std::vector<Darts::DoubleArray::result_pair_type> results(
-            kMaxTrieResultsSize);
-        trie_results_size_ = 0;
-        for (const auto& p : *pieces) {
-            const size_t num_nodes = trie_->commonPrefixSearch(
-                p.first.data(), results.data(), results.size(), p.first.size());
-            trie_results_size_ = std::max(trie_results_size_, static_cast<int>(num_nodes));
-        }
-
-        if (trie_results_size_ == 0)
-            status_ = NO_ENTRY_FOUND;
-    }
-
-    // Non-virtual (inlined) implementation for faster execution.
-    inline float GetScoreInlined(int id) const {
-        return piece_score_pairs[id].second;
-    }
-
-    inline bool IsUnusedInlined(int id) const {
-        return false;  // TODO
-    }
-
-    inline bool IsUserDefinedInlined(int id) const {
-        return false;  // TODO
-    }
-
-    inline size_t OneCharLen(const char* src) const {
-        return "\1\1\1\1\1\1\1\1\1\1\1\1\2\2\3\4"[(*src & 0xFF) >> 4];
-    }
-
-    // The optimized Viterbi encode.
-    // Main differences from the original function:
-    // 1. Memorizes the best path at each postion so far,
-    // 2. No need to store the Lattice nodes,
-    // 3. Works in utf-8 directly,
-    // 4. Defines a new struct with fewer fields than Lattice,
-    // 5. Does not depend on `class Lattice` nor call `SetSentence()`,
-    // `PopulateNodes()`, or `Viterbi()`. It does everything in one function.
-    // For detailed explanations please see the comments inside the function body.
-    EncodeResult EncodeOptimized(const std::string& normalized) const {
-        // An optimized Viterbi algorithm for unigram language models. Benchmarking
-        // results show that it generates almost identical outputs and achieves 2.1x
-        // speedup on average for 102 languages compared to the original
-        // implementation. It's based on the following three ideas:
-        //
-        // 1. Because it uses the *unigram* model:
-        //     best_score(x1, x2, ... xt) = best_score(x1, x2, ... x{t-1}) + score(xt)
-        // Deciding the best path (and score) can be decoupled into two isolated
-        // terms: (a) the best path ended before the last token `best_score(x1, x2, ...)`
-        // x{t-1})`, and (b) the last token and its `score(xt)`. The two terms are
-        // not related to each other at all.
-        //
-        // Therefore, we can compute once and store the *best_path ending at
-        // each character position*. In this way, when we know best_path_ends_at[M],
-        // we can reuse it to compute all the best_path_ends_at_[...] where the last
-        // token starts at the same character position M.
-        //
-        // This improves the time complexity from O(n*k*k) to O(n*k) because it
-        // eliminates the extra loop of recomputing the best path ending at the same
-        // position, where n is the input length and k is the maximum number of tokens
-        // that can be recognized starting at each position.
-        //
-        // 2. Again, because it uses the *unigram* model, we don't need to actually
-        // store the lattice nodes. We still recognize all the tokens and lattice
-        // nodes from the input, but along identifying them, we use and discard them
-        // on the fly. There is no need to actually store them for best path Viterbi
-        // decoding. The only thing we need to store is the best_path ending at
-        // each character position.
-        //
-        // This improvement reduces the things needed to store in memory from O(n*k)
-        // to O(n), where n is the input length and k is the maximum number of tokens
-        // that can be recognized starting at each position.
-        //
-        // It also avoids the need of dynamic-size lattice node pool, because the
-        // number of things to store is fixed as n.
-        //
-        // 3. SentencePiece is designed to work with unicode, taking utf-8 encoding
-        // inputs. In the original implementation, the lattice positions are based on
-        // unicode positions. A mapping from unicode position to the utf-8 position is
-        // maintained to recover the utf-8 string piece.
-        //
-        // We found that it is sufficient and beneficial to directly work with utf-8
-        // positions:
-        //
-        // Firstly, it saves the conversion and mapping between unicode positions and
-        // utf-8 positions.
-        //
-        // Secondly, it reduces the number of fields we need to maintain in the
-        // node/path structure. Specifically, there are 8 fields defined in
-        // `Lattice::Node` used by the original encoder, but here in the optimized
-        // encoder we only need to define 3 fields in `BestPathNode`.
-
-        if (status() != OK || normalized.empty()) {
-            return {};
-        }
-        // Represents the last node of the best path.
-        struct BestPathNode {
-            int id = -1;  // The vocab id. (maybe -1 for UNK)
-            float best_path_score =
-                0;  // The total score of the best path ending at this node.
-            int starts_at =
-                -1;  // The starting position (in utf-8) of this node. The entire best
-                     // path can be constructed by backtracking along this link.
-        };
-        const int size        = static_cast<int>(normalized.size());
-        const float unk_score = min_score() - kUnkPenalty;
-        // The ends are exclusive.
-        std::vector<BestPathNode> best_path_ends_at(size + 1);
-        // Generate lattice on-the-fly (not stored) and update best_path_ends_at.
-        int starts_at = 0;
-        while (starts_at < size) {
-            std::size_t node_pos = 0;
-            std::size_t key_pos  = starts_at;
-            const auto best_path_score_till_here =
-                best_path_ends_at[starts_at].best_path_score;
-            bool has_single_node = false;
-            const int mblen =
-                std::min<int>(static_cast<int>(OneCharLen(normalized.data() + starts_at)),
-                              size - starts_at);
-            while (key_pos < size) {
-                const int ret =
-                    trie_->traverse(normalized.data(), node_pos, key_pos, key_pos + 1);
-                if (ret == -2)
-                    break;
-                if (ret >= 0) {
-                    if (IsUnusedInlined(ret))
-                        continue;
-                    // Update the best path node.
-                    auto& target_node = best_path_ends_at[key_pos];
-                    const auto length = (key_pos - starts_at);
-                    // User defined symbol receives extra bonus to always be selected.
-                    const auto score = IsUserDefinedInlined(ret)
-                                           ? (length * max_score_ - 0.1)
-                                           : GetScoreInlined(ret);
-                    const auto candidate_best_path_score =
-                        score + best_path_score_till_here;
-                    if (target_node.starts_at == -1 ||
-                        candidate_best_path_score > target_node.best_path_score) {
-                        target_node.best_path_score = static_cast<float>(candidate_best_path_score);
-                        target_node.starts_at       = starts_at;
-                        target_node.id              = ret;
-                    }
-                    if (!has_single_node && length == mblen) {
-                        has_single_node = true;
-                    }
-                }
-            }
-            if (!has_single_node) {
-                auto& target_node = best_path_ends_at[starts_at + mblen];
-                const auto candidate_best_path_score =
-                    unk_score + best_path_score_till_here;
-                if (target_node.starts_at == -1 ||
-                    candidate_best_path_score > target_node.best_path_score) {
-                    target_node.best_path_score = candidate_best_path_score;
-                    target_node.starts_at       = starts_at;
-                    target_node.id              = unk_id_;
-                }
-            }
-            // Move by one unicode character.
-            starts_at += mblen;
-        }
-        // Backtrack to identify the best path.
-        EncodeResult results;
-        int ends_at = size;
-        while (ends_at > 0) {
-            const auto& node = best_path_ends_at[ends_at];
-            results.emplace_back(
-                normalized.substr(node.starts_at, ends_at - node.starts_at), node.id);
-            ends_at = node.starts_at;
-        }
-        std::reverse(results.begin(), results.end());
-        return results;
-    }
-
-public:
-    explicit T5UniGramTokenizer(bool is_umt5 = false) {
-        if (is_umt5) {
-            InitializePieces(load_umt5_tokenizer_json());
-        } else {
-            InitializePieces(load_t5_tokenizer_json());
-        }
-
-        min_score_ = FLT_MAX;
-        max_score_ = FLT_MIN;
-
-        std::vector<std::pair<std::string, int>> pieces;
-        for (int i = 0; i < piece_score_pairs.size(); i++) {
-            const auto& sp = piece_score_pairs[i];
-
-            min_score_ = std::min(min_score_, sp.second);
-            max_score_ = std::max(max_score_, sp.second);
-
-            pieces.emplace_back(sp.first, i);
-        }
-
-        BuildTrie(&pieces);
-    }
-    ~T5UniGramTokenizer(){};
-
-    std::string Normalize(const std::string& input) const {
-        // Ref: https://github.com/huggingface/tokenizers/blob/1ff56c0c70b045f0cd82da1af9ac08cd4c7a6f9f/bindings/python/py_src/tokenizers/implementations/sentencepiece_unigram.py#L29
-        // TODO: nmt-nfkc
-        std::string normalized = std::regex_replace(input, std::regex(" {2,}"), " ");
-        return normalized;
-    }
-
-    std::vector<int> Encode(const std::string& input, bool append_eos_if_not_present = true) const {
-        std::string normalized = Normalize(input);
-        normalized             = pre_tokenizer.tokenize(normalized);
-        EncodeResult result    = EncodeOptimized(normalized);
-        if (result.size() > 0 && append_eos_if_not_present) {
-            auto item = result[result.size() - 1];
-            if (item.first != eos_token_) {
-                result.emplace_back(eos_token_, eos_id_);
-            }
-        }
-        std::vector<int> tokens;
-        for (auto item : result) {
-            tokens.push_back(item.second);
-        }
-        return tokens;
-    }
-
-    void pad_tokens(std::vector<int>& tokens,
-                    std::vector<float>& weights,
-                    std::vector<float>* attention_mask,
-                    size_t max_length = 0,
-                    bool padding      = false) {
-        if (max_length > 0 && padding) {
-            size_t orig_token_num = tokens.size() - 1;
-            size_t n              = static_cast<size_t>(std::ceil(orig_token_num * 1.0 / (max_length - 1)));
-            if (n == 0) {
-                n = 1;
-            }
-            size_t length = max_length * n;
-            LOG_DEBUG("token length: %llu", length);
-            std::vector<int> new_tokens;
-            std::vector<float> new_weights;
-            std::vector<float> new_attention_mask;
-            int token_idx = 0;
-            for (int i = 0; i < length; i++) {
-                if (token_idx >= orig_token_num) {
-                    break;
-                }
-                if (attention_mask != nullptr) {
-                    new_attention_mask.push_back(0.0);
-                }
-                if (i % max_length == max_length - 1) {
-                    new_tokens.push_back(eos_id_);
-                    new_weights.push_back(1.0);
-                } else {
-                    new_tokens.push_back(tokens[token_idx]);
-                    new_weights.push_back(weights[token_idx]);
-                    token_idx++;
-                }
-            }
-
-            new_tokens.push_back(eos_id_);
-            new_weights.push_back(1.0);
-            if (attention_mask != nullptr) {
-                new_attention_mask.push_back(0.0);
-            }
-
-            tokens  = new_tokens;
-            weights = new_weights;
-            if (attention_mask != nullptr) {
-                *attention_mask = new_attention_mask;
-            }
-
-            if (padding) {
-                int pad_token_id = pad_id_;
-                tokens.insert(tokens.end(), length - tokens.size(), pad_token_id);
-                weights.insert(weights.end(), length - weights.size(), 1.0);
-                if (attention_mask != nullptr) {
-                    // maybe keep some padding tokens unmasked?
-                    attention_mask->insert(attention_mask->end(), length - attention_mask->size(), -HUGE_VALF);
-                }
-            }
-        }
-    }
-
-    // Returns the minimum score in sentence pieces.
-    // min_score() - 10 is used for the cost of unknown sentence.
-    float min_score() const { return min_score_; }
-
-    // Returns the maximum score in sentence pieces.
-    // max_score() is used for the cost of user defined symbols.
-    float max_score() const { return max_score_; }
-
-    Status status() const { return status_; }
-};
+#include "tokenizers/t5_unigram_tokenizer.h"
 
