feat: add max-vram based segmented param offload (#1476)

Co-authored-by: Stéphane du Hamel <stephduh@live.fr>
Co-authored-by: Cyberhan123 <255542417@qq.com>
Co-authored-by: leejet <leejet714@gmail.com>

.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/**",
       ]
 
@@ -64,7 +66,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Dependencies
         id: depends
@@ -127,7 +129,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Dependencies
         id: depends
@@ -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:
@@ -205,7 +208,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Get commit hash
         id: commit
@@ -239,6 +242,7 @@ jobs:
         id: build-push
         uses: docker/build-push-action@v6
         with:
+          context: .
           platforms: linux/amd64
           push: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
           file: Dockerfile.${{ matrix.variant }}
@@ -264,7 +268,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Dependencies
         id: depends
@@ -345,7 +349,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Install cuda-toolkit
         id: cuda-toolkit
@@ -460,7 +464,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Cache ROCm Installation
         id: cache-rocm
@@ -573,7 +577,7 @@ jobs:
       - name: Setup pnpm
         uses: pnpm/action-setup@v4
         with:
-          version: 9
+          version: 10.15.1
 
       - name: Free disk space
         run: |

.gitmodules

@@ -3,4 +3,10 @@
 	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
+[submodule "thirdparty/libwebm"]
+	path = thirdparty/libwebm
+	url = https://github.com/webmproject/libwebm.git

CMakeLists.txt

@@ -11,6 +11,10 @@ endif()
 if (MSVC)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
     add_compile_definitions(_SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING)
+    add_compile_options(
+        $<$<COMPILE_LANGUAGE:C>:/MP>
+        $<$<COMPILE_LANGUAGE:CXX>:/MP>
+    )
 endif()
 
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
@@ -22,6 +26,26 @@ else()
     set(SD_STANDALONE OFF)
 endif()
 
+set(SD_SUBMODULE_WEBP FALSE)
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/thirdparty/libwebp/CMakeLists.txt")
+    set(SD_SUBMODULE_WEBP TRUE)
+endif()
+if(SD_SUBMODULE_WEBP)
+    set(SD_WEBP_DEFAULT ON)
+else()
+    set(SD_WEBP_DEFAULT ${SD_USE_SYSTEM_WEBP})
+endif()
+
+set(SD_SUBMODULE_WEBM FALSE)
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/thirdparty/libwebm/CMakeLists.txt")
+    set(SD_SUBMODULE_WEBM TRUE)
+endif()
+if(SD_SUBMODULE_WEBM)
+    set(SD_WEBM_DEFAULT ON)
+else()
+    set(SD_WEBM_DEFAULT ${SD_USE_SYSTEM_WEBM})
+endif()
+
 #
 # Option list
 #
@@ -29,6 +53,10 @@ endif()
 # general
 #option(SD_BUILD_TESTS                "sd: build tests"    ${SD_STANDALONE})
 option(SD_BUILD_EXAMPLES             "sd: build examples" ${SD_STANDALONE})
+option(SD_WEBP                       "sd: enable WebP image I/O support" ${SD_WEBP_DEFAULT})
+option(SD_USE_SYSTEM_WEBP            "sd: link against system libwebp" OFF)
+option(SD_WEBM                       "sd: enable WebM video output support" ${SD_WEBM_DEFAULT})
+option(SD_USE_SYSTEM_WEBM            "sd: link against system libwebm" OFF)
 option(SD_CUDA                       "sd: cuda backend" OFF)
 option(SD_HIPBLAS                    "sd: rocm backend" OFF)
 option(SD_METAL                      "sd: metal backend" OFF)
@@ -44,47 +72,94 @@ option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF
 if(SD_CUDA)
     message("-- Use CUDA as backend stable-diffusion")
     set(GGML_CUDA ON)
-    add_definitions(-DSD_USE_CUDA)
 endif()
 
 if(SD_METAL)
     message("-- Use Metal as backend stable-diffusion")
     set(GGML_METAL ON)
-    add_definitions(-DSD_USE_METAL)
 endif()
 
 if (SD_VULKAN)
     message("-- Use Vulkan as backend stable-diffusion")
     set(GGML_VULKAN ON)
-    add_definitions(-DSD_USE_VULKAN)
 endif ()
 
 if (SD_OPENCL)
     message("-- Use OpenCL as backend stable-diffusion")
     set(GGML_OPENCL ON)
-    add_definitions(-DSD_USE_OPENCL)
 endif ()
 
 if (SD_HIPBLAS)
     message("-- Use HIPBLAS as backend stable-diffusion")
     set(GGML_HIP ON)
-    add_definitions(-DSD_USE_CUDA)
 endif ()
 
 if(SD_MUSA)
     message("-- Use MUSA as backend stable-diffusion")
     set(GGML_MUSA ON)
-    add_definitions(-DSD_USE_CUDA)
+endif()
+
+if(SD_WEBP)
+    if(NOT SD_SUBMODULE_WEBP AND NOT SD_USE_SYSTEM_WEBP)
+        message(FATAL_ERROR "WebP support enabled but no source found.
+          Either initialize the submodule:\n  git submodule update --init thirdparty/libwebp\n\n"
+          "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBP=ON")
+    endif()
+    if(SD_USE_SYSTEM_WEBP)
+        find_package(WebP REQUIRED)
+        add_library(webp ALIAS WebP::webp)
+        # libwebp CMake target naming is not consistent across versions/distros.
+        # Some export WebP::libwebpmux, others export WebP::webpmux.
+        if(TARGET WebP::libwebpmux)
+            add_library(libwebpmux ALIAS WebP::libwebpmux)
+        elseif(TARGET WebP::webpmux)
+            add_library(libwebpmux ALIAS WebP::webpmux)
+        else()
+            message(FATAL_ERROR
+                "Could not find a compatible webpmux target in system WebP package. "
+                "Expected WebP::libwebpmux or WebP::webpmux."
+            )
+        endif()
+    endif()
+endif()
+
+if(SD_WEBM)
+    if(NOT SD_WEBP)
+        message(FATAL_ERROR "SD_WEBM requires SD_WEBP because WebM output reuses libwebp VP8 encoding.")
+    endif()
+    if(NOT SD_SUBMODULE_WEBM AND NOT SD_USE_SYSTEM_WEBM)
+        message(FATAL_ERROR "WebM support enabled but no source found.
+          Either initialize the submodule:\n  git submodule update --init thirdparty/libwebm\n\n"
+          "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBM=ON")
+    endif()
+    if(SD_USE_SYSTEM_WEBM)
+        find_path(WEBM_INCLUDE_DIR
+            NAMES mkvmuxer/mkvmuxer.h mkvparser/mkvparser.h common/webmids.h
+            PATH_SUFFIXES webm
+            REQUIRED)
+        find_library(WEBM_LIBRARY
+            NAMES webm libwebm
+            REQUIRED)
+
+        add_library(webm UNKNOWN IMPORTED)
+        set_target_properties(webm PROPERTIES
+            IMPORTED_LOCATION "${WEBM_LIBRARY}"
+            INTERFACE_INCLUDE_DIRECTORIES "${WEBM_INCLUDE_DIR}")
+    endif()
 endif()
 
 set(SD_LIB stable-diffusion)
 
-file(GLOB SD_LIB_SOURCES
+file(GLOB SD_LIB_SOURCES CONFIGURE_DEPENDS
     "src/*.h"
     "src/*.cpp"
     "src/*.hpp"
-    "src/vocab/*.h"
-    "src/vocab/*.cpp"
+    "src/model_io/*.h"
+    "src/model_io/*.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)
@@ -141,7 +216,6 @@ if(SD_SYCL)
     message("-- Use SYCL as backend stable-diffusion")
     set(GGML_SYCL ON)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
-    add_definitions(-DSD_USE_SYCL)
     # disable fast-math on host, see:
     # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html
     if (WIN32)
@@ -177,7 +251,7 @@ 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)
@@ -73,9 +77,10 @@ API and command-line option may change frequently.***
   - OpenCL
   - SYCL
 - Supported weight formats
-  - Pytorch checkpoint (`.ckpt` or `.pth`)
+  - Pytorch checkpoint (`.ckpt` or `.pth` or `.pt`)
   - Safetensors (`.safetensors`)
   - GGUF (`.gguf`)
+- Convert mode supports converting model weights to `.gguf` or `.safetensors`
 - Supported platforms
     - Linux
     - Mac OS
@@ -93,6 +98,7 @@ 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 +147,7 @@ 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/build.md

@@ -16,6 +16,26 @@ git submodule init
 git submodule update
 ```
 
+## WebP and WebM Support in Examples
+
+The example applications (`examples/cli` and `examples/server`) use `libwebp` to support WebP image I/O, and `examples/cli` can also use `libwebm` for `.webm` video output. Both are enabled by default. WebM output currently reuses `libwebp` to encode each frame as VP8 before muxing with `libwebm`.
+
+If you do not want WebP/WebM support, you can disable them at configure time:
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_WEBP=OFF -DSD_WEBM=OFF
+cmake --build . --config Release
+```
+
+If the submodules are not available, you can also link against system packages instead:
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_USE_SYSTEM_WEBP=ON -DSD_USE_SYSTEM_WEBM=ON
+cmake --build . --config Release
+```
+
 ## Build (CPU only)
 
 If you don't have a GPU or CUDA installed, you can build a CPU-only version.

docs/caching.md

@@ -131,8 +131,6 @@ sd-cli -m model.safetensors -p "a cat" --cache-mode spectrum
 | `warmup` | Steps to always compute before caching starts | 4 |
 | `stop` | Stop caching at this fraction of total steps | 0.9 |
 
-```
-
 ### Performance Tips
 
 - Start with default thresholds and adjust based on output quality

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" />

docs/flux2.md

@@ -8,6 +8,8 @@
     - gguf: https://huggingface.co/city96/FLUX.2-dev-gguf/tree/main
 - Download vae
     - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download FLUX.2-small-decoder (full_encoder_small_decoder.safetensors) as an alternative VAE option
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/tree/main
 - Download Mistral-Small-3.2-24B-Instruct-2506-GGUF
     - gguf: https://huggingface.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/tree/main
 
@@ -31,6 +33,8 @@
     - gguf: https://huggingface.co/leejet/FLUX.2-klein-base-4B-GGUF/tree/main
 - Download vae
     - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download FLUX.2-small-decoder (full_encoder_small_decoder.safetensors) as an alternative VAE option
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/tree/main
 - Download Qwen3 4b
     - safetensors: https://huggingface.co/Comfy-Org/flux2-klein-4B/tree/main/split_files/text_encoders
     - gguf: https://huggingface.co/unsloth/Qwen3-4B-GGUF/tree/main

examples/cli/CMakeLists.txt

@@ -1,6 +1,20 @@
 set(TARGET sd-cli)
 
-add_executable(${TARGET} main.cpp)
+add_executable(${TARGET}
+    ../common/common.cpp
+    ../common/log.cpp
+    ../common/media_io.cpp
+    image_metadata.cpp
+    main.cpp
+)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE stable-diffusion ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)
+target_link_libraries(${TARGET} PRIVATE stable-diffusion zip ${CMAKE_THREAD_LIBS_INIT})
+if(SD_WEBP)
+    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBP)
+    target_link_libraries(${TARGET} PRIVATE webp libwebpmux)
+endif()
+if(SD_WEBM)
+    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBM)
+    target_link_libraries(${TARGET} PRIVATE webm)
+endif()
+target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)

examples/cli/README.md

@@ -4,21 +4,29 @@
 usage: ./bin/sd-cli  [options]
 
 CLI Options:
-  -o, --output <string>       path to write result image to. you can use printf-style %d format specifiers for image sequences (default:
-                              ./output.png) (eg. output_%03d.png)
-  --preview-path <string>     path to write preview image to (default: ./preview.png)
-  --preview-interval <int>    interval in denoising steps between consecutive updates of the image preview file (default is 1, meaning updating at
-                              every step)
-  --output-begin-idx <int>    starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)
-  --canny                     apply canny preprocessor (edge detection)
-  --convert-name              convert tensor name (for convert mode)
-  -v, --verbose               print extra info
-  --color                     colors the logging tags according to level
-  --taesd-preview-only        prevents usage of taesd for decoding the final image. (for use with --preview tae)
-  --preview-noisy             enables previewing noisy inputs of the models rather than the denoised outputs
-  -M, --mode                  run mode, one of [img_gen, vid_gen, upscale, convert], default: img_gen
-  --preview                   preview method. must be one of the following [none, proj, tae, vae] (default is none)
-  -h, --help                  show this help message and exit
+  -o, --output <string>         path to write result image to. you can use printf-style %d format specifiers for image
+                                sequences (default: ./output.png) (eg. output_%03d.png). Single-file video outputs
+                                support .avi, .webm, and animated .webp
+  --image <string>              path to the image to inspect (for metadata mode)
+  --metadata-format <string>    metadata output format, one of [text, json] (default: text)
+  --preview-path <string>       path to write preview image to (default: ./preview.png). Multi-frame previews support
+                                .avi, .webm, and animated .webp
+  --preview-interval <int>      interval in denoising steps between consecutive updates of the image preview file
+                                (default is 1, meaning updating at every step)
+  --output-begin-idx <int>      starting index for output image sequence, must be non-negative (default 0 if specified
+                                %d in output path, 1 otherwise)
+  --canny                       apply canny preprocessor (edge detection)
+  --convert-name                convert tensor name (for convert mode)
+  -v, --verbose                 print extra info
+  --color                       colors the logging tags according to level
+  --taesd-preview-only          prevents usage of taesd for decoding the final image. (for use with --preview tae)
+  --preview-noisy               enables previewing noisy inputs of the models rather than the denoised outputs
+  --metadata-raw                include raw hex previews for unparsed metadata payloads
+  --metadata-brief              truncate long metadata text values in text output
+  --metadata-all                include structural/container entries such as IHDR, IDAT, and non-metadata JPEG segments
+  -M, --mode                    run mode, one of [img_gen, vid_gen, upscale, convert, metadata], default: img_gen
+  --preview                     preview method. must be one of the following [none, proj, tae, vae] (default is none)
+  -h, --help                    show this help message and exit
 
 Context Options:
   -m, --model <string>                     path to full model
@@ -26,7 +34,8 @@ Context Options:
   --clip_g <string>                        path to the clip-g text encoder
   --clip_vision <string>                   path to the clip-vision encoder
   --t5xxl <string>                         path to the t5xxl text encoder
-  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image, mistral-small3.2 for flux2, ...)
+  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image,
+                                           mistral-small3.2 for flux2, ...)
   --llm_vision <string>                    path to the llm vit
   --qwen2vl <string>                       alias of --llm. Deprecated.
   --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
@@ -38,16 +47,18 @@ Context Options:
   --control-net <string>                   path to control net model
   --embd-dir <string>                      embeddings directory
   --lora-model-dir <string>                lora model directory
+  --hires-upscalers-dir <string>           highres fix upscaler model directory
   --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
   --photo-maker <string>                   path to PHOTOMAKER model
   --upscale-model <string>                 path to esrgan model.
-  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0, then threads will be set to the number of
-                                           CPU physical cores
+  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0,
+                                           then threads will be set to the number of CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
-  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-  --vae-tiling                             process vae in tiles to reduce memory usage
+  --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
+                                           graph splitting
   --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
-  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
+  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM
+                                           when needed
   --mmap                                   whether to memory-map model
   --control-net-cpu                        keep controlnet in cpu (for low vram)
   --clip-on-cpu                            keep clip in cpu (for low vram)
@@ -62,20 +73,19 @@ Context Options:
   --chroma-disable-dit-mask                disable dit mask for chroma
   --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --chroma-enable-t5-mask                  enable t5 mask for chroma
-  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
-                                           type of the weight file
+  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K,
+                                           q4_K). If not specified, the default is the type of the weight file
   --rng                                    RNG, one of [std_default, cuda, cpu], default: cuda(sd-webui), cpu(comfyui)
   --sampler-rng                            sampler RNG, one of [std_default, cuda, cpu]. If not specified, use --rng
-  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow, flux2_flow]
-  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is auto. In auto mode, if the model weights
-                                           contain any quantized parameters, the at_runtime mode will be used; otherwise,
-                                           immediately will be used.The immediately mode may have precision and
-                                           compatibility issues with quantized parameters, but it usually offers faster inference
-                                           speed and, in some cases, lower memory usage. The at_runtime mode, on the
-                                           other hand, is exactly the opposite.
-  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1
-                                           (overrides --vae-tile-size)
+  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow,
+                                           flux2_flow]
+  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is
+                                           auto. In auto mode, if the model weights contain any quantized parameters,
+                                           the at_runtime mode will be used; otherwise, immediately will be used.The
+                                           immediately mode may have precision and compatibility issues with quantized
+                                           parameters, but it usually offers faster inference speed and, in some cases,
+                                           lower memory usage. The at_runtime mode, on the other hand, is exactly the
+                                           opposite.
 
 Generation Options:
   -p, --prompt <string>                    the prompt to render
@@ -84,66 +94,106 @@ Generation Options:
   --end-img <string>                       path to the end image, required by flf2v
   --mask <string>                          path to the mask image
   --control-image <string>                 path to control image, control net
-  --control-video <string>                 path to control video frames, It must be a directory path. The video frames inside should be stored as images in
-                                           lexicographical (character) order. For example, if the control video path is
-                                           `frames`, the directory contain images such as 00.png, 01.png, ... etc.
+  --control-video <string>                 path to control video frames, It must be a directory path. The video frames
+                                           inside should be stored as images in lexicographical (character) order. For
+                                           example, if the control video path is `frames`, the directory contain images
+                                           such as 00.png, 01.png, ... etc.
   --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
   --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
+  --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
+                                           (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
+                                           antialiased), or a model name under --hires-upscalers-dir (default: Latent)
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)
   --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
-  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1). <= 0 represents unspecified,
-                                           will be 1 for SD1.x, 2 for SD2.x
+  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer
+                                           (default: -1). <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
   -b, --batch-count <int>                  batch count
   --video-frames <int>                     video frames (default: 1)
   --fps <int>                              fps (default: 24)
-  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for NitroSD-Realism around 250 and 500 for
-                                           NitroSD-Vibrant
+  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for
+                                           NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
   --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
   --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
+  --hires-width <int>                      highres fix target width, 0 to use --hires-scale (default: 0)
+  --hires-height <int>                     highres fix target height, 0 to use --hires-scale (default: 0)
+  --hires-steps <int>                      highres fix second pass sample steps, 0 to reuse --steps (default: 0)
+  --hires-upscale-tile-size <int>          highres fix upscaler tile size, reserved for model-backed upscalers (default:
+                                           128)
   --cfg-scale <float>                      unconditional guidance scale: (default: 7.0)
-  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same
+                                           as --cfg-scale)
   --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
-  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means disabled, a value of 2.5 is nice for sd3.5
-                                           medium
+  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means
+                                           disabled, a value of 2.5 is nice for sd3.5 medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            eta in DDIM, only for DDIM and TCD (default: 0)
+  --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)
-  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
-  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
+                                           (default: same as --cfg-scale)
+  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input
+                                           (default: 3.5)
+  --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) eta in DDIM, only for DDIM and TCD (default: 0)
+  --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
-  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if `--high-noise-steps` is set to -1
+  --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
+  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if
+                                           `--high-noise-steps` is set to -1
   --vace-strength <float>                  wan vace strength
-  --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
+  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --hires-scale <float>                    highres fix scale when target size is not set (default: 2.0)
+  --hires-denoising-strength <float>       highres fix second pass denoising strength (default: 0.7)
+  --increase-ref-index                     automatically increase the indices of references images based on the order
+                                           they are listed (starting with 1).
   --disable-auto-resize-ref-image          disable auto resize of ref images
+  --disable-image-metadata                 do not embed generation metadata on image files
+  --vae-tiling                             process vae in tiles to reduce memory usage
+  --hires                                  enable highres fix
   -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
-                                           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,
-                                           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
-  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
+  --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,
+                                           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, 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
+  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
+                                           "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level),
-                                           'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET),
+                                           'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT
+                                           Chebyshev+Taylor forecasting)
   --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
-                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=;
-                                           spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=. Examples:
-                                           "threshold=0.25" or "threshold=1.5,reset=0" or "w=0.4,window=2"
-  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
+                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit:
+                                           Fn=,Bn=,threshold=,warmup=; spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=.
+                                           Examples: "threshold=0.25" or "threshold=1.5,reset=0"
+  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g.,
+                                           "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
   --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
+  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
+  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size
+                                           if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
+```
+
+Metadata mode inspects PNG/JPEG container metadata without loading any model:
+
+```bash
+./bin/sd-cli -M metadata --image ./output.png
+./bin/sd-cli -M metadata --image ./output.jpg --metadata-format json
+./bin/sd-cli -M metadata --image ./output.png --metadata-raw
+./bin/sd-cli -M metadata --image ./output.png --metadata-all
 ```

examples/cli/avi_writer.h

@@ -1,217 +0,0 @@
-#ifndef __AVI_WRITER_H__
-#define __AVI_WRITER_H__
-
-#include <cstdint>
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-
-#include "stable-diffusion.h"
-
-#ifndef INCLUDE_STB_IMAGE_WRITE_H
-#include "stb_image_write.h"
-#endif
-
-typedef struct {
-    uint32_t offset;
-    uint32_t size;
-} avi_index_entry;
-
-// Write 32-bit little-endian integer
-void write_u32_le(FILE* f, uint32_t val) {
-    fwrite(&val, 4, 1, f);
-}
-
-// Write 16-bit little-endian integer
-void write_u16_le(FILE* f, uint16_t val) {
-    fwrite(&val, 2, 1, f);
-}
-
-/**
- * Create an MJPG AVI file from an array of sd_image_t images.
- * Images are encoded to JPEG using stb_image_write.
- *
- * @param filename Output AVI file name.
- * @param images Array of input images.
- * @param num_images Number of images in the array.
- * @param fps Frames per second for the video.
- * @param quality JPEG quality (0-100).
- * @return 0 on success, -1 on failure.
- */
-int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int num_images, int fps, int quality = 90) {
-    if (num_images == 0) {
-        fprintf(stderr, "Error: Image array is empty.\n");
-        return -1;
-    }
-
-    FILE* f = fopen(filename, "wb");
-    if (!f) {
-        perror("Error opening file for writing");
-        return -1;
-    }
-
-    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);
-        fclose(f);
-        return -1;
-    }
-
-    // --- RIFF AVI Header ---
-    fwrite("RIFF", 4, 1, f);
-    long riff_size_pos = ftell(f);
-    write_u32_le(f, 0);  // Placeholder for file size
-    fwrite("AVI ", 4, 1, f);
-
-    // 'hdrl' LIST (header list)
-    fwrite("LIST", 4, 1, f);
-    write_u32_le(f, 4 + 8 + 56 + 8 + 4 + 8 + 56 + 8 + 40);
-    fwrite("hdrl", 4, 1, f);
-
-    // 'avih' chunk (AVI main header)
-    fwrite("avih", 4, 1, f);
-    write_u32_le(f, 56);
-    write_u32_le(f, 1000000 / fps);       // Microseconds per frame
-    write_u32_le(f, 0);                   // Max bytes per second
-    write_u32_le(f, 0);                   // Padding granularity
-    write_u32_le(f, 0x110);               // Flags (HASINDEX | ISINTERLEAVED)
-    write_u32_le(f, num_images);          // Total frames
-    write_u32_le(f, 0);                   // Initial frames
-    write_u32_le(f, 1);                   // Number of streams
-    write_u32_le(f, width * height * 3);  // Suggested buffer size
-    write_u32_le(f, width);
-    write_u32_le(f, height);
-    write_u32_le(f, 0);  // Reserved
-    write_u32_le(f, 0);  // Reserved
-    write_u32_le(f, 0);  // Reserved
-    write_u32_le(f, 0);  // Reserved
-
-    // 'strl' LIST (stream list)
-    fwrite("LIST", 4, 1, f);
-    write_u32_le(f, 4 + 8 + 56 + 8 + 40);
-    fwrite("strl", 4, 1, f);
-
-    // 'strh' chunk (stream header)
-    fwrite("strh", 4, 1, f);
-    write_u32_le(f, 56);
-    fwrite("vids", 4, 1, f);              // Stream type: video
-    fwrite("MJPG", 4, 1, f);              // Codec: Motion JPEG
-    write_u32_le(f, 0);                   // Flags
-    write_u16_le(f, 0);                   // Priority
-    write_u16_le(f, 0);                   // Language
-    write_u32_le(f, 0);                   // Initial frames
-    write_u32_le(f, 1);                   // Scale
-    write_u32_le(f, fps);                 // Rate
-    write_u32_le(f, 0);                   // Start
-    write_u32_le(f, num_images);          // Length
-    write_u32_le(f, width * height * 3);  // Suggested buffer size
-    write_u32_le(f, (uint32_t)-1);        // Quality
-    write_u32_le(f, 0);                   // Sample size
-    write_u16_le(f, 0);                   // rcFrame.left
-    write_u16_le(f, 0);                   // rcFrame.top
-    write_u16_le(f, 0);                   // rcFrame.right
-    write_u16_le(f, 0);                   // rcFrame.bottom
-
-    // 'strf' chunk (stream format: BITMAPINFOHEADER)
-    fwrite("strf", 4, 1, f);
-    write_u32_le(f, 40);
-    write_u32_le(f, 40);  // biSize
-    write_u32_le(f, width);
-    write_u32_le(f, height);
-    write_u16_le(f, 1);                   // biPlanes
-    write_u16_le(f, 24);                  // biBitCount
-    fwrite("MJPG", 4, 1, f);              // biCompression (FOURCC)
-    write_u32_le(f, width * height * 3);  // biSizeImage
-    write_u32_le(f, 0);                   // XPelsPerMeter
-    write_u32_le(f, 0);                   // YPelsPerMeter
-    write_u32_le(f, 0);                   // Colors used
-    write_u32_le(f, 0);                   // Colors important
-
-    // 'movi' LIST (video frames)
-    // long movi_list_pos = ftell(f);
-    fwrite("LIST", 4, 1, f);
-    long movi_size_pos = ftell(f);
-    write_u32_le(f, 0);  // Placeholder for movi size
-    fwrite("movi", 4, 1, f);
-
-    avi_index_entry* index = (avi_index_entry*)malloc(sizeof(avi_index_entry) * num_images);
-    if (!index) {
-        fclose(f);
-        return -1;
-    }
-
-    // Encode and write each frame as JPEG
-    struct {
-        uint8_t* buf;
-        size_t size;
-    } jpeg_data;
-
-    for (int i = 0; i < num_images; i++) {
-        jpeg_data.buf  = nullptr;
-        jpeg_data.size = 0;
-
-        // Callback function to collect JPEG data into memory
-        auto write_to_buf = [](void* context, void* data, int size) {
-            auto jd = (decltype(jpeg_data)*)context;
-            jd->buf = (uint8_t*)realloc(jd->buf, jd->size + size);
-            memcpy(jd->buf + jd->size, data, size);
-            jd->size += size;
-        };
-
-        // Encode to JPEG in memory
-        stbi_write_jpg_to_func(
-            write_to_buf,
-            &jpeg_data,
-            images[i].width,
-            images[i].height,
-            channels,
-            images[i].data,
-            quality);
-
-        // Write '00dc' chunk (video frame)
-        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.buf, 1, jpeg_data.size, f);
-
-        // Align to even byte size
-        if (jpeg_data.size % 2)
-            fputc(0, f);
-
-        free(jpeg_data.buf);
-    }
-
-    // Finalize 'movi' size
-    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);
-
-    // Write 'idx1' index
-    fwrite("idx1", 4, 1, f);
-    write_u32_le(f, 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);
-    }
-
-    // Finalize RIFF 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);
-
-    fclose(f);
-    free(index);
-
-    return 0;
-}
-
-#endif  // __AVI_WRITER_H__

examples/cli/image_metadata.h

@@ -0,0 +1,21 @@
+#pragma once
+
+#include <iosfwd>
+#include <string>
+
+enum class MetadataOutputFormat {
+    TEXT,
+    JSON,
+};
+
+struct MetadataReadOptions {
+    MetadataOutputFormat output_format = MetadataOutputFormat::TEXT;
+    bool include_raw                   = false;
+    bool brief                         = false;
+    bool include_structural            = false;
+};
+
+bool print_image_metadata(const std::string& image_path,
+                          const MetadataReadOptions& options,
+                          std::ostream& out,
+                          std::string& error);

examples/cli/main.cpp

@@ -15,9 +15,12 @@
 // #include "preprocessing.hpp"
 #include "stable-diffusion.h"
 
-#include "common/common.hpp"
+#include "common/common.h"
+#include "common/media_io.h"
+#include "common/resource_owners.hpp"
+#include "image_metadata.h"
 
-#include "avi_writer.h"
+namespace fs = std::filesystem;
 
 const char* previews_str[] = {
     "none",
@@ -32,6 +35,8 @@ struct SDCliParams {
     SDMode mode             = IMG_GEN;
     std::string output_path = "output.png";
     int output_begin_idx    = -1;
+    std::string image_path;
+    std::string metadata_format = "text";
 
     bool verbose          = false;
     bool canny_preprocess = false;
@@ -44,6 +49,9 @@ struct SDCliParams {
     bool taesd_preview       = false;
     bool preview_noisy       = false;
     bool color               = false;
+    bool metadata_raw        = false;
+    bool metadata_brief      = false;
+    bool metadata_all        = false;
 
     bool normal_exit = false;
 
@@ -53,11 +61,19 @@ struct SDCliParams {
         options.string_options = {
             {"-o",
              "--output",
-             "path to write result image to. you can use printf-style %d format specifiers for image sequences (default: ./output.png) (eg. output_%03d.png)",
+             "path to write result image to. you can use printf-style %d format specifiers for image sequences (default: ./output.png) (eg. output_%03d.png). Single-file video outputs support .avi, .webm, and animated .webp",
              &output_path},
+            {"",
+             "--image",
+             "path to the image to inspect (for metadata mode)",
+             &image_path},
+            {"",
+             "--metadata-format",
+             "metadata output format, one of [text, json] (default: text)",
+             &metadata_format},
             {"",
              "--preview-path",
-             "path to write preview image to (default: ./preview.png)",
+             "path to write preview image to (default: ./preview.png). Multi-frame previews support .avi, .webm, and animated .webp",
              &preview_path},
         };
 
@@ -97,6 +113,18 @@ struct SDCliParams {
              "--preview-noisy",
              "enables previewing noisy inputs of the models rather than the denoised outputs",
              true, &preview_noisy},
+            {"",
+             "--metadata-raw",
+             "include raw hex previews for unparsed metadata payloads",
+             true, &metadata_raw},
+            {"",
+             "--metadata-brief",
+             "truncate long metadata text values in text output",
+             true, &metadata_brief},
+            {"",
+             "--metadata-all",
+             "include structural/container entries such as IHDR, IDAT, and non-metadata JPEG segments",
+             true, &metadata_all},
 
         };
 
@@ -149,7 +177,7 @@ struct SDCliParams {
         options.manual_options = {
             {"-M",
              "--mode",
-             "run mode, one of [img_gen, vid_gen, upscale, convert], default: img_gen",
+             "run mode, one of [img_gen, vid_gen, upscale, convert, metadata], default: img_gen",
              on_mode_arg},
             {"",
              "--preview",
@@ -164,12 +192,7 @@ struct SDCliParams {
         return options;
     };
 
-    bool process_and_check() {
-        if (output_path.length() == 0) {
-            LOG_ERROR("error: the following arguments are required: output_path");
-            return false;
-        }
-
+    bool resolve() {
         if (mode == CONVERT) {
             if (output_path == "output.png") {
                 output_path = "output.gguf";
@@ -178,11 +201,43 @@ struct SDCliParams {
         return true;
     }
 