 class T5LayerNorm : public UnaryBlock {
 protected:
@@ -937,18 +494,17 @@ struct T5Embedder {
         for (const auto& item : parsed_attention) {
             const std::string& curr_text = item.first;
             float curr_weight            = item.second;
-            std::vector<int> curr_tokens = tokenizer.Encode(curr_text, false);
+            std::vector<int> curr_tokens = tokenizer.encode(curr_text);
             tokens.insert(tokens.end(), curr_tokens.begin(), curr_tokens.end());
             weights.insert(weights.end(), curr_tokens.size(), curr_weight);
         }
 
-        int EOS_TOKEN_ID = 1;
-        tokens.push_back(EOS_TOKEN_ID);
-        weights.push_back(1.0);
-
         std::vector<float> attention_mask;
 
-        tokenizer.pad_tokens(tokens, weights, &attention_mask, max_length, padding);
+        tokenizer.pad_tokens(tokens, &weights, &attention_mask, padding ? max_length : 0, padding ? max_length : 100000000, padding);
+        for (auto& mask_value : attention_mask) {
+            mask_value = mask_value > 0.0f ? 0.0f : -HUGE_VALF;
+        }
 
         // for (int i = 0; i < tokens.size(); i++) {
         //     std::cout << tokens[i] << ":" << weights[i] << ", ";

src/tokenizers/bpe_tokenizer.cpp

@@ -0,0 +1,189 @@
+#include "bpe_tokenizer.h"
+
+#include <algorithm>
+#include <sstream>
+
+#include "tokenize_util.h"
+#include "util.h"
+
+std::vector<std::pair<int, std::u32string>> BPETokenizer::bytes_to_unicode() {
+    std::vector<std::pair<int, std::u32string>> byte_unicode_pairs;
+    std::set<int> byte_set;
+    for (int b = static_cast<int>('!'); b <= static_cast<int>('~'); ++b) {
+        byte_set.insert(b);
+        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
+    }
+    for (int b = 161; b <= 172; ++b) {
+        byte_set.insert(b);
+        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
+    }
+    for (int b = 174; b <= 255; ++b) {
+        byte_set.insert(b);
+        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
+    }
+    int n = 0;
+    for (int b = 0; b < 256; ++b) {
+        if (byte_set.find(b) == byte_set.end()) {
+            byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(n + 256)));
+            ++n;
+        }
+    }
+    return byte_unicode_pairs;
+}
+
+std::vector<std::string> BPETokenizer::token_split(const std::string& text) const {
+    return ::token_split(text);
+}
+
+std::vector<std::u32string> BPETokenizer::split_utf32(const std::string& text, char32_t delimiter) {
+    std::vector<std::u32string> result;
+    size_t start              = 0;
+    size_t pos                = 0;
+    std::u32string utf32_text = utf8_to_utf32(text);
+    while ((pos = utf32_text.find(delimiter, start)) != std::u32string::npos) {
+        result.push_back(utf32_text.substr(start, pos - start));
+        start = pos + 1;
+    }
+    return result;
+}
+
+static std::set<std::pair<std::u32string, std::u32string>> get_pairs(const std::vector<std::u32string>& subwords) {
+    std::set<std::pair<std::u32string, std::u32string>> pairs;
+    if (subwords.empty()) {
+        return pairs;
+    }
+
+    std::u32string prev_subword = subwords[0];
+    for (int i = 1; i < static_cast<int>(subwords.size()); i++) {
+        std::u32string subword = subwords[i];
+        std::pair<std::u32string, std::u32string> pair(prev_subword, subword);
+        pairs.insert(pair);
+        prev_subword = subword;
+    }
+    return pairs;
+}
+
+std::vector<std::u32string> BPETokenizer::bpe(const std::u32string& token) const {
+    std::vector<std::u32string> word;
+
+    for (int i = 0; i < static_cast<int>(token.size()) - 1; i++) {
+        word.emplace_back(1, token[i]);
+    }
+    word.push_back(token.substr(token.size() - 1) + utf8_to_utf32(end_of_word_suffix));
+
+    std::set<std::pair<std::u32string, std::u32string>> pairs = get_pairs(word);
+
+    if (pairs.empty()) {
+        return {token + utf8_to_utf32(end_of_word_suffix)};
+    }
+
+    while (true) {
+        auto min_pair_iter = std::min_element(pairs.begin(),
+                                              pairs.end(),
+                                              [&](const std::pair<std::u32string, std::u32string>& a,
+                                                  const std::pair<std::u32string, std::u32string>& b) {
+                                                  if (bpe_ranks.find(a) == bpe_ranks.end()) {
+                                                      return false;
+                                                  } else if (bpe_ranks.find(b) == bpe_ranks.end()) {
+                                                      return true;
+                                                  }
+                                                  return bpe_ranks.at(a) < bpe_ranks.at(b);
+                                              });
+
+        const std::pair<std::u32string, std::u32string>& bigram = *min_pair_iter;
+
+        if (bpe_ranks.find(bigram) == bpe_ranks.end()) {
+            break;
+        }
+
+        std::u32string first  = bigram.first;
+        std::u32string second = bigram.second;
+        std::vector<std::u32string> new_word;
+        int32_t i = 0;
+
+        while (i < static_cast<int32_t>(word.size())) {
+            auto it = std::find(word.begin() + i, word.end(), first);
+            if (it == word.end()) {
+                new_word.insert(new_word.end(), word.begin() + i, word.end());
+                break;
+            }
+            new_word.insert(new_word.end(), word.begin() + i, it);
+            i = static_cast<int32_t>(std::distance(word.begin(), it));
+
+            if (word[i] == first && i < static_cast<int32_t>(word.size()) - 1 && word[i + 1] == second) {
+                new_word.push_back(first + second);
+                i += 2;
+            } else {
+                new_word.push_back(word[i]);
+                i += 1;
+            }
+        }
+
+        word = new_word;
+
+        if (word.size() == 1) {
+            break;
+        }
+        pairs = get_pairs(word);
+    }
+
+    return word;
+}
+
+std::vector<int> BPETokenizer::encode(const std::string& text, on_new_token_cb_t on_new_token_cb) {
+    std::string normalized_text = normalize(text);
+    std::vector<int32_t> bpe_tokens;
+    std::vector<std::string> token_strs;
+
+    auto splited_texts = split_with_special_tokens(normalized_text, special_tokens);
+
+    for (auto& splited_text : splited_texts) {
+        if (is_special_token(splited_text)) {
+            if (on_new_token_cb != nullptr) {
+                bool skip = on_new_token_cb(splited_text, bpe_tokens);
+                if (skip) {
+                    token_strs.push_back(splited_text);
+                    continue;
+                }
+            }
+            bpe_tokens.push_back(encoder[utf8_to_utf32(splited_text)]);
+            token_strs.push_back(splited_text);
+            continue;
+        }
+        auto tokens = token_split(splited_text);
+        for (auto& token : tokens) {
+            if (on_new_token_cb != nullptr) {
+                bool skip = on_new_token_cb(token, bpe_tokens);
+                if (skip) {
+                    token_strs.push_back(splited_text);
+                    continue;
+                }
+            }
+
+            std::string token_str = token;
+            std::u32string utf32_token;
+            for (int i = 0; i < static_cast<int>(token_str.length()); i++) {
+                unsigned char b = token_str[i];
+                utf32_token += byte_encoder[b];
+            }
+            auto bpe_strs = bpe(utf32_token);
+            for (auto bpe_str : bpe_strs) {
+                bpe_tokens.push_back(encoder[bpe_str]);
+                token_strs.push_back(utf32_to_utf8(bpe_str));
+            }
+        }
+    }
+
+    std::stringstream ss;
+    ss << "[";
+    for (auto token : token_strs) {
+        ss << "\"" << token << "\", ";
+    }
+    ss << "]";
+    LOG_DEBUG("split prompt \"%s\" to tokens %s", text.c_str(), ss.str().c_str());
+    return bpe_tokens;
+}
+
+std::string BPETokenizer::decode_token(int token_id) const {
+    return utf32_to_utf8(decoder.at(token_id));
+}

src/tokenizers/bpe_tokenizer.h

@@ -0,0 +1,40 @@
+#ifndef __SD_TOKENIZERS_BPE_TOKENIZER_H__
+#define __SD_TOKENIZERS_BPE_TOKENIZER_H__
+
+#include <cstddef>
+#include <cstdint>
+#include <functional>
+#include <map>
+#include <regex>
+#include <set>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "tokenizer.h"
+
+class BPETokenizer : public Tokenizer {
+protected:
+    std::map<int, std::u32string> byte_encoder;
+    std::map<std::u32string, int> byte_decoder;
+    std::map<std::u32string, int> encoder;
+    std::map<int, std::u32string> decoder;
+    std::map<std::pair<std::u32string, std::u32string>, int> bpe_ranks;
+    int encoder_len = 0;
+    int bpe_len     = 0;
+
+protected:
+    static std::vector<std::pair<int, std::u32string>> bytes_to_unicode();
+    static std::vector<std::u32string> split_utf32(const std::string& text, char32_t delimiter = U'\n');
+    virtual std::vector<std::string> token_split(const std::string& text) const;
+    std::vector<std::u32string> bpe(const std::u32string& token) const;
+    std::string decode_token(int token_id) const override;
+
+public:
+    BPETokenizer()          = default;
+    virtual ~BPETokenizer() = default;
+
+    std::vector<int> encode(const std::string& text, on_new_token_cb_t on_new_token_cb = nullptr) override;
+};
+
+#endif  // __SD_TOKENIZERS_BPE_TOKENIZER_H__