+    bool validate() {
+        if (mode != METADATA) {
+            if (output_path.length() == 0) {
+                LOG_ERROR("error: the following arguments are required: output_path");
+                return false;
+            }
+        } else {
+            if (image_path.empty()) {
+                LOG_ERROR("error: metadata mode needs an image path (--image)");
+                return false;
+            }
+            if (metadata_format != "text" && metadata_format != "json") {
+                LOG_ERROR("error: invalid metadata format %s, must be one of [text, json]",
+                          metadata_format.c_str());
+                return false;
+            }
+        }
+        return true;
+    }
+
+    bool resolve_and_validate() {
+        if (!resolve()) {
+            return false;
+        }
+        if (!validate()) {
+            return false;
+        }
+        return true;
+    }
+
     std::string to_string() const {
         std::ostringstream oss;
         oss << "SDCliParams {\n"
             << "  mode: " << modes_str[mode] << ",\n"
             << "  output_path: \"" << output_path << "\",\n"
+            << "  image_path: \"" << image_path << "\",\n"
+            << "  metadata_format: \"" << metadata_format << "\",\n"
             << "  verbose: " << (verbose ? "true" : "false") << ",\n"
             << "  color: " << (color ? "true" : "false") << ",\n"
             << "  canny_preprocess: " << (canny_preprocess ? "true" : "false") << ",\n"
@@ -192,7 +247,10 @@ struct SDCliParams {
             << "  preview_path: \"" << preview_path << "\",\n"
             << "  preview_fps: " << preview_fps << ",\n"
             << "  taesd_preview: " << (taesd_preview ? "true" : "false") << ",\n"
-            << "  preview_noisy: " << (preview_noisy ? "true" : "false") << "\n"
+            << "  preview_noisy: " << (preview_noisy ? "true" : "false") << ",\n"
+            << "  metadata_raw: " << (metadata_raw ? "true" : "false") << ",\n"
+            << "  metadata_brief: " << (metadata_brief ? "true" : "false") << ",\n"
+            << "  metadata_all: " << (metadata_all ? "true" : "false") << "\n"
             << "}";
         return oss.str();
     }
@@ -217,78 +275,27 @@ void parse_args(int argc, const char** argv, SDCliParams& cli_params, SDContextP
         exit(cli_params.normal_exit ? 0 : 1);
     }
 
-    if (!cli_params.process_and_check() ||
-        !ctx_params.process_and_check(cli_params.mode) ||
-        !gen_params.process_and_check(cli_params.mode, ctx_params.lora_model_dir)) {
+    bool valid = cli_params.resolve_and_validate();
+    if (valid && cli_params.mode != METADATA) {
+        valid = ctx_params.resolve_and_validate(cli_params.mode) &&
+                gen_params.resolve_and_validate(cli_params.mode,
+                                                ctx_params.lora_model_dir,
+                                                ctx_params.hires_upscalers_dir);
+    }
+
+    if (!valid) {
         print_usage(argc, argv, options_vec);
         exit(1);
     }
 }
 
-std::string get_image_params(const SDCliParams& cli_params, const SDContextParams& ctx_params, const SDGenerationParams& gen_params, int64_t seed) {
-    std::string parameter_string = gen_params.prompt_with_lora + "\n";
-    if (gen_params.negative_prompt.size() != 0) {
-        parameter_string += "Negative prompt: " + gen_params.negative_prompt + "\n";
-    }
-    parameter_string += "Steps: " + std::to_string(gen_params.sample_params.sample_steps) + ", ";
-    parameter_string += "CFG scale: " + std::to_string(gen_params.sample_params.guidance.txt_cfg) + ", ";
-    if (gen_params.sample_params.guidance.slg.scale != 0 && gen_params.skip_layers.size() != 0) {
-        parameter_string += "SLG scale: " + std::to_string(gen_params.sample_params.guidance.txt_cfg) + ", ";
-        parameter_string += "Skip layers: [";
-        for (const auto& layer : gen_params.skip_layers) {
-            parameter_string += std::to_string(layer) + ", ";
-        }
-        parameter_string += "], ";
-        parameter_string += "Skip layer start: " + std::to_string(gen_params.sample_params.guidance.slg.layer_start) + ", ";
-        parameter_string += "Skip layer end: " + std::to_string(gen_params.sample_params.guidance.slg.layer_end) + ", ";
-    }
-    parameter_string += "Guidance: " + std::to_string(gen_params.sample_params.guidance.distilled_guidance) + ", ";
-    parameter_string += "Eta: " + std::to_string(gen_params.sample_params.eta) + ", ";
-    parameter_string += "Seed: " + std::to_string(seed) + ", ";
-    parameter_string += "Size: " + std::to_string(gen_params.get_resolved_width()) + "x" + std::to_string(gen_params.get_resolved_height()) + ", ";
-    parameter_string += "Model: " + sd_basename(ctx_params.model_path) + ", ";
-    parameter_string += "RNG: " + std::string(sd_rng_type_name(ctx_params.rng_type)) + ", ";
-    if (ctx_params.sampler_rng_type != RNG_TYPE_COUNT) {
-        parameter_string += "Sampler RNG: " + std::string(sd_rng_type_name(ctx_params.sampler_rng_type)) + ", ";
-    }
-    parameter_string += "Sampler: " + std::string(sd_sample_method_name(gen_params.sample_params.sample_method));
-    if (!gen_params.custom_sigmas.empty()) {
-        parameter_string += ", Custom Sigmas: [";
-        for (size_t i = 0; i < gen_params.custom_sigmas.size(); ++i) {
-            std::ostringstream oss;
-            oss << std::fixed << std::setprecision(4) << gen_params.custom_sigmas[i];
-            parameter_string += oss.str() + (i == gen_params.custom_sigmas.size() - 1 ? "" : ", ");
-        }
-        parameter_string += "]";
-    } else if (gen_params.sample_params.scheduler != SCHEDULER_COUNT) {  // Only show schedule if not using custom sigmas
-        parameter_string += " " + std::string(sd_scheduler_name(gen_params.sample_params.scheduler));
-    }
-    parameter_string += ", ";
-    for (const auto& te : {ctx_params.clip_l_path, ctx_params.clip_g_path, ctx_params.t5xxl_path, ctx_params.llm_path, ctx_params.llm_vision_path}) {
-        if (!te.empty()) {
-            parameter_string += "TE: " + sd_basename(te) + ", ";
-        }
-    }
-    if (!ctx_params.diffusion_model_path.empty()) {
-        parameter_string += "Unet: " + sd_basename(ctx_params.diffusion_model_path) + ", ";
-    }
-    if (!ctx_params.vae_path.empty()) {
-        parameter_string += "VAE: " + sd_basename(ctx_params.vae_path) + ", ";
-    }
-    if (gen_params.clip_skip != -1) {
-        parameter_string += "Clip skip: " + std::to_string(gen_params.clip_skip) + ", ";
-    }
-    parameter_string += "Version: stable-diffusion.cpp";
-    return parameter_string;
-}
-
 void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
     SDCliParams* cli_params = (SDCliParams*)data;
     log_print(level, log, cli_params->verbose, cli_params->color);
 }
 
 bool load_images_from_dir(const std::string dir,
-                          std::vector<sd_image_t>& images,
+                          std::vector<SDImageOwner>& images,
                           int expected_width  = 0,
                           int expected_height = 0,
                           int max_image_num   = 0,
@@ -315,7 +322,7 @@ bool load_images_from_dir(const std::string dir,
         std::string ext  = entry.path().extension().string();
         std::transform(ext.begin(), ext.end(), ext.begin(), ::tolower);
 
-        if (ext == ".jpg" || ext == ".jpeg" || ext == ".png" || ext == ".bmp") {
+        if (ext == ".jpg" || ext == ".jpeg" || ext == ".png" || ext == ".bmp" || ext == ".webp") {
             LOG_DEBUG("load image %zu from '%s'", images.size(), path.c_str());
             int width             = 0;
             int height            = 0;
@@ -325,12 +332,12 @@ bool load_images_from_dir(const std::string dir,
                 return false;
             }
 
-            images.push_back({(uint32_t)width,
-                              (uint32_t)height,
-                              3,
-                              image_buffer});
+            images.emplace_back(sd_image_t{(uint32_t)width,
+                                           (uint32_t)height,
+                                           3,
+                                           image_buffer});
 
-            if (max_image_num > 0 && images.size() >= max_image_num) {
+            if (max_image_num > 0 && static_cast<int>(images.size()) >= max_image_num) {
                 break;
             }
         }
@@ -345,9 +352,17 @@ void step_callback(int step, int frame_count, sd_image_t* image, bool is_noisy,
     // is_noisy is set to true if the preview corresponds to noisy latents, false if it's denoised latents
     // unused in this app, it will either be always noisy or always denoised here
     if (frame_count == 1) {
-        stbi_write_png(cli_params->preview_path.c_str(), image->width, image->height, image->channel, image->data, 0);
+        if (!write_image_to_file(cli_params->preview_path,
+                                 image->data,
+                                 image->width,
+                                 image->height,
+                                 image->channel)) {
+            LOG_ERROR("save preview image to '%s' failed", cli_params->preview_path.c_str());
+        }
     } else {
-        create_mjpg_avi_from_sd_images(cli_params->preview_path.c_str(), image, frame_count, cli_params->preview_fps);
+        if (create_video_from_sd_images(cli_params->preview_path.c_str(), image, frame_count, cli_params->preview_fps) != 0) {
+            LOG_ERROR("save preview video to '%s' failed", cli_params->preview_path.c_str());
+        }
     }
 }
 
@@ -397,9 +412,13 @@ bool save_results(const SDCliParams& cli_params,
 
     std::string ext_lower = ext.string();
     std::transform(ext_lower.begin(), ext_lower.end(), ext_lower.begin(), ::tolower);
-    bool is_jpg = (ext_lower == ".jpg" || ext_lower == ".jpeg" || ext_lower == ".jpe");
+    const EncodedImageFormat output_format = encoded_image_format_from_path(out_path.string());
     if (!ext.empty()) {
-        if (is_jpg || ext_lower == ".png") {
+        if (output_format == EncodedImageFormat::JPEG ||
+            output_format == EncodedImageFormat::PNG ||
+            output_format == EncodedImageFormat::WEBP ||
+            ext_lower == ".avi" ||
+            ext_lower == ".webm") {
             base_path.replace_extension();
         }
     }
@@ -414,21 +433,19 @@ bool save_results(const SDCliParams& cli_params,
         if (!img.data)
             return false;
 
-        std::string params = get_image_params(cli_params, ctx_params, gen_params, gen_params.seed + idx);
-        int ok             = 0;
-        if (is_jpg) {
-            ok = stbi_write_jpg(path.string().c_str(), img.width, img.height, img.channel, img.data, 90, params.c_str());
-        } else {
-            ok = stbi_write_png(path.string().c_str(), img.width, img.height, img.channel, img.data, 0, params.c_str());
-        }
+        const int64_t metadata_seed = cli_params.mode == VID_GEN ? gen_params.seed : gen_params.seed + idx;
+        std::string params          = gen_params.embed_image_metadata
+                                          ? get_image_params(ctx_params, gen_params, metadata_seed, cli_params.mode)
+                                          : "";
+        const bool ok               = write_image_to_file(path.string(), img.data, img.width, img.height, img.channel, params, 90);
         LOG_INFO("save result image %d to '%s' (%s)", idx, path.string().c_str(), ok ? "success" : "failure");
-        return ok != 0;
+        return ok;
     };
 
     int sucessful_reults = 0;
 
     if (std::regex_search(cli_params.output_path, format_specifier_regex)) {
-        if (!is_jpg && ext_lower != ".png")
+        if (output_format == EncodedImageFormat::UNKNOWN)
             ext = ".png";
         fs::path pattern = base_path;
         pattern += ext;
@@ -444,20 +461,20 @@ bool save_results(const SDCliParams& cli_params,
     }
 
     if (cli_params.mode == VID_GEN && num_results > 1) {
-        if (ext_lower != ".avi")
+        if (ext_lower != ".avi" && ext_lower != ".webp" && ext_lower != ".webm")
             ext = ".avi";
         fs::path video_path = base_path;
         video_path += ext;
-        if (create_mjpg_avi_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps) == 0) {
-            LOG_INFO("save result MJPG AVI video to '%s'", video_path.string().c_str());
+        if (create_video_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps) == 0) {
+            LOG_INFO("save result video to '%s'", video_path.string().c_str());
             return true;
         } else {
-            LOG_ERROR("Failed to save result MPG AVI video to '%s'", video_path.string().c_str());
+            LOG_ERROR("Failed to save result video to '%s'", video_path.string().c_str());
             return false;
         }
     }
 
-    if (!is_jpg && ext_lower != ".png")
+    if (output_format == EncodedImageFormat::UNKNOWN)
         ext = ".png";
 
     for (int i = 0; i < num_results; ++i) {
@@ -485,6 +502,27 @@ int main(int argc, const char* argv[]) {
     SDGenerationParams gen_params;
 
     parse_args(argc, argv, cli_params, ctx_params, gen_params);
+    sd_set_log_callback(sd_log_cb, (void*)&cli_params);
+    log_verbose = cli_params.verbose;
+    log_color   = cli_params.color;
+
+    if (cli_params.mode == METADATA) {
+        MetadataReadOptions options;
+        options.output_format      = cli_params.metadata_format == "json"
+                                         ? MetadataOutputFormat::JSON
+                                         : MetadataOutputFormat::TEXT;
+        options.include_raw        = cli_params.metadata_raw;
+        options.brief              = cli_params.metadata_brief;
+        options.include_structural = cli_params.metadata_all;
+
+        std::string error;
+        if (!print_image_metadata(cli_params.image_path, options, std::cout, error)) {
+            LOG_ERROR("%s", error.c_str());
+            return 1;
+        }
+        return 0;
+    }
+
     if (gen_params.video_frames > 4) {
         size_t last_dot_pos   = cli_params.preview_path.find_last_of(".");
         std::string base_path = cli_params.preview_path;
@@ -502,9 +540,6 @@ int main(int argc, const char* argv[]) {
     if (cli_params.preview_method == PREVIEW_PROJ)
         cli_params.preview_fps /= 4;
 
-    sd_set_log_callback(sd_log_cb, (void*)&cli_params);
-    log_verbose = cli_params.verbose;
-    log_color   = cli_params.color;
     sd_set_preview_callback(step_callback,
                             cli_params.preview_method,
                             cli_params.preview_interval,
@@ -540,39 +575,10 @@ int main(int argc, const char* argv[]) {
         }
     }
 
-    bool vae_decode_only     = true;
-    sd_image_t init_image    = {0, 0, 3, nullptr};
-    sd_image_t end_image     = {0, 0, 3, nullptr};
-    sd_image_t control_image = {0, 0, 3, nullptr};
-    sd_image_t mask_image    = {0, 0, 1, nullptr};
-    std::vector<sd_image_t> ref_images;
-    std::vector<sd_image_t> pmid_images;
-    std::vector<sd_image_t> control_frames;
-
-    auto release_all_resources = [&]() {
-        free(init_image.data);
-        free(end_image.data);
-        free(control_image.data);
-        free(mask_image.data);
-        for (auto image : ref_images) {
-            free(image.data);
-            image.data = nullptr;
-        }
-        ref_images.clear();
-        for (auto image : pmid_images) {
-            free(image.data);
-            image.data = nullptr;
-        }
-        pmid_images.clear();
-        for (auto image : control_frames) {
-            free(image.data);
-            image.data = nullptr;
-        }
-        control_frames.clear();
-    };
+    bool vae_decode_only = true;
 
     auto load_image_and_update_size = [&](const std::string& path,
-                                          sd_image_t& image,
+                                          SDImageOwner& image,
                                           bool resize_image    = true,
                                           int expected_channel = 3) -> bool {
         int expected_width  = 0;
@@ -582,74 +588,73 @@ int main(int argc, const char* argv[]) {
             expected_height = gen_params.height;
         }
 
-        if (!load_sd_image_from_file(&image, path.c_str(), expected_width, expected_height, expected_channel)) {
+        if (!load_sd_image_from_file(image.put(), path.c_str(), expected_width, expected_height, expected_channel)) {
             LOG_ERROR("load image from '%s' failed", path.c_str());
-            release_all_resources();
             return false;
         }
 
-        gen_params.set_width_and_height_if_unset(image.width, image.height);
+        gen_params.set_width_and_height_if_unset(image.get().width, image.get().height);
         return true;
     };
 
     if (gen_params.init_image_path.size() > 0) {
         vae_decode_only = false;
-        if (!load_image_and_update_size(gen_params.init_image_path, init_image)) {
+        if (!load_image_and_update_size(gen_params.init_image_path, gen_params.init_image)) {
             return 1;
         }
     }
 
     if (gen_params.end_image_path.size() > 0) {
         vae_decode_only = false;
-        if (!load_image_and_update_size(gen_params.init_image_path, end_image)) {
+        if (!load_image_and_update_size(gen_params.end_image_path, gen_params.end_image)) {
             return 1;
         }
     }
 
     if (gen_params.ref_image_paths.size() > 0) {
         vae_decode_only = false;
+        gen_params.ref_images.clear();
         for (auto& path : gen_params.ref_image_paths) {
-            sd_image_t ref_image = {0, 0, 3, nullptr};
+            SDImageOwner ref_image({0, 0, 3, nullptr});
             if (!load_image_and_update_size(path, ref_image, false)) {
                 return 1;
             }
-            ref_images.push_back(ref_image);
+            gen_params.ref_images.push_back(std::move(ref_image));
         }
     }
 
     if (gen_params.mask_image_path.size() > 0) {
-        if (!load_sd_image_from_file(&mask_image,
+        if (!load_sd_image_from_file(gen_params.mask_image.put(),
                                      gen_params.mask_image_path.c_str(),
                                      gen_params.get_resolved_width(),
                                      gen_params.get_resolved_height(),
                                      1)) {
             LOG_ERROR("load image from '%s' failed", gen_params.mask_image_path.c_str());
-            release_all_resources();
             return 1;
         }
     } else {
-        mask_image.data = (uint8_t*)malloc(gen_params.get_resolved_width() * gen_params.get_resolved_height());
-        if (mask_image.data == nullptr) {
+        sd_image_t generated_mask = {0, 0, 1, nullptr};
+        generated_mask.data       = (uint8_t*)malloc(gen_params.get_resolved_width() * gen_params.get_resolved_height());
+        if (generated_mask.data == nullptr) {
             LOG_ERROR("malloc mask image failed");
-            release_all_resources();
             return 1;
         }
-        mask_image.width  = gen_params.get_resolved_width();
-        mask_image.height = gen_params.get_resolved_height();
-        memset(mask_image.data, 255, gen_params.get_resolved_width() * gen_params.get_resolved_height());
+        generated_mask.width  = gen_params.get_resolved_width();
+        generated_mask.height = gen_params.get_resolved_height();
+        memset(generated_mask.data, 255, gen_params.get_resolved_width() * gen_params.get_resolved_height());
+        gen_params.mask_image.reset(generated_mask);
     }
 
     if (gen_params.control_image_path.size() > 0) {
-        if (!load_sd_image_from_file(&control_image,
+        if (!load_sd_image_from_file(gen_params.control_image.put(),
                                      gen_params.control_image_path.c_str(),
                                      gen_params.get_resolved_width(),
                                      gen_params.get_resolved_height())) {
             LOG_ERROR("load image from '%s' failed", gen_params.control_image_path.c_str());
-            release_all_resources();
             return 1;
         }
         if (cli_params.canny_preprocess) {  // apply preprocessor
-            preprocess_canny(control_image,
+            preprocess_canny(gen_params.control_image.get(),
                              0.08f,
                              0.08f,
                              0.8f,
@@ -659,25 +664,25 @@ int main(int argc, const char* argv[]) {
     }
 
     if (!gen_params.control_video_path.empty()) {
+        gen_params.control_frames.clear();
         if (!load_images_from_dir(gen_params.control_video_path,
-                                  control_frames,
+                                  gen_params.control_frames,
                                   gen_params.get_resolved_width(),
                                   gen_params.get_resolved_height(),
                                   gen_params.video_frames,
                                   cli_params.verbose)) {
-            release_all_resources();
             return 1;
         }
     }
 
     if (!gen_params.pm_id_images_dir.empty()) {
+        gen_params.pm_id_images.clear();
         if (!load_images_from_dir(gen_params.pm_id_images_dir,
-                                  pmid_images,
+                                  gen_params.pm_id_images,
                                   0,
                                   0,
                                   0,
                                   cli_params.verbose)) {
-            release_all_resources();
             return 1;
         }
     }
@@ -686,119 +691,65 @@ int main(int argc, const char* argv[]) {
         vae_decode_only = false;
     }
 
+    if (gen_params.hires_enabled &&
+        (gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_MODEL ||
+         gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_LANCZOS ||
+         gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_NEAREST)) {
+        vae_decode_only = false;
+    }
+
     sd_ctx_params_t sd_ctx_params = ctx_params.to_sd_ctx_params_t(vae_decode_only, true, cli_params.taesd_preview);
 
-    sd_image_t* results = nullptr;
-    int num_results     = 0;
+    SDImageVec results;
+    int num_results = 0;
 
     if (cli_params.mode == UPSCALE) {
         num_results = 1;
-        results     = (sd_image_t*)calloc(num_results, sizeof(sd_image_t));
-        if (results == nullptr) {
-            LOG_INFO("failed to allocate results array");
-            release_all_resources();
-            return 1;
-        }
-
-        results[0]      = init_image;
-        init_image.data = nullptr;
+        results.push_back(gen_params.init_image.release());
     } else {
-        sd_ctx_t* sd_ctx = new_sd_ctx(&sd_ctx_params);
+        SDCtxPtr sd_ctx(new_sd_ctx(&sd_ctx_params));
 
         if (sd_ctx == nullptr) {
             LOG_INFO("new_sd_ctx_t failed");
-            release_all_resources();
             return 1;
         }
 
         if (gen_params.sample_params.sample_method == SAMPLE_METHOD_COUNT) {
-            gen_params.sample_params.sample_method = sd_get_default_sample_method(sd_ctx);
+            gen_params.sample_params.sample_method = sd_get_default_sample_method(sd_ctx.get());
         }
 
         if (gen_params.high_noise_sample_params.sample_method == SAMPLE_METHOD_COUNT) {
-            gen_params.high_noise_sample_params.sample_method = sd_get_default_sample_method(sd_ctx);
+            gen_params.high_noise_sample_params.sample_method = sd_get_default_sample_method(sd_ctx.get());
         }
 
         if (gen_params.sample_params.scheduler == SCHEDULER_COUNT) {
-            gen_params.sample_params.scheduler = sd_get_default_scheduler(sd_ctx, gen_params.sample_params.sample_method);
+            gen_params.sample_params.scheduler = sd_get_default_scheduler(sd_ctx.get(), gen_params.sample_params.sample_method);
         }
 
         if (cli_params.mode == IMG_GEN) {
-            sd_img_gen_params_t img_gen_params = {
-                gen_params.lora_vec.data(),
-                static_cast<uint32_t>(gen_params.lora_vec.size()),
-                gen_params.prompt.c_str(),
-                gen_params.negative_prompt.c_str(),
-                gen_params.clip_skip,
-                init_image,
-                ref_images.data(),
-                (int)ref_images.size(),
-                gen_params.auto_resize_ref_image,
-                gen_params.increase_ref_index,
-                mask_image,
-                gen_params.get_resolved_width(),
-                gen_params.get_resolved_height(),
-                gen_params.sample_params,
-                gen_params.strength,
-                gen_params.seed,
-                gen_params.batch_count,
-                control_image,
-                gen_params.control_strength,
-                {
-                    pmid_images.data(),
-                    (int)pmid_images.size(),
-                    gen_params.pm_id_embed_path.c_str(),
-                    gen_params.pm_style_strength,
-                },  // pm_params
-                ctx_params.vae_tiling_params,
-                gen_params.cache_params,
-            };
+            sd_img_gen_params_t img_gen_params = gen_params.to_sd_img_gen_params_t();
 
-            results     = generate_image(sd_ctx, &img_gen_params);
             num_results = gen_params.batch_count;
+            results.adopt(generate_image(sd_ctx.get(), &img_gen_params), num_results);
         } else if (cli_params.mode == VID_GEN) {
-            sd_vid_gen_params_t vid_gen_params = {
-                gen_params.lora_vec.data(),
-                static_cast<uint32_t>(gen_params.lora_vec.size()),
-                gen_params.prompt.c_str(),
-                gen_params.negative_prompt.c_str(),
-                gen_params.clip_skip,
-                init_image,
-                end_image,
-                control_frames.data(),
-                (int)control_frames.size(),
-                gen_params.get_resolved_width(),
-                gen_params.get_resolved_height(),
-                gen_params.sample_params,
-                gen_params.high_noise_sample_params,
-                gen_params.moe_boundary,
-                gen_params.strength,
-                gen_params.seed,
-                gen_params.video_frames,
-                gen_params.vace_strength,
-                ctx_params.vae_tiling_params,
-                gen_params.cache_params,
-            };
-
-            results = generate_video(sd_ctx, &vid_gen_params, &num_results);
+            sd_vid_gen_params_t vid_gen_params = gen_params.to_sd_vid_gen_params_t();
+            sd_image_t* generated_video        = generate_video(sd_ctx.get(), &vid_gen_params, &num_results);
+            results.adopt(generated_video, num_results);
         }
 
-        if (results == nullptr) {
+        if (!results) {
             LOG_ERROR("generate failed");
-            free_sd_ctx(sd_ctx);
             return 1;
         }
-
-        free_sd_ctx(sd_ctx);
     }
 
     int upscale_factor = 4;  // unused for RealESRGAN_x4plus_anime_6B.pth
     if (ctx_params.esrgan_path.size() > 0 && gen_params.upscale_repeats > 0) {
-        upscaler_ctx_t* upscaler_ctx = new_upscaler_ctx(ctx_params.esrgan_path.c_str(),
-                                                        ctx_params.offload_params_to_cpu,
-                                                        ctx_params.diffusion_conv_direct,
-                                                        ctx_params.n_threads,
-                                                        gen_params.upscale_tile_size);
+        UpscalerCtxPtr upscaler_ctx(new_upscaler_ctx(ctx_params.esrgan_path.c_str(),
+                                                     ctx_params.offload_params_to_cpu,
+                                                     ctx_params.diffusion_conv_direct,
+                                                     ctx_params.n_threads,
+                                                     gen_params.upscale_tile_size));
 
         if (upscaler_ctx == nullptr) {
             LOG_ERROR("new_upscaler_ctx failed");
@@ -807,32 +758,24 @@ int main(int argc, const char* argv[]) {
                 if (results[i].data == nullptr) {
                     continue;
                 }
-                sd_image_t current_image = results[i];
+                SDImageOwner current_image(results[i]);
+                results[i] = {0, 0, 0, nullptr};
                 for (int u = 0; u < gen_params.upscale_repeats; ++u) {
-                    sd_image_t upscaled_image = upscale(upscaler_ctx, current_image, upscale_factor);
-                    if (upscaled_image.data == nullptr) {
+                    SDImageOwner upscaled_image(upscale(upscaler_ctx.get(), current_image.get(), upscale_factor));
+                    if (upscaled_image.get().data == nullptr) {
                         LOG_ERROR("upscale failed");
                         break;
                     }
-                    free(current_image.data);
-                    current_image = upscaled_image;
+                    current_image = std::move(upscaled_image);
                 }
-                results[i] = current_image;  // Set the final upscaled image as the result
+                results[i] = current_image.release();  // Set the final upscaled image as the result
             }
         }
     }
 
-    if (!save_results(cli_params, ctx_params, gen_params, results, num_results)) {
+    if (!save_results(cli_params, ctx_params, gen_params, results.data(), num_results)) {
         return 1;
     }
 
-    for (int i = 0; i < num_results; i++) {
-        free(results[i].data);
-        results[i].data = nullptr;
-    }
-    free(results);
-
-    release_all_resources();
-
     return 0;
 }

examples/common/common.h

@@ -0,0 +1,262 @@
+#ifndef __EXAMPLES_COMMON_COMMON_H__
+#define __EXAMPLES_COMMON_COMMON_H__
+
+#include <cmath>
+#include <cstdint>
+#include <functional>
+#include <map>
+#include <string>
+#include <vector>
+
+#include "log.h"
+#include "resource_owners.hpp"
+#include "stable-diffusion.h"
+
+#define SAFE_STR(s) ((s) ? (s) : "")
+#define BOOL_STR(b) ((b) ? "true" : "false")
+
+extern const char* const modes_str[];
+#define SD_ALL_MODES_STR "img_gen, vid_gen, convert, upscale, metadata"
+
+enum SDMode {
+    IMG_GEN,
+    VID_GEN,
+    CONVERT,
+    UPSCALE,
+    METADATA,
+    MODE_COUNT
+};
+
+struct StringOption {
+    std::string short_name;
+    std::string long_name;
+    std::string desc;
+    std::string* target;
+};
+
+struct IntOption {
+    std::string short_name;
+    std::string long_name;
+    std::string desc;
+    int* target;
+};
+
+struct FloatOption {
+    std::string short_name;
+    std::string long_name;
+    std::string desc;
+    float* target;
+};
+
+struct BoolOption {
+    std::string short_name;
+    std::string long_name;
+    std::string desc;
+    bool keep_true;
+    bool* target;
+};
+
+struct ManualOption {
+    std::string short_name;
+    std::string long_name;
+    std::string desc;
+    std::function<int(int argc, const char** argv, int index)> cb;
+};
+
+struct ArgOptions {
+    std::vector<StringOption> string_options;
+    std::vector<IntOption> int_options;
+    std::vector<FloatOption> float_options;
+    std::vector<BoolOption> bool_options;
+    std::vector<ManualOption> manual_options;
+
+    static std::string wrap_text(const std::string& text, size_t width, size_t indent);
+    void print() const;
+};
+
+bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& options_list);
+bool decode_base64_image(const std::string& encoded_input,
+                         int target_channels,
+                         int expected_width,
+                         int expected_height,
+                         SDImageOwner& out_image);
+
+struct SDContextParams {
+    int n_threads = -1;
+    std::string model_path;
+    std::string clip_l_path;
+    std::string clip_g_path;
+    std::string clip_vision_path;
+    std::string t5xxl_path;
+    std::string llm_path;
+    std::string llm_vision_path;
+    std::string diffusion_model_path;
+    std::string high_noise_diffusion_model_path;
+    std::string vae_path;
+    std::string taesd_path;
+    std::string esrgan_path;
+    std::string control_net_path;
+    std::string embedding_dir;
+    std::string photo_maker_path;
+    sd_type_t wtype = SD_TYPE_COUNT;
+    std::string tensor_type_rules;
+    std::string lora_model_dir = ".";
+    std::string hires_upscalers_dir;
+
+    std::map<std::string, std::string> embedding_map;
+    std::vector<sd_embedding_t> embedding_vec;
+
+    rng_type_t rng_type         = CUDA_RNG;
+    rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
+    bool offload_params_to_cpu  = false;
+    float max_vram              = 0.f;
+    bool enable_mmap            = false;
+    bool control_net_cpu        = false;
+    bool clip_on_cpu            = false;
+    bool vae_on_cpu             = false;
+    bool flash_attn             = false;
+    bool diffusion_flash_attn   = false;
+    bool diffusion_conv_direct  = false;
+    bool vae_conv_direct        = false;
+
+    bool circular   = false;
+    bool circular_x = false;
+    bool circular_y = false;
+
+    bool chroma_use_dit_mask = true;
+    bool chroma_use_t5_mask  = false;
+    int chroma_t5_mask_pad   = 1;
+
+    bool qwen_image_zero_cond_t = false;
+
+    prediction_t prediction           = PREDICTION_COUNT;
+    lora_apply_mode_t lora_apply_mode = LORA_APPLY_AUTO;
+
+    bool force_sdxl_vae_conv_scale = false;
+
+    float flow_shift = INFINITY;
+    ArgOptions get_options();
+    void build_embedding_map();
+    bool resolve(SDMode mode);
+    bool validate(SDMode mode);
+    bool resolve_and_validate(SDMode mode);
+    std::string to_string() const;
+    sd_ctx_params_t to_sd_ctx_params_t(bool vae_decode_only, bool free_params_immediately, bool taesd_preview);
+};
+
+struct SDGenerationParams {
+    // User-facing input fields.
+    std::string prompt;
+    std::string negative_prompt;
+    int clip_skip              = -1;  // <= 0 represents unspecified
+    int width                  = -1;
+    int height                 = -1;
+    int batch_count            = 1;
+    int64_t seed               = 42;
+    float strength             = 0.75f;
+    float control_strength     = 0.9f;
+    bool auto_resize_ref_image = true;
+    bool increase_ref_index    = false;
+    bool embed_image_metadata  = true;
+
+    std::string init_image_path;
+    std::string end_image_path;
+    std::string mask_image_path;
+    std::string control_image_path;
+    std::vector<std::string> ref_image_paths;
+    std::string control_video_path;
+
+    sd_sample_params_t sample_params;
+    sd_sample_params_t high_noise_sample_params;
+    std::vector<int> skip_layers            = {7, 8, 9};
+    std::vector<int> high_noise_skip_layers = {7, 8, 9};
+
+    std::vector<float> custom_sigmas;
+
+    std::string cache_mode;
+    std::string cache_option;
+    std::string scm_mask;
+    bool scm_policy_dynamic = true;
+    sd_cache_params_t cache_params{};
+
+    float moe_boundary                   = 0.875f;
+    int video_frames                     = 1;
+    int fps                              = 16;
+    float vace_strength                  = 1.f;
+    sd_tiling_params_t vae_tiling_params = {false, 0, 0, 0.5f, 0.0f, 0.0f};
+
+    std::string pm_id_images_dir;
+    std::string pm_id_embed_path;
+    float pm_style_strength = 20.f;
+
+    int upscale_repeats   = 1;
+    int upscale_tile_size = 128;
+
+    bool hires_enabled         = false;
+    std::string hires_upscaler = "Latent";
+    std::string hires_upscaler_model_path;
+    float hires_scale              = 2.f;
+    int hires_width                = 0;
+    int hires_height               = 0;
+    int hires_steps                = 0;
+    float hires_denoising_strength = 0.7f;
+    int hires_upscale_tile_size    = 128;
+
+    std::map<std::string, float> lora_map;
+    std::map<std::string, float> high_noise_lora_map;
+
+    // Derived and normalized fields.
+    std::string prompt_with_lora;  // for metadata record only
+    std::vector<sd_lora_t> lora_vec;
+    sd_hires_upscaler_t resolved_hires_upscaler;
+
+    // Owned execution payload.
+    SDImageOwner init_image;
+    SDImageOwner end_image;
+    std::vector<SDImageOwner> ref_images;
+    SDImageOwner mask_image;
+    SDImageOwner control_image;
+    std::vector<SDImageOwner> pm_id_images;
+    std::vector<SDImageOwner> control_frames;
+
+    // Backing storage for sd_img_gen_params_t view fields.
+    std::vector<sd_image_t> ref_image_views;
+    std::vector<sd_image_t> pm_id_image_views;
+    std::vector<sd_image_t> control_frame_views;
+
+    SDGenerationParams();
+    SDGenerationParams(const SDGenerationParams& other)                = default;
+    SDGenerationParams& operator=(const SDGenerationParams& other)     = default;
+    SDGenerationParams(SDGenerationParams&& other) noexcept            = default;
+    SDGenerationParams& operator=(SDGenerationParams&& other) noexcept = default;
+    ArgOptions get_options();
+    bool from_json_str(const std::string& json_str,
+                       const std::function<std::string(const std::string&)>& lora_path_resolver = {});
+    bool initialize_cache_params();
+    void extract_and_remove_lora(const std::string& lora_model_dir);
+    bool width_and_height_are_set() const;
+    void set_width_and_height_if_unset(int w, int h);
+    int get_resolved_width() const;
+    int get_resolved_height() const;
+    bool resolve(const std::string& lora_model_dir, const std::string& hires_upscalers_dir, bool strict = false);
+    bool validate(SDMode mode);
+    bool resolve_and_validate(SDMode mode,
+                              const std::string& lora_model_dir,
+                              const std::string& hires_upscalers_dir,
+                              bool strict = false);
+    sd_img_gen_params_t to_sd_img_gen_params_t();
+    sd_vid_gen_params_t to_sd_vid_gen_params_t();
+    std::string to_string() const;
+};
+
+std::string version_string();
+std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
+                                            const SDGenerationParams& gen_params,
+                                            int64_t seed,
+                                            SDMode mode = IMG_GEN);
+std::string get_image_params(const SDContextParams& ctx_params,
+                             const SDGenerationParams& gen_params,
+                             int64_t seed,
+                             SDMode mode = IMG_GEN);
+
+#endif  // __EXAMPLES_COMMON_COMMON_H__

examples/common/log.cpp

@@ -0,0 +1,115 @@
+#include "log.h"
+
+#include <vector>
+
+bool log_verbose = false;
+bool log_color   = false;
+
+std::string sd_basename(const std::string& path) {
+    size_t pos = path.find_last_of('/');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    pos = path.find_last_of('\\');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    return path;
+}
+
+void print_utf8(FILE* stream, const char* utf8) {
+    if (!utf8) {
+        return;
+    }
+
+#ifdef _WIN32
+    HANDLE h = (stream == stderr)
+                   ? GetStdHandle(STD_ERROR_HANDLE)
+                   : GetStdHandle(STD_OUTPUT_HANDLE);
+
+    DWORD mode;
+    BOOL is_console = GetConsoleMode(h, &mode);
+
+    if (is_console) {
+        int wlen = MultiByteToWideChar(CP_UTF8, 0, utf8, -1, NULL, 0);
+        if (wlen <= 0) {
+            return;
+        }
+
+        std::vector<wchar_t> wbuf(static_cast<size_t>(wlen));
+
+        MultiByteToWideChar(CP_UTF8, 0, utf8, -1, wbuf.data(), wlen);
+
+        DWORD written;
+        WriteConsoleW(h, wbuf.data(), wlen - 1, &written, NULL);
+    } else {
+        DWORD written;
+        WriteFile(h, utf8, (DWORD)strlen(utf8), &written, NULL);
+    }
+#else
+    fputs(utf8, stream);
+#endif
+}
+
+void log_print(enum sd_log_level_t level, const char* log, bool verbose, bool color) {
+    int tag_color;
+    const char* level_str;
+    FILE* out_stream = (level == SD_LOG_ERROR) ? stderr : stdout;
+
+    if (!log || (!verbose && level <= SD_LOG_DEBUG)) {
+        return;
+    }
+
+    switch (level) {
+        case SD_LOG_DEBUG:
+            tag_color = 37;
+            level_str = "DEBUG";
+            break;
+        case SD_LOG_INFO:
+            tag_color = 34;
+            level_str = "INFO";
+            break;
+        case SD_LOG_WARN:
+            tag_color = 35;
+            level_str = "WARN";
+            break;
+        case SD_LOG_ERROR:
+            tag_color = 31;
+            level_str = "ERROR";
+            break;
+        default:
+            tag_color = 33;
+            level_str = "?????";
+            break;
+    }
+
+    if (color) {
+        fprintf(out_stream, "\033[%d;1m[%-5s]\033[0m ", tag_color, level_str);
+    } else {
+        fprintf(out_stream, "[%-5s] ", level_str);
+    }
+    print_utf8(out_stream, log);
+    fflush(out_stream);
+}
+
+void example_log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...) {
+    constexpr size_t LOG_BUFFER_SIZE = 4096;
+
+    va_list args;
+    va_start(args, format);
+
+    static char log_buffer[LOG_BUFFER_SIZE + 1];
+    int written = snprintf(log_buffer, LOG_BUFFER_SIZE, "%s:%-4d - ", sd_basename(file).c_str(), line);
+
+    if (written >= 0 && written < static_cast<int>(LOG_BUFFER_SIZE)) {
+        vsnprintf(log_buffer + written, LOG_BUFFER_SIZE - written, format, args);
+    }
+    size_t len = strlen(log_buffer);
+    if (len == 0 || log_buffer[len - 1] != '\n') {
+        strncat(log_buffer, "\n", LOG_BUFFER_SIZE - len);
+    }
+
+    log_print(level, log_buffer, log_verbose, log_color);
+
+    va_end(args);
+}

examples/common/log.h

@@ -0,0 +1,32 @@
+#ifndef __EXAMPLE_LOG_H__
+#define __EXAMPLE_LOG_H__
+
+#include <cstdarg>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <string>
+
+#if defined(_WIN32)
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#endif  // _WIN32
+
+#include "stable-diffusion.h"
+
+extern bool log_verbose;
+extern bool log_color;
+
+std::string sd_basename(const std::string& path);
+void print_utf8(FILE* stream, const char* utf8);
+void log_print(sd_log_level_t level, const char* log, bool verbose, bool color);
+void example_log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...);
+
+#define LOG_DEBUG(format, ...) example_log_printf(SD_LOG_DEBUG, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_INFO(format, ...) example_log_printf(SD_LOG_INFO, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_WARN(format, ...) example_log_printf(SD_LOG_WARN, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_ERROR(format, ...) example_log_printf(SD_LOG_ERROR, __FILE__, __LINE__, format, ##__VA_ARGS__)
+
+#endif  // __EXAMPLE_LOG_H__

examples/common/media_io.h

@@ -0,0 +1,101 @@
+#ifndef __MEDIA_IO_H__
+#define __MEDIA_IO_H__
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "stable-diffusion.h"
+
+enum class EncodedImageFormat {
+    JPEG,
+    PNG,
+    WEBP,
+    UNKNOWN,
+};
+
+EncodedImageFormat encoded_image_format_from_path(const std::string& path);
+
+std::vector<uint8_t> encode_image_to_vector(EncodedImageFormat format,
+                                            const uint8_t* image,
+                                            int width,
+                                            int height,
+                                            int channels,
+                                            const std::string& parameters = "",
+                                            int quality                   = 90);
+
+bool write_image_to_file(const std::string& path,
+                         const uint8_t* image,
+                         int width,
+                         int height,
+                         int channels,
+                         const std::string& parameters = "",
+                         int quality                   = 90);
+
+uint8_t* load_image_from_file(const char* image_path,
+                              int& width,
+                              int& height,
+                              int expected_width   = 0,
+                              int expected_height  = 0,
+                              int expected_channel = 3);
+
+bool load_sd_image_from_file(sd_image_t* image,
+                             const char* image_path,
+                             int expected_width   = 0,
+                             int expected_height  = 0,
+                             int expected_channel = 3);
+
+uint8_t* load_image_from_memory(const char* image_bytes,
+                                int len,
+                                int& width,
+                                int& height,
+                                int expected_width   = 0,
+                                int expected_height  = 0,
+                                int expected_channel = 3);
+
+int create_mjpg_avi_from_sd_images(const char* filename,
+                                   sd_image_t* images,
+                                   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,
+                                        sd_image_t* images,
+                                        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
+int create_webm_from_sd_images(const char* filename,
+                               sd_image_t* images,
+                               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,
+                                sd_image_t* images,
+                                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/common/resource_owners.hpp

@@ -0,0 +1,236 @@
+#ifndef __EXAMPLE_RESOURCE_OWNERS_H__
+#define __EXAMPLE_RESOURCE_OWNERS_H__
+
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include "stable-diffusion.h"
+
+struct FreeDeleter {
+    void operator()(void* ptr) const {
+        free(ptr);
+    }
+};
+
+struct FileCloser {
+    void operator()(FILE* file) const {
+        if (file != nullptr) {
+            fclose(file);
+        }
+    }
+};
+
+struct SDCtxDeleter {
+    void operator()(sd_ctx_t* ctx) const {
+        if (ctx != nullptr) {
+            free_sd_ctx(ctx);
+        }
+    }
+};
+
+struct UpscalerCtxDeleter {
+    void operator()(upscaler_ctx_t* ctx) const {
+        if (ctx != nullptr) {
+            free_upscaler_ctx(ctx);
+        }
+    }
+};
+
+template <typename T>
+using FreeUniquePtr = std::unique_ptr<T, FreeDeleter>;
+
+using FilePtr        = std::unique_ptr<FILE, FileCloser>;
+using SDCtxPtr       = std::unique_ptr<sd_ctx_t, SDCtxDeleter>;
+using UpscalerCtxPtr = std::unique_ptr<upscaler_ctx_t, UpscalerCtxDeleter>;
+
+class SDImageOwner {
+private:
+    static sd_image_t copy_image(const sd_image_t& image) {
+        if (image.data == nullptr) {
+            return {image.width, image.height, image.channel, nullptr};
+        }
+
+        const size_t byte_count = static_cast<size_t>(image.width) * image.height * image.channel;
+        uint8_t* raw_copy       = static_cast<uint8_t*>(malloc(byte_count));
+        if (raw_copy == nullptr) {
+            return {0, 0, 0, nullptr};
+        }
+
+        std::memcpy(raw_copy, image.data, byte_count);
+        return {image.width, image.height, image.channel, raw_copy};
+    }
+
+    sd_image_t image_ = {0, 0, 0, nullptr};
+
+public:
+    SDImageOwner() = default;
+    explicit SDImageOwner(sd_image_t image)
+        : image_(image) {
+    }
+
+    SDImageOwner(const SDImageOwner& other)
+        : image_(copy_image(other.image_)) {
+    }
+
+    SDImageOwner& operator=(const SDImageOwner& other) {
+        if (this != &other) {
+            reset(copy_image(other.image_));
+        }
+        return *this;
+    }
+
+    SDImageOwner(SDImageOwner&& other) noexcept
+        : image_(other.release()) {
+    }
+
+    SDImageOwner& operator=(SDImageOwner&& other) noexcept {
+        if (this != &other) {
+            reset();
+            image_ = other.release();
+        }
+        return *this;
+    }
+
+    ~SDImageOwner() {
+        reset();
+    }
+
+    sd_image_t* put() {
+        if (image_.data != nullptr) {
+            free(image_.data);
+            image_.data = nullptr;
+        }
+        image_.width   = 0;
+        image_.height  = 0;
+        image_.channel = 0;
+        return &image_;
+    }
+
+    sd_image_t& get() {
+        return image_;
+    }
+
+    const sd_image_t& get() const {
+        return image_;
+    }
+
+    sd_image_t release() {
+        sd_image_t image = image_;
+        image_           = {0, 0, 0, nullptr};
+        return image;
+    }
+
+    void reset(sd_image_t image = {0, 0, 0, nullptr}) {
+        if (image_.data != nullptr) {
+            free(image_.data);
+        }
+        image_ = image;
+    }
+};
+
+class SDImageVec {
+private:
+    std::vector<sd_image_t> images_;
+
+public:
+    SDImageVec() = default;
+
+    SDImageVec(const SDImageVec&)            = delete;
+    SDImageVec& operator=(const SDImageVec&) = delete;
+
+    SDImageVec(SDImageVec&& other) noexcept
+        : images_(std::move(other.images_)) {
+    }
+
+    SDImageVec& operator=(SDImageVec&& other) noexcept {
+        if (this != &other) {
+            clear();
+            images_ = std::move(other.images_);
+        }
+        return *this;
+    }
+
+    ~SDImageVec() {
+        clear();
+    }
+
+    void push_back(sd_image_t image) {
+        images_.push_back(image);
+    }
+
+    void push_back(SDImageOwner&& image) {
+        images_.push_back(image.release());
+    }
+
+    void reserve(size_t count) {
+        images_.reserve(count);
+    }
+
+    void adopt(sd_image_t* images, int count) {
+        clear();
+        if (images == nullptr || count <= 0) {
+            free(images);
+            return;
+        }
+
+        images_.reserve(static_cast<size_t>(count));
+        for (int i = 0; i < count; ++i) {
+            images_.push_back(images[i]);
+        }
+        free(images);
+    }
+
+    size_t size() const {
+        return images_.size();
+    }
+
+    bool empty() const {
+        return images_.empty();
+    }
+
+    int count() const {
+        return static_cast<int>(images_.size());
+    }
+
+    explicit operator bool() const {
+        return !images_.empty();
+    }
+
+    sd_image_t* data() {
+        return images_.data();
+    }
+
+    const sd_image_t* data() const {
+        return images_.data();
+    }
+
+    sd_image_t& operator[](size_t index) {
+        return images_[index];
+    }
+
+    const sd_image_t& operator[](size_t index) const {
+        return images_[index];
+    }
+
+    std::vector<sd_image_t>& raw() {
+        return images_;
+    }
+
+    const std::vector<sd_image_t>& raw() const {
+        return images_;
+    }
+
+    void clear() {
+        for (sd_image_t& image : images_) {
+            free(image.data);
+            image.data = nullptr;
+        }
+        images_.clear();
+    }
+};
+
+#endif  // __EXAMPLE_RESOURCE_OWNERS_H__

examples/server/CMakeLists.txt

@@ -50,13 +50,30 @@ 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}")
 endif()
 
-add_executable(${TARGET} main.cpp)
+add_executable(${TARGET}
+    ../common/common.cpp
+    ../common/log.cpp
+    ../common/media_io.cpp
+    main.cpp
+    runtime.cpp
+    async_jobs.cpp
+    routes_index.cpp
+    routes_openai.cpp
+    routes_sdapi.cpp
+    routes_sdcpp.cpp
+)
 
 if(HAVE_FRONTEND_BUILD)
     add_dependencies(${TARGET} ${TARGET}_frontend)
@@ -70,4 +87,18 @@ endif()
 
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE stable-diffusion ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)
+if(SD_WEBP)
+    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBP)
+    target_link_libraries(${TARGET} PRIVATE webp libwebpmux)
+endif()
+if(SD_WEBM)
+    target_compile_definitions(${TARGET} PRIVATE SD_USE_WEBM)
+    target_link_libraries(${TARGET} PRIVATE webm)
+endif()
+
+# due to httplib; it contains a pragma for MSVC, but other things need explicit flags
+if(WIN32 AND NOT MSVC)
+    target_link_libraries(${TARGET} PRIVATE ws2_32)
+endif()
+
+target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)

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**.
 
@@ -93,11 +123,11 @@ In this case, the server will load and serve the specified `index.html` file ins
 usage: ./bin/sd-server  [options]
 
 Svr Options:
-  -l, --listen-ip <string>      server listen ip (default: 127.0.0.1)        
+  -l, --listen-ip <string>      server listen ip (default: 127.0.0.1)
   --serve-html-path <string>    path to HTML file to serve at root (optional)
   --listen-port <int>           server listen port (default: 1234)
   -v, --verbose                 print extra info
-  --color                       colors the logging tags according to level   
+  --color                       colors the logging tags according to level
   -h, --help                    show this help message and exit
 
 Context Options:
@@ -106,7 +136,8 @@ Context Options:
   --clip_g <string>                        path to the clip-g text encoder
   --clip_vision <string>                   path to the clip-vision encoder
   --t5xxl <string>                         path to the t5xxl text encoder
-  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image, mistral-small3.2 for flux2, ...)
+  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image,
+                                           mistral-small3.2 for flux2, ...)
   --llm_vision <string>                    path to the llm vit
   --qwen2vl <string>                       alias of --llm. Deprecated.
   --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
@@ -118,16 +149,18 @@ Context Options:
   --control-net <string>                   path to control net model
   --embd-dir <string>                      embeddings directory
   --lora-model-dir <string>                lora model directory
+  --hires-upscalers-dir <string>           highres fix upscaler model directory
   --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
   --photo-maker <string>                   path to PHOTOMAKER model
   --upscale-model <string>                 path to esrgan model.
-  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0, then threads will be set to the number of
-                                           CPU physical cores
+  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0,
+                                           then threads will be set to the number of CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
-  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-  --vae-tiling                             process vae in tiles to reduce memory usage
+  --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
+                                           graph splitting
   --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
-  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
+  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM
+                                           when needed
   --mmap                                   whether to memory-map model
   --control-net-cpu                        keep controlnet in cpu (for low vram)
   --clip-on-cpu                            keep clip in cpu (for low vram)
@@ -142,20 +175,19 @@ Context Options:
   --chroma-disable-dit-mask                disable dit mask for chroma
   --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --chroma-enable-t5-mask                  enable t5 mask for chroma
-  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
-                                           type of the weight file
+  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K,
+                                           q4_K). If not specified, the default is the type of the weight file
   --rng                                    RNG, one of [std_default, cuda, cpu], default: cuda(sd-webui), cpu(comfyui)
   --sampler-rng                            sampler RNG, one of [std_default, cuda, cpu]. If not specified, use --rng
-  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow, flux2_flow]
-  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is auto. In auto mode, if the model weights
-                                           contain any quantized parameters, the at_runtime mode will be used; otherwise,
-                                           immediately will be used.The immediately mode may have precision and
-                                           compatibility issues with quantized parameters, but it usually offers faster inference
-                                           speed and, in some cases, lower memory usage. The at_runtime mode, on the
-                                           other hand, is exactly the opposite.
-  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1
-                                           (overrides --vae-tile-size)
+  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow,
+                                           flux2_flow]
+  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is
+                                           auto. In auto mode, if the model weights contain any quantized parameters,
+                                           the at_runtime mode will be used; otherwise, immediately will be used.The
+                                           immediately mode may have precision and compatibility issues with quantized
+                                           parameters, but it usually offers faster inference speed and, in some cases,
+                                           lower memory usage. The at_runtime mode, on the other hand, is exactly the
+                                           opposite.
 