src/tokenizers/clip_tokenizer.cpp

@@ -0,0 +1,116 @@
+#include "clip_tokenizer.h"
+
+#include <algorithm>
+#include <cctype>
+#include <cmath>
+#include <regex>
+#include <set>
+
+#include "ggml.h"
+#include "tokenize_util.h"
+#include "util.h"
+#include "vocab/vocab.h"
+
+CLIPTokenizer::CLIPTokenizer(int pad_token_id, const std::string& merges_utf8_str) {
+    UNK_TOKEN = "<|endoftext|>";
+    BOS_TOKEN = "<|startoftext|>";
+    EOS_TOKEN = "<|endoftext|>";
+    PAD_TOKEN = "<|endoftext|>";
+
+    UNK_TOKEN_ID = 49407;
+    BOS_TOKEN_ID = 49406;
+    EOS_TOKEN_ID = 49407;
+    PAD_TOKEN_ID = pad_token_id;
+
+    end_of_word_suffix = "</w>";
+    add_bos_token      = true;
+    add_eos_token      = true;
+
+    if (merges_utf8_str.size() > 0) {
+        load_from_merges(merges_utf8_str);
+    } else {
+        load_from_merges(load_clip_merges());
+    }
+    add_special_token("<|startoftext|>");
+    add_special_token("<|endoftext|>");
+}
+
+void CLIPTokenizer::load_from_merges(const std::string& merges_utf8_str) {
+    auto byte_unicode_pairs = bytes_to_unicode();
+    byte_encoder            = std::map<int, std::u32string>(byte_unicode_pairs.begin(), byte_unicode_pairs.end());
+    for (auto& pair : byte_unicode_pairs) {
+        byte_decoder[pair.second] = pair.first;
+    }
+
+    std::vector<std::u32string> merges = split_utf32(merges_utf8_str);
+    GGML_ASSERT(merges.size() == 48895);
+    merges = std::vector<std::u32string>(merges.begin() + 1, merges.end());
+    std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
+    for (const auto& merge : merges) {
+        size_t space_pos = merge.find(' ');
+        merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
+    }
+    std::vector<std::u32string> vocab;
+    for (const auto& pair : byte_unicode_pairs) {
+        vocab.push_back(pair.second);
+    }
+    for (const auto& pair : byte_unicode_pairs) {
+        vocab.push_back(pair.second + utf8_to_utf32("</w>"));
+    }
+    for (const auto& merge : merge_pairs) {
+        vocab.push_back(merge.first + merge.second);
+    }
+    vocab.push_back(utf8_to_utf32("<|startoftext|>"));
+    vocab.push_back(utf8_to_utf32("<|endoftext|>"));
+    LOG_DEBUG("vocab size: %llu", vocab.size());
+    int i = 0;
+    for (const auto& token : vocab) {
+        encoder[token] = i;
+        decoder[i]     = token;
+        i++;
+    }
+    encoder_len = i;
+
+    int rank = 0;
+    for (const auto& merge : merge_pairs) {
+        bpe_ranks[merge] = rank++;
+    }
+    bpe_len = rank;
+}
+
+static std::string strip(const std::string& str) {
+    std::string::size_type start = str.find_first_not_of(" \t\n\r\v\f");
+    std::string::size_type end   = str.find_last_not_of(" \t\n\r\v\f");
+
+    if (start == std::string::npos) {
+        return "";
+    }
+
+    return str.substr(start, end - start + 1);
+}
+
+static std::string whitespace_clean(const std::string& text) {
+    auto result = std::regex_replace(text, std::regex(R"(\s+)"), " ");
+    result      = strip(result);
+    return result;
+}
+
+std::string CLIPTokenizer::normalize(const std::string& text) const {
+    auto normalized_text = whitespace_clean(text);
+    std::transform(normalized_text.begin(), normalized_text.end(), normalized_text.begin(), [](unsigned char c) { return static_cast<char>(std::tolower(c)); });
+    return normalized_text;
+}
+
+std::vector<std::string> CLIPTokenizer::token_split(const std::string& text) const {
+    std::regex clip_pat(R"('s|'t|'re|'ve|'m|'ll|'d|[[:alpha:]]+|[[:digit:]]|[^[:space:][:alpha:][:digit:]]+)",
+                        std::regex::icase);
+    std::sregex_iterator iter(text.begin(), text.end(), clip_pat);
+    std::sregex_iterator end;
+
+    std::vector<std::string> result;
+    for (; iter != end; ++iter) {
+        result.emplace_back(iter->str());
+    }
+
+    return result;
+}

src/tokenizers/clip_tokenizer.h

@@ -0,0 +1,20 @@
+#ifndef __SD_TOKENIZERS_CLIP_TOKENIZER_H__
+#define __SD_TOKENIZERS_CLIP_TOKENIZER_H__
+
+#include <cstddef>
+#include <string>
+#include <vector>
+
+#include "bpe_tokenizer.h"
+
+class CLIPTokenizer : public BPETokenizer {
+protected:
+    void load_from_merges(const std::string& merges_utf8_str);
+    std::string normalize(const std::string& text) const override;
+    std::vector<std::string> token_split(const std::string& text) const override;
+
+public:
+    explicit CLIPTokenizer(int pad_token_id = 49407, const std::string& merges_utf8_str = "");
+};
+
+#endif  // __SD_TOKENIZERS_CLIP_TOKENIZER_H__

src/tokenizers/mistral_tokenizer.cpp

@@ -0,0 +1,89 @@
+#include "mistral_tokenizer.h"
+
+#include "ggml.h"
+#include "json.hpp"
+#include "util.h"
+#include "vocab/vocab.h"
+
+void MistralTokenizer::load_from_merges(const std::string& merges_utf8_str, const std::string& vocab_utf8_str) {
+    nlohmann::json vocab;
+
+    try {
+        vocab = nlohmann::json::parse(vocab_utf8_str);
+    } catch (const nlohmann::json::parse_error&) {
+        GGML_ABORT("invalid vocab json str");
+    }
+    for (const auto& [key, value] : vocab.items()) {
+        std::u32string token = utf8_to_utf32(key);
+        int i                = value;
+        encoder[token]       = i;
+        decoder[i]           = token;
+    }
+    encoder_len = static_cast<int>(vocab.size());
+    LOG_DEBUG("vocab size: %d", encoder_len);
+
+    auto byte_unicode_pairs = bytes_to_unicode();
+    byte_encoder            = std::map<int, std::u32string>(byte_unicode_pairs.begin(), byte_unicode_pairs.end());
+    for (auto& pair : byte_unicode_pairs) {
+        byte_decoder[pair.second] = pair.first;
+    }
+    std::vector<std::u32string> merges = split_utf32(merges_utf8_str);
+    LOG_DEBUG("merges size %llu", merges.size());
+    std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
+    for (const auto& merge : merges) {
+        size_t space_pos = merge.find(' ');
+        merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
+    }
+
+    int rank = 0;
+    for (const auto& merge : merge_pairs) {
+        bpe_ranks[merge] = rank++;
+    }
+    bpe_len = rank;
+}
+
+MistralTokenizer::MistralTokenizer(const std::string& merges_utf8_str, const std::string& vocab_utf8_str) {
+    add_bos_token = true;
+
+    UNK_TOKEN = "<unk>";
+    BOS_TOKEN = "<s>";
+    EOS_TOKEN = "</s>";
+    PAD_TOKEN = "<pad>";
+
+    UNK_TOKEN_ID = 0;
+    BOS_TOKEN_ID = 1;
+    EOS_TOKEN_ID = 2;
+    PAD_TOKEN_ID = 11;
+
+    special_tokens = {
+        "<unk>",
+        "<s>",
+        "</s>",
+        "[INST]",
+        "[/INST]",
+        "[AVAILABLE_TOOLS]",
+        "[/AVAILABLE_TOOLS]",
+        "[TOOL_RESULTS]",
+        "[/TOOL_RESULTS]",
+        "[TOOL_CALLS]",
+        "[IMG]",
+        "<pad>",
+        "[IMG_BREAK]",
+        "[IMG_END]",
+        "[PREFIX]",
+        "[MIDDLE]",
+        "[SUFFIX]",
+        "[SYSTEM_PROMPT]",
+        "[/SYSTEM_PROMPT]",
+        "[TOOL_CONTENT]",
+    };
+    for (int i = 20; i < 1000; i++) {
+        special_tokens.push_back("<SPECIAL_" + std::to_string(i) + ">");
+    }
+
+    if (merges_utf8_str.size() > 0 && vocab_utf8_str.size() > 0) {
+        load_from_merges(merges_utf8_str, vocab_utf8_str);
+    } else {
+        load_from_merges(load_mistral_merges(), load_mistral_vocab_json());
+    }
+}

src/tokenizers/mistral_tokenizer.h

@@ -0,0 +1,16 @@
+#ifndef __SD_TOKENIZERS_MISTRAL_TOKENIZER_H__
+#define __SD_TOKENIZERS_MISTRAL_TOKENIZER_H__
+
+#include <string>
+
+#include "bpe_tokenizer.h"
+
+class MistralTokenizer : public BPETokenizer {
+protected:
+    void load_from_merges(const std::string& merges_utf8_str, const std::string& vocab_utf8_str);
+
+public:
+    explicit MistralTokenizer(const std::string& merges_utf8_str = "", const std::string& vocab_utf8_str = "");
+};
+
+#endif  // __SD_TOKENIZERS_MISTRAL_TOKENIZER_H__