 Default Generation Options:
   -p, --prompt <string>                    the prompt to render
@@ -164,64 +196,97 @@ Default Generation Options:
   --end-img <string>                       path to the end image, required by flf2v
   --mask <string>                          path to the mask image
   --control-image <string>                 path to control image, control net
-  --control-video <string>                 path to control video frames, It must be a directory path. The video frames inside should be stored as images in
-                                           lexicographical (character) order. For example, if the control video path is
-                                           `frames`, the directory contain images such as 00.png, 01.png, ... etc.
+  --control-video <string>                 path to control video frames, It must be a directory path. The video frames
+                                           inside should be stored as images in lexicographical (character) order. For
+                                           example, if the control video path is `frames`, the directory contain images
+                                           such as 00.png, 01.png, ... etc.
   --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
   --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
+  --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
+                                           (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
+                                           antialiased), or a model name under --hires-upscalers-dir (default: Latent)
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)
   --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
-  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1). <= 0 represents unspecified,
-                                           will be 1 for SD1.x, 2 for SD2.x
+  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer
+                                           (default: -1). <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
   -b, --batch-count <int>                  batch count
   --video-frames <int>                     video frames (default: 1)
   --fps <int>                              fps (default: 24)
-  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for NitroSD-Realism around 250 and 500 for
-                                           NitroSD-Vibrant
+  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for
+                                           NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
   --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
   --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
+  --hires-width <int>                      highres fix target width, 0 to use --hires-scale (default: 0)
+  --hires-height <int>                     highres fix target height, 0 to use --hires-scale (default: 0)
+  --hires-steps <int>                      highres fix second pass sample steps, 0 to reuse --steps (default: 0)
+  --hires-upscale-tile-size <int>          highres fix upscaler tile size, reserved for model-backed upscalers (default:
+                                           128)
   --cfg-scale <float>                      unconditional guidance scale: (default: 7.0)
-  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same
+                                           as --cfg-scale)
   --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
-  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means disabled, a value of 2.5 is nice for sd3.5
-                                           medium
+  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means
+                                           disabled, a value of 2.5 is nice for sd3.5 medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            eta in DDIM, only for DDIM and TCD (default: 0)
+  --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)
-  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
-  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
+                                           (default: same as --cfg-scale)
+  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input
+                                           (default: 3.5)
+  --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) eta in DDIM, only for DDIM and TCD (default: 0)
+  --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
-  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if `--high-noise-steps` is set to -1
+  --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
+  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if
+                                           `--high-noise-steps` is set to -1
   --vace-strength <float>                  wan vace strength
-  --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
+  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --hires-scale <float>                    highres fix scale when target size is not set (default: 2.0)
+  --hires-denoising-strength <float>       highres fix second pass denoising strength (default: 0.7)
+  --increase-ref-index                     automatically increase the indices of references images based on the order
+                                           they are listed (starting with 1).
   --disable-auto-resize-ref-image          disable auto resize of ref images
+  --disable-image-metadata                 do not embed generation metadata on image files
+  --vae-tiling                             process vae in tiles to reduce memory usage
+  --hires                                  enable highres fix
   -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
-                                           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,
-                                           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
-  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
+  --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,
+                                           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, 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
+  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
+                                           "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET),
+                                           'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT
+                                           Chebyshev+Taylor forecasting)
   --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
-                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=. Examples:
-                                           "threshold=0.25" or "threshold=1.5,reset=0"
-  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
+                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit:
+                                           Fn=,Bn=,threshold=,warmup=; spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=.
+                                           Examples: "threshold=0.25" or "threshold=1.5,reset=0"
+  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g.,
+                                           "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
   --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
+  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
+  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size
+                                           if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
 ```

examples/server/async_jobs.cpp

@@ -0,0 +1,349 @@
+// Extracted from main.cpp during server refactor.
+
+#include "async_jobs.h"
+
+#include <iomanip>
+#include <sstream>
+
+#include "common/log.h"
+#include "common/media_io.h"
+#include "common/resource_owners.hpp"
+
+const char* async_job_kind_name(AsyncJobKind kind) {
+    switch (kind) {
+        case AsyncJobKind::ImgGen:
+            return "img_gen";
+        case AsyncJobKind::VidGen:
+            return "vid_gen";
+        default:
+            return "img_gen";
+    }
+}
+
+const char* async_job_status_name(AsyncJobStatus status) {
+    switch (status) {
+        case AsyncJobStatus::Queued:
+            return "queued";
+        case AsyncJobStatus::Generating:
+            return "generating";
+        case AsyncJobStatus::Completed:
+            return "completed";
+        case AsyncJobStatus::Failed:
+            return "failed";
+        case AsyncJobStatus::Cancelled:
+            return "cancelled";
+        default:
+            return "failed";
+    }
+}
+
+void purge_expired_jobs(AsyncJobManager& manager) {
+    const int64_t now = unix_timestamp_now();
+
+    for (auto it = manager.expired_jobs.begin(); it != manager.expired_jobs.end();) {
+        if (it->second <= now) {
+            it = manager.expired_jobs.erase(it);
+        } else {
+            ++it;
+        }
+    }
+
+    for (auto it = manager.jobs.begin(); it != manager.jobs.end();) {
+        const auto& job = it->second;
+        if (job->completed_at == 0) {
+            ++it;
+            continue;
+        }
+
+        int64_t ttl_seconds = job->status == AsyncJobStatus::Completed
+                                  ? manager.completed_ttl_seconds
+                                  : manager.failed_ttl_seconds;
+        if (now - job->completed_at >= ttl_seconds) {
+            manager.expired_jobs[job->id] = now + std::max<int64_t>(ttl_seconds, 60);
+            it                            = manager.jobs.erase(it);
+        } else {
+            ++it;
+        }
+    }
+}
+
+size_t count_pending_jobs(const AsyncJobManager& manager) {
+    size_t pending = 0;
+    for (const auto& entry : manager.jobs) {
+        if (entry.second->status == AsyncJobStatus::Queued ||
+            entry.second->status == AsyncJobStatus::Generating) {
+            ++pending;
+        }
+    }
+    return pending;
+}
+
+std::string make_async_job_id(AsyncJobManager& manager) {
+    std::ostringstream oss;
+    oss << "job_" << std::hex << unix_timestamp_now() << "_" << std::setw(8)
+        << std::setfill('0') << manager.next_id++;
+    return oss.str();
+}
+
+bool cancel_queued_job(AsyncJobManager& manager, AsyncGenerationJob& job) {
+    auto new_end = std::remove(manager.queue.begin(), manager.queue.end(), job.id);
+    if (new_end == manager.queue.end()) {
+        return false;
+    }
+
+    manager.queue.erase(new_end, manager.queue.end());
+    job.status       = AsyncJobStatus::Cancelled;
+    job.completed_at = unix_timestamp_now();
+    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;
+    job.error_code         = "cancelled";
+    job.error_message      = "job cancelled by client";
+    return true;
+}
+
+json make_async_job_json(const AsyncJobManager& manager, const AsyncGenerationJob& job) {
+    json result;
+    result["id"]             = job.id;
+    result["kind"]           = async_job_kind_name(job.kind);
+    result["status"]         = async_job_status_name(job.status);
+    result["created"]        = job.created_at;
+    result["started"]        = job.started_at == 0 ? json(nullptr) : json(job.started_at);
+    result["completed"]      = job.completed_at == 0 ? json(nullptr) : json(job.completed_at);
+    result["queue_position"] = 0;
+
+    if (job.status == AsyncJobStatus::Queued) {
+        size_t position = 1;
+        for (const auto& queued_id : manager.queue) {
+            if (queued_id == job.id) {
+                result["queue_position"] = position;
+                break;
+            }
+            ++position;
+        }
+    }
+
+    if (job.status == AsyncJobStatus::Completed) {
+        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["error"] = nullptr;
+    } else if (job.status == AsyncJobStatus::Failed ||
+               job.status == AsyncJobStatus::Cancelled) {
+        result["result"] = nullptr;
+        result["error"]  = {
+             {"code",
+             job.error_code.empty()
+                  ? (job.status == AsyncJobStatus::Cancelled ? "cancelled" : "generation_failed")
+                  : job.error_code},
+             {"message", job.error_message},
+        };
+    } else {
+        result["result"] = nullptr;
+        result["error"]  = nullptr;
+    }
+
+    return result;
+}
+
+bool execute_img_gen_job(ServerRuntime& runtime,
+                         AsyncGenerationJob& job,
+                         std::vector<std::string>& output_images,
+                         std::string& error_message) {
+    sd_img_gen_params_t params = job.img_gen.to_sd_img_gen_params_t();
+
+    SDImageVec results;
+
+    {
+        std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
+        sd_image_t* raw_results = generate_image(runtime.sd_ctx, &params);
+        results.adopt(raw_results, params.batch_count);
+    }
+
+    const int num_results = results.count();
+    if (num_results <= 0) {
+        error_message = "generate_image returned no results";
+        return false;
+    }
+
+    EncodedImageFormat encoded_format = EncodedImageFormat::PNG;
+    if (job.img_gen.output_format == "jpeg") {
+        encoded_format = EncodedImageFormat::JPEG;
+    } else if (job.img_gen.output_format == "webp") {
+        encoded_format = EncodedImageFormat::WEBP;
+    }
+
+    for (int i = 0; i < num_results; ++i) {
+        if (results[i].data == nullptr) {
+            continue;
+        }
+
+        const std::string metadata = job.img_gen.gen_params.embed_image_metadata
+                                         ? get_image_params(*runtime.ctx_params,
+                                                            job.img_gen.gen_params,
+                                                            job.img_gen.gen_params.seed + i)
+                                         : "";
+        auto image_bytes           = encode_image_to_vector(encoded_format,
+                                                            results[i].data,
+                                                            results[i].width,
+                                                            results[i].height,
+                                                            results[i].channel,
+                                                            metadata,
+                                                            job.img_gen.output_compression);
+        if (image_bytes.empty()) {
+            continue;
+        }
+        output_images.push_back(base64_encode(image_bytes));
+    }
+
+    if (output_images.empty()) {
+        error_message = "generate_image returned empty encoded outputs";
+        return false;
+    }
+
+    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;
+
+    while (true) {
+        std::shared_ptr<AsyncGenerationJob> job;
+        {
+            std::unique_lock<std::mutex> lock(manager.mutex);
+            manager.cv.wait(lock, [&]() { return manager.stop || !manager.queue.empty(); });
+
+            if (manager.stop && manager.queue.empty()) {
+                break;
+            }
+
+            purge_expired_jobs(manager);
+            if (manager.queue.empty()) {
+                continue;
+            }
+
+            const std::string job_id = manager.queue.front();
+            manager.queue.pop_front();
+
+            auto it = manager.jobs.find(job_id);
+            if (it == manager.jobs.end()) {
+                continue;
+            }
+
+            job             = it->second;
+            job->status     = AsyncJobStatus::Generating;
+            job->started_at = unix_timestamp_now();
+        }
+
+        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";
+        }
+
+        {
+            std::lock_guard<std::mutex> lock(manager.mutex);
+            auto it = manager.jobs.find(job->id);
+            if (it == manager.jobs.end()) {
+                continue;
+            }
+
+            job->completed_at = unix_timestamp_now();
+            if (ok) {
+                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 {
+                job->status        = AsyncJobStatus::Failed;
+                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

@@ -0,0 +1,78 @@
+#pragma once
+
+#include <condition_variable>
+#include <cstdint>
+#include <deque>
+#include <memory>
+#include <mutex>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+
+#include "runtime.h"
+
+enum class AsyncJobKind {
+    ImgGen,
+    VidGen,
+};
+
+enum class AsyncJobStatus {
+    Queued,
+    Generating,
+    Completed,
+    Failed,
+    Cancelled,
+};
+
+const char* async_job_kind_name(AsyncJobKind kind);
+const char* async_job_status_name(AsyncJobStatus status);
+
+struct AsyncGenerationJob {
+    std::string id;
+    AsyncJobKind kind     = AsyncJobKind::ImgGen;
+    AsyncJobStatus status = AsyncJobStatus::Queued;
+    int64_t created_at    = unix_timestamp_now();
+    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;
+};
+
+struct AsyncJobManager {
+    std::mutex mutex;
+    std::condition_variable cv;
+    std::unordered_map<std::string, std::shared_ptr<AsyncGenerationJob>> jobs;
+    std::unordered_map<std::string, int64_t> expired_jobs;
+    std::deque<std::string> queue;
+    uint64_t next_id              = 0;
+    bool stop                     = false;
+    size_t max_pending_jobs       = 64;
+    int64_t completed_ttl_seconds = 600;
+    int64_t failed_ttl_seconds    = 600;
+};
+
+void purge_expired_jobs(AsyncJobManager& manager);
+size_t count_pending_jobs(const AsyncJobManager& manager);
+std::string make_async_job_id(AsyncJobManager& manager);
+bool cancel_queued_job(AsyncJobManager& manager, AsyncGenerationJob& job);
+json make_async_job_json(const AsyncJobManager& manager, const AsyncGenerationJob& job);
+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/frontend

@@ -1 +1 @@
-Subproject commit 1a34176cd6d39ad3a226b2b69047e71f6797f6bc
+Subproject commit 797ccf80825cc035508ba9b599b2a21953e7f835

examples/server/routes.h

@@ -0,0 +1,11 @@
+#pragma once
+
+#include <string>
+
+#include "httplib.h"
+#include "runtime.h"
+
+void register_index_endpoints(httplib::Server& svr, const SDSvrParams& svr_params, const std::string& index_html);
+void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt);
+void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt);
+void register_sdcpp_api_endpoints(httplib::Server& svr, ServerRuntime& rt);

examples/server/routes_index.cpp

@@ -0,0 +1,22 @@
+#include "routes.h"
+
+#include <fstream>
+#include <iterator>
+
+void register_index_endpoints(httplib::Server& svr, const SDSvrParams& svr_params, const std::string& index_html) {
+    const std::string serve_html_path = svr_params.serve_html_path;
+    svr.Get("/", [serve_html_path, index_html](const httplib::Request&, httplib::Response& res) {
+        if (!serve_html_path.empty()) {
+            std::ifstream file(serve_html_path);
+            if (file) {
+                std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+                res.set_content(content, "text/html");
+            } else {
+                res.status = 500;
+                res.set_content("Error: Unable to read HTML file", "text/plain");
+            }
+        } else {
+            res.set_content(index_html, "text/html");
+        }
+    });
+}

examples/server/routes_openai.cpp

@@ -0,0 +1,388 @@
+#include "routes.h"
+
+#include <algorithm>
+#include <ctime>
+#include <regex>
+
+#include "common/common.h"
+#include "common/media_io.h"
+#include "common/resource_owners.hpp"
+
+static std::string extract_and_remove_sd_cpp_extra_args(std::string& text) {
+    std::regex re("<sd_cpp_extra_args>(.*?)</sd_cpp_extra_args>");
+    std::smatch match;
+
+    std::string extracted;
+    if (std::regex_search(text, match, re)) {
+        extracted = match[1].str();
+        text      = std::regex_replace(text, re, "");
+    }
+    return extracted;
+}
+
+static bool build_openai_generation_request(const httplib::Request& req,
+                                            ServerRuntime& runtime,
+                                            ImgGenJobRequest& request,
+                                            std::string& error_message) {
+    if (req.body.empty()) {
+        error_message = "empty body";
+        return false;
+    }
+
+    json j                    = json::parse(req.body);
+    std::string prompt        = j.value("prompt", "");
+    int n                     = std::max(1, j.value("n", 1));
+    std::string size          = j.value("size", "");
+    std::string output_format = j.value("output_format", "png");
+    int output_compression    = j.value("output_compression", 100);
+    int width                 = runtime.default_gen_params->width > 0 ? runtime.default_gen_params->width : 512;
+    int height                = runtime.default_gen_params->width > 0 ? runtime.default_gen_params->height : 512;
+    if (!size.empty()) {
+        auto pos = size.find('x');
+        if (pos != std::string::npos) {
+            try {
+                width  = std::stoi(size.substr(0, pos));
+                height = std::stoi(size.substr(pos + 1));
+            } catch (...) {
+            }
+        }
+    }
+
+    if (prompt.empty()) {
+        error_message = "prompt required";
+        return false;
+    }
+
+    request.gen_params = *runtime.default_gen_params;
+    if (!assign_output_options(request, output_format, output_compression, true, error_message)) {
+        return false;
+    }
+
+    request.gen_params.prompt      = prompt;
+    request.gen_params.width       = width;
+    request.gen_params.height      = height;
+    request.gen_params.batch_count = n;
+
+    std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(request.gen_params.prompt);
+    if (!sd_cpp_extra_args_str.empty() && !request.gen_params.from_json_str(sd_cpp_extra_args_str)) {
+        error_message = "invalid sd_cpp_extra_args";
+        return false;
+    }
+
+    // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
+        error_message = "invalid params";
+        return false;
+    }
+    return true;
+}
+
+static bool build_openai_edit_request(const httplib::Request& req,
+                                      ServerRuntime& runtime,
+                                      ImgGenJobRequest& request,
+                                      std::string& error_message) {
+    if (!req.is_multipart_form_data()) {
+        error_message = "Content-Type must be multipart/form-data";
+        return false;
+    }
+
+    std::string prompt = req.form.get_field("prompt");
+    if (prompt.empty()) {
+        error_message = "prompt required";
+        return false;
+    }
+
+    size_t image_count    = req.form.get_file_count("image[]");
+    bool has_legacy_image = req.form.has_file("image");
+    if (image_count == 0 && !has_legacy_image) {
+        error_message = "at least one image[] required";
+        return false;
+    }
+
+    std::vector<std::vector<uint8_t>> images_bytes;
+    for (size_t i = 0; i < image_count; ++i) {
+        auto file = req.form.get_file("image[]", i);
+        images_bytes.emplace_back(file.content.begin(), file.content.end());
+    }
+    if (image_count == 0 && has_legacy_image) {
+        auto file = req.form.get_file("image");
+        images_bytes.emplace_back(file.content.begin(), file.content.end());
+    }
+
+    std::vector<uint8_t> mask_bytes;
+    if (req.form.has_file("mask")) {
+        auto file = req.form.get_file("mask");
+        mask_bytes.assign(file.content.begin(), file.content.end());
+    }
+
+    int n = 1;
+    if (req.form.has_field("n")) {
+        try {
+            n = std::stoi(req.form.get_field("n"));
+        } catch (...) {
+        }
+    }
+
+    std::string size = req.form.get_field("size");
+    int width        = -1;
+    int height       = -1;
+    if (!size.empty()) {
+        auto pos = size.find('x');
+        if (pos != std::string::npos) {
+            try {
+                width  = std::stoi(size.substr(0, pos));
+                height = std::stoi(size.substr(pos + 1));
+            } catch (...) {
+            }
+        }
+    }
+
+    std::string output_format = req.form.has_field("output_format")
+                                    ? req.form.get_field("output_format")
+                                    : "png";
+
+    int output_compression = 100;
+    try {
+        output_compression = std::stoi(req.form.get_field("output_compression"));
+    } catch (...) {
+    }
+
+    request.gen_params = *runtime.default_gen_params;
+    if (!assign_output_options(request, output_format, output_compression, false, error_message)) {
+        return false;
+    }
+
+    request.gen_params.prompt      = prompt;
+    request.gen_params.width       = width;
+    request.gen_params.height      = height;
+    request.gen_params.batch_count = n;
+
+    for (auto& bytes : images_bytes) {
+        int img_w           = 0;
+        int img_h           = 0;
+        uint8_t* raw_pixels = load_image_from_memory(
+            reinterpret_cast<const char*>(bytes.data()),
+            static_cast<int>(bytes.size()),
+            img_w, img_h,
+            width, height, 3);
+        if (raw_pixels == nullptr) {
+            continue;
+        }
+
+        SDImageOwner image_owner({(uint32_t)img_w, (uint32_t)img_h, 3, raw_pixels});
+        request.gen_params.set_width_and_height_if_unset(image_owner.get().width, image_owner.get().height);
+        request.gen_params.ref_images.push_back(std::move(image_owner));
+    }
+
+    if (!request.gen_params.ref_images.empty()) {
+        request.gen_params.init_image = request.gen_params.ref_images.front();
+    }
+
+    if (!mask_bytes.empty()) {
+        int expected_width  = 0;
+        int expected_height = 0;
+        if (request.gen_params.width_and_height_are_set()) {
+            expected_width  = request.gen_params.width;
+            expected_height = request.gen_params.height;
+        }
+        int mask_w = 0;
+        int mask_h = 0;
+
+        uint8_t* mask_raw = load_image_from_memory(
+            reinterpret_cast<const char*>(mask_bytes.data()),
+            static_cast<int>(mask_bytes.size()),
+            mask_w, mask_h,
+            expected_width, expected_height, 1);
+        request.gen_params.mask_image.reset({(uint32_t)mask_w, (uint32_t)mask_h, 1, mask_raw});
+        const sd_image_t& mask_image = request.gen_params.mask_image.get();
+        request.gen_params.set_width_and_height_if_unset(mask_image.width, mask_image.height);
+    } else {
+        request.gen_params.mask_image.reset({
+            (uint32_t)request.gen_params.get_resolved_width(),
+            (uint32_t)request.gen_params.get_resolved_height(),
+            1,
+            nullptr,
+        });
+    }
+
+    std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(request.gen_params.prompt);
+    if (!sd_cpp_extra_args_str.empty() && !request.gen_params.from_json_str(sd_cpp_extra_args_str)) {
+        error_message = "invalid sd_cpp_extra_args";
+        return false;
+    }
+
+    // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
+        error_message = "invalid params";
+        return false;
+    }
+
+    return true;
+}
+
+static bool execute_sync_img_gen_request(ServerRuntime& runtime,
+                                         ImgGenJobRequest& request,
+                                         SDImageVec& results,
+                                         std::string& error_message) {
+    sd_img_gen_params_t img_gen_params = request.to_sd_img_gen_params_t();
+    int num_results                    = 0;
+
+    {
+        std::lock_guard<std::mutex> lock(*runtime.sd_ctx_mutex);
+        sd_image_t* raw_results = generate_image(runtime.sd_ctx, &img_gen_params);
+        num_results             = request.gen_params.batch_count;
+        results.adopt(raw_results, num_results);
+    }
+
+    if (results.empty()) {
+        error_message = "generate_image returned no results";
+        return false;
+    }
+    return true;
+}
+
+void register_openai_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
+    ServerRuntime* runtime = &rt;
+
+    svr.Get("/v1/models", [runtime](const httplib::Request&, httplib::Response& res) {
+        json r;
+        r["data"] = json::array();
+        r["data"].push_back({{"id", "sd-cpp-local"}, {"object", "model"}, {"owned_by", "local"}});
+        res.set_content(r.dump(), "application/json");
+    });
+
+    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)) {
+                res.status = 400;
+                res.set_content(json({{"error", error_message}}).dump(), "application/json");
+                return;
+            }
+
+            LOG_DEBUG("%s\n", request.gen_params.to_string().c_str());
+
+            SDImageVec results;
+            if (!execute_sync_img_gen_request(*runtime, request, results, error_message)) {
+                res.status = 500;
+                res.set_content(json({{"error", error_message}}).dump(), "application/json");
+                return;
+            }
+
+            json out;
+            out["created"]       = static_cast<long long>(std::time(nullptr));
+            out["data"]          = json::array();
+            out["output_format"] = request.output_format;
+
+            for (int i = 0; i < request.gen_params.batch_count; ++i) {
+                if (results[i].data == nullptr) {
+                    continue;
+                }
+                std::string params = request.gen_params.embed_image_metadata
+                                         ? get_image_params(*runtime->ctx_params,
+                                                            request.gen_params,
+                                                            request.gen_params.seed + i)
+                                         : "";
+                auto image_bytes   = encode_image_to_vector(request.output_format == "jpeg"
+                                                                ? EncodedImageFormat::JPEG
+                                                            : request.output_format == "webp"
+                                                                ? EncodedImageFormat::WEBP
+                                                                : EncodedImageFormat::PNG,
+                                                          results[i].data,
+                                                          results[i].width,
+                                                          results[i].height,
+                                                          results[i].channel,
+                                                          params,
+                                                          request.output_compression);
+                if (image_bytes.empty()) {
+                    LOG_ERROR("write image to mem failed");
+                    continue;
+                }
+
+                json item;
+                item["b64_json"] = base64_encode(image_bytes);
+                out["data"].push_back(item);
+            }
+
+            res.set_content(out.dump(), "application/json");
+            res.status = 200;
+
+        } catch (const std::exception& e) {
+            res.status = 500;
+            json err;
+            err["error"]   = "server_error";
+            err["message"] = e.what();
+            res.set_content(err.dump(), "application/json");
+        }
+    });
+
+    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)) {
+                res.status = 400;
+                res.set_content(json({{"error", error_message}}).dump(), "application/json");
+                return;
+            }
+
+            LOG_DEBUG("%s\n", request.gen_params.to_string().c_str());
+
+            SDImageVec results;
+            if (!execute_sync_img_gen_request(*runtime, request, results, error_message)) {
+                res.status = 500;
+                res.set_content(json({{"error", error_message}}).dump(), "application/json");
+                return;
+            }
+
+            json out;
+            out["created"]       = static_cast<long long>(std::time(nullptr));
+            out["data"]          = json::array();
+            out["output_format"] = request.output_format;
+
+            for (int i = 0; i < request.gen_params.batch_count; ++i) {
+                if (results[i].data == nullptr) {
+                    continue;
+                }
+                std::string params = request.gen_params.embed_image_metadata
+                                         ? get_image_params(*runtime->ctx_params,
+                                                            request.gen_params,
+                                                            request.gen_params.seed + i)
+                                         : "";
+                auto image_bytes   = encode_image_to_vector(request.output_format == "jpeg" ? EncodedImageFormat::JPEG : EncodedImageFormat::PNG,
+                                                          results[i].data,
+                                                          results[i].width,
+                                                          results[i].height,
+                                                          results[i].channel,
+                                                          params,
+                                                          request.output_compression);
+                json item;
+                item["b64_json"] = base64_encode(image_bytes);
+                out["data"].push_back(item);
+            }
+
+            res.set_content(out.dump(), "application/json");
+            res.status = 200;
+
+        } catch (const std::exception& e) {
+            res.status = 500;
+            json err;
+            err["error"]   = "server_error";
+            err["message"] = e.what();
+            res.set_content(err.dump(), "application/json");
+        }
+    });
+}

examples/server/routes_sdapi.cpp

@@ -0,0 +1,469 @@
+#include "routes.h"
+
+#include <algorithm>
+#include <cctype>
+#include <cstring>
+#include <regex>
+#include <string_view>
+#include <unordered_map>
+
+#include "common/common.h"
+#include "common/media_io.h"
+#include "common/resource_owners.hpp"
+
+namespace fs = std::filesystem;
+
+static std::string extract_and_remove_sd_cpp_extra_args(std::string& text) {
+    std::regex re("<sd_cpp_extra_args>(.*?)</sd_cpp_extra_args>");
+    std::smatch match;
+
+    std::string extracted;
+    if (std::regex_search(text, match, re)) {
+        extracted = match[1].str();
+        text      = std::regex_replace(text, re, "");
+    }
+    return extracted;
+}
+
+static fs::path resolve_display_model_path(const ServerRuntime& runtime) {
+    const auto& ctx = *runtime.ctx_params;
+    if (!ctx.model_path.empty()) {
+        return fs::path(ctx.model_path);
+    }
+    if (!ctx.diffusion_model_path.empty()) {
+        return fs::path(ctx.diffusion_model_path);
+    }
+    return {};
+}
+
+static std::string lower_ascii(std::string value) {
+    std::transform(value.begin(), value.end(), value.begin(), [](unsigned char c) {
+        return static_cast<char>(std::tolower(c));
+    });
+    return value;
+}
+
+static enum sample_method_t get_sdapi_sample_method(std::string name) {
+    enum sample_method_t result = str_to_sample_method(name.c_str());
+    if (result != SAMPLE_METHOD_COUNT) {
+        return result;
+    }
+
+    name = lower_ascii(name);
+    static const std::unordered_map<std::string_view, sample_method_t> hardcoded{
+        {"euler a", EULER_A_SAMPLE_METHOD},
+        {"k_euler_a", EULER_A_SAMPLE_METHOD},
+        {"euler", EULER_SAMPLE_METHOD},
+        {"k_euler", EULER_SAMPLE_METHOD},
+        {"heun", HEUN_SAMPLE_METHOD},
+        {"k_heun", HEUN_SAMPLE_METHOD},
+        {"dpm2", DPM2_SAMPLE_METHOD},
+        {"k_dpm_2", DPM2_SAMPLE_METHOD},
+        {"lcm", LCM_SAMPLE_METHOD},
+        {"ddim", DDIM_TRAILING_SAMPLE_METHOD},
+        {"dpm++ 2m", DPMPP2M_SAMPLE_METHOD},
+        {"k_dpmpp_2m", DPMPP2M_SAMPLE_METHOD},
+        {"res multistep", RES_MULTISTEP_SAMPLE_METHOD},
+        {"k_res_multistep", RES_MULTISTEP_SAMPLE_METHOD},
+        {"res 2s", RES_2S_SAMPLE_METHOD},
+        {"k_res_2s", RES_2S_SAMPLE_METHOD},
+    };
+    auto it = hardcoded.find(name);
+    return it != hardcoded.end() ? it->second : SAMPLE_METHOD_COUNT;
+}
+
+static void assign_solid_mask(SDImageOwner& mask_owner, int width, int height) {
+    const size_t pixel_count = static_cast<size_t>(width) * static_cast<size_t>(height);
+    uint8_t* raw_mask        = static_cast<uint8_t*>(malloc(pixel_count));
+    if (raw_mask == nullptr) {
+        mask_owner.reset({0, 0, 1, nullptr});
+        return;
+    }
+    std::memset(raw_mask, 255, pixel_count);
+    mask_owner.reset({(uint32_t)width, (uint32_t)height, 1, raw_mask});
+}
+
+static bool build_sdapi_img_gen_request(const json& j,
+                                        ServerRuntime& runtime,
+                                        bool img2img,
+                                        ImgGenJobRequest& request,
+                                        std::string& error_message) {
+    std::string prompt          = j.value("prompt", "");
+    std::string negative_prompt = j.value("negative_prompt", "");
+    int width                   = j.value("width", 512);
+    int height                  = j.value("height", 512);
+    int steps                   = j.value("steps", runtime.default_gen_params->sample_params.sample_steps);
+    float cfg_scale             = j.value("cfg_scale", runtime.default_gen_params->sample_params.guidance.txt_cfg);
+    int64_t seed                = j.value("seed", -1);
+    int batch_size              = j.value("batch_size", 1);
+    int clip_skip               = j.value("clip_skip", -1);
+    std::string sampler_name    = j.value("sampler_name", "");
+    std::string scheduler_name  = j.value("scheduler", "");
+
+    if (width <= 0 || height <= 0) {
+        error_message = "width and height must be positive";
+        return false;
+    }
+
+    if (prompt.empty()) {
+        error_message = "prompt required";
+        return false;
+    }
+
+    request.gen_params = *runtime.default_gen_params;
+
+    request.gen_params.prompt                         = prompt;
+    request.gen_params.negative_prompt                = negative_prompt;
+    request.gen_params.seed                           = seed;
+    request.gen_params.sample_params.sample_steps     = steps;
+    request.gen_params.batch_count                    = batch_size;
+    request.gen_params.sample_params.guidance.txt_cfg = cfg_scale;
+    request.gen_params.width                          = j.value("width", -1);
+    request.gen_params.height                         = j.value("height", -1);
+
+    if (!img2img && j.value("enable_hr", false)) {
+        request.gen_params.hires_enabled = true;
+        request.gen_params.hires_scale   = j.value("hr_scale", request.gen_params.hires_scale);
+        request.gen_params.hires_width   = j.value("hr_resize_x", request.gen_params.hires_width);
+        request.gen_params.hires_height  = j.value("hr_resize_y", request.gen_params.hires_height);
+        request.gen_params.hires_steps   = j.value("hr_steps", request.gen_params.hires_steps);
+        request.gen_params.hires_denoising_strength =
+            j.value("denoising_strength", request.gen_params.hires_denoising_strength);
+
+        request.gen_params.hires_upscaler = j.value("hr_upscaler", request.gen_params.hires_upscaler);
+    }
+
+    std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(request.gen_params.prompt);
+    if (!sd_cpp_extra_args_str.empty() && !request.gen_params.from_json_str(sd_cpp_extra_args_str)) {
+        error_message = "invalid sd_cpp_extra_args";
+        return false;
+    }
+
+    if (clip_skip > 0) {
+        request.gen_params.clip_skip = clip_skip;
+    }
+
+    enum sample_method_t sample_method = get_sdapi_sample_method(sampler_name);
+    if (sample_method != SAMPLE_METHOD_COUNT) {
+        request.gen_params.sample_params.sample_method = sample_method;
+    }
+
+    enum scheduler_t scheduler = str_to_scheduler(scheduler_name.c_str());
+    if (scheduler != SCHEDULER_COUNT) {
+        request.gen_params.sample_params.scheduler = scheduler;
+    }
+
+    if (j.contains("lora") && j["lora"].is_array()) {
+        request.gen_params.lora_map.clear();
+        request.gen_params.high_noise_lora_map.clear();
+
+        for (const auto& item : j["lora"]) {
+            if (!item.is_object()) {
+                continue;
+            }
+
+            std::string path   = item.value("path", "");
+            float multiplier   = item.value("multiplier", 1.0f);
+            bool is_high_noise = item.value("is_high_noise", false);
+
+            if (path.empty()) {
+                error_message = "lora.path required";
+                return false;
+            }
+
+            std::string fullpath = get_lora_full_path(runtime, path);
+            if (fullpath.empty()) {
+                error_message = "invalid lora path: " + path;
+                return false;
+            }
+
+            if (is_high_noise) {
+                request.gen_params.high_noise_lora_map[fullpath] += multiplier;
+            } else {
+                request.gen_params.lora_map[fullpath] += multiplier;
+            }
+        }
+    }
+
+    if (img2img) {
+        const int expected_width  = request.gen_params.width_and_height_are_set() ? request.gen_params.width : 0;
+        const int expected_height = request.gen_params.width_and_height_are_set() ? request.gen_params.height : 0;
+
+        if (j.contains("init_images") && j["init_images"].is_array() && !j["init_images"].empty()) {
+            if (decode_base64_image(j["init_images"][0].get<std::string>(),
+                                    3,
+                                    expected_width,
+                                    expected_height,
+                                    request.gen_params.init_image)) {
+                const sd_image_t& image = request.gen_params.init_image.get();
+                request.gen_params.set_width_and_height_if_unset(image.width, image.height);
+            }
+        }
+
+        if (j.contains("mask") && j["mask"].is_string()) {
+            if (decode_base64_image(j["mask"].get<std::string>(),
+                                    1,
+                                    expected_width,
+                                    expected_height,
+                                    request.gen_params.mask_image)) {
+                const sd_image_t& image = request.gen_params.mask_image.get();
+                request.gen_params.set_width_and_height_if_unset(image.width, image.height);
+            }
+            sd_image_t& mask_image      = request.gen_params.mask_image.get();
+            bool inpainting_mask_invert = j.value("inpainting_mask_invert", 0) != 0;
+            if (inpainting_mask_invert && mask_image.data != nullptr) {
+                for (uint32_t i = 0; i < mask_image.width * mask_image.height; ++i) {
+                    mask_image.data[i] = 255 - mask_image.data[i];
+                }
+            }
+        } else {
+            const int resolved_width  = request.gen_params.get_resolved_width();
+            const int resolved_height = request.gen_params.get_resolved_height();
+            assign_solid_mask(request.gen_params.mask_image, resolved_width, resolved_height);
+        }
+
+        float denoising_strength = j.value("denoising_strength", -1.f);
+        if (denoising_strength >= 0.f) {
+            request.gen_params.strength = std::min(denoising_strength, 1.0f);
+        }
+    }
+
+    if (j.contains("extra_images") && j["extra_images"].is_array()) {
+        for (const auto& extra_image : j["extra_images"]) {
+            if (!extra_image.is_string()) {
+                continue;
+            }
+            SDImageOwner image_owner;
+            if (decode_base64_image(extra_image.get<std::string>(),
+                                    3,
+                                    request.gen_params.width_and_height_are_set() ? request.gen_params.width : 0,
+                                    request.gen_params.width_and_height_are_set() ? request.gen_params.height : 0,
+                                    image_owner)) {
+                const sd_image_t& image = image_owner.get();
+                request.gen_params.set_width_and_height_if_unset(image.width, image.height);
+                request.gen_params.ref_images.push_back(std::move(image_owner));
+            }
+        }
+    }
+
+    // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
+        error_message = "invalid params";
+        return false;
+    }
+
+    return true;
+}
+
+void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
+    ServerRuntime* runtime = &rt;
+
+    auto sdapi_any2img = [runtime](const httplib::Request& req, httplib::Response& res, bool img2img) {
+        try {
+            if (req.body.empty()) {
+                res.status = 400;
+                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;
+            std::string error_message;
+            if (!build_sdapi_img_gen_request(j, *runtime, img2img, request, error_message)) {
+                res.status = 400;
+                res.set_content(json({{"error", error_message}}).dump(), "application/json");
+                return;
+            }
+
+            LOG_DEBUG("%s\n", request.gen_params.to_string().c_str());
+
+            sd_img_gen_params_t img_gen_params = request.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, &img_gen_params);
+                num_results             = request.gen_params.batch_count;
+                results.adopt(raw_results, num_results);
+            }
+
+            if (results.empty()) {
+                res.status = 500;
+                res.set_content(R"({"error":"generate_image returned no results"})", "application/json");
+                return;
+            }
+
+            json out;
+            out["images"]     = json::array();
+            out["parameters"] = j;
+            out["info"]       = "";
+
+            for (int i = 0; i < num_results; ++i) {
+                if (results[i].data == nullptr) {
+                    continue;
+                }
+
+                std::string params = request.gen_params.embed_image_metadata
+                                         ? get_image_params(*runtime->ctx_params,
+                                                            request.gen_params,
+                                                            request.gen_params.seed + i)
+                                         : "";
+                auto image_bytes   = encode_image_to_vector(EncodedImageFormat::PNG,
+                                                            results[i].data,
+                                                            results[i].width,
+                                                            results[i].height,
+                                                            results[i].channel,
+                                                            params);
+
+                if (image_bytes.empty()) {
+                    LOG_ERROR("write image to mem failed");
+                    continue;
+                }
+
+                out["images"].push_back(base64_encode(image_bytes));
+            }
+
+            res.set_content(out.dump(), "application/json");
+            res.status = 200;
+
+        } catch (const std::exception& e) {
+            res.status = 500;
+            json err;
+            err["error"]   = "server_error";
+            err["message"] = e.what();
+            res.set_content(err.dump(), "application/json");
+        }
+    };
+
+    svr.Post("/sdapi/v1/txt2img", [sdapi_any2img](const httplib::Request& req, httplib::Response& res) {
+        sdapi_any2img(req, res, false);
+    });
+
+    svr.Post("/sdapi/v1/img2img", [sdapi_any2img](const httplib::Request& req, httplib::Response& res) {
+        sdapi_any2img(req, res, true);
+    });
+
+    svr.Get("/sdapi/v1/loras", [runtime](const httplib::Request&, httplib::Response& res) {
+        refresh_lora_cache(*runtime);
+
+        json result = json::array();
+        {
+            std::lock_guard<std::mutex> lock(*runtime->lora_mutex);
+            for (const auto& e : *runtime->lora_cache) {
+                json item;
+                item["name"] = e.name;
+                item["path"] = e.path;
+                result.push_back(item);
+            }
+        }
+
+        res.set_content(result.dump(), "application/json");
+    });
+
+    svr.Get("/sdapi/v1/upscalers", [runtime](const httplib::Request&, httplib::Response& res) {
+        refresh_upscaler_cache(*runtime);
+
+        auto make_builtin = [](const char* name) {
+            json item;
+            item["name"]       = name;
+            item["model_name"] = nullptr;
+            item["model_path"] = nullptr;
+            item["model_url"]  = nullptr;
+            item["scale"]      = 4;
+            return item;
+        };
+
+        json result = json::array();
+        result.push_back(make_builtin("None"));
+        result.push_back(make_builtin("Lanczos"));
+        result.push_back(make_builtin("Nearest"));
+
+        {
+            std::lock_guard<std::mutex> lock(*runtime->upscaler_mutex);
+            for (const auto& e : *runtime->upscaler_cache) {
+                json item;
+                item["name"]       = e.name;
+                item["model_name"] = e.model_name;
+                item["model_path"] = e.fullpath;
+                item["model_url"]  = nullptr;
+                item["scale"]      = e.scale;
+                result.push_back(item);
+            }
+        }
+
+        res.set_content(result.dump(), "application/json");
+    });
+
+    svr.Get("/sdapi/v1/latent-upscale-modes", [](const httplib::Request&, httplib::Response& res) {
+        json result = json::array({
+            {{"name", "Latent"}},
+            {{"name", "Latent (nearest)"}},
+            {{"name", "Latent (nearest-exact)"}},
+            {{"name", "Latent (antialiased)"}},
+            {{"name", "Latent (bicubic)"}},
+            {{"name", "Latent (bicubic antialiased)"}},
+        });
+        res.set_content(result.dump(), "application/json");
+    });
+
+    svr.Get("/sdapi/v1/samplers", [runtime](const httplib::Request&, httplib::Response& res) {
+        std::vector<std::string> sampler_names;
+        sampler_names.push_back("default");
+        for (int i = 0; i < SAMPLE_METHOD_COUNT; i++) {
+            sampler_names.push_back(sd_sample_method_name((sample_method_t)i));
+        }
+        json r = json::array();
+        for (auto name : sampler_names) {
+            json entry;
+            entry["name"]    = name;
+            entry["aliases"] = json::array({name});
+            entry["options"] = json::object();
+            r.push_back(entry);
+        }
+        res.set_content(r.dump(), "application/json");
+    });
+
+    svr.Get("/sdapi/v1/schedulers", [runtime](const httplib::Request&, httplib::Response& res) {
+        std::vector<std::string> scheduler_names;
+        scheduler_names.push_back("default");
+        for (int i = 0; i < SCHEDULER_COUNT; i++) {
+            scheduler_names.push_back(sd_scheduler_name((scheduler_t)i));
+        }
+        json r = json::array();
+        for (auto name : scheduler_names) {
+            json entry;
+            entry["name"]  = name;
+            entry["label"] = name;
+            r.push_back(entry);
+        }
+        res.set_content(r.dump(), "application/json");
+    });
+
+    svr.Get("/sdapi/v1/sd-models", [runtime](const httplib::Request&, httplib::Response& res) {
+        fs::path model_path = resolve_display_model_path(*runtime);
+        json entry;
+        entry["title"]      = model_path.stem();
+        entry["model_name"] = model_path.stem();
+        entry["filename"]   = model_path.filename();
+        entry["hash"]       = "8888888888";
+        entry["sha256"]     = "8888888888888888888888888888888888888888888888888888888888888888";
+        entry["config"]     = nullptr;
+        json r              = json::array();
+        r.push_back(entry);
+        res.set_content(r.dump(), "application/json");
+    });
+
+    svr.Get("/sdapi/v1/options", [runtime](const httplib::Request&, httplib::Response& res) {
+        fs::path model_path = resolve_display_model_path(*runtime);
+        json r;
+        r["samples_format"]      = "png";
+        r["sd_model_checkpoint"] = model_path.stem();
+        res.set_content(r.dump(), "application/json");
+    });
+}

examples/server/routes_sdcpp.cpp

@@ -0,0 +1,588 @@
+#include "routes.h"
+
+#include <algorithm>
+#include <cmath>
+#include <filesystem>
+
+#include "async_jobs.h"
+#include "common/common.h"
+
+namespace fs = std::filesystem;
+
+static bool parse_cache_mode(const std::string& mode_str, sd_cache_mode_t& mode_out) {
+    if (mode_str == "disabled") {
+        mode_out = SD_CACHE_DISABLED;
+        return true;
+    }
+    if (mode_str == "easycache") {
+        mode_out = SD_CACHE_EASYCACHE;
+        return true;
+    }
+    if (mode_str == "ucache") {
+        mode_out = SD_CACHE_UCACHE;
+        return true;
+    }
+    if (mode_str == "dbcache") {
+        mode_out = SD_CACHE_DBCACHE;
+        return true;
+    }
+    if (mode_str == "taylorseer") {
+        mode_out = SD_CACHE_TAYLORSEER;
+        return true;
+    }
+    if (mode_str == "cache-dit") {
+        mode_out = SD_CACHE_CACHE_DIT;
+        return true;
+    }
+    if (mode_str == "spectrum") {
+        mode_out = SD_CACHE_SPECTRUM;
+        return true;
+    }
+    return false;
+}
+
+static json finite_number_or_null(float value) {
+    return std::isfinite(value) ? json(value) : json(nullptr);
+}
+
+static const char* capability_scheduler_name(enum scheduler_t scheduler) {
+    return scheduler < SCHEDULER_COUNT ? sd_scheduler_name(scheduler) : "default";
+}
+
+static const char* capability_sample_method_name(enum sample_method_t sample_method) {
+    return sample_method < SAMPLE_METHOD_COUNT ? sd_sample_method_name(sample_method) : "default";
+}
+
+static json make_vae_tiling_json(const sd_tiling_params_t& params) {
+    return {
+        {"enabled", params.enabled},
+        {"tile_size_x", params.tile_size_x},
+        {"tile_size_y", params.tile_size_y},
+        {"target_overlap", params.target_overlap},
+        {"rel_size_x", params.rel_size_x},
+        {"rel_size_y", params.rel_size_y},
+    };
+}
+
+static fs::path resolve_display_model_path(const ServerRuntime& runtime) {
+    const auto& ctx = *runtime.ctx_params;
+    if (!ctx.model_path.empty()) {
+        return fs::path(ctx.model_path);
+    }
+    if (!ctx.diffusion_model_path.empty()) {
+        return fs::path(ctx.diffusion_model_path);
+    }
+    return {};
+}
+
+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},
+              }},
+         }},
+    };
+}
+
+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},
+        {"width", defaults.width > 0 ? defaults.width : 512},
+        {"height", defaults.height > 0 ? defaults.height : 512},
+        {"strength", defaults.strength},
+        {"seed", defaults.seed},
+        {"batch_count", defaults.batch_count},
+        {"auto_resize_ref_image", defaults.auto_resize_ref_image},
+        {"increase_ref_index", defaults.increase_ref_index},
+        {"control_strength", defaults.control_strength},
+        {"sample_params", make_sample_params_json(defaults.sample_params, defaults.skip_layers)},
+        {"hires",
+         {
+             {"enabled", defaults.hires_enabled},
+             {"upscaler", defaults.hires_upscaler},
+             {"scale", defaults.hires_scale},
+             {"target_width", defaults.hires_width},
+             {"target_height", defaults.hires_height},
+             {"steps", defaults.hires_steps},
+             {"denoising_strength", defaults.hires_denoising_strength},
+             {"upscale_tile_size", defaults.hires_upscale_tile_size},
+         }},
+        {"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},
+    };
+}
+
+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},
+    };
+}
+
+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},
+        {"hires", 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);
+    refresh_upscaler_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 available_upscalers  = 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},
+            });
+        }
+    }
+
+    available_upscalers.push_back({
+        {"name", "None"},
+    });
+    available_upscalers.push_back({
+        {"name", "Lanczos"},
+    });
+    available_upscalers.push_back({
+        {"name", "Nearest"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (nearest)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (nearest-exact)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (antialiased)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (bicubic)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (bicubic antialiased)"},
+    });
+    {
+        std::lock_guard<std::mutex> lock(*runtime.upscaler_mutex);
+        for (const auto& entry : *runtime.upscaler_cache) {
+            available_upscalers.push_back({
+                {"name", entry.name},
+            });
+        }
+    }
+
+    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},
+    };
+    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;
+    result["upscalers"]              = available_upscalers;
+    return result;
+}
+
+static bool parse_img_gen_request(const json& body,
+                                  ServerRuntime& runtime,
+                                  ImgGenJobRequest& 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", "png");
+    int output_compression    = body.value("output_compression", 100);
+    if (!assign_output_options(request, output_format, output_compression, true, error_message)) {
+        return false;
+    }
+    // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
+        error_message = "invalid generation parameters";
+        return false;
+    }
+    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, "", runtime.ctx_params->hires_upscalers_dir, true)) {
+        error_message = "invalid generation parameters";
+        return false;
+    }
+    return true;
+}
+
+void register_sdcpp_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
+    ServerRuntime* runtime = &rt;
+
+    svr.Get("/sdcpp/v1/capabilities", [runtime](const httplib::Request&, httplib::Response& res) {
+        res.status = 200;
+        res.set_content(make_capabilities_json(*runtime).dump(), "application/json");
+    });
+
+    svr.Post("/sdcpp/v1/img_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, 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;
+            std::string error_message;
+            if (!parse_img_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::ImgGen;
+            job->status                             = AsyncJobStatus::Queued;
+            job->created_at                         = unix_timestamp_now();
+            job->img_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.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) {
+        AsyncJobManager& manager = *runtime->async_job_manager;
+        std::lock_guard<std::mutex> lock(manager.mutex);
+        purge_expired_jobs(manager);
+
+        std::string job_id = req.matches[1];
+        auto it            = manager.jobs.find(job_id);
+        if (it == manager.jobs.end()) {
+            if (manager.expired_jobs.find(job_id) != manager.expired_jobs.end()) {
+                res.status = 410;
+                res.set_content(R"({"error":"job expired"})", "application/json");
+            } else {
+                res.status = 404;
+                res.set_content(R"({"error":"job not found"})", "application/json");
+            }
+            return;
+        }
+
+        res.status = 200;
+        res.set_content(make_async_job_json(manager, *it->second).dump(), "application/json");
+    });
+
+    svr.Post(R"(/sdcpp/v1/jobs/([A-Za-z0-9_\-]+)/cancel)", [runtime](const httplib::Request& req, httplib::Response& res) {
+        AsyncJobManager& manager = *runtime->async_job_manager;
+        std::lock_guard<std::mutex> lock(manager.mutex);
+        purge_expired_jobs(manager);
+
+        std::string job_id = req.matches[1];
+        auto it            = manager.jobs.find(job_id);
+        if (it == manager.jobs.end()) {
+            if (manager.expired_jobs.find(job_id) != manager.expired_jobs.end()) {
+                res.status = 410;
+                res.set_content(R"({"error":"job expired"})", "application/json");
+            } else {
+                res.status = 404;
+                res.set_content(R"({"error":"job not found"})", "application/json");
+            }
+            return;
+        }
+
+        auto& job = *it->second;
+        if (job.status == AsyncJobStatus::Queued) {
+            if (!cancel_queued_job(manager, job)) {
+                res.status = 409;
+                res.set_content(R"({"error":"job queue state changed before cancellation"})", "application/json");
+                return;
+            }
+            res.status = 200;
+            res.set_content(make_async_job_json(manager, job).dump(), "application/json");
+            return;
+        }
+
+        if (job.status == AsyncJobStatus::Generating) {
+            res.status = 409;
+            res.set_content(R"({"error":"job is currently generating and cannot be interrupted yet"})", "application/json");
+            return;
+        }
+
+        res.status = 200;
+        res.set_content(make_async_job_json(manager, job).dump(), "application/json");
+    });
+}

examples/server/runtime.cpp

@@ -0,0 +1,332 @@
+#include "runtime.h"
+
+#include <algorithm>
+#include <cctype>
+#include <chrono>
+#include <cstdlib>
+#include <filesystem>
+#include <mutex>
+#include <regex>
+#include <sstream>
+
+#include "common/common.h"
+#include "common/log.h"
+
+namespace fs = std::filesystem;
+
+static std::string lower_ascii(std::string value) {
+    std::transform(value.begin(), value.end(), value.begin(), [](unsigned char c) {
+        return static_cast<char>(std::tolower(c));
+    });
+    return value;
+}
+
+static bool is_supported_model_ext(const fs::path& p) {
+    auto ext = lower_ascii(p.extension().string());
+    return ext == ".gguf" || ext == ".pt" || ext == ".pth" || ext == ".safetensors";
+}
+
+static const std::string k_base64_chars =
+    "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+    "abcdefghijklmnopqrstuvwxyz"
+    "0123456789+/";
+
+std::string base64_encode(const std::vector<uint8_t>& bytes) {
+    std::string ret;
+    int val  = 0;
+    int valb = -6;
+    for (uint8_t c : bytes) {
+        val = (val << 8) + c;
+        valb += 8;
+        while (valb >= 0) {
+            ret.push_back(k_base64_chars[(val >> valb) & 0x3F]);
+            valb -= 6;
+        }
+    }
+    if (valb > -6) {
+        ret.push_back(k_base64_chars[((val << 8) >> (valb + 8)) & 0x3F]);
+    }
+    while (ret.size() % 4) {
+        ret.push_back('=');
+    }
+    return ret;
+}
+
+std::string normalize_output_format(std::string output_format) {
+    std::transform(output_format.begin(), output_format.end(), output_format.begin(),
+                   [](unsigned char c) { return static_cast<char>(std::tolower(c)); });
+    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,
+                           bool allow_webp,
+                           std::string& error_message) {
+    request.output_format      = normalize_output_format(std::move(output_format));
+    request.output_compression = std::clamp(output_compression, 0, 100);
+
+    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 = "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;
+
+    options.string_options = {
+        {"-l", "--listen-ip", "server listen ip (default: 127.0.0.1)", &listen_ip},
+        {"", "--serve-html-path", "path to HTML file to serve at root (optional)", &serve_html_path},
+    };
+
+    options.int_options = {
+        {"", "--listen-port", "server listen port (default: 1234)", &listen_port},
+    };
+
+    options.bool_options = {
+        {"-v", "--verbose", "print extra info", true, &verbose},
+        {"", "--color", "colors the logging tags according to level", true, &color},
+    };
+
+    auto on_help_arg = [&](int, const char**, int) {
+        normal_exit = true;
+        return -1;
+    };
+
+    options.manual_options = {
+        {"-h", "--help", "show this help message and exit", on_help_arg},
+    };
+    return options;
+}
+
+bool SDSvrParams::validate() {
+    if (listen_ip.empty()) {
+        LOG_ERROR("error: the following arguments are required: listen_ip");
+        return false;
+    }
+
+    if (listen_port < 0 || listen_port > 65535) {
+        LOG_ERROR("error: listen_port should be in the range [0, 65535]");
+        return false;
+    }
+
+    if (!serve_html_path.empty() && !fs::exists(serve_html_path)) {
+        LOG_ERROR("error: serve_html_path file does not exist: %s", serve_html_path.c_str());
+        return false;
+    }
+    return true;
+}
+
+bool SDSvrParams::resolve_and_validate() {
+    if (!validate()) {
+        return false;
+    }
+    return true;
+}
+
+std::string SDSvrParams::to_string() const {
+    std::ostringstream oss;
+    oss << "SDSvrParams {\n"
+        << "  listen_ip: " << listen_ip << ",\n"
+        << "  listen_port: \"" << listen_port << "\",\n"
+        << "  serve_html_path: \"" << serve_html_path << "\",\n"
+        << "}";
+    return oss.str();
+}
+
+void refresh_lora_cache(ServerRuntime& rt) {
+    std::vector<LoraEntry> new_cache;
+
+    fs::path lora_dir = rt.ctx_params->lora_model_dir;
+    if (fs::exists(lora_dir) && fs::is_directory(lora_dir)) {
+        for (auto& entry : fs::recursive_directory_iterator(lora_dir)) {
+            if (!entry.is_regular_file()) {
+                continue;
+            }
+            const fs::path& p = entry.path();
+            if (!is_supported_model_ext(p)) {
+                continue;
+            }
+
+            LoraEntry lora_entry;
+            lora_entry.name     = p.stem().u8string();
+            lora_entry.fullpath = p.u8string();
+            std::string rel     = p.lexically_relative(lora_dir).u8string();
+            std::replace(rel.begin(), rel.end(), '\\', '/');
+            lora_entry.path = rel;
+
+            new_cache.push_back(std::move(lora_entry));
+        }
+    }
+
+    std::sort(new_cache.begin(), new_cache.end(), [](const LoraEntry& a, const LoraEntry& b) {
+        return a.path < b.path;
+    });
+
+    {
+        std::lock_guard<std::mutex> lock(*rt.lora_mutex);
+        *rt.lora_cache = std::move(new_cache);
+    }
+}
+
+std::string get_lora_full_path(ServerRuntime& rt, const std::string& path) {
+    std::lock_guard<std::mutex> lock(*rt.lora_mutex);
+    auto it = std::find_if(rt.lora_cache->begin(), rt.lora_cache->end(),
+                           [&](const LoraEntry& entry) { return entry.path == path; });
+    return it != rt.lora_cache->end() ? it->fullpath : "";
+}
+
+void refresh_upscaler_cache(ServerRuntime& rt) {
+    std::vector<UpscalerEntry> new_cache;
+
+    fs::path upscaler_dir = rt.ctx_params->hires_upscalers_dir;
+    if (fs::exists(upscaler_dir) && fs::is_directory(upscaler_dir)) {
+        for (auto& entry : fs::directory_iterator(upscaler_dir)) {
+            if (!entry.is_regular_file()) {
+                continue;
+            }
+            const fs::path& p = entry.path();
+            if (!is_supported_model_ext(p)) {
+                continue;
+            }
+
+            UpscalerEntry upscaler_entry;
+            upscaler_entry.name       = p.stem().u8string();
+            upscaler_entry.fullpath   = fs::absolute(p).lexically_normal().u8string();
+            upscaler_entry.model_name = "ESRGAN_4x";
+            upscaler_entry.path       = p.filename().u8string();
+
+            new_cache.push_back(std::move(upscaler_entry));
+        }
+    }
+
+    std::sort(new_cache.begin(), new_cache.end(), [](const UpscalerEntry& a, const UpscalerEntry& b) {
+        return a.name < b.name;
+    });
+
+    {
+        std::lock_guard<std::mutex> lock(*rt.upscaler_mutex);
+        *rt.upscaler_cache = std::move(new_cache);
+    }
+}
+
+int64_t unix_timestamp_now() {
+    return std::chrono::duration_cast<std::chrono::seconds>(
+               std::chrono::system_clock::now().time_since_epoch())
+        .count();
+}

examples/server/runtime.h

@@ -0,0 +1,100 @@
+#pragma once
+
+#include <algorithm>
+#include <cstdint>
+#include <mutex>
+#include <string>
+#include <vector>
+
+#include <json.hpp>
+#include "common/common.h"
+#include "common/resource_owners.hpp"
+#include "stable-diffusion.h"
+
+using json = nlohmann::json;
+
+struct ArgOptions;
+struct SDContextParams;
+struct AsyncJobManager;
+
+struct SDSvrParams {
+    std::string listen_ip = "127.0.0.1";
+    int listen_port       = 1234;
+    std::string serve_html_path;
+    bool normal_exit = false;
+    bool verbose     = false;
+    bool color       = false;
+
+    ArgOptions get_options();
+    bool validate();
+    bool resolve_and_validate();
+    std::string to_string() const;
+};
+
+struct LoraEntry {
+    std::string name;
+    std::string path;
+    std::string fullpath;
+};
+
+struct UpscalerEntry {
+    std::string name;
+    std::string path;
+    std::string fullpath;
+    std::string model_name;
+    int scale = 4;
+};
+
+struct ServerRuntime {
+    sd_ctx_t* sd_ctx;
+    std::mutex* sd_ctx_mutex;
+    const SDSvrParams* svr_params;
+    const SDContextParams* ctx_params;
+    const SDGenerationParams* default_gen_params;
+    std::vector<LoraEntry>* lora_cache;
+    std::mutex* lora_mutex;
+    std::vector<UpscalerEntry>* upscaler_cache;
+    std::mutex* upscaler_mutex;
+    AsyncJobManager* async_job_manager;
+};
+
+struct ImgGenJobRequest {
+    SDGenerationParams gen_params;
+    std::string output_format = "png";
+    int output_compression    = 100;
+
+    sd_img_gen_params_t to_sd_img_gen_params_t() {
+        return gen_params.to_sd_img_gen_params_t();
+    }
+};
+
+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);
+void refresh_upscaler_cache(ServerRuntime& rt);
+int64_t unix_timestamp_now();

format-code.sh

@@ -1,4 +1,6 @@
-for f in src/*.cpp src/*.h src/*.hpp src/vocab/*.h src/vocab/*.cpp examples/cli/*.cpp examples/common/*.hpp examples/cli/*.h examples/server/*.cpp; do
+for f in src/*.cpp src/*.h src/*.hpp src/tokenizers/*.h src/tokenizers/*.cpp src/tokenizers/vocab/*.h src/tokenizers/vocab/*.cpp \
+         src/model_io/*.h src/model_io/*.cpp examples/cli/*.cpp examples/cli/*.h examples/server/*.cpp \
+         examples/common/*.hpp examples/common/*.h examples/common/*.cpp; do
   [[ "$f" == vocab* ]] && continue
   echo "formatting '$f'"
   # if [ "$f" != "stable-diffusion.h" ]; then

ggml

@@ -1 +1 @@
-Subproject commit a8db410a252c8c8f2d120c6f2e7133ebe032f35d
+Subproject commit 404fcb9d7c96989569e68c9e7881ee3465a05c50

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
 };
 
@@ -120,7 +121,8 @@ enum sd_type_t {
     // SD_TYPE_IQ4_NL_4_8 = 37,
     // SD_TYPE_IQ4_NL_8_8 = 38,
     SD_TYPE_MXFP4 = 39,  // MXFP4 (1 block)
-    SD_TYPE_COUNT = 40,
+    SD_TYPE_NVFP4 = 40,  // NVFP4 (4 blocks, E4M3 scale)
+    SD_TYPE_COUNT = 41,
 };
 
 enum sd_log_level_t {
@@ -201,6 +203,7 @@ typedef struct {
     bool chroma_use_t5_mask;
     int chroma_t5_mask_pad;
     bool qwen_image_zero_cond_t;
+    float max_vram;
 } sd_ctx_params_t;
 
 typedef struct {
@@ -287,6 +290,32 @@ typedef struct {
     const char* path;
 } sd_lora_t;
 
+enum sd_hires_upscaler_t {
+    SD_HIRES_UPSCALER_NONE,
+    SD_HIRES_UPSCALER_LATENT,
+    SD_HIRES_UPSCALER_LATENT_NEAREST,
+    SD_HIRES_UPSCALER_LATENT_NEAREST_EXACT,
+    SD_HIRES_UPSCALER_LATENT_ANTIALIASED,
+    SD_HIRES_UPSCALER_LATENT_BICUBIC,
+    SD_HIRES_UPSCALER_LATENT_BICUBIC_ANTIALIASED,
+    SD_HIRES_UPSCALER_LANCZOS,
+    SD_HIRES_UPSCALER_NEAREST,
+    SD_HIRES_UPSCALER_MODEL,
+    SD_HIRES_UPSCALER_COUNT,
+};
+
+typedef struct {
+    bool enabled;
+    enum sd_hires_upscaler_t upscaler;
+    const char* model_path;
+    float scale;
+    int target_width;
+    int target_height;
+    int steps;
+    float denoising_strength;
+    int upscale_tile_size;
+} sd_hires_params_t;
+
 typedef struct {
     const sd_lora_t* loras;
     uint32_t lora_count;
@@ -310,6 +339,7 @@ typedef struct {
     sd_pm_params_t pm_params;
     sd_tiling_params_t vae_tiling_params;
     sd_cache_params_t cache;
+    sd_hires_params_t hires;
 } sd_img_gen_params_t;
 
 typedef struct {
@@ -346,6 +376,8 @@ 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);
@@ -361,8 +393,11 @@ SD_API const char* sd_preview_name(enum preview_t preview);
 SD_API enum preview_t str_to_preview(const char* str);
 SD_API const char* sd_lora_apply_mode_name(enum lora_apply_mode_t mode);
 SD_API enum lora_apply_mode_t str_to_lora_apply_mode(const char* str);
+SD_API const char* sd_hires_upscaler_name(enum sd_hires_upscaler_t upscaler);
+SD_API enum sd_hires_upscaler_t str_to_sd_hires_upscaler(const char* str);
 
 SD_API void sd_cache_params_init(sd_cache_params_t* cache_params);
+SD_API void sd_hires_params_init(sd_hires_params_t* hires_params);
 
 SD_API void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params);
 SD_API char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params);

src/anima.hpp

@@ -499,9 +499,15 @@ namespace Anima {
                 encoder_hidden_states = adapted_context;
             }
 
+            sd::ggml_graph_cut::mark_graph_cut(x, "anima.prelude", "x");
+            sd::ggml_graph_cut::mark_graph_cut(embedded_timestep, "anima.prelude", "embedded_timestep");
+            sd::ggml_graph_cut::mark_graph_cut(temb, "anima.prelude", "temb");
+            sd::ggml_graph_cut::mark_graph_cut(encoder_hidden_states, "anima.prelude", "context");
+
             for (int i = 0; i < num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<TransformerBlock>(blocks["blocks." + std::to_string(i)]);
                 x          = block->forward(ctx, x, encoder_hidden_states, embedded_timestep, temb, image_pe);
+                sd::ggml_graph_cut::mark_graph_cut(x, "anima.blocks." + std::to_string(i), "x");
             }
 
             x = final_layer->forward(ctx, x, embedded_timestep, temb);  // [N, h*w, ph*pw*C]
@@ -602,20 +608,19 @@ namespace Anima {
             return Rope::embed_nd(ids, bs, axis_thetas, axes_dim);
         }
 
-        ggml_cgraph* build_graph(ggml_tensor* x,
-                                 ggml_tensor* timesteps,
-                                 ggml_tensor* context,
-                                 ggml_tensor* t5_ids     = nullptr,
-                                 ggml_tensor* t5_weights = nullptr) {
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor    = {},
+                                 const sd::Tensor<int32_t>& t5_ids_tensor   = {},
+                                 const sd::Tensor<float>& t5_weights_tensor = {}) {
+            ggml_tensor* x          = make_input(x_tensor);
+            ggml_tensor* timesteps  = make_input(timesteps_tensor);
+            ggml_tensor* context    = make_optional_input(context_tensor);
+            ggml_tensor* t5_ids     = make_optional_input(t5_ids_tensor);
+            ggml_tensor* t5_weights = make_optional_input(t5_weights_tensor);
             GGML_ASSERT(x->ne[3] == 1);
             ggml_cgraph* gf = new_graph_custom(ANIMA_GRAPH_SIZE);
 
-            x          = to_backend(x);
-            timesteps  = to_backend(timesteps);
-            context    = to_backend(context);
-            t5_ids     = to_backend(t5_ids);
-            t5_weights = to_backend(t5_weights);
-
             int64_t pad_h = (net.patch_size - x->ne[1] % net.patch_size) % net.patch_size;