src/tokenizers/qwen2_tokenizer.cpp

@@ -0,0 +1,91 @@
+#include "qwen2_tokenizer.h"
+
+#include "util.h"
+#include "vocab/vocab.h"
+
+void Qwen2Tokenizer::load_from_merges(const std::string& merges_utf8_str) {
+    auto byte_unicode_pairs = bytes_to_unicode();
+    byte_encoder            = std::map<int, std::u32string>(byte_unicode_pairs.begin(), byte_unicode_pairs.end());
+    for (auto& pair : byte_unicode_pairs) {
+        byte_decoder[pair.second] = pair.first;
+    }
+
+    std::vector<std::u32string> merges = split_utf32(merges_utf8_str);
+    LOG_DEBUG("merges size %llu", merges.size());
+    std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
+    for (const auto& merge : merges) {
+        size_t space_pos = merge.find(' ');
+        merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
+    }
+
+    std::vector<std::u32string> tokens;
+    for (const auto& pair : byte_unicode_pairs) {
+        tokens.push_back(pair.second);
+    }
+    for (const auto& merge : merge_pairs) {
+        tokens.push_back(merge.first + merge.second);
+    }
+    for (auto& special_token : special_tokens) {
+        tokens.push_back(utf8_to_utf32(special_token));
+    }
+
+    int i = 0;
+    for (const auto& token : tokens) {
+        encoder[token] = i;
+        decoder[i]     = token;
+        i++;
+    }
+    encoder_len = i;
+    LOG_DEBUG("vocab size: %d", encoder_len);
+
+    int rank = 0;
+    for (const auto& merge : merge_pairs) {
+        bpe_ranks[merge] = rank++;
+    }
+    bpe_len = rank;
+}
+
+Qwen2Tokenizer::Qwen2Tokenizer(const std::string& merges_utf8_str) {
+    UNK_TOKEN = "<|endoftext|>";
+    EOS_TOKEN = "<|endoftext|>";
+    PAD_TOKEN = "<|endoftext|>";
+
+    UNK_TOKEN_ID = 151643;
+    EOS_TOKEN_ID = 151643;
+    PAD_TOKEN_ID = 151643;
+
+    special_tokens = {
+        "<|endoftext|>",
+        "<|im_start|>",
+        "<|im_end|>",
+        "<|object_ref_start|>",
+        "<|object_ref_end|>",
+        "<|box_start|>",
+        "<|box_end|>",
+        "<|quad_start|>",
+        "<|quad_end|>",
+        "<|vision_start|>",
+        "<|vision_end|>",
+        "<|vision_pad|>",
+        "<|image_pad|>",
+        "<|video_pad|>",
+        "<tool_call>",
+        "</tool_call>",
+        "<|fim_prefix|>",
+        "<|fim_middle|>",
+        "<|fim_suffix|>",
+        "<|fim_pad|>",
+        "<|repo_name|>",
+        "<|file_sep|>",
+        "<tool_response>",
+        "</tool_response>",
+        "<think>",
+        "</think>",
+    };
+
+    if (merges_utf8_str.size() > 0) {
+        load_from_merges(merges_utf8_str);
+    } else {
+        load_from_merges(load_qwen2_merges());
+    }
+}

src/tokenizers/qwen2_tokenizer.h

@@ -0,0 +1,16 @@
+#ifndef __SD_TOKENIZERS_QWEN2_TOKENIZER_H__
+#define __SD_TOKENIZERS_QWEN2_TOKENIZER_H__
+
+#include <string>
+
+#include "bpe_tokenizer.h"
+
+class Qwen2Tokenizer : public BPETokenizer {
+protected:
+    void load_from_merges(const std::string& merges_utf8_str);
+
+public:
+    explicit Qwen2Tokenizer(const std::string& merges_utf8_str = "");
+};
+
+#endif  // __SD_TOKENIZERS_QWEN2_TOKENIZER_H__