             int64_t pad_w = (net.patch_size - x->ne[0] % net.patch_size) % net.patch_size;
             int64_t h_pad = x->ne[1] + pad_h;
@@ -667,18 +672,16 @@ namespace Anima {
             return gf;
         }
 
-        bool compute(int n_threads,
-                     ggml_tensor* x,
-                     ggml_tensor* timesteps,
-                     ggml_tensor* context,
-                     ggml_tensor* t5_ids      = nullptr,
-                     ggml_tensor* t5_weights  = nullptr,
-                     ggml_tensor** output     = nullptr,
-                     ggml_context* output_ctx = nullptr) {
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context    = {},
+                                  const sd::Tensor<int32_t>& t5_ids   = {},
+                                  const sd::Tensor<float>& t5_weights = {}) {
             auto get_graph = [&]() -> ggml_cgraph* {
                 return build_graph(x, timesteps, context, t5_ids, t5_weights);
             };
-            return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
         }
     };
 }  // namespace Anima

src/auto_encoder_kl.hpp

@@ -1,4 +1,4 @@
-#ifndef __AUTO_ENCODER_KL_HPP__
+#ifndef __AUTO_ENCODER_KL_HPP__
 #define __AUTO_ENCODER_KL_HPP__
 
 #include "vae.hpp"
@@ -328,6 +328,7 @@ public:
         auto conv_out    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_out"]);
 
         auto h = conv_in->forward(ctx, x);  // [N, ch, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.prelude", "h");
 
         // downsampling
         size_t num_resolutions = ch_mult.size();
@@ -337,12 +338,14 @@ public:
                 auto down_block  = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
 
                 h = down_block->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.down." + std::to_string(i) + ".block." + std::to_string(j), "h");
             }
             if (i != num_resolutions - 1) {
                 std::string name = "down." + std::to_string(i) + ".downsample";
                 auto down_sample = std::dynamic_pointer_cast<DownSampleBlock>(blocks[name]);
 
                 h = down_sample->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.down." + std::to_string(i) + ".downsample", "h");
             }
         }
 
@@ -350,6 +353,7 @@ public:
         h = mid_block_1->forward(ctx, h);
         h = mid_attn_1->forward(ctx, h);
         h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.mid", "h");
 
         // end
         h = norm_out->forward(ctx, h);
@@ -450,6 +454,7 @@ public:
 
         // conv_in
         auto h = conv_in->forward(ctx, z);  // [N, block_in, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.prelude", "h");
 
         // middle
         h = mid_block_1->forward(ctx, h);
@@ -457,6 +462,7 @@ public:
 
         h = mid_attn_1->forward(ctx, h);
         h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.mid", "h");
 
         // upsampling
         int num_resolutions = static_cast<int>(ch_mult.size());
@@ -466,12 +472,14 @@ public:
                 auto up_block    = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
 
                 h = up_block->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.up." + std::to_string(i) + ".block." + std::to_string(j), "h");
             }
             if (i != 0) {
                 std::string name = "up." + std::to_string(i) + ".upsample";
                 auto up_sample   = std::dynamic_pointer_cast<UpSampleBlock>(blocks[name]);
 
                 h = up_sample->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.up." + std::to_string(i) + ".upsample", "h");
             }
         }
 
@@ -501,14 +509,39 @@ protected:
         bool double_z            = true;
     } dd_config;
 
+    static std::string get_tensor_name(const std::string& prefix, const std::string& name) {
+        return prefix.empty() ? name : prefix + "." + name;
+    }
+
+    void detect_decoder_ch(const String2TensorStorage& tensor_storage_map,
+                           const std::string& prefix,
+                           int& decoder_ch) {
+        auto conv_in_iter = tensor_storage_map.find(get_tensor_name(prefix, "decoder.conv_in.weight"));
+        if (conv_in_iter != tensor_storage_map.end() && conv_in_iter->second.n_dims >= 4 && conv_in_iter->second.ne[3] > 0) {
+            int last_ch_mult             = dd_config.ch_mult.back();
+            int64_t conv_in_out_channels = conv_in_iter->second.ne[3];
+            if (last_ch_mult > 0 && conv_in_out_channels % last_ch_mult == 0) {
+                decoder_ch = static_cast<int>(conv_in_out_channels / last_ch_mult);
+                LOG_INFO("vae decoder: ch = %d", decoder_ch);
+            } else {
+                LOG_WARN("vae decoder: failed to infer ch from %s (%" PRId64 " / %d)",
+                         get_tensor_name(prefix, "decoder.conv_in.weight").c_str(),
+                         conv_in_out_channels,
+                         last_ch_mult);
+            }
+        }
+    }
+
 public:
-    AutoEncoderKLModel(SDVersion version          = VERSION_SD1,
-                       bool decode_only           = true,
-                       bool use_linear_projection = false,
-                       bool use_video_decoder     = false)
+    AutoEncoderKLModel(SDVersion version                              = VERSION_SD1,
+                       bool decode_only                               = true,
+                       bool use_linear_projection                     = false,
+                       bool use_video_decoder                         = false,
+                       const String2TensorStorage& tensor_storage_map = {},
+                       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 {
@@ -519,7 +552,9 @@ public:
         if (use_video_decoder) {
             use_quant = false;
         }
-        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new Decoder(dd_config.ch,
+        int decoder_ch = dd_config.ch;
+        detect_decoder_ch(tensor_storage_map, prefix, decoder_ch);
+        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new Decoder(decoder_ch,
                                                                    dd_config.out_ch,
                                                                    dd_config.ch_mult,
                                                                    dd_config.num_res_blocks,
@@ -551,7 +586,7 @@ 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;
 
@@ -572,6 +607,7 @@ public:
         if (use_quant) {
             auto post_quant_conv = std::dynamic_pointer_cast<Conv2d>(blocks["post_quant_conv"]);
             z                    = post_quant_conv->forward(ctx, z);  // [N, z_channels, h, w]
+            // sd::ggml_graph_cut::mark_graph_cut(z, "vae.decode.prelude", "z");
         }
         auto decoder = std::dynamic_pointer_cast<Decoder>(blocks["decoder"]);
 
@@ -589,8 +625,9 @@ public:
         if (use_quant) {
             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]
+            // sd::ggml_graph_cut::mark_graph_cut(z, "vae.encode.final", "z");
         }
-        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]
@@ -613,7 +650,7 @@ 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;
@@ -646,7 +683,7 @@ 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;
         }
@@ -662,7 +699,7 @@ struct AutoEncoderKL : public VAE {
                 break;
             }
         }
-        ae = AutoEncoderKLModel(version, decode_only, use_linear_projection, use_video_decoder);
+        ae = AutoEncoderKLModel(version, decode_only, use_linear_projection, use_video_decoder, tensor_storage_map, prefix);
         ae.init(params_ctx, tensor_storage_map, prefix);
     }
 
@@ -685,10 +722,9 @@ struct AutoEncoderKL : public VAE {
         ae.get_param_tensors(tensors, prefix);
     }
 
-    ggml_cgraph* build_graph(ggml_tensor* z, bool decode_graph) {
+    ggml_cgraph* build_graph(const sd::Tensor<float>& z_tensor, bool decode_graph) {
         ggml_cgraph* gf = ggml_new_graph(compute_ctx);
-
-        z = to_backend(z);
+        ggml_tensor* z  = make_input(z_tensor);
 
         auto runner_ctx = get_context();
 
@@ -699,184 +735,100 @@ struct AutoEncoderKL : public VAE {
         return gf;
     }
 
-    bool _compute(const int n_threads,
-                  ggml_tensor* z,
-                  bool decode_graph,
-                  ggml_tensor** output,
-                  ggml_context* output_ctx = nullptr) override {
+    sd::Tensor<float> _compute(const int n_threads,
+                               const sd::Tensor<float>& z,
+                               bool decode_graph) override {
         GGML_ASSERT(!decode_only || decode_graph);
         auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(z, decode_graph);
         };
-        // ggml_set_f32(z, 0.5f);
-        // print_ggml_tensor(z);
-        return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), z.dim());
     }
 
-    ggml_tensor* gaussian_latent_sample(ggml_context* work_ctx, ggml_tensor* moments, std::shared_ptr<RNG> rng) {
+    sd::Tensor<float> gaussian_latent_sample(const sd::Tensor<float>& moments, std::shared_ptr<RNG> rng) {
         // ldm.modules.distributions.distributions.DiagonalGaussianDistribution.sample
-        ggml_tensor* latents = ggml_new_tensor_4d(work_ctx, moments->type, moments->ne[0], moments->ne[1], moments->ne[2] / 2, moments->ne[3]);
-        ggml_tensor* noise   = ggml_dup_tensor(work_ctx, latents);
-        ggml_ext_im_set_randn_f32(noise, rng);
-        {
-            float mean   = 0;
-            float logvar = 0;
-            float value  = 0;
-            float std_   = 0;
-            for (int i = 0; i < latents->ne[3]; i++) {
-                for (int j = 0; j < latents->ne[2]; j++) {
-                    for (int k = 0; k < latents->ne[1]; k++) {
-                        for (int l = 0; l < latents->ne[0]; l++) {
-                            mean   = ggml_ext_tensor_get_f32(moments, l, k, j, i);
-                            logvar = ggml_ext_tensor_get_f32(moments, l, k, j + (int)latents->ne[2], i);
-                            logvar = std::max(-30.0f, std::min(logvar, 20.0f));
-                            std_   = std::exp(0.5f * logvar);
-                            value  = mean + std_ * ggml_ext_tensor_get_f32(noise, l, k, j, i);
-                            // printf("%d %d %d %d -> %f\n", i, j, k, l, value);
-                            ggml_ext_tensor_set_f32(latents, value, l, k, j, i);
-                        }
-                    }
-                }
-            }
-        }
+        auto chunks               = sd::ops::chunk(moments, 2, 2);
+        const auto& mean          = chunks[0];
+        const auto& logvar        = chunks[1];
+        sd::Tensor<float> stddev  = sd::ops::exp(0.5f * sd::ops::clamp(logvar, -30.0f, 20.0f));
+        sd::Tensor<float> noise   = sd::Tensor<float>::randn_like(mean, rng);
+        sd::Tensor<float> latents = mean + stddev * noise;
         return latents;
     }
 
-    ggml_tensor* vae_output_to_latents(ggml_context* work_ctx, ggml_tensor* vae_output, std::shared_ptr<RNG> rng) {
-        if (sd_version_is_flux2(version)) {
+    sd::Tensor<float> vae_output_to_latents(const sd::Tensor<float>& vae_output, std::shared_ptr<RNG> rng) override {
+        if (sd_version_uses_flux2_vae(version)) {
             return vae_output;
         } else if (version == VERSION_SD1_PIX2PIX) {
-            return ggml_view_3d(work_ctx,
-                                vae_output,
-                                vae_output->ne[0],
-                                vae_output->ne[1],
-                                vae_output->ne[2] / 2,
-                                vae_output->nb[1],
-                                vae_output->nb[2],
-                                0);
+            return sd::ops::chunk(vae_output, 2, 2)[0];
         } else {
-            return gaussian_latent_sample(work_ctx, vae_output, rng);
+            return gaussian_latent_sample(vae_output, rng);
         }
     }
 
-    void get_latents_mean_std_vec(ggml_tensor* latents, int channel_dim, std::vector<float>& latents_mean_vec, std::vector<float>& latents_std_vec) {
-        // flux2
-        if (sd_version_is_flux2(version)) {
-            GGML_ASSERT(latents->ne[channel_dim] == 128);
-            latents_mean_vec = {-0.0676f, -0.0715f, -0.0753f, -0.0745f, 0.0223f, 0.0180f, 0.0142f, 0.0184f,
-                                -0.0001f, -0.0063f, -0.0002f, -0.0031f, -0.0272f, -0.0281f, -0.0276f, -0.0290f,
-                                -0.0769f, -0.0672f, -0.0902f, -0.0892f, 0.0168f, 0.0152f, 0.0079f, 0.0086f,
-                                0.0083f, 0.0015f, 0.0003f, -0.0043f, -0.0439f, -0.0419f, -0.0438f, -0.0431f,
-                                -0.0102f, -0.0132f, -0.0066f, -0.0048f, -0.0311f, -0.0306f, -0.0279f, -0.0180f,
-                                0.0030f, 0.0015f, 0.0126f, 0.0145f, 0.0347f, 0.0338f, 0.0337f, 0.0283f,
-                                0.0020f, 0.0047f, 0.0047f, 0.0050f, 0.0123f, 0.0081f, 0.0081f, 0.0146f,
-                                0.0681f, 0.0679f, 0.0767f, 0.0732f, -0.0462f, -0.0474f, -0.0392f, -0.0511f,
-                                -0.0528f, -0.0477f, -0.0470f, -0.0517f, -0.0317f, -0.0316f, -0.0345f, -0.0283f,
-                                0.0510f, 0.0445f, 0.0578f, 0.0458f, -0.0412f, -0.0458f, -0.0487f, -0.0467f,
-                                -0.0088f, -0.0106f, -0.0088f, -0.0046f, -0.0376f, -0.0432f, -0.0436f, -0.0499f,
-                                0.0118f, 0.0166f, 0.0203f, 0.0279f, 0.0113f, 0.0129f, 0.0016f, 0.0072f,
-                                -0.0118f, -0.0018f, -0.0141f, -0.0054f, -0.0091f, -0.0138f, -0.0145f, -0.0187f,
-                                0.0323f, 0.0305f, 0.0259f, 0.0300f, 0.0540f, 0.0614f, 0.0495f, 0.0590f,
-                                -0.0511f, -0.0603f, -0.0478f, -0.0524f, -0.0227f, -0.0274f, -0.0154f, -0.0255f,
-                                -0.0572f, -0.0565f, -0.0518f, -0.0496f, 0.0116f, 0.0054f, 0.0163f, 0.0104f};
-            latents_std_vec  = {
-                 1.8029f, 1.7786f, 1.7868f, 1.7837f, 1.7717f, 1.7590f, 1.7610f, 1.7479f,
-                 1.7336f, 1.7373f, 1.7340f, 1.7343f, 1.8626f, 1.8527f, 1.8629f, 1.8589f,
-                 1.7593f, 1.7526f, 1.7556f, 1.7583f, 1.7363f, 1.7400f, 1.7355f, 1.7394f,
-                 1.7342f, 1.7246f, 1.7392f, 1.7304f, 1.7551f, 1.7513f, 1.7559f, 1.7488f,
-                 1.8449f, 1.8454f, 1.8550f, 1.8535f, 1.8240f, 1.7813f, 1.7854f, 1.7945f,
-                 1.8047f, 1.7876f, 1.7695f, 1.7676f, 1.7782f, 1.7667f, 1.7925f, 1.7848f,
-                 1.7579f, 1.7407f, 1.7483f, 1.7368f, 1.7961f, 1.7998f, 1.7920f, 1.7925f,
-                 1.7780f, 1.7747f, 1.7727f, 1.7749f, 1.7526f, 1.7447f, 1.7657f, 1.7495f,
-                 1.7775f, 1.7720f, 1.7813f, 1.7813f, 1.8162f, 1.8013f, 1.8023f, 1.8033f,
-                 1.7527f, 1.7331f, 1.7563f, 1.7482f, 1.7610f, 1.7507f, 1.7681f, 1.7613f,
-                 1.7665f, 1.7545f, 1.7828f, 1.7726f, 1.7896f, 1.7999f, 1.7864f, 1.7760f,
-                 1.7613f, 1.7625f, 1.7560f, 1.7577f, 1.7783f, 1.7671f, 1.7810f, 1.7799f,
-                 1.7201f, 1.7068f, 1.7265f, 1.7091f, 1.7793f, 1.7578f, 1.7502f, 1.7455f,
-                 1.7587f, 1.7500f, 1.7525f, 1.7362f, 1.7616f, 1.7572f, 1.7444f, 1.7430f,
-                 1.7509f, 1.7610f, 1.7634f, 1.7612f, 1.7254f, 1.7135f, 1.7321f, 1.7226f,
-                 1.7664f, 1.7624f, 1.7718f, 1.7664f, 1.7457f, 1.7441f, 1.7569f, 1.7530f};
+    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_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];
+
+            auto mean_tensor = sd::Tensor<float>::from_vector({-0.0676f, -0.0715f, -0.0753f, -0.0745f, 0.0223f, 0.0180f, 0.0142f, 0.0184f,
+                                                               -0.0001f, -0.0063f, -0.0002f, -0.0031f, -0.0272f, -0.0281f, -0.0276f, -0.0290f,
+                                                               -0.0769f, -0.0672f, -0.0902f, -0.0892f, 0.0168f, 0.0152f, 0.0079f, 0.0086f,
+                                                               0.0083f, 0.0015f, 0.0003f, -0.0043f, -0.0439f, -0.0419f, -0.0438f, -0.0431f,
+                                                               -0.0102f, -0.0132f, -0.0066f, -0.0048f, -0.0311f, -0.0306f, -0.0279f, -0.0180f,
+                                                               0.0030f, 0.0015f, 0.0126f, 0.0145f, 0.0347f, 0.0338f, 0.0337f, 0.0283f,
+                                                               0.0020f, 0.0047f, 0.0047f, 0.0050f, 0.0123f, 0.0081f, 0.0081f, 0.0146f,
+                                                               0.0681f, 0.0679f, 0.0767f, 0.0732f, -0.0462f, -0.0474f, -0.0392f, -0.0511f,
+                                                               -0.0528f, -0.0477f, -0.0470f, -0.0517f, -0.0317f, -0.0316f, -0.0345f, -0.0283f,
+                                                               0.0510f, 0.0445f, 0.0578f, 0.0458f, -0.0412f, -0.0458f, -0.0487f, -0.0467f,
+                                                               -0.0088f, -0.0106f, -0.0088f, -0.0046f, -0.0376f, -0.0432f, -0.0436f, -0.0499f,
+                                                               0.0118f, 0.0166f, 0.0203f, 0.0279f, 0.0113f, 0.0129f, 0.0016f, 0.0072f,
+                                                               -0.0118f, -0.0018f, -0.0141f, -0.0054f, -0.0091f, -0.0138f, -0.0145f, -0.0187f,
+                                                               0.0323f, 0.0305f, 0.0259f, 0.0300f, 0.0540f, 0.0614f, 0.0495f, 0.0590f,
+                                                               -0.0511f, -0.0603f, -0.0478f, -0.0524f, -0.0227f, -0.0274f, -0.0154f, -0.0255f,
+                                                               -0.0572f, -0.0565f, -0.0518f, -0.0496f, 0.0116f, 0.0054f, 0.0163f, 0.0104f});
+            mean_tensor.reshape_(stats_shape);
+            auto std_tensor = sd::Tensor<float>::from_vector({1.8029f, 1.7786f, 1.7868f, 1.7837f, 1.7717f, 1.7590f, 1.7610f, 1.7479f,
+                                                              1.7336f, 1.7373f, 1.7340f, 1.7343f, 1.8626f, 1.8527f, 1.8629f, 1.8589f,
+                                                              1.7593f, 1.7526f, 1.7556f, 1.7583f, 1.7363f, 1.7400f, 1.7355f, 1.7394f,
+                                                              1.7342f, 1.7246f, 1.7392f, 1.7304f, 1.7551f, 1.7513f, 1.7559f, 1.7488f,
+                                                              1.8449f, 1.8454f, 1.8550f, 1.8535f, 1.8240f, 1.7813f, 1.7854f, 1.7945f,
+                                                              1.8047f, 1.7876f, 1.7695f, 1.7676f, 1.7782f, 1.7667f, 1.7925f, 1.7848f,
+                                                              1.7579f, 1.7407f, 1.7483f, 1.7368f, 1.7961f, 1.7998f, 1.7920f, 1.7925f,
+                                                              1.7780f, 1.7747f, 1.7727f, 1.7749f, 1.7526f, 1.7447f, 1.7657f, 1.7495f,
+                                                              1.7775f, 1.7720f, 1.7813f, 1.7813f, 1.8162f, 1.8013f, 1.8023f, 1.8033f,
+                                                              1.7527f, 1.7331f, 1.7563f, 1.7482f, 1.7610f, 1.7507f, 1.7681f, 1.7613f,
+                                                              1.7665f, 1.7545f, 1.7828f, 1.7726f, 1.7896f, 1.7999f, 1.7864f, 1.7760f,
+                                                              1.7613f, 1.7625f, 1.7560f, 1.7577f, 1.7783f, 1.7671f, 1.7810f, 1.7799f,
+                                                              1.7201f, 1.7068f, 1.7265f, 1.7091f, 1.7793f, 1.7578f, 1.7502f, 1.7455f,
+                                                              1.7587f, 1.7500f, 1.7525f, 1.7362f, 1.7616f, 1.7572f, 1.7444f, 1.7430f,
+                                                              1.7509f, 1.7610f, 1.7634f, 1.7612f, 1.7254f, 1.7135f, 1.7321f, 1.7226f,
+                                                              1.7664f, 1.7624f, 1.7718f, 1.7664f, 1.7457f, 1.7441f, 1.7569f, 1.7530f});
+            std_tensor.reshape_(stats_shape);
+            return {std::move(mean_tensor), std::move(std_tensor)};
         } else {
             GGML_ABORT("unknown version %d", version);
         }
     }
 
-    ggml_tensor* diffusion_to_vae_latents(ggml_context* work_ctx, ggml_tensor* latents) {
-        ggml_tensor* vae_latents = ggml_dup(work_ctx, latents);
-        if (sd_version_is_flux2(version)) {
-            int channel_dim = 2;
-            std::vector<float> latents_mean_vec;
-            std::vector<float> latents_std_vec;
-            get_latents_mean_std_vec(latents, channel_dim, latents_mean_vec, latents_std_vec);
-
-            float mean;
-            float std_;
-            for (int i = 0; i < latents->ne[3]; i++) {
-                if (channel_dim == 3) {
-                    mean = latents_mean_vec[i];
-                    std_ = latents_std_vec[i];
-                }
-                for (int j = 0; j < latents->ne[2]; j++) {
-                    if (channel_dim == 2) {
-                        mean = latents_mean_vec[j];
-                        std_ = latents_std_vec[j];
-                    }
-                    for (int k = 0; k < latents->ne[1]; k++) {
-                        for (int l = 0; l < latents->ne[0]; l++) {
-                            float value = ggml_ext_tensor_get_f32(latents, l, k, j, i);
-                            value       = value * std_ / scale_factor + mean;
-                            ggml_ext_tensor_set_f32(vae_latents, value, l, k, j, i);
-                        }
-                    }
-                }
-            }
-        } else {
-            ggml_ext_tensor_iter(latents, [&](ggml_tensor* latents, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
-                float value = ggml_ext_tensor_get_f32(latents, i0, i1, i2, i3);
-                value       = (value / scale_factor) + shift_factor;
-                ggml_ext_tensor_set_f32(vae_latents, value, i0, i1, i2, i3);
-            });
+    sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
+        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;
         }
-        return vae_latents;
+        return (latents / scale_factor) + shift_factor;
     }
 
-    ggml_tensor* vae_to_diffuison_latents(ggml_context* work_ctx, ggml_tensor* latents) {
-        ggml_tensor* diffusion_latents = ggml_dup(work_ctx, latents);
-        if (sd_version_is_flux2(version)) {
-            int channel_dim = 2;
-            std::vector<float> latents_mean_vec;
-            std::vector<float> latents_std_vec;
-            get_latents_mean_std_vec(latents, channel_dim, latents_mean_vec, latents_std_vec);
-
-            float mean;
-            float std_;
-            for (int i = 0; i < latents->ne[3]; i++) {
-                if (channel_dim == 3) {
-                    mean = latents_mean_vec[i];
-                    std_ = latents_std_vec[i];
-                }
-                for (int j = 0; j < latents->ne[2]; j++) {
-                    if (channel_dim == 2) {
-                        mean = latents_mean_vec[j];
-                        std_ = latents_std_vec[j];
-                    }
-                    for (int k = 0; k < latents->ne[1]; k++) {
-                        for (int l = 0; l < latents->ne[0]; l++) {
-                            float value = ggml_ext_tensor_get_f32(latents, l, k, j, i);
-                            value       = (value - mean) * scale_factor / std_;
-                            ggml_ext_tensor_set_f32(diffusion_latents, value, l, k, j, i);
-                        }
-                    }
-                }
-            }
-        } else {
-            ggml_ext_tensor_iter(latents, [&](ggml_tensor* latents, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
-                float value = ggml_ext_tensor_get_f32(latents, i0, i1, i2, i3);
-                value       = (value - shift_factor) * scale_factor;
-                ggml_ext_tensor_set_f32(diffusion_latents, value, i0, i1, i2, i3);
-            });
+    sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents) override {
+        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;
         }
-        return diffusion_latents;
+        return (latents - shift_factor) * scale_factor;
     }
 
     int get_encoder_output_channels(int input_channels) {
@@ -889,24 +841,26 @@ struct AutoEncoderKL : public VAE {
         params.mem_buffer = nullptr;
         params.no_alloc   = false;
 
-        ggml_context* work_ctx = ggml_init(params);
-        GGML_ASSERT(work_ctx != nullptr);
+        ggml_context* ctx = ggml_init(params);
+        GGML_ASSERT(ctx != nullptr);
 
         {
             // CPU, x{1, 3, 64, 64}: Pass
             // CUDA, x{1, 3, 64, 64}: Pass, but sill get wrong result for some image, may be due to interlnal nan
             // CPU, x{2, 3, 64, 64}: Wrong result
             // CUDA, x{2, 3, 64, 64}: Wrong result, and different from CPU result
-            auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 64, 64, 3, 2);
-            ggml_set_f32(x, 0.5f);
-            print_ggml_tensor(x);
-            ggml_tensor* out = nullptr;
+            sd::Tensor<float> x({64, 64, 3, 2});
+            x.fill_(0.5f);
+            print_sd_tensor(x);
+            sd::Tensor<float> out;
 
-            int64_t t0 = ggml_time_ms();
-            _compute(8, x, false, &out, work_ctx);
-            int64_t t1 = ggml_time_ms();
+            int64_t t0   = ggml_time_ms();
+            auto out_opt = _compute(8, x, false);
+            int64_t t1   = ggml_time_ms();
 
-            print_ggml_tensor(out);
+            GGML_ASSERT(!out_opt.empty());
+            out = std::move(out_opt);
+            print_sd_tensor(out);
             LOG_DEBUG("encode test done in %lldms", t1 - t0);
         }
 
@@ -915,16 +869,18 @@ struct AutoEncoderKL : public VAE {
             // CUDA, z{1, 4, 8, 8}: Pass
             // CPU, z{3, 4, 8, 8}: Wrong result
             // CUDA, z{3, 4, 8, 8}: Wrong result, and different from CPU result
-            auto z = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 8, 8, 4, 1);
-            ggml_set_f32(z, 0.5f);
-            print_ggml_tensor(z);
-            ggml_tensor* out = nullptr;
+            sd::Tensor<float> z({8, 8, 4, 1});
+            z.fill_(0.5f);
+            print_sd_tensor(z);
+            sd::Tensor<float> out;
 
-            int64_t t0 = ggml_time_ms();
-            _compute(8, z, true, &out, work_ctx);
-            int64_t t1 = ggml_time_ms();
+            int64_t t0   = ggml_time_ms();
+            auto out_opt = _compute(8, z, true);
+            int64_t t1   = ggml_time_ms();
 
-            print_ggml_tensor(out);
+            GGML_ASSERT(!out_opt.empty());
+            out = std::move(out_opt);
+            print_sd_tensor(out);
             LOG_DEBUG("decode test done in %lldms", t1 - t0);
         }
     };

src/cache_dit.hpp

@@ -8,7 +8,9 @@
 #include <unordered_map>
 #include <vector>
 
+#include "condition_cache_utils.hpp"
 #include "ggml_extend.hpp"
+#include "tensor.hpp"
 
 struct DBCacheConfig {
     bool enabled                        = false;
@@ -771,35 +773,37 @@ struct CacheDitConditionState {
         return it != cache_diffs.end() && !it->second.diff.empty();
     }
 
-    void update_cache(const void* cond, const float* input, const float* output, size_t size) {
+    void update_cache(const void* cond, const sd::Tensor<float>& input, const sd::Tensor<float>& output) {
         CacheEntry& entry = cache_diffs[cond];
-        entry.diff.resize(size);
-        for (size_t i = 0; i < size; i++) {
-            entry.diff[i] = output[i] - input[i];
+        if (!sd::store_condition_cache_diff(&entry.diff, input, output)) {
+            entry.prev_input.clear();
+            entry.prev_output.clear();
+            entry.has_prev = false;
+            return;
         }
 
+        size_t size              = static_cast<size_t>(output.numel());
+        const float* input_data  = input.data();
+        const float* output_data = output.data();
         entry.prev_input.resize(size);
         entry.prev_output.resize(size);
         for (size_t i = 0; i < size; i++) {
-            entry.prev_input[i]  = input[i];
-            entry.prev_output[i] = output[i];
+            entry.prev_input[i]  = input_data[i];
+            entry.prev_output[i] = output_data[i];
         }
         entry.has_prev = true;
     }
 
-    void apply_cache(const void* cond, const float* input, float* output, size_t size) {
+    void apply_cache(const void* cond,
+                     const sd::Tensor<float>& input,
+                     sd::Tensor<float>* output) {
         auto it = cache_diffs.find(cond);
         if (it == cache_diffs.end() || it->second.diff.empty())
             return;
-        if (it->second.diff.size() != size)
-            return;
-
-        for (size_t i = 0; i < size; i++) {
-            output[i] = input[i] + it->second.diff[i];
-        }
+        sd::apply_condition_cache_diff(it->second.diff, input, output);
     }
 
-    bool before_condition(const void* cond, ggml_tensor* input, ggml_tensor* output, float sigma, int step_index) {
+    bool before_condition(const void* cond, const sd::Tensor<float>& input, sd::Tensor<float>* output, float sigma, int step_index) {
         if (!enabled() || step_index < 0)
             return false;
 
@@ -819,8 +823,7 @@ struct CacheDitConditionState {
 
         if (skip_current_step) {
             if (has_cache(cond)) {
-                apply_cache(cond, (float*)input->data, (float*)output->data,
-                            static_cast<size_t>(ggml_nelements(output)));
+                apply_cache(cond, input, output);
                 return true;
             }
             return false;
@@ -833,13 +836,13 @@ struct CacheDitConditionState {
         if (it == cache_diffs.end() || !it->second.has_prev)
             return false;
 
-        size_t ne = static_cast<size_t>(ggml_nelements(input));
+        size_t ne = static_cast<size_t>(input.numel());
         if (it->second.prev_input.size() != ne)
             return false;
 
-        float* input_data = (float*)input->data;
-        float diff        = CacheDitState::calculate_residual_diff(
-                   it->second.prev_input.data(), input_data, ne);
+        const float* input_data = input.data();
+        float diff              = CacheDitState::calculate_residual_diff(
+                         it->second.prev_input.data(), input_data, ne);
 
         float effective_threshold = config.residual_diff_threshold;
         if (config.Fn_compute_blocks > 0) {
@@ -859,7 +862,7 @@ struct CacheDitConditionState {
             cached_steps.push_back(current_step_index);
             continuous_cached_steps++;
             accumulated_residual_diff += diff;
-            apply_cache(cond, input_data, (float*)output->data, ne);
+            apply_cache(cond, input, output);
             return true;
         }
 
@@ -867,15 +870,14 @@ struct CacheDitConditionState {
         return false;
     }
 
-    void after_condition(const void* cond, ggml_tensor* input, ggml_tensor* output) {
+    void after_condition(const void* cond, const sd::Tensor<float>& input, const sd::Tensor<float>& output) {
         if (!step_is_active())
             return;
 
-        size_t ne = static_cast<size_t>(ggml_nelements(output));
-        update_cache(cond, (float*)input->data, (float*)output->data, ne);
+        update_cache(cond, input, output);
 
         if (cond == anchor_condition && taylor_config.enabled) {
-            taylor_state.update_derivatives((float*)output->data, ne, current_step_index);
+            taylor_state.update_derivatives(output.data(), static_cast<size_t>(output.numel()), current_step_index);
         }
     }
 

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 ================================================*/
 
@@ -543,8 +95,9 @@ public:
 
     ggml_tensor* forward(GGMLRunnerContext* ctx,
                          ggml_tensor* x,
-                         ggml_tensor* mask = nullptr,
-                         int clip_skip     = -1) {
+                         ggml_tensor* mask                   = nullptr,
+                         int clip_skip                       = -1,
+                         const std::string& graph_cut_prefix = "") {
         // x: [N, n_token, d_model]
         int layer_idx = n_layer - 1;
         // LOG_DEBUG("clip_skip %d", clip_skip);
@@ -560,6 +113,9 @@ public:
             std::string name = "layers." + std::to_string(i);
             auto layer       = std::dynamic_pointer_cast<CLIPLayer>(blocks[name]);
             x                = layer->forward(ctx, x, mask);  // [N, n_token, d_model]
+            if (!graph_cut_prefix.empty()) {
+                sd::ggml_graph_cut::mark_graph_cut(x, graph_cut_prefix + ".layers." + std::to_string(i), "x");
+            }
             // LOG_DEBUG("layer %d", i);
         }
         return x;
@@ -752,7 +308,8 @@ public:
         auto final_layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["final_layer_norm"]);
 
         auto x = embeddings->forward(ctx, input_ids, tkn_embeddings);  // [N, n_token, hidden_size]
-        x      = encoder->forward(ctx, x, mask, return_pooled ? -1 : clip_skip);
+        sd::ggml_graph_cut::mark_graph_cut(x, "clip_text.prelude", "x");
+        x = encoder->forward(ctx, x, mask, return_pooled ? -1 : clip_skip, "clip_text");
         if (return_pooled || with_final_ln) {
             x = final_layer_norm->forward(ctx, x);
         }
@@ -816,7 +373,8 @@ public:
 
         auto x = embeddings->forward(ctx, pixel_values);  // [N, num_positions, embed_dim]
         x      = pre_layernorm->forward(ctx, x);
-        x      = encoder->forward(ctx, x, nullptr, clip_skip);
+        sd::ggml_graph_cut::mark_graph_cut(x, "clip_vision.prelude", "x");
+        x = encoder->forward(ctx, x, nullptr, clip_skip, "clip_vision");
 
         auto last_hidden_state = x;
 
@@ -957,15 +515,14 @@ struct CLIPTextModelRunner : public GGMLRunner {
         return model.forward(ctx, input_ids, embeddings, mask, max_token_idx, return_pooled, clip_skip);
     }
 
-    ggml_cgraph* build_graph(ggml_tensor* input_ids,
+    ggml_cgraph* build_graph(const sd::Tensor<int32_t>& input_ids_tensor,
                              int num_custom_embeddings    = 0,
                              void* custom_embeddings_data = nullptr,
                              size_t max_token_idx         = 0,
                              bool return_pooled           = false,
                              int clip_skip                = -1) {
-        ggml_cgraph* gf = new_graph_custom(2048);
-
-        input_ids = to_backend(input_ids);
+        ggml_cgraph* gf        = new_graph_custom(2048);
+        ggml_tensor* input_ids = make_input(input_ids_tensor);
 
         ggml_tensor* embeddings = nullptr;
 
@@ -1004,19 +561,21 @@ struct CLIPTextModelRunner : public GGMLRunner {
         return gf;
     }
 
-    bool compute(const int n_threads,
-                 ggml_tensor* input_ids,
-                 int num_custom_embeddings,
-                 void* custom_embeddings_data,
-                 size_t max_token_idx,
-                 bool return_pooled,
-                 int clip_skip,
-                 ggml_tensor** output,
-                 ggml_context* output_ctx = nullptr) {
+    sd::Tensor<float> compute(const int n_threads,
+                              const sd::Tensor<int32_t>& input_ids,
+                              int num_custom_embeddings,
+                              void* custom_embeddings_data,
+                              size_t max_token_idx,
+                              bool return_pooled,
+                              int clip_skip) {
         auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled, clip_skip);
         };
-        return GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
+        auto result = GGMLRunner::compute<float>(get_graph, n_threads, true);
+        if (return_pooled) {
+            return take_or_empty(std::move(result));
+        }
+        return restore_trailing_singleton_dims(std::move(result), 3);
     }
 };
 

src/common_block.hpp

@@ -1,7 +1,9 @@
 #ifndef __COMMON_BLOCK_HPP__
 #define __COMMON_BLOCK_HPP__
 
+#include "ggml-backend.h"
 #include "ggml_extend.hpp"
+#include "util.h"
 
 class DownSampleBlock : public GGMLBlock {
 protected:
@@ -248,9 +250,6 @@ public:
         float scale         = 1.f;
         if (precision_fix) {
             scale = 1.f / 128.f;
-#ifdef SD_USE_VULKAN
-            force_prec_f32 = true;
-#endif
         }
         // The purpose of the scale here is to prevent NaN issues in certain situations.
         // For example, when using Vulkan without enabling force_prec_f32,
@@ -264,6 +263,9 @@ public:
 
         auto net_0 = std::dynamic_pointer_cast<UnaryBlock>(blocks["net.0"]);
         auto net_2 = std::dynamic_pointer_cast<Linear>(blocks["net.2"]);
+        if (sd_backend_is(ctx->backend, "Vulkan")) {
+            net_2->set_force_prec_f32(true);
+        }
 
         x = net_0->forward(ctx, x);  // [ne3, ne2, ne1, inner_dim]
         x = net_2->forward(ctx, x);  // [ne3, ne2, ne1, dim_out]
@@ -277,6 +279,7 @@ 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 +291,11 @@ 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 +320,16 @@ 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/common_dit.hpp

@@ -4,11 +4,11 @@
 #include "ggml_extend.hpp"
 
 namespace DiT {
-    ggml_tensor* patchify(ggml_context* ctx,
-                          ggml_tensor* x,
-                          int pw,
-                          int ph,
-                          bool patch_last = true) {
+    inline ggml_tensor* patchify(ggml_context* ctx,
+                                 ggml_tensor* x,
+                                 int pw,
+                                 int ph,
+                                 bool patch_last = true) {
         // x: [N, C, H, W]
         // return: [N, h*w, C*ph*pw] if patch_last else [N, h*w, ph*pw*C]
         int64_t N = x->ne[3];
@@ -33,13 +33,13 @@ namespace DiT {
         return x;
     }
 
-    ggml_tensor* unpatchify(ggml_context* ctx,
-                            ggml_tensor* x,
-                            int64_t h,
-                            int64_t w,
-                            int ph,
-                            int pw,
-                            bool patch_last = true) {
+    inline ggml_tensor* unpatchify(ggml_context* ctx,
+                                   ggml_tensor* x,
+                                   int64_t h,
+                                   int64_t w,
+                                   int ph,
+                                   int pw,
+                                   bool patch_last = true) {
         // x: [N, h*w, C*ph*pw] if patch_last else [N, h*w, ph*pw*C]
         // return: [N, C, H, W]
         int64_t N = x->ne[2];
@@ -64,10 +64,10 @@ namespace DiT {
         return x;
     }
 
-    ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
-                                   ggml_tensor* x,
-                                   int ph,
-                                   int pw) {
+    inline ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
+                                          ggml_tensor* x,
+                                          int ph,
+                                          int pw) {
         int64_t W = x->ne[0];
         int64_t H = x->ne[1];
 
@@ -77,23 +77,23 @@ namespace DiT {
         return x;
     }
 
-    ggml_tensor* pad_and_patchify(GGMLRunnerContext* ctx,
-                                  ggml_tensor* x,
-                                  int ph,
-                                  int pw,
-                                  bool patch_last = true) {
+    inline ggml_tensor* pad_and_patchify(GGMLRunnerContext* ctx,
+                                         ggml_tensor* x,
+                                         int ph,
+                                         int pw,
+                                         bool patch_last = true) {
         x = pad_to_patch_size(ctx, x, ph, pw);
         x = patchify(ctx->ggml_ctx, x, ph, pw, patch_last);
         return x;
     }
 
-    ggml_tensor* unpatchify_and_crop(ggml_context* ctx,
-                                     ggml_tensor* x,
-                                     int64_t H,
-                                     int64_t W,
-                                     int ph,
-                                     int pw,
-                                     bool patch_last = true) {
+    inline ggml_tensor* unpatchify_and_crop(ggml_context* ctx,
+                                            ggml_tensor* x,
+                                            int64_t H,
+                                            int64_t W,
+                                            int ph,
+                                            int pw,
+                                            bool patch_last = true) {
         int pad_h = (ph - H % ph) % ph;
         int pad_w = (pw - W % pw) % pw;
         int64_t h = ((H + pad_h) / ph);
@@ -105,4 +105,4 @@ namespace DiT {
     }
 }  // namespace DiT
 
-#endif  // __COMMON_DIT_HPP__
+#endif  // __COMMON_DIT_HPP__

src/condition_cache_utils.hpp

@@ -0,0 +1,64 @@
+#ifndef __CONDITION_CACHE_UTILS_HPP__
+#define __CONDITION_CACHE_UTILS_HPP__
+
+#include <vector>
+
+#include "tensor.hpp"
+
+namespace sd {
+
+    inline bool store_condition_cache_diff(std::vector<float>* diff,
+                                           const sd::Tensor<float>& input,
+                                           const sd::Tensor<float>& output) {
+        if (diff == nullptr || input.empty() || output.empty()) {
+            return false;
+        }
+
+        size_t input_size  = static_cast<size_t>(input.numel());
+        size_t output_size = static_cast<size_t>(output.numel());
+        if (input_size == 0 || input_size != output_size) {
+            diff->clear();
+            return false;
+        }
+
+        const float* input_data  = input.data();
+        const float* output_data = output.data();
+        if (input_data == nullptr || output_data == nullptr) {
+            diff->clear();
+            return false;
+        }
+
+        diff->resize(output_size);
+        for (size_t i = 0; i < output_size; ++i) {
+            (*diff)[i] = output_data[i] - input_data[i];
+        }
+        return true;
+    }
+
+    inline bool apply_condition_cache_diff(const std::vector<float>& diff,
+                                           const sd::Tensor<float>& input,
+                                           sd::Tensor<float>* output) {
+        if (output == nullptr || input.empty() || diff.empty()) {
+            return false;
+        }
+
+        size_t input_size = static_cast<size_t>(input.numel());
+        if (input_size == 0 || diff.size() != input_size) {
+            return false;
+        }
+
+        *output            = input;
+        float* output_data = output->data();
+        if (output_data == nullptr) {
+            return false;
+        }
+
+        for (size_t i = 0; i < input_size; ++i) {
+            output_data[i] += diff[i];
+        }
+        return true;
+    }
+
+}  // namespace sd
+
+#endif  // __CONDITION_CACHE_UTILS_HPP__

src/control.hpp

@@ -310,11 +310,13 @@ struct ControlNet : public GGMLRunner {
     SDVersion version = VERSION_SD1;
     ControlNetBlock control_net;
 
-    ggml_backend_buffer_t control_buffer = nullptr;  // keep control output tensors in backend memory
+    ggml_backend_buffer_t control_buffer = nullptr;
     ggml_context* control_ctx            = nullptr;
-    std::vector<ggml_tensor*> controls;  // (12 input block outputs, 1 middle block output) SD 1.5
-    ggml_tensor* guided_hint = nullptr;  // guided_hint cache, for faster inference
-    bool guided_hint_cached  = false;
+    std::vector<ggml_tensor*> control_outputs_ggml;
+    ggml_tensor* guided_hint_output_ggml = nullptr;
+    std::vector<sd::Tensor<float>> controls;
+    sd::Tensor<float> guided_hint;
+    bool guided_hint_cached = false;
 
     ControlNet(ggml_backend_t backend,
                bool offload_params_to_cpu,
@@ -335,16 +337,16 @@ struct ControlNet : public GGMLRunner {
         params.no_alloc   = true;
         control_ctx       = ggml_init(params);
 
-        controls.resize(outs.size() - 1);
+        control_outputs_ggml.resize(outs.size() - 1);
 
         size_t control_buffer_size = 0;
 
-        guided_hint = ggml_dup_tensor(control_ctx, outs[0]);
-        control_buffer_size += ggml_nbytes(guided_hint);
+        guided_hint_output_ggml = ggml_dup_tensor(control_ctx, outs[0]);
+        control_buffer_size += ggml_nbytes(guided_hint_output_ggml);
 
         for (int i = 0; i < outs.size() - 1; i++) {
-            controls[i] = ggml_dup_tensor(control_ctx, outs[i + 1]);
-            control_buffer_size += ggml_nbytes(controls[i]);
+            control_outputs_ggml[i] = ggml_dup_tensor(control_ctx, outs[i + 1]);
+            control_buffer_size += ggml_nbytes(control_outputs_ggml[i]);
         }
 
         control_buffer = ggml_backend_alloc_ctx_tensors(control_ctx, runtime_backend);
@@ -361,8 +363,10 @@ struct ControlNet : public GGMLRunner {
             ggml_free(control_ctx);
             control_ctx = nullptr;
         }
-        guided_hint        = nullptr;
-        guided_hint_cached = false;
+        guided_hint_output_ggml = nullptr;
+        guided_hint_cached      = false;
+        guided_hint             = {};
+        control_outputs_ggml.clear();
         controls.clear();
     }
 
@@ -374,29 +378,33 @@ struct ControlNet : public GGMLRunner {
         control_net.get_param_tensors(tensors, prefix);
     }
 
-    ggml_cgraph* build_graph(ggml_tensor* x,
-                             ggml_tensor* hint,
-                             ggml_tensor* timesteps,
-                             ggml_tensor* context,
-                             ggml_tensor* y = nullptr) {
+    ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                             const sd::Tensor<float>& hint_tensor,
+                             const sd::Tensor<float>& timesteps_tensor,
+                             const sd::Tensor<float>& context_tensor = {},
+                             const sd::Tensor<float>& y_tensor       = {}) {
         ggml_cgraph* gf = new_graph_custom(CONTROL_NET_GRAPH_SIZE);
 
-        x = to_backend(x);
-        if (guided_hint_cached) {
-            hint = nullptr;
+        ggml_tensor* x         = make_input(x_tensor);
+        ggml_tensor* hint      = nullptr;
+        ggml_tensor* timesteps = make_input(timesteps_tensor);
+        ggml_tensor* context   = make_optional_input(context_tensor);
+        ggml_tensor* y         = make_optional_input(y_tensor);
+
+        ggml_tensor* guided_hint_input = nullptr;
+        if (guided_hint_cached && !guided_hint.empty()) {
+            guided_hint_input = make_input(guided_hint);
+            hint              = nullptr;
         } else {
-            hint = to_backend(hint);
+            hint = make_input(hint_tensor);
         }
-        context   = to_backend(context);
-        y         = to_backend(y);
-        timesteps = to_backend(timesteps);
 
         auto runner_ctx = get_context();
 
         auto outs = control_net.forward(&runner_ctx,
                                         x,
                                         hint,
-                                        guided_hint_cached ? guided_hint : nullptr,
+                                        guided_hint_input,
                                         timesteps,
                                         context,
                                         y);
@@ -405,22 +413,20 @@ struct ControlNet : public GGMLRunner {
             alloc_control_ctx(outs);
         }
 
-        ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[0], guided_hint));
+        ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[0], guided_hint_output_ggml));
         for (int i = 0; i < outs.size() - 1; i++) {
-            ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[i + 1], controls[i]));
+            ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[i + 1], control_outputs_ggml[i]));
         }
 
         return gf;
     }
 
-    bool compute(int n_threads,
-                 ggml_tensor* x,
-                 ggml_tensor* hint,
-                 ggml_tensor* timesteps,
-                 ggml_tensor* context,
-                 ggml_tensor* y,
-                 ggml_tensor** output     = nullptr,
-                 ggml_context* output_ctx = nullptr) {
+    std::optional<std::vector<sd::Tensor<float>>> compute(int n_threads,
+                                                          const sd::Tensor<float>& x,
+                                                          const sd::Tensor<float>& hint,
+                                                          const sd::Tensor<float>& timesteps,
+                                                          const sd::Tensor<float>& context = {},
+                                                          const sd::Tensor<float>& y       = {}) {
         // x: [N, in_channels, h, w]
         // timesteps: [N, ]
         // context: [N, max_position, hidden_size]([N, 77, 768]) or [1, max_position, hidden_size]
@@ -429,12 +435,24 @@ struct ControlNet : public GGMLRunner {
             return build_graph(x, hint, timesteps, context, y);
         };
 
-        bool res = GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
-        if (res) {
-            // cache guided_hint
-            guided_hint_cached = true;
+        auto compute_result = GGMLRunner::compute<float>(get_graph, n_threads, false);
+        if (!compute_result.has_value()) {
+            return std::nullopt;
         }
-        return res;
+
+        if (guided_hint_output_ggml != nullptr) {
+            guided_hint = restore_trailing_singleton_dims(sd::make_sd_tensor_from_ggml<float>(guided_hint_output_ggml),
+                                                          4);
+        }
+        controls.clear();
+        controls.reserve(control_outputs_ggml.size());
+        for (ggml_tensor* control : control_outputs_ggml) {
+            auto control_host = restore_trailing_singleton_dims(sd::make_sd_tensor_from_ggml<float>(control), 4);
+            GGML_ASSERT(!control_host.empty());
+            controls.push_back(std::move(control_host));
+        }
+        guided_hint_cached = true;
+        return controls;
     }
 
     bool load_from_file(const std::string& file_path, int n_threads) {
@@ -462,4 +480,4 @@ struct ControlNet : public GGMLRunner {
     }
 };
 
-#endif  // __CONTROL_HPP__
+#endif  // __CONTROL_HPP__

src/convert.cpp

@@ -0,0 +1,138 @@
+#include <cstring>
+#include <mutex>
+#include <regex>
+#include <vector>
+
+#include "model.h"
+#include "model_io/gguf_io.h"
+#include "model_io/safetensors_io.h"
+#include "util.h"
+
+#include "ggml-cpu.h"
+
+static ggml_type get_export_tensor_type(ModelLoader& model_loader,
+                                        const TensorStorage& tensor_storage,
+                                        ggml_type type,
+                                        const TensorTypeRules& tensor_type_rules) {
+    const std::string& name = tensor_storage.name;
+    ggml_type tensor_type   = tensor_storage.type;
+    ggml_type dst_type      = type;
+
+    for (const auto& tensor_type_rule : tensor_type_rules) {
+        std::regex pattern(tensor_type_rule.first);
+        if (std::regex_search(name, pattern)) {
+            dst_type = tensor_type_rule.second;
+            break;
+        }
+    }
+
+    if (model_loader.tensor_should_be_converted(tensor_storage, dst_type)) {
+        tensor_type = dst_type;
+    }
+
+    return tensor_type;
+}
+
+static bool load_tensors_for_export(ModelLoader& model_loader,
+                                    ggml_context* ggml_ctx,
+                                    ggml_type type,
+                                    const TensorTypeRules& tensor_type_rules,
+                                    std::vector<TensorWriteInfo>& tensors) {
+    std::mutex tensor_mutex;
+    auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
+        const std::string& name = tensor_storage.name;
+        ggml_type tensor_type   = get_export_tensor_type(model_loader, tensor_storage, type, tensor_type_rules);
+
+        std::lock_guard<std::mutex> lock(tensor_mutex);
+        ggml_tensor* tensor = ggml_new_tensor(ggml_ctx, tensor_type, tensor_storage.n_dims, tensor_storage.ne);
+        if (tensor == nullptr) {
+            LOG_ERROR("ggml_new_tensor failed");
+            return false;
+        }
+        ggml_set_name(tensor, name.c_str());
+
+        if (!tensor->data) {
+            GGML_ASSERT(ggml_nelements(tensor) == 0);
+            // Avoid crashing writers by setting a dummy pointer for zero-sized tensors.
+            LOG_DEBUG("setting dummy pointer for zero-sized tensor %s", name.c_str());
+            tensor->data = ggml_get_mem_buffer(ggml_ctx);
+        }
+
+        TensorWriteInfo write_info;
+        write_info.tensor = tensor;
+        write_info.n_dims = tensor_storage.n_dims;
+        for (int i = 0; i < tensor_storage.n_dims; ++i) {
+            write_info.ne[i] = tensor_storage.ne[i];
+        }
+
+        *dst_tensor = tensor;
+        tensors.push_back(std::move(write_info));
+
+        return true;
+    };
+
+    bool success = model_loader.load_tensors(on_new_tensor_cb);
+    LOG_INFO("load tensors done");
+    return success;
+}
+
+bool convert(const char* input_path,
+             const char* vae_path,
+             const char* output_path,
+             sd_type_t output_type,
+             const char* tensor_type_rules,
+             bool convert_name) {
+    ModelLoader model_loader;
+
+    if (!model_loader.init_from_file(input_path)) {
+        LOG_ERROR("init model loader from file failed: '%s'", input_path);
+        return false;
+    }
+
+    if (vae_path != nullptr && strlen(vae_path) > 0) {
+        if (!model_loader.init_from_file(vae_path, "vae.")) {
+            LOG_ERROR("init model loader from file failed: '%s'", vae_path);
+            return false;
+        }
+    }
+    if (convert_name) {
+        model_loader.convert_tensors_name();
+    }
+
+    ggml_type type             = (ggml_type)output_type;
+    bool output_is_safetensors = ends_with(output_path, ".safetensors");
+    TensorTypeRules type_rules = parse_tensor_type_rules(tensor_type_rules);
+
+    auto backend    = ggml_backend_cpu_init();
+    size_t mem_size = 1 * 1024 * 1024;  // for padding
+    mem_size += model_loader.get_tensor_storage_map().size() * ggml_tensor_overhead();
+    mem_size += model_loader.get_params_mem_size(backend, type);
+    LOG_INFO("model tensors mem size: %.2fMB", mem_size / 1024.f / 1024.f);
+    ggml_context* ggml_ctx = ggml_init({mem_size, nullptr, false});
+
+    if (ggml_ctx == nullptr) {
+        LOG_ERROR("ggml_init failed for converter");
+        ggml_backend_free(backend);
+        return false;
+    }
+
+    std::vector<TensorWriteInfo> tensors;
+    bool success = load_tensors_for_export(model_loader, ggml_ctx, type, type_rules, tensors);
+    ggml_backend_free(backend);
+
+    std::string error;
+    if (success) {
+        if (output_is_safetensors) {
+            success = write_safetensors_file(output_path, tensors, &error);
+        } else {
+            success = write_gguf_file(output_path, tensors, &error);
+        }
+    }
+
+    if (!success && !error.empty()) {
+        LOG_ERROR("%s", error.c_str());
+    }
+
+    ggml_free(ggml_ctx);
+    return success;
+}

src/diffusion_model.hpp

@@ -1,37 +1,46 @@
 #ifndef __DIFFUSION_MODEL_H__
 #define __DIFFUSION_MODEL_H__
 
+#include <optional>
 #include "anima.hpp"
+#include "ernie_image.hpp"
 #include "flux.hpp"
 #include "mmdit.hpp"
 #include "qwen_image.hpp"
+#include "tensor_ggml.hpp"
 #include "unet.hpp"
 #include "wan.hpp"
 #include "z_image.hpp"
 
 struct DiffusionParams {
-    ggml_tensor* x                        = nullptr;
-    ggml_tensor* timesteps                = nullptr;
-    ggml_tensor* context                  = nullptr;
-    ggml_tensor* c_concat                 = nullptr;
-    ggml_tensor* y                        = nullptr;
-    ggml_tensor* guidance                 = nullptr;
-    std::vector<ggml_tensor*> ref_latents = {};
-    bool increase_ref_index               = false;
-    int num_video_frames                  = -1;
-    std::vector<ggml_tensor*> controls    = {};
-    float control_strength                = 0.f;
-    ggml_tensor* vace_context             = nullptr;
-    float vace_strength                   = 1.f;
-    std::vector<int> skip_layers          = {};
+    const sd::Tensor<float>* x                        = nullptr;
+    const sd::Tensor<float>* timesteps                = nullptr;
+    const sd::Tensor<float>* context                  = nullptr;
+    const sd::Tensor<float>* c_concat                 = nullptr;
+    const sd::Tensor<float>* y                        = nullptr;
+    const sd::Tensor<int32_t>* t5_ids                 = nullptr;
+    const sd::Tensor<float>* t5_weights               = nullptr;
+    const sd::Tensor<float>* guidance                 = nullptr;
+    const std::vector<sd::Tensor<float>>* ref_latents = nullptr;
+    bool increase_ref_index                           = false;
+    int num_video_frames                              = -1;
+    const std::vector<sd::Tensor<float>>* controls    = nullptr;
+    float control_strength                            = 0.f;
+    const sd::Tensor<float>* vace_context             = nullptr;
+    float vace_strength                               = 1.f;
+    const std::vector<int>* skip_layers               = nullptr;
 };
 
+template <typename T>
+static inline const sd::Tensor<T>& tensor_or_empty(const sd::Tensor<T>* tensor) {
+    static const sd::Tensor<T> kEmpty;
+    return tensor != nullptr ? *tensor : kEmpty;
+}
+
 struct DiffusionModel {
     virtual std::string get_desc()                                               = 0;
-    virtual bool compute(int n_threads,
-                         DiffusionParams diffusion_params,
-                         ggml_tensor** output     = nullptr,
-                         ggml_context* output_ctx = nullptr)                     = 0;
+    virtual sd::Tensor<float> compute(int n_threads,
+                                      const DiffusionParams& diffusion_params)   = 0;
     virtual void alloc_params_buffer()                                           = 0;
     virtual void free_params_buffer()                                            = 0;
     virtual void free_compute_buffer()                                           = 0;
@@ -40,6 +49,7 @@ struct DiffusionModel {
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter){};
     virtual int64_t get_adm_in_channels()                            = 0;
     virtual void set_flash_attention_enabled(bool enabled)           = 0;
+    virtual void set_max_graph_vram_bytes(size_t max_vram_bytes)     = 0;
     virtual void set_circular_axes(bool circular_x, bool circular_y) = 0;
 };
 
@@ -89,23 +99,28 @@ struct UNetModel : public DiffusionModel {
         unet.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        unet.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         unet.set_circular_axes(circular_x, circular_y);
     }
 
-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 ggml_tensor** output     = nullptr,
-                 ggml_context* output_ctx = nullptr) override {
+    sd::Tensor<float> compute(int n_threads,
+                              const DiffusionParams& diffusion_params) override {
+        GGML_ASSERT(diffusion_params.x != nullptr);
+        GGML_ASSERT(diffusion_params.timesteps != nullptr);
+        static const std::vector<sd::Tensor<float>> empty_controls;
         return unet.compute(n_threads,
-                            diffusion_params.x,
-                            diffusion_params.timesteps,
-                            diffusion_params.context,
-                            diffusion_params.c_concat,
-                            diffusion_params.y,
+                            *diffusion_params.x,
+                            *diffusion_params.timesteps,
+                            tensor_or_empty(diffusion_params.context),
+                            tensor_or_empty(diffusion_params.c_concat),
+                            tensor_or_empty(diffusion_params.y),
                             diffusion_params.num_video_frames,
-                            diffusion_params.controls,
-                            diffusion_params.control_strength, output, output_ctx);
+                            diffusion_params.controls ? *diffusion_params.controls : empty_controls,
+                            diffusion_params.control_strength);
     }
 };
 
@@ -154,22 +169,25 @@ struct MMDiTModel : public DiffusionModel {
         mmdit.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        mmdit.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         mmdit.set_circular_axes(circular_x, circular_y);
     }
 
-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 ggml_tensor** output     = nullptr,
-                 ggml_context* output_ctx = nullptr) override {
+    sd::Tensor<float> compute(int n_threads,
+                              const DiffusionParams& diffusion_params) override {
+        GGML_ASSERT(diffusion_params.x != nullptr);
+        GGML_ASSERT(diffusion_params.timesteps != nullptr);
+        static const std::vector<int> empty_skip_layers;
         return mmdit.compute(n_threads,
-                             diffusion_params.x,
-                             diffusion_params.timesteps,
-                             diffusion_params.context,
-                             diffusion_params.y,
-                             output,
-                             output_ctx,
-                             diffusion_params.skip_layers);
+                             *diffusion_params.x,
+                             *diffusion_params.timesteps,
+                             tensor_or_empty(diffusion_params.context),
+                             tensor_or_empty(diffusion_params.y),
+                             diffusion_params.skip_layers ? *diffusion_params.skip_layers : empty_skip_layers);
     }
 };
 
@@ -220,26 +238,30 @@ struct FluxModel : public DiffusionModel {
         flux.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        flux.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         flux.set_circular_axes(circular_x, circular_y);
     }
 
-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 ggml_tensor** output     = nullptr,
-                 ggml_context* output_ctx = nullptr) override {
+    sd::Tensor<float> compute(int n_threads,
+                              const DiffusionParams& diffusion_params) override {
+        GGML_ASSERT(diffusion_params.x != nullptr);
+        GGML_ASSERT(diffusion_params.timesteps != nullptr);
+        static const std::vector<sd::Tensor<float>> empty_ref_latents;
+        static const std::vector<int> empty_skip_layers;
         return flux.compute(n_threads,
-                            diffusion_params.x,
-                            diffusion_params.timesteps,
-                            diffusion_params.context,
-                            diffusion_params.c_concat,
-                            diffusion_params.y,
-                            diffusion_params.guidance,
-                            diffusion_params.ref_latents,
+                            *diffusion_params.x,
+                            *diffusion_params.timesteps,
+                            tensor_or_empty(diffusion_params.context),
+                            tensor_or_empty(diffusion_params.c_concat),
+                            tensor_or_empty(diffusion_params.y),
+                            tensor_or_empty(diffusion_params.guidance),
+                            diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
                             diffusion_params.increase_ref_index,
-                            output,
-                            output_ctx,
-                            diffusion_params.skip_layers);
+                            diffusion_params.skip_layers ? *diffusion_params.skip_layers : empty_skip_layers);
     }
 };
 
@@ -290,22 +312,24 @@ struct AnimaModel : public DiffusionModel {
         anima.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        anima.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         anima.set_circular_axes(circular_x, circular_y);
     }
 
-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 ggml_tensor** output     = nullptr,
-                 ggml_context* output_ctx = nullptr) override {
+    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 anima.compute(n_threads,
-                             diffusion_params.x,
-                             diffusion_params.timesteps,
-                             diffusion_params.context,
-                             diffusion_params.c_concat,
-                             diffusion_params.y,
-                             output,
-                             output_ctx);
+                             *diffusion_params.x,
+                             *diffusion_params.timesteps,
+                             tensor_or_empty(diffusion_params.context),
+                             tensor_or_empty(diffusion_params.t5_ids),
+                             tensor_or_empty(diffusion_params.t5_weights));
     }
 };
 
@@ -357,25 +381,27 @@ struct WanModel : public DiffusionModel {
         wan.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        wan.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         wan.set_circular_axes(circular_x, circular_y);
     }
 
-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 ggml_tensor** output     = nullptr,
-                 ggml_context* output_ctx = nullptr) override {
+    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 wan.compute(n_threads,
-                           diffusion_params.x,
-                           diffusion_params.timesteps,
-                           diffusion_params.context,
-                           diffusion_params.y,
-                           diffusion_params.c_concat,
-                           nullptr,
-                           diffusion_params.vace_context,
-                           diffusion_params.vace_strength,
-                           output,
-                           output_ctx);
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           tensor_or_empty(diffusion_params.y),
+                           tensor_or_empty(diffusion_params.c_concat),
+                           sd::Tensor<float>(),
+                           tensor_or_empty(diffusion_params.vace_context),
+                           diffusion_params.vace_strength);
     }
 };
 
@@ -428,22 +454,25 @@ struct QwenImageModel : public DiffusionModel {
         qwen_image.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        qwen_image.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         qwen_image.set_circular_axes(circular_x, circular_y);
     }
 
-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 ggml_tensor** output     = nullptr,
-                 ggml_context* output_ctx = nullptr) override {
+    sd::Tensor<float> compute(int n_threads,
+                              const DiffusionParams& diffusion_params) override {
+        GGML_ASSERT(diffusion_params.x != nullptr);
+        GGML_ASSERT(diffusion_params.timesteps != nullptr);
+        static const std::vector<sd::Tensor<float>> empty_ref_latents;
         return qwen_image.compute(n_threads,
-                                  diffusion_params.x,
-                                  diffusion_params.timesteps,
-                                  diffusion_params.context,
-                                  diffusion_params.ref_latents,
-                                  true,  // increase_ref_index
-                                  output,
-                                  output_ctx);
+                                  *diffusion_params.x,
+                                  *diffusion_params.timesteps,
+                                  tensor_or_empty(diffusion_params.context),
+                                  diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
+                                  true);
     }
 };
 
@@ -495,22 +524,91 @@ struct ZImageModel : public DiffusionModel {
         z_image.set_flash_attention_enabled(enabled);
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        z_image.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_circular_axes(bool circular_x, bool circular_y) override {
         z_image.set_circular_axes(circular_x, circular_y);
     }
 
-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 ggml_tensor** output     = nullptr,
-                 ggml_context* output_ctx = nullptr) override {
+    sd::Tensor<float> compute(int n_threads,
+                              const DiffusionParams& diffusion_params) override {
+        GGML_ASSERT(diffusion_params.x != nullptr);
+        GGML_ASSERT(diffusion_params.timesteps != nullptr);
+        static const std::vector<sd::Tensor<float>> empty_ref_latents;
         return z_image.compute(n_threads,
-                               diffusion_params.x,
-                               diffusion_params.timesteps,
-                               diffusion_params.context,
-                               diffusion_params.ref_latents,
-                               true,  // increase_ref_index
-                               output,
-                               output_ctx);
+                               *diffusion_params.x,
+                               *diffusion_params.timesteps,
+                               tensor_or_empty(diffusion_params.context),
+                               diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
+                               true);
+    }
+};
+
+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_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        ernie_image.set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
+    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));
     }
 };
 

src/easycache.hpp

@@ -1,10 +1,15 @@
+#ifndef __EASYCACHE_HPP__
+#define __EASYCACHE_HPP__
+
 #include <cmath>
 #include <limits>
 #include <unordered_map>
 #include <vector>
 
+#include "condition_cache_utils.hpp"
 #include "denoiser.hpp"
 #include "ggml_extend.hpp"
+#include "tensor.hpp"
 
 struct EasyCacheConfig {
     bool enabled          = false;
@@ -19,15 +24,15 @@ struct EasyCacheCacheEntry {
 
 struct EasyCacheState {
     EasyCacheConfig config;
-    Denoiser* denoiser                  = nullptr;
-    float start_sigma                   = std::numeric_limits<float>::max();
-    float end_sigma                     = 0.0f;
-    bool initialized                    = false;
-    bool initial_step                   = true;
-    bool skip_current_step              = false;
-    bool step_active                    = false;
-    const SDCondition* anchor_condition = nullptr;
-    std::unordered_map<const SDCondition*, EasyCacheCacheEntry> cache_diffs;
+    Denoiser* denoiser           = nullptr;
+    float start_sigma            = std::numeric_limits<float>::max();
+    float end_sigma              = 0.0f;
+    bool initialized             = false;
+    bool initial_step            = true;
+    bool skip_current_step       = false;
+    bool step_active             = false;
+    const void* anchor_condition = nullptr;
+    std::unordered_map<const void*, EasyCacheCacheEntry> cache_diffs;
     std::vector<float> prev_input;
     std::vector<float> prev_output;
     float output_prev_norm                = 0.0f;
@@ -120,41 +125,30 @@ struct EasyCacheState {
         return enabled() && step_active && skip_current_step;
     }
 
-    bool has_cache(const SDCondition* cond) const {
+    bool has_cache(const void* cond) const {
         auto it = cache_diffs.find(cond);
         return it != cache_diffs.end() && !it->second.diff.empty();
     }
 
-    void update_cache(const SDCondition* cond, ggml_tensor* input, ggml_tensor* output) {
+    void update_cache(const void* cond, const sd::Tensor<float>& input, const sd::Tensor<float>& output) {
         EasyCacheCacheEntry& entry = cache_diffs[cond];
-        size_t ne                  = static_cast<size_t>(ggml_nelements(output));
-        entry.diff.resize(ne);
-        float* out_data = (float*)output->data;
-        float* in_data  = (float*)input->data;
-        for (size_t i = 0; i < ne; ++i) {
-            entry.diff[i] = out_data[i] - in_data[i];
-        }
+        sd::store_condition_cache_diff(&entry.diff, input, output);
     }
 
-    void apply_cache(const SDCondition* cond, ggml_tensor* input, ggml_tensor* output) {
+    void apply_cache(const void* cond, const sd::Tensor<float>& input, sd::Tensor<float>* output) {
         auto it = cache_diffs.find(cond);
         if (it == cache_diffs.end() || it->second.diff.empty()) {
             return;
         }
-        copy_ggml_tensor(output, input);
-        float* out_data                = (float*)output->data;
-        const std::vector<float>& diff = it->second.diff;
-        for (size_t i = 0; i < diff.size(); ++i) {
-            out_data[i] += diff[i];
-        }
+        sd::apply_condition_cache_diff(it->second.diff, input, output);
     }
 
-    bool before_condition(const SDCondition* cond,
-                          ggml_tensor* input,
-                          ggml_tensor* output,
+    bool before_condition(const void* cond,
+                          const sd::Tensor<float>& input,
+                          sd::Tensor<float>* output,
                           float sigma,
                           int step_index) {
-        if (!enabled() || step_index < 0) {
+        if (!enabled() || step_index < 0 || output == nullptr) {
             return false;
         }
         if (step_index != current_step_index) {
@@ -181,12 +175,12 @@ struct EasyCacheState {
         if (!has_prev_input || !has_prev_output || !has_cache(cond)) {
             return false;
         }
-        size_t ne = static_cast<size_t>(ggml_nelements(input));
+        size_t ne = static_cast<size_t>(input.numel());
         if (prev_input.size() != ne) {
             return false;
         }
-        float* input_data = (float*)input->data;
-        last_input_change = 0.0f;
+        const float* input_data = input.data();
+        last_input_change       = 0.0f;
         for (size_t i = 0; i < ne; ++i) {
             last_input_change += std::fabs(input_data[i] - prev_input[i]);
         }
@@ -211,7 +205,7 @@ struct EasyCacheState {
         return false;
     }
 
-    void after_condition(const SDCondition* cond, ggml_tensor* input, ggml_tensor* output) {
+    void after_condition(const void* cond, const sd::Tensor<float>& input, const sd::Tensor<float>& output) {
         if (!step_is_active()) {
             return;
         }
@@ -220,16 +214,16 @@ struct EasyCacheState {
             return;
         }
 
-        size_t ne      = static_cast<size_t>(ggml_nelements(input));
-        float* in_data = (float*)input->data;
+        size_t ne            = static_cast<size_t>(input.numel());
+        const float* in_data = input.data();
         prev_input.resize(ne);
         for (size_t i = 0; i < ne; ++i) {
             prev_input[i] = in_data[i];
         }
         has_prev_input = true;
 
-        float* out_data     = (float*)output->data;