src/tokenizers/t5_unigram_tokenizer.cpp

@@ -0,0 +1,339 @@
+#include "t5_unigram_tokenizer.h"
+
+#include <algorithm>
+#include <cfloat>
+#include <cmath>
+#include <regex>
+#include <sstream>
+
+#include "json.hpp"
+#include "tokenize_util.h"
+#include "util.h"
+#include "vocab/vocab.h"
+
+// Port from: https://github.com/google/sentencepiece/blob/master/src/unigram_model.h
+// and https://github.com/google/sentencepiece/blob/master/src/unigram_model.h.
+// Original License: https://github.com/google/sentencepiece/blob/master/LICENSE
+//
+// Since tokenization is not the bottleneck in SD, performance was not a major consideration
+// during the migration.
+
+MetaspacePreTokenizer::MetaspacePreTokenizer(const std::string replacement, bool add_prefix_space)
+    : replacement(replacement), add_prefix_space(add_prefix_space) {}
+
+std::string MetaspacePreTokenizer::tokenize(const std::string& input) const {
+    std::string tokens;
+    std::stringstream ss(input);
+
+    if (add_prefix_space) {
+        tokens += replacement;
+    }
+
+    std::string token;
+    bool first_token = true;
+    while (std::getline(ss, token, ' ')) {
+        if (!first_token) {
+            tokens += replacement + token;
+        } else {
+            tokens += token;
+        }
+
+        first_token = false;
+    }
+
+    return tokens;
+}
+
+void T5UniGramTokenizer::InitializePieces(const std::string& json_str) {
+    nlohmann::json data;
+
+    try {
+        data = nlohmann::json::parse(json_str);
+    } catch (const nlohmann::json::parse_error&) {
+        status_ = INVLIAD_JSON;
+        return;
+    }
+    if (!data.contains("model")) {
+        status_ = INVLIAD_JSON;
+        return;
+    }
+    nlohmann::json model = data["model"];
+    if (!model.contains("vocab")) {
+        status_ = INVLIAD_JSON;
+        return;
+    }
+    if (model.contains("unk_id")) {
+        UNK_TOKEN_ID = model["unk_id"];
+    }
+
+    replacement      = data["pre_tokenizer"]["replacement"];
+    add_prefix_space = data["pre_tokenizer"]["add_prefix_space"];
+
+    pre_tokenizer = MetaspacePreTokenizer(replacement, add_prefix_space);
+
+    for (const auto& item : model["vocab"]) {
+        if (item.size() != 2 || !item[0].is_string() || !item[1].is_number_float()) {
+            status_ = INVLIAD_JSON;
+            return;
+        }
+        std::string piece = item[0];
+        if (piece.empty()) {
+            piece = "<empty_token>";
+        }
+        float score = item[1];
+        piece_score_pairs.emplace_back(piece, score);
+    }
+}
+
+void T5UniGramTokenizer::BuildTrie(std::vector<std::pair<std::string, int>>* pieces) {
+    if (status_ != OK) {
+        return;
+    }
+
+    if (pieces->empty()) {
+        status_ = NO_PIECES_LOADED;
+        return;
+    }
+
+    std::sort(pieces->begin(), pieces->end());
+
+    std::vector<const char*> key(pieces->size());
+    std::vector<int> value(pieces->size());
+    for (size_t i = 0; i < pieces->size(); ++i) {
+        key[i]   = (*pieces)[i].first.data();
+        value[i] = (*pieces)[i].second;
+    }
+
+    trie_ = std::unique_ptr<Darts::DoubleArray>(new Darts::DoubleArray());
+    if (trie_->build(key.size(), const_cast<char**>(&key[0]), nullptr, &value[0]) != 0) {
+        status_ = BUILD_DOUBLE_ARRAY_FAILED;
+        return;
+    }
+
+    const int kMaxTrieResultsSize = 1024;
+    std::vector<Darts::DoubleArray::result_pair_type> results(kMaxTrieResultsSize);
+    trie_results_size_ = 0;
+    for (const auto& p : *pieces) {
+        const size_t num_nodes = trie_->commonPrefixSearch(
+            p.first.data(), results.data(), results.size(), p.first.size());
+        trie_results_size_ = std::max(trie_results_size_, static_cast<int>(num_nodes));
+    }
+
+    if (trie_results_size_ == 0) {
+        status_ = NO_ENTRY_FOUND;
+    }
+}
+
+float T5UniGramTokenizer::GetScoreInlined(int id) const {
+    return piece_score_pairs[id].second;
+}
+
+bool T5UniGramTokenizer::IsUnusedInlined(int id) const {
+    (void)id;
+    return false;
+}
+
+bool T5UniGramTokenizer::IsUserDefinedInlined(int id) const {
+    (void)id;
+    return false;
+}
+
+size_t T5UniGramTokenizer::OneCharLen(const char* src) const {
+    return "\1\1\1\1\1\1\1\1\1\1\1\1\2\2\3\4"[(*src & 0xFF) >> 4];
+}
+
+EncodeResult T5UniGramTokenizer::EncodeOptimized(const std::string& normalized) const {
+    if (status() != OK || normalized.empty()) {
+        return {};
+    }
+
+    struct BestPathNode {
+        int id                = -1;
+        float best_path_score = 0;
+        int starts_at         = -1;
+    };
+
+    const int size        = static_cast<int>(normalized.size());
+    const float unk_score = min_score() - kUnkPenalty;
+    std::vector<BestPathNode> best_path_ends_at(size + 1);
+
+    int starts_at = 0;
+    while (starts_at < size) {
+        std::size_t node_pos                 = 0;
+        std::size_t key_pos                  = starts_at;
+        const auto best_path_score_till_here = best_path_ends_at[starts_at].best_path_score;
+        bool has_single_node                 = false;
+        const int mblen                      = std::min<int>(static_cast<int>(OneCharLen(normalized.data() + starts_at)), size - starts_at);
+        while (key_pos < static_cast<size_t>(size)) {
+            const int ret = trie_->traverse(normalized.data(), node_pos, key_pos, key_pos + 1);
+            if (ret == -2) {
+                break;
+            }
+            if (ret >= 0) {
+                if (IsUnusedInlined(ret)) {
+                    continue;
+                }
+                auto& target_node                    = best_path_ends_at[key_pos];
+                const auto length                    = static_cast<int>(key_pos - starts_at);
+                const auto score                     = IsUserDefinedInlined(ret) ? (length * max_score_ - 0.1f) : GetScoreInlined(ret);
+                const auto candidate_best_path_score = score + best_path_score_till_here;
+                if (target_node.starts_at == -1 || candidate_best_path_score > target_node.best_path_score) {
+                    target_node.best_path_score = static_cast<float>(candidate_best_path_score);
+                    target_node.starts_at       = starts_at;
+                    target_node.id              = ret;
+                }
+                if (!has_single_node && length == mblen) {
+                    has_single_node = true;
+                }
+            }
+        }
+        if (!has_single_node) {
+            auto& target_node                    = best_path_ends_at[starts_at + mblen];
+            const auto candidate_best_path_score = unk_score + best_path_score_till_here;
+            if (target_node.starts_at == -1 || candidate_best_path_score > target_node.best_path_score) {
+                target_node.best_path_score = candidate_best_path_score;
+                target_node.starts_at       = starts_at;
+                target_node.id              = UNK_TOKEN_ID;
+            }
+        }
+        starts_at += mblen;
+    }
+
+    EncodeResult results;
+    int ends_at = size;
+    while (ends_at > 0) {
+        const auto& node = best_path_ends_at[ends_at];
+        results.emplace_back(normalized.substr(node.starts_at, ends_at - node.starts_at), node.id);
+        ends_at = node.starts_at;
+    }
+    std::reverse(results.begin(), results.end());
+    return results;
+}
+
+T5UniGramTokenizer::T5UniGramTokenizer(bool is_umt5) {
+    add_bos_token = false;
+    add_eos_token = true;
+
+    if (is_umt5) {
+        PAD_TOKEN_ID = 0;
+        EOS_TOKEN_ID = 1;
+        BOS_TOKEN_ID = 2;
+        UNK_TOKEN_ID = 3;
+
+        PAD_TOKEN = "<pad>";
+        EOS_TOKEN = "</s>";
+        BOS_TOKEN = "<s>";
+        UNK_TOKEN = "<unk>";
+    } else {
+        PAD_TOKEN_ID = 0;
+        EOS_TOKEN_ID = 1;
+        UNK_TOKEN_ID = 2;
+
+        PAD_TOKEN = "<pad>";
+        EOS_TOKEN = "</s>";
+        UNK_TOKEN = "<unk>";
+    }
+
+    special_tokens = {
+        "<pad>",
+        "</s>",
+        "<unk>",
+    };
+
+    if (is_umt5) {
+        special_tokens.push_back("<s>");
+    }
+
+    if (is_umt5) {
+        InitializePieces(load_umt5_tokenizer_json());
+    } else {
+        InitializePieces(load_t5_tokenizer_json());
+    }
+
+    min_score_ = FLT_MAX;
+    max_score_ = FLT_MIN;
+
+    std::vector<std::pair<std::string, int>> pieces;
+    for (int i = 0; i < static_cast<int>(piece_score_pairs.size()); i++) {
+        const auto& sp = piece_score_pairs[i];
+
+        min_score_ = std::min(min_score_, sp.second);
+        max_score_ = std::max(max_score_, sp.second);
+
+        pieces.emplace_back(sp.first, i);
+    }
+
+    BuildTrie(&pieces);
+}
+
+T5UniGramTokenizer::~T5UniGramTokenizer() = default;
+
+std::string T5UniGramTokenizer::decode_token(int token_id) const {
+    if (token_id < 0 || token_id >= static_cast<int>(piece_score_pairs.size())) {
+        return "";
+    }
+
+    const std::string& piece = piece_score_pairs[token_id].first;
+    if (piece == "<empty_token>") {
+        return "";
+    }
+    return piece;
+}
+
+std::string T5UniGramTokenizer::normalize(const std::string& input) const {
+    // Ref: https://github.com/huggingface/tokenizers/blob/1ff56c0c70b045f0cd82da1af9ac08cd4c7a6f9f/bindings/python/py_src/tokenizers/implementations/sentencepiece_unigram.py#L29
+    // TODO: nmt-nfkc
+    std::string normalized = std::regex_replace(input, std::regex(" {2,}"), " ");
+    return normalized;
+}
+
+std::vector<int> T5UniGramTokenizer::encode(const std::string& input, on_new_token_cb_t on_new_token_cb) {
+    std::vector<int32_t> tokens;
+    std::vector<std::string> token_strs;
+    std::string normalized = normalize(input);
+    auto splited_texts     = split_with_special_tokens(normalized, special_tokens);
+    if (splited_texts.empty()) {
+        splited_texts.push_back(normalized);  // for empty string
+    }
+
+    for (auto& splited_text : splited_texts) {
+        if (is_special_token(splited_text)) {
+            if (on_new_token_cb != nullptr) {
+                bool skip = on_new_token_cb(splited_text, tokens);
+                if (skip) {
+                    token_strs.push_back(splited_text);
+                    continue;
+                }
+            }
+
+            if (splited_text == UNK_TOKEN) {
+                tokens.push_back(UNK_TOKEN_ID);
+                token_strs.push_back(UNK_TOKEN);
+            } else if (splited_text == EOS_TOKEN) {
+                tokens.push_back(EOS_TOKEN_ID);
+                token_strs.push_back(EOS_TOKEN);
+            } else if (splited_text == PAD_TOKEN) {
+                tokens.push_back(PAD_TOKEN_ID);
+                token_strs.push_back(PAD_TOKEN);
+            }
+            continue;
+        }
+
+        std::string pretokenized = pre_tokenizer.tokenize(splited_text);
+        EncodeResult result      = EncodeOptimized(pretokenized);
+        for (const auto& item : result) {
+            tokens.push_back(item.second);
+            token_strs.push_back(item.first);
+        }
+    }
+
+    std::stringstream ss;
+    ss << "[";
+    for (const auto& token_str : token_strs) {
+        ss << "\"" << token_str << "\", ";
+    }
+    ss << "]";
+    LOG_DEBUG("split prompt \"%s\" to tokens %s", input.c_str(), ss.str().c_str());
+
+    return tokens;
+}

src/tokenizers/t5_unigram_tokenizer.h

@@ -0,0 +1,70 @@
+#ifndef __SD_TOKENIZERS_T5_UNIGRAM_TOKENIZER_H__
+#define __SD_TOKENIZERS_T5_UNIGRAM_TOKENIZER_H__
+
+#include <cstddef>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "darts.h"
+#include "tokenizer.h"
+
+class MetaspacePreTokenizer {
+private:
+    std::string replacement;
+    bool add_prefix_space;
+
+public:
+    MetaspacePreTokenizer(const std::string replacement = " ", bool add_prefix_space = true);
+
+    std::string tokenize(const std::string& input) const;
+};
+
+using EncodeResult = std::vector<std::pair<std::string, int>>;
+
+class T5UniGramTokenizer : public Tokenizer {
+public:
+    enum Status {
+        OK,
+        NO_PIECES_LOADED,
+        NO_ENTRY_FOUND,
+        BUILD_DOUBLE_ARRAY_FAILED,
+        PIECE_ALREADY_DEFINED,
+        INVLIAD_JSON
+    };
+
+protected:
+    MetaspacePreTokenizer pre_tokenizer;
+    std::vector<std::pair<std::string, float>> piece_score_pairs;
+    float min_score_ = 0.0f;
+    float max_score_ = 0.0f;
+    std::unique_ptr<Darts::DoubleArray> trie_;
+    int trie_results_size_ = 0;
+    Status status_         = OK;
+    float kUnkPenalty      = 10.0f;
+    std::string replacement;
+    bool add_prefix_space = true;
+
+    void InitializePieces(const std::string& json_str);
+    void BuildTrie(std::vector<std::pair<std::string, int>>* pieces);
+    float GetScoreInlined(int id) const;
+    bool IsUnusedInlined(int id) const;
+    bool IsUserDefinedInlined(int id) const;
+    size_t OneCharLen(const char* src) const;
+    EncodeResult EncodeOptimized(const std::string& normalized) const;
+
+    float min_score() const { return min_score_; }
+    float max_score() const { return max_score_; }
+    Status status() const { return status_; }
+    std::string decode_token(int token_id) const override;
+    std::string normalize(const std::string& input) const override;
+
+public:
+    explicit T5UniGramTokenizer(bool is_umt5 = false);
+    ~T5UniGramTokenizer();
+
+    std::vector<int> encode(const std::string& input, on_new_token_cb_t on_new_token_cb = nullptr) override;
+};
+
+#endif  // __SD_TOKENIZERS_T5_UNIGRAM_TOKENIZER_H__

src/tokenizers/tokenize_util.h

@@ -1,5 +1,5 @@
-#ifndef __TOKENIZE_UTIL__
-#define __TOKENIZE_UTIL__
+#ifndef __SD_TOKENIZERS_BPE_TOKENIZE_UTIL_H__
+#define __SD_TOKENIZERS_BPE_TOKENIZE_UTIL_H__
 
 #include <string>
 #include <vector>
@@ -7,4 +7,4 @@
 std::vector<std::string> token_split(const std::string& text);
 std::vector<std::string> split_with_special_tokens(const std::string& text, const std::vector<std::string>& special_tokens);
 
-#endif  // __TOKENIZE_UTIL__
+#endif  // __SD_TOKENIZERS_BPE_TOKENIZE_UTIL_H__

src/tokenizers/tokenizer.cpp

@@ -0,0 +1,222 @@
+#include "tokenizer.h"
+
+#include <algorithm>
+#include <cmath>
+#include <regex>
+
+#include "util.h"
+
+void Tokenizer::add_special_token(const std::string& token) {
+    special_tokens.push_back(token);
+}
+
+bool Tokenizer::is_special_token(const std::string& token) const {
+    for (const auto& special_token : special_tokens) {
+        if (special_token == token) {
+            return true;
+        }
+    }
+    return false;
+}
+
+std::string Tokenizer::normalize(const std::string& text) const {
+    return text;
+}
+
+std::vector<int> Tokenizer::tokenize(const std::string& text,
+                                     on_new_token_cb_t on_new_token_cb,
+                                     bool padding,
+                                     size_t min_length,
+                                     size_t max_length,
+                                     bool allow_overflow_expand) {
+    std::vector<int> tokens = encode(text, on_new_token_cb);
+    if (padding) {
+        pad_tokens(tokens, nullptr, nullptr, min_length, max_length, allow_overflow_expand);
+    }
+    return tokens;
+}
+
+void Tokenizer::pad_tokens(std::vector<int>& tokens,
+                           std::vector<float>* weights,
+                           std::vector<float>* mask,
+                           size_t min_length,
+                           size_t max_length,
+                           bool allow_overflow_expand) {
+    const bool use_weights = weights != nullptr;
+    const bool use_mask    = mask != nullptr;
+
+    if (use_weights && tokens.size() != weights->size()) {
+        LOG_ERROR("tokens size != weights size");
+        return;
+    }
+
+    const size_t bos_count           = add_bos_token ? 1 : 0;
+    const size_t eos_count           = add_eos_token ? 1 : 0;
+    const size_t special_token_count = bos_count + eos_count;
+
+    auto build_sequence = [&](size_t begin,
+                              size_t count,
+                              size_t target_length,
+                              std::vector<int>& out_tokens,
+                              std::vector<float>& out_weights,
+                              std::vector<float>& out_mask) {
+        const size_t base_length  = count + special_token_count;
+        const size_t final_length = std::max(target_length, base_length);
+
+        out_tokens.clear();
+        out_weights.clear();
+        out_mask.clear();
+
+        out_tokens.reserve(final_length);
+        if (use_weights) {
+            out_weights.reserve(final_length);
+        }
+        if (use_mask) {
+            out_mask.reserve(final_length);
+        }
+
+        if (add_bos_token) {
+            out_tokens.push_back(BOS_TOKEN_ID);
+            if (use_weights) {
+                out_weights.push_back(1.0f);
+            }
+            if (use_mask) {
+                out_mask.push_back(1.0f);
+            }
+        }
+
+        for (size_t i = 0; i < count; ++i) {
+            out_tokens.push_back(tokens[begin + i]);
+            if (use_weights) {
+                out_weights.push_back((*weights)[begin + i]);
+            }
+            if (use_mask) {
+                out_mask.push_back(1.0f);
+            }
+        }
+
+        if (add_eos_token) {
+            out_tokens.push_back(EOS_TOKEN_ID);
+            if (use_weights) {
+                out_weights.push_back(1.0f);
+            }
+            if (use_mask) {
+                out_mask.push_back(1.0f);
+            }
+        }
+
+        if (final_length > out_tokens.size()) {
+            const size_t pad_count = final_length - out_tokens.size();
+            if (pad_left) {
+                out_tokens.insert(out_tokens.begin(), pad_count, PAD_TOKEN_ID);
+
+                if (use_weights) {
+                    out_weights.insert(out_weights.begin(), pad_count, 1.0f);
+                }
+                if (use_mask) {
+                    out_mask.insert(out_mask.begin(), pad_count, 0.0f);
+                }
+            } else {
+                out_tokens.insert(out_tokens.end(), pad_count, PAD_TOKEN_ID);
+
+                if (use_weights) {
+                    out_weights.insert(out_weights.end(), pad_count, 1.0f);
+                }
+                if (use_mask) {
+                    out_mask.insert(out_mask.end(), pad_count, 0.0f);
+                }
+            }
+        }
+    };
+
+    const size_t single_length    = std::max(min_length, tokens.size() + special_token_count);
+    const bool exceeds_max_length = max_length > 0 && single_length > max_length;
+
+    std::vector<int> new_tokens;
+    std::vector<float> new_weights;
+    std::vector<float> new_mask;
+
+    if (!exceeds_max_length) {
+        build_sequence(0, tokens.size(), min_length, new_tokens, new_weights, new_mask);
+    } else if (!allow_overflow_expand) {
+        build_sequence(0, tokens.size(), 0, new_tokens, new_weights, new_mask);
+
+        new_tokens.resize(max_length);
+        if (use_weights) {
+            new_weights.resize(max_length);
+        }
+        if (use_mask) {
+            new_mask.resize(max_length);
+        }
+
+        if (add_eos_token && !new_tokens.empty()) {
+            new_tokens.back() = EOS_TOKEN_ID;
+            if (use_weights) {
+                new_weights.back() = 1.0f;
+            }
+            if (use_mask) {
+                new_mask.back() = 1.0f;
+            }
+        }
+    } else if (min_length > special_token_count) {
+        const size_t tokens_per_chunk = min_length - special_token_count;
+        size_t offset                 = 0;
+
+        while (offset < tokens.size()) {
+            const size_t remaining = tokens.size() - offset;
+            const size_t take      = std::min(tokens_per_chunk, remaining);
+
+            std::vector<int> chunk_tokens;
+            std::vector<float> chunk_weights;
+            std::vector<float> chunk_mask;
+
+            build_sequence(offset, take, min_length, chunk_tokens, chunk_weights, chunk_mask);
+
+            new_tokens.insert(new_tokens.end(), chunk_tokens.begin(), chunk_tokens.end());
+            if (use_weights) {
+                new_weights.insert(new_weights.end(), chunk_weights.begin(), chunk_weights.end());
+            }
+            if (use_mask) {
+                new_mask.insert(new_mask.end(), chunk_mask.begin(), chunk_mask.end());
+            }
+
+            offset += take;
+        }
+    } else {
+        build_sequence(0, tokens.size(), min_length, new_tokens, new_weights, new_mask);
+    }
+
+    tokens = std::move(new_tokens);
+    if (use_weights) {
+        *weights = std::move(new_weights);
+    }
+    if (use_mask) {
+        *mask = std::move(new_mask);
+    }
+}
+
+static std::string clean_up_tokenization(std::string& text) {
+    std::regex pattern(R"( ,)");
+    return std::regex_replace(text, pattern, ",");
+}
+
+std::string Tokenizer::decode(const std::vector<int>& tokens) const {
+    std::string text;
+
+    for (int token_id : tokens) {
+        if (token_id == BOS_TOKEN_ID || token_id == EOS_TOKEN_ID || token_id == PAD_TOKEN_ID) {
+            continue;
+        }
+
+        std::string piece = decode_token(token_id);
+        if (!end_of_word_suffix.empty() && ends_with(piece, end_of_word_suffix)) {
+            piece.erase(piece.size() - end_of_word_suffix.size());
+            text += piece + " ";
+        } else {
+            text += piece;
+        }
+    }
+
+    text = clean_up_tokenization(text);
+    return trim(text);
+}