-        float output_change = 0.0f;
+        const float* out_data = output.data();
+        float output_change   = 0.0f;
         if (has_prev_output && prev_output.size() == ne) {
             for (size_t i = 0; i < ne; ++i) {
                 output_change += std::fabs(out_data[i] - prev_output[i]);
@@ -262,4 +256,6 @@ struct EasyCacheState {
         cumulative_change_rate = 0.0f;
         has_last_input_change  = false;
     }
-};
+};
+
+#endif

src/ernie_image.hpp

@@ -0,0 +1,441 @@
+#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]
+            sd::ggml_graph_cut::mark_graph_cut(hidden_states, "ernie_image.prelude", "hidden_states");
+            // sd::ggml_graph_cut::mark_graph_cut(mod_params, "ernie_image.prelude", "mod_params");
+            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);
+                sd::ggml_graph_cut::mark_graph_cut(hidden_states, "ernie_image.layers." + std::to_string(i), "hidden_states");
+            }
+
+            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/esrgan.hpp

@@ -124,27 +124,33 @@ public:
         auto conv_hr    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_hr"]);
         auto conv_last  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_last"]);
 
-        auto feat      = conv_first->forward(ctx, x);
+        auto feat = conv_first->forward(ctx, x);
+        sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.prelude", "feat");
         auto body_feat = feat;
         for (int i = 0; i < num_block; i++) {
             std::string name = "body." + std::to_string(i);
             auto block       = std::dynamic_pointer_cast<RRDB>(blocks[name]);
 
             body_feat = block->forward(ctx, body_feat);
+            sd::ggml_graph_cut::mark_graph_cut(body_feat, "esrgan.body." + std::to_string(i), "feat");
         }
         body_feat = conv_body->forward(ctx, body_feat);
         feat      = ggml_add(ctx->ggml_ctx, feat, body_feat);
+        sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.body.out", "feat");
         // upsample
         if (scale >= 2) {
             auto conv_up1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up1"]);
             feat          = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx->ggml_ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+            sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.up1", "feat");
             if (scale == 4) {
                 auto conv_up2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up2"]);
                 feat          = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx->ggml_ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+                sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.up2", "feat");
             }
         }
         // for all scales
         auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat)));
+        sd::ggml_graph_cut::mark_graph_cut(out, "esrgan.final", "out");
         return out;
     }
 };
@@ -341,12 +347,12 @@ struct ESRGAN : public GGMLRunner {
         return success;
     }
 
-    ggml_cgraph* build_graph(ggml_tensor* x) {
+    ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor) {
         if (!rrdb_net)
             return nullptr;
         constexpr int kGraphNodes = 1 << 16;  // 65k
         ggml_cgraph* gf           = new_graph_custom(kGraphNodes);
-        x                         = to_backend(x);
+        ggml_tensor* x            = make_input(x_tensor);
 
         auto runner_ctx  = get_context();
         ggml_tensor* out = rrdb_net->forward(&runner_ctx, x);
@@ -354,15 +360,12 @@ struct ESRGAN : public GGMLRunner {
         return gf;
     }
 
-    bool compute(const int n_threads,
-                 ggml_tensor* x,
-                 ggml_tensor** output,
-                 ggml_context* output_ctx = nullptr) {
-        auto get_graph = [&]() -> ggml_cgraph* {
-            return build_graph(x);
-        };
-        return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+    sd::Tensor<float> compute(const int n_threads,
+                              const sd::Tensor<float>& x) {
+        auto get_graph = [&]() -> ggml_cgraph* { return build_graph(x); };
+        auto result    = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+        return result;
     }
 };
 
-#endif  // __ESRGAN_HPP__
+#endif  // __ESRGAN_HPP__

src/flux.hpp

@@ -928,6 +928,9 @@ namespace Flux {
             }
 
             txt = txt_in->forward(ctx, txt);
+            sd::ggml_graph_cut::mark_graph_cut(img, "flux.prelude", "img");
+            sd::ggml_graph_cut::mark_graph_cut(txt, "flux.prelude", "txt");
+            sd::ggml_graph_cut::mark_graph_cut(vec, "flux.prelude", "vec");
 
             for (int i = 0; i < params.depth; i++) {
                 if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i) != skip_layers.end()) {
@@ -939,6 +942,8 @@ namespace Flux {
                 auto img_txt = block->forward(ctx, img, txt, vec, pe, txt_img_mask, ds_img_mods, ds_txt_mods);
                 img          = img_txt.first;   // [N, n_img_token, hidden_size]
                 txt          = img_txt.second;  // [N, n_txt_token, hidden_size]
+                sd::ggml_graph_cut::mark_graph_cut(img, "flux.double_blocks." + std::to_string(i), "img");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "flux.double_blocks." + std::to_string(i), "txt");
             }
 
             auto txt_img = ggml_concat(ctx->ggml_ctx, txt, img, 1);  // [N, n_txt_token + n_img_token, hidden_size]
@@ -949,6 +954,7 @@ namespace Flux {
                 auto block = std::dynamic_pointer_cast<SingleStreamBlock>(blocks["single_blocks." + std::to_string(i)]);
 
                 txt_img = block->forward(ctx, txt_img, vec, pe, txt_img_mask, ss_mods);
+                sd::ggml_graph_cut::mark_graph_cut(txt_img, "flux.single_blocks." + std::to_string(i), "txt_img");
             }
 
             img = ggml_view_3d(ctx->ggml_ctx,
@@ -1178,6 +1184,7 @@ namespace Flux {
         std::vector<float> pe_vec;
         std::vector<float> mod_index_arange_vec;
         std::vector<float> dct_vec;
+        sd::Tensor<float> guidance_tensor;
         SDVersion version;
         bool use_mask = false;
 
@@ -1353,29 +1360,42 @@ namespace Flux {
             return dct;
         }
 
-        ggml_cgraph* build_graph(ggml_tensor* x,
-                                 ggml_tensor* timesteps,
-                                 ggml_tensor* context,
-                                 ggml_tensor* c_concat,
-                                 ggml_tensor* y,
-                                 ggml_tensor* guidance,
-                                 std::vector<ggml_tensor*> ref_latents = {},
-                                 bool increase_ref_index               = false,
-                                 std::vector<int> skip_layers          = {}) {
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor                  = {},
+                                 const sd::Tensor<float>& c_concat_tensor                 = {},
+                                 const sd::Tensor<float>& y_tensor                        = {},
+                                 const sd::Tensor<float>& guidance_tensor                 = {},
+                                 const std::vector<sd::Tensor<float>>& ref_latents_tensor = {},
+                                 bool increase_ref_index                                  = false,
+                                 std::vector<int> skip_layers                             = {}) {
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timesteps = make_input(timesteps_tensor);
+            ggml_tensor* context   = make_optional_input(context_tensor);
+            ggml_tensor* c_concat  = make_optional_input(c_concat_tensor);
+            ggml_tensor* y         = make_optional_input(y_tensor);
+            if (flux_params.guidance_embed || flux_params.is_chroma) {
+                if (!guidance_tensor.empty()) {
+                    this->guidance_tensor = guidance_tensor;
+                    if (flux_params.is_chroma) {
+                        this->guidance_tensor.fill_(0.f);
+                    }
+                }
+            }
+            ggml_tensor* guidance = make_optional_input(this->guidance_tensor);
+            std::vector<ggml_tensor*> ref_latents;
+            ref_latents.reserve(ref_latents_tensor.size());
+            for (const auto& ref_latent_tensor : ref_latents_tensor) {
+                ref_latents.push_back(make_input(ref_latent_tensor));
+            }
+
             GGML_ASSERT(x->ne[3] == 1);
             ggml_cgraph* gf = new_graph_custom(FLUX_GRAPH_SIZE);
 
             ggml_tensor* mod_index_arange = nullptr;
             ggml_tensor* dct              = nullptr;  // for chroma radiance
 
-            x       = to_backend(x);
-            context = to_backend(context);
-            if (c_concat != nullptr) {
-                c_concat = to_backend(c_concat);
-            }
             if (flux_params.is_chroma) {
-                guidance = ggml_set_f32(guidance, 0);
-
                 if (!use_mask) {
                     y = nullptr;
                 }
@@ -1385,16 +1405,6 @@ namespace Flux {
                 mod_index_arange     = ggml_new_tensor_1d(compute_ctx, GGML_TYPE_F32, mod_index_arange_vec.size());
                 set_backend_tensor_data(mod_index_arange, mod_index_arange_vec.data());
             }
-            y = to_backend(y);
-
-            timesteps = to_backend(timesteps);
-            if (flux_params.guidance_embed || flux_params.is_chroma) {
-                guidance = to_backend(guidance);
-            }
-            for (int i = 0; i < ref_latents.size(); i++) {
-                ref_latents[i] = to_backend(ref_latents[i]);
-            }
-
             std::set<int> txt_arange_dims;
             if (sd_version_is_flux2(version)) {
                 txt_arange_dims    = {3};
@@ -1455,18 +1465,16 @@ namespace Flux {
             return gf;
         }
 
-        bool compute(int n_threads,
-                     ggml_tensor* x,
-                     ggml_tensor* timesteps,
-                     ggml_tensor* context,
-                     ggml_tensor* c_concat,
-                     ggml_tensor* y,
-                     ggml_tensor* guidance,
-                     std::vector<ggml_tensor*> ref_latents = {},
-                     bool increase_ref_index               = false,
-                     ggml_tensor** output                  = nullptr,
-                     ggml_context* output_ctx              = nullptr,
-                     std::vector<int> skip_layers          = std::vector<int>()) {
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context                  = {},
+                                  const sd::Tensor<float>& c_concat                 = {},
+                                  const sd::Tensor<float>& y                        = {},
+                                  const sd::Tensor<float>& guidance                 = {},
+                                  const std::vector<sd::Tensor<float>>& ref_latents = {},
+                                  bool increase_ref_index                           = false,
+                                  std::vector<int> skip_layers                      = std::vector<int>()) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
@@ -1476,7 +1484,8 @@ namespace Flux {
                 return build_graph(x, timesteps, context, c_concat, y, guidance, ref_latents, increase_ref_index, skip_layers);
             };
 
-            return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+            auto result = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+            return result;
         }
 
         void test() {
@@ -1485,41 +1494,51 @@ namespace Flux {
             params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
-            ggml_context* work_ctx = ggml_init(params);
-            GGML_ASSERT(work_ctx != nullptr);
+            ggml_context* ctx = ggml_init(params);
+            GGML_ASSERT(ctx != nullptr);
 
             {
                 // cpu f16:
                 // cuda f16: nan
                 // cuda q8_0: pass
-                auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 16, 16, 128, 1);
+                sd::Tensor<float> x({16, 16, 128, 1});
                 // ggml_set_f32(x, 0.01f);
-                // auto x = load_tensor_from_file(work_ctx, "chroma_x.bin");
+                // auto x = load_tensor_from_file(ctx, "chroma_x.bin");
                 // print_ggml_tensor(x);
 
                 std::vector<float> timesteps_vec(1, 1.f);
-                auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
+                auto timesteps = sd::Tensor<float>::from_vector(timesteps_vec);
 
                 std::vector<float> guidance_vec(1, 0.f);
-                auto guidance = vector_to_ggml_tensor(work_ctx, guidance_vec);
+                auto guidance = sd::Tensor<float>::from_vector(guidance_vec);
 
-                auto context = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 15360, 256, 1);
+                sd::Tensor<float> context({15360, 256, 1});
                 // ggml_set_f32(context, 0.01f);
-                // auto context = load_tensor_from_file(work_ctx, "chroma_context.bin");
+                // auto context = load_tensor_from_file(ctx, "chroma_context.bin");
                 // print_ggml_tensor(context);
 
-                // auto y = ggml_new_tensor_2d(work_ctx, GGML_TYPE_F32, 768, 1);
+                // auto y = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 768, 1);
                 // ggml_set_f32(y, 0.01f);
                 auto y = nullptr;
                 // print_ggml_tensor(y);
 
-                ggml_tensor* out = nullptr;
+                sd::Tensor<float> out;
 
-                int64_t t0 = ggml_time_ms();
-                compute(8, x, timesteps, context, nullptr, y, guidance, {}, false, &out, work_ctx);
-                int64_t t1 = ggml_time_ms();
+                int64_t t0   = ggml_time_ms();
+                auto out_opt = compute(8,
+                                       x,
+                                       timesteps,
+                                       context,
+                                       {},
+                                       {},
+                                       guidance,
+                                       {},
+                                       false);
+                int64_t t1   = ggml_time_ms();
 
-                print_ggml_tensor(out);
+                GGML_ASSERT(!out_opt.empty());
+                out = std::move(out_opt);
+                print_sd_tensor(out);
                 LOG_DEBUG("flux test done in %lldms", t1 - t0);
             }
         }

src/ggml_extend_backend.hpp

@@ -0,0 +1,298 @@
+#ifndef __GGML_EXTEND_BACKEND_HPP__
+#define __GGML_EXTEND_BACKEND_HPP__
+
+#include <cstring>
+#include <mutex>
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+#ifndef __STATIC_INLINE__
+#define __STATIC_INLINE__ static inline
+#endif
+
+inline void ggml_backend_load_all_once() {
+    // If the registry already has devices and the CPU backend is present,
+    // assume either static registration or explicit host-side preloading has
+    // completed and avoid rescanning the default paths.
+    if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
+        return;
+    }
+    // In dynamic-backend mode the backend modules are discovered at runtime,
+    // so we must load them before asking for the CPU backend or its proc table.
+    // If the host preloaded only a subset of backends, allow one default-path
+    // scan so missing modules can still be discovered.
+    static std::once_flag once;
+    std::call_once(once, []() {
+        if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
+            return;
+        }
+        ggml_backend_load_all();
+    });
+}
+
+// Do not gate this branch on GGML_CPU or GGML_CPU_ALL_VARIANTS:
+// those are CMake options used to configure ggml itself, but they are not
+// exported as PUBLIC compile definitions to stable-diffusion in backend-DL mode.
+// In practice, this target can reliably see GGML_BACKEND_DL, but not whether
+// the CPU backend was compiled as a loadable module. We therefore use runtime
+// backend discovery instead of compile-time assumptions.
+
+__STATIC_INLINE__ ggml_backend_reg_t ggml_backend_cpu_reg() {
+    ggml_backend_reg_t reg = ggml_backend_reg_by_name("CPU");
+    if (reg != nullptr) {
+        return reg;
+    }
+
+    ggml_backend_load_all_once();
+    return ggml_backend_reg_by_name("CPU");
+}
+
+__STATIC_INLINE__ ggml_backend_reg_t ggml_backend_reg_from_backend(ggml_backend_t backend) {
+    if (backend != nullptr) {
+        ggml_backend_dev_t device = ggml_backend_get_device(backend);
+        if (device != nullptr) {
+            return ggml_backend_dev_backend_reg(device);
+        }
+    }
+
+    return ggml_backend_cpu_reg();
+}
+
+__STATIC_INLINE__ ggml_backend_t ggml_backend_cpu_init() {
+    ggml_backend_t backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
+    if (backend != nullptr) {
+        return backend;
+    }
+
+    ggml_backend_load_all_once();
+    return ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
+}
+
+__STATIC_INLINE__ bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    if (backend == nullptr) {
+        return false;
+    }
+
+    ggml_backend_dev_t device = ggml_backend_get_device(backend);
+    if (device != nullptr) {
+        return ggml_backend_dev_type(device) == GGML_BACKEND_DEVICE_TYPE_CPU;
+    }
+
+    const char* backend_name = ggml_backend_name(backend);
+    return backend_name != nullptr && std::strcmp(backend_name, "CPU") == 0;
+}
+
+__STATIC_INLINE__ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
+    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
+    if (reg == nullptr) {
+        return;
+    }
+
+    auto fn = reinterpret_cast<ggml_backend_set_n_threads_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads"));
+    if (fn != nullptr) {
+        fn(backend_cpu, n_threads);
+    }
+}
+
+using __ggml_backend_cpu_set_threadpool_t = void (*)(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
+
+__STATIC_INLINE__ void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
+    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
+    if (reg == nullptr) {
+        return;
+    }
+
+    auto fn = reinterpret_cast<__ggml_backend_cpu_set_threadpool_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool"));
+    if (fn != nullptr) {
+        fn(backend_cpu, threadpool);
+    }
+}
+
+__STATIC_INLINE__ void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void* abort_callback_data) {
+    ggml_backend_reg_t reg = ggml_backend_reg_from_backend(backend_cpu);
+    if (reg == nullptr) {
+        return;
+    }
+
+    auto fn = reinterpret_cast<ggml_backend_set_abort_callback_t>(ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback"));
+    if (fn != nullptr) {
+        fn(backend_cpu, abort_callback, abort_callback_data);
+    }
+}
+
+__STATIC_INLINE__ ggml_backend_buffer_t ggml_backend_tensor_buffer(const struct ggml_tensor* tensor) {
+    if (tensor == nullptr) {
+        return nullptr;
+    }
+
+    return tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+}
+
+__STATIC_INLINE__ bool ggml_backend_tensor_is_host_accessible(const struct ggml_tensor* tensor) {
+    if (tensor == nullptr || tensor->data == nullptr) {
+        return false;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_tensor_buffer(tensor);
+    return buffer == nullptr || ggml_backend_buffer_is_host(buffer);
+}
+
+__STATIC_INLINE__ size_t ggml_backend_tensor_offset(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
+    return (size_t)(i0 * tensor->nb[0] + i1 * tensor->nb[1] + i2 * tensor->nb[2] + i3 * tensor->nb[3]);
+}
+
+template <typename T>
+__STATIC_INLINE__ void ggml_backend_tensor_write_scalar(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, T value) {
+    const size_t offset = ggml_backend_tensor_offset(tensor, i0, i1, i2, i3);
+
+    if (ggml_backend_tensor_is_host_accessible(tensor)) {
+        auto* dst = reinterpret_cast<T*>(reinterpret_cast<char*>(tensor->data) + offset);
+        *dst      = value;
+        return;
+    }
+
+    ggml_backend_tensor_set(const_cast<struct ggml_tensor*>(tensor), &value, offset, sizeof(T));
+}
+
+__STATIC_INLINE__ void ggml_set_f32_nd(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, float value) {
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int8_t>(value));
+            break;
+        case GGML_TYPE_I16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int16_t>(value));
+            break;
+        case GGML_TYPE_I32:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int32_t>(value));
+            break;
+        case GGML_TYPE_F16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_fp16(value));
+            break;
+        case GGML_TYPE_BF16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_bf16(value));
+            break;
+        case GGML_TYPE_F32:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, value);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+}
+
+__STATIC_INLINE__ void ggml_set_f32_1d(const struct ggml_tensor* tensor, int i, float value) {
+    if (!ggml_is_contiguous(tensor)) {
+        int64_t id[4] = {0, 0, 0, 0};
+        ggml_unravel_index(tensor, i, &id[0], &id[1], &id[2], &id[3]);
+        ggml_set_f32_nd(tensor, id[0], id[1], id[2], id[3], value);
+        return;
+    }
+
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int8_t>(value));
+            break;
+        case GGML_TYPE_I16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int16_t>(value));
+            break;
+        case GGML_TYPE_I32:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int32_t>(value));
+            break;
+        case GGML_TYPE_F16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_fp16(value));
+            break;
+        case GGML_TYPE_BF16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_bf16(value));
+            break;
+        case GGML_TYPE_F32:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, value);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+}
+
+__STATIC_INLINE__ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context* ctx, struct ggml_cgraph* cgraph, int n_threads) {
+    (void)ctx;
+
+    // The legacy ggml_graph_compute_with_ctx() symbol lives in ggml-cpu, but
+    // the backend proc table does not expose it in GGML_BACKEND_DL mode.
+    // Recreate the old behavior by initializing the CPU backend explicitly and
+    // executing the graph through the generic backend API.
+    ggml_backend_t backend = ggml_backend_cpu_init();
+    if (backend == nullptr) {
+        return GGML_STATUS_ALLOC_FAILED;
+    }
+
+    ggml_backend_cpu_set_n_threads(backend, n_threads);
+
+    const enum ggml_status status = ggml_backend_graph_compute(backend, cgraph);
+    ggml_backend_free(backend);
+
+    return status;
+}
+
+__STATIC_INLINE__ ggml_tensor* ggml_set_f32(struct ggml_tensor* tensor, float value) {
+    GGML_ASSERT(tensor != nullptr);
+
+    if (ggml_backend_tensor_is_host_accessible(tensor) && ggml_is_contiguous(tensor)) {
+        const int64_t nelements = ggml_nelements(tensor);
+
+        switch (tensor->type) {
+            case GGML_TYPE_I8: {
+                auto* data     = reinterpret_cast<int8_t*>(tensor->data);
+                const int8_t v = static_cast<int8_t>(value);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = v;
+                }
+            } break;
+            case GGML_TYPE_I16: {
+                auto* data      = reinterpret_cast<int16_t*>(tensor->data);
+                const int16_t v = static_cast<int16_t>(value);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = v;
+                }
+            } break;
+            case GGML_TYPE_I32: {
+                auto* data      = reinterpret_cast<int32_t*>(tensor->data);
+                const int32_t v = static_cast<int32_t>(value);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = v;
+                }
+            } break;
+            case GGML_TYPE_F16: {
+                auto* data          = reinterpret_cast<ggml_fp16_t*>(tensor->data);
+                const ggml_fp16_t v = ggml_fp32_to_fp16(value);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = v;
+                }
+            } break;
+            case GGML_TYPE_BF16: {
+                auto* data          = reinterpret_cast<ggml_bf16_t*>(tensor->data);
+                const ggml_bf16_t v = ggml_fp32_to_bf16(value);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = v;
+                }
+            } break;
+            case GGML_TYPE_F32: {
+                auto* data = reinterpret_cast<float*>(tensor->data);
+                for (int64_t i = 0; i < nelements; ++i) {
+                    data[i] = value;
+                }
+            } break;
+            default:
+                GGML_ABORT("fatal error");
+        }
+
+        return tensor;
+    }
+
+    const int64_t nelements = ggml_nelements(tensor);
+    for (int64_t i = 0; i < nelements; ++i) {
+        ggml_set_f32_1d(tensor, static_cast<int>(i), value);
+    }
+
+    return tensor;
+}
+
+#endif

src/ggml_graph_cut.cpp

@@ -0,0 +1,676 @@
+#include "ggml_graph_cut.h"
+
+#include <algorithm>
+#include <cstring>
+#include <map>
+#include <set>
+#include <sstream>
+#include <stack>
+#include <unordered_map>
+
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include "util.h"
+
+#include "../ggml/src/ggml-impl.h"
+
+namespace sd::ggml_graph_cut {
+
+    static std::string graph_cut_tensor_display_name(const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return "<null>";
+        }
+        if (tensor->name[0] != '\0') {
+            return tensor->name;
+        }
+        return sd_format("<tensor@%p>", (const void*)tensor);
+    }
+
+    static int graph_leaf_index(ggml_cgraph* gf, const ggml_tensor* tensor) {
+        GGML_ASSERT(gf != nullptr);
+        GGML_ASSERT(tensor != nullptr);
+        for (int i = 0; i < gf->n_leafs; ++i) {
+            if (gf->leafs[i] == tensor) {
+                return i;
+            }
+        }
+        return -1;
+    }
+
+    static bool is_params_tensor(const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                                 const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return false;
+        }
+        return params_tensor_set.find(tensor) != params_tensor_set.end();
+    }
+
+    static Plan::InputShape input_shape(const ggml_tensor* tensor) {
+        Plan::InputShape shape;
+        if (tensor == nullptr) {
+            return shape;
+        }
+        shape.type = tensor->type;
+        for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+            shape.ne[static_cast<size_t>(i)] = tensor->ne[i];
+        }
+        return shape;
+    }
+
+    static size_t graph_cut_segment_vram_bytes(const Segment& segment) {
+        return segment.compute_buffer_size +
+               segment.input_param_bytes +
+               segment.input_previous_cut_bytes +
+               segment.output_bytes;
+    }
+
+    static Segment make_segment_seed(const Plan& plan,
+                                     size_t start_segment_index,
+                                     size_t end_segment_index) {
+        GGML_ASSERT(start_segment_index < plan.segments.size());
+        GGML_ASSERT(end_segment_index < plan.segments.size());
+        GGML_ASSERT(start_segment_index <= end_segment_index);
+
+        Segment seed;
+        const auto& start_segment  = plan.segments[start_segment_index];
+        const auto& target_segment = plan.segments[end_segment_index];
+        std::unordered_set<int> seen_output_node_indices;
+        for (size_t seg_idx = start_segment_index; seg_idx <= end_segment_index; ++seg_idx) {
+            for (int output_node_index : plan.segments[seg_idx].output_node_indices) {
+                if (seen_output_node_indices.insert(output_node_index).second) {
+                    seed.output_node_indices.push_back(output_node_index);
+                }
+            }
+        }
+        if (start_segment_index == end_segment_index) {
+            seed.group_name = target_segment.group_name;
+        } else {
+            seed.group_name = sd_format("%s..%s",
+                                        start_segment.group_name.c_str(),
+                                        target_segment.group_name.c_str());
+        }
+        return seed;
+    }
+
+    static void build_segment(ggml_cgraph* gf,
+                              Plan& plan,
+                              Segment& segment,
+                              const std::unordered_map<const ggml_tensor*, int>& producer_index,
+                              std::unordered_set<int>& available_cut_output_node_indices,
+                              ggml_backend_t backend,
+                              const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                              const char* log_desc) {
+        std::set<int> internal_nodes;
+        std::unordered_set<const ggml_tensor*> input_seen;
+        std::vector<Segment::InputRef> input_refs;
+
+        std::stack<ggml_tensor*> work_stack;
+        for (int output_node_index : segment.output_node_indices) {
+            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
+            if (output != nullptr) {
+                work_stack.push(output);
+            }
+        }
+
+        while (!work_stack.empty()) {
+            ggml_tensor* tensor = work_stack.top();
+            work_stack.pop();
+
+            if (tensor == nullptr) {
+                continue;
+            }
+
+            auto producer_it = producer_index.find(tensor);
+            if (producer_it == producer_index.end()) {
+                if (input_seen.insert(tensor).second) {
+                    Segment::InputRef input_ref;
+                    input_ref.type         = is_params_tensor(params_tensor_set, tensor) ? Segment::INPUT_PARAM : Segment::INPUT_EXTERNAL;
+                    input_ref.display_name = graph_cut_tensor_display_name(tensor);
+                    input_ref.leaf_index   = graph_leaf_index(gf, tensor);
+                    input_refs.push_back(std::move(input_ref));
+                }
+                continue;
+            }
+
+            int node_idx = producer_it->second;
+            if (available_cut_output_node_indices.find(node_idx) != available_cut_output_node_indices.end()) {
+                if (input_seen.insert(tensor).second) {
+                    Segment::InputRef input_ref;
+                    input_ref.type         = Segment::INPUT_PREVIOUS_CUT;
+                    input_ref.display_name = graph_cut_tensor_display_name(tensor);
+                    input_ref.node_index   = node_idx;
+                    input_refs.push_back(std::move(input_ref));
+                }
+                continue;
+            }
+
+            if (!internal_nodes.insert(node_idx).second) {
+                continue;
+            }
+
+            ggml_tensor* node = ggml_graph_node(gf, node_idx);
+            for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
+                if (node->src[src_idx] != nullptr) {
+                    work_stack.push(node->src[src_idx]);
+                }
+            }
+        }
+
+        if (!internal_nodes.empty()) {
+            segment.internal_node_indices.assign(internal_nodes.begin(), internal_nodes.end());
+        }
+
+        std::sort(input_refs.begin(),
+                  input_refs.end(),
+                  [](const Segment::InputRef& a, const Segment::InputRef& b) {
+                      if (a.type != b.type) {
+                          return a.type < b.type;
+                      }
+                      return a.display_name < b.display_name;
+                  });
+        segment.input_refs = input_refs;
+        for (const auto& input : input_refs) {
+            ggml_tensor* current_input = input_tensor(gf, input);
+            size_t tensor_bytes        = current_input == nullptr
+                                             ? 0
+                                             : (input.type == Segment::INPUT_PREVIOUS_CUT
+                                                    ? cache_tensor_bytes(current_input)
+                                                    : ggml_nbytes(current_input));
+            switch (input.type) {
+                case Segment::INPUT_PREVIOUS_CUT:
+                    segment.input_previous_cut_bytes += tensor_bytes;
+                    break;
+                case Segment::INPUT_PARAM:
+                    segment.input_param_bytes += tensor_bytes;
+                    break;
+                case Segment::INPUT_EXTERNAL:
+                default:
+                    segment.input_external_bytes += tensor_bytes;
+                    break;
+            }
+        }
+        for (int output_node_index : segment.output_node_indices) {
+            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
+            segment.output_bytes += cache_tensor_bytes(output);
+        }
+        segment.compute_buffer_size = measure_segment_compute_buffer(backend, gf, segment, log_desc);
+
+        for (int output_node_index : segment.output_node_indices) {
+            available_cut_output_node_indices.insert(output_node_index);
+        }
+        plan.segments.push_back(std::move(segment));
+    }
+
+    bool is_graph_cut_tensor(const ggml_tensor* tensor) {
+        if (tensor == nullptr || tensor->name[0] == '\0') {
+            return false;
+        }
+        return std::strncmp(tensor->name, GGML_RUNNER_CUT_PREFIX, std::strlen(GGML_RUNNER_CUT_PREFIX)) == 0;
+    }
+
+    std::string make_graph_cut_name(const std::string& group, const std::string& output) {
+        return std::string(GGML_RUNNER_CUT_PREFIX) + group + "|" + output;
+    }
+
+    void mark_graph_cut(ggml_tensor* tensor, const std::string& group, const std::string& output) {
+        if (tensor == nullptr) {
+            return;
+        }
+        auto name = make_graph_cut_name(group, output);
+        ggml_set_name(tensor, name.c_str());
+    }
+
+    int leaf_count(ggml_cgraph* gf) {
+        GGML_ASSERT(gf != nullptr);
+        return gf->n_leafs;
+    }
+
+    ggml_tensor* leaf_tensor(ggml_cgraph* gf, int leaf_index) {
+        GGML_ASSERT(gf != nullptr);
+        if (leaf_index < 0 || leaf_index >= gf->n_leafs) {
+            return nullptr;
+        }
+        return gf->leafs[leaf_index];
+    }
+
+    ggml_backend_buffer_t tensor_buffer(const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return nullptr;
+        }
+        return tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+    }
+
+    ggml_tensor* cache_source_tensor(ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return nullptr;
+        }
+        return tensor->view_src ? tensor->view_src : tensor;
+    }
+
+    size_t cache_tensor_bytes(const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return 0;
+        }
+        const ggml_tensor* cache_src = tensor->view_src ? tensor->view_src : tensor;
+        return ggml_nbytes(cache_src);
+    }
+
+    bool plan_matches_graph(ggml_cgraph* gf, const Plan& plan) {
+        GGML_ASSERT(gf != nullptr);
+        if (ggml_graph_n_nodes(gf) != plan.n_nodes || gf->n_leafs != plan.n_leafs) {
+            return false;
+        }
+        for (const auto& input_shape_ref : plan.input_shapes) {
+            if (input_shape_ref.leaf_index < 0 || input_shape_ref.leaf_index >= gf->n_leafs) {
+                return false;
+            }
+            ggml_tensor* leaf = gf->leafs[input_shape_ref.leaf_index];
+            if (leaf == nullptr || input_shape_ref.type != leaf->type) {
+                return false;
+            }
+            for (int d = 0; d < GGML_MAX_DIMS; ++d) {
+                if (input_shape_ref.ne[static_cast<size_t>(d)] != leaf->ne[d]) {
+                    return false;
+                }
+            }
+        }
+        return true;
+    }
+
+    ggml_tensor* output_tensor(ggml_cgraph* gf, const Segment& segment, size_t output_index) {
+        GGML_ASSERT(gf != nullptr);
+        if (output_index >= segment.output_node_indices.size()) {
+            return nullptr;
+        }
+        int node_index = segment.output_node_indices[output_index];
+        if (node_index < 0 || node_index >= ggml_graph_n_nodes(gf)) {
+            return nullptr;
+        }
+        return ggml_graph_node(gf, node_index);
+    }
+
+    ggml_tensor* input_tensor(ggml_cgraph* gf, const Segment::InputRef& input_ref) {
+        GGML_ASSERT(gf != nullptr);
+        if (input_ref.type == Segment::INPUT_PREVIOUS_CUT) {
+            if (input_ref.node_index < 0 || input_ref.node_index >= ggml_graph_n_nodes(gf)) {
+                return nullptr;
+            }
+            return ggml_graph_node(gf, input_ref.node_index);
+        }
+        if (input_ref.leaf_index < 0 || input_ref.leaf_index >= gf->n_leafs) {
+            return nullptr;
+        }
+        return leaf_tensor(gf, input_ref.leaf_index);
+    }
+
+    std::vector<ggml_tensor*> param_tensors(ggml_cgraph* gf, const Segment& segment) {
+        GGML_ASSERT(gf != nullptr);
+        std::vector<ggml_tensor*> tensors;
+        std::unordered_set<ggml_tensor*> seen_tensors;
+        tensors.reserve(segment.input_refs.size());
+        seen_tensors.reserve(segment.input_refs.size());
+        for (const auto& input_ref : segment.input_refs) {
+            if (input_ref.type != Segment::INPUT_PARAM) {
+                continue;
+            }
+            ggml_tensor* tensor = input_tensor(gf, input_ref);
+            if (tensor == nullptr) {
+                continue;
+            }
+            if (seen_tensors.insert(tensor).second) {
+                tensors.push_back(tensor);
+            }
+        }
+        return tensors