src/tokenizers/tokenizer.h

@@ -0,0 +1,53 @@
+#ifndef __SD_TOKENIZERS_TOKENIZER_H__
+#define __SD_TOKENIZERS_TOKENIZER_H__
+
+#include <cstddef>
+#include <cstdint>
+#include <functional>
+#include <string>
+#include <vector>
+
+using on_new_token_cb_t = std::function<bool(std::string&, std::vector<int32_t>&)>;
+
+class Tokenizer {
+protected:
+    std::vector<std::string> special_tokens;
+    bool add_bos_token = false;
+    bool add_eos_token = false;
+    bool pad_left      = false;
+    std::string end_of_word_suffix;
+
+    virtual std::string decode_token(int token_id) const = 0;
+    virtual std::string normalize(const std::string& text) const;
+
+public:
+    std::string UNK_TOKEN;
+    std::string BOS_TOKEN;
+    std::string EOS_TOKEN;
+    std::string PAD_TOKEN;
+    int UNK_TOKEN_ID = 0;
+    int BOS_TOKEN_ID = 0;
+    int EOS_TOKEN_ID = 0;
+    int PAD_TOKEN_ID = 0;
+
+    virtual ~Tokenizer() = default;
+
+    void add_special_token(const std::string& token);
+    bool is_special_token(const std::string& token) const;
+    virtual std::vector<int> encode(const std::string& text, on_new_token_cb_t on_new_token_cb = nullptr) = 0;
+    std::vector<int> tokenize(const std::string& text,
+                              on_new_token_cb_t on_new_token_cb = nullptr,
+                              bool padding                      = false,
+                              size_t min_length                 = 0,
+                              size_t max_length                 = 100000000,
+                              bool allow_overflow_expand        = false);
+    void pad_tokens(std::vector<int>& tokens,
+                    std::vector<float>* weights,
+                    std::vector<float>* mask,
+                    size_t min_length          = 0,
+                    size_t max_length          = 100000000,
+                    bool allow_overflow_expand = false);
+    std::string decode(const std::vector<int>& tokens) const;
+};
+
+#endif  // __SD_TOKENIZERS_TOKENIZER_H__

src/tokenizers/vocab/vocab.h

@@ -1,5 +1,5 @@
-#ifndef __VOCAB_H__
-#define __VOCAB_H__
+#ifndef __SD_TOKENIZERS_VOCAB_VOCAB_H__
+#define __SD_TOKENIZERS_VOCAB_VOCAB_H__
 
 #include <string>
 
@@ -10,4 +10,4 @@ std::string load_mistral_vocab_json();
 std::string load_t5_tokenizer_json();
 std::string load_umt5_tokenizer_json();
 
-#endif  // __VOCAB_H__
+#endif  // __SD_TOKENIZERS_VOCAB_VOCAB_H__

src/unet.hpp

@@ -217,11 +217,11 @@ public:
         } else if (sd_version_is_unet_edit(version)) {
             in_channels = 8;
         }
-        if (version == VERSION_SD1_TINY_UNET || version == VERSION_SD2_TINY_UNET || version == VERSION_SDXS) {
+        if (version == VERSION_SD1_TINY_UNET || version == VERSION_SD2_TINY_UNET || version == VERSION_SDXS_512_DS || version == VERSION_SDXS_09) {
             num_res_blocks = 1;
             channel_mult   = {1, 2, 4};
             tiny_unet      = true;
-            if (version == VERSION_SDXS) {
+            if (version == VERSION_SDXS_512_DS) {
                 attention_resolutions = {4, 2};  // here just like SDXL
             }
         }
@@ -264,6 +264,10 @@ public:
             if (version == VERSION_SVD) {
                 return new SpatialVideoTransformer(in_channels, n_head, d_head, depth, context_dim, use_linear_projection);
             } else {
+                if (version == VERSION_SDXS_09 && n_head == 5) {
+                    n_head = 1;    // to carry a special case of sdxs_09 into CrossAttentionLayer,
+                    d_head = 320;  // works as long the product remains equal (5*64 == 1*320)
+                }
                 return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim, use_linear_projection);
             }
         };

src/vae.hpp

@@ -69,7 +69,7 @@ public:
         int scale_factor = 8;
         if (version == VERSION_WAN2_2_TI2V) {
             scale_factor = 16;
-        } else if (sd_version_is_flux2(version)) {
+        } else if (sd_version_uses_flux2_vae(version)) {
             scale_factor = 16;
         } else if (version == VERSION_CHROMA_RADIANCE) {
             scale_factor = 1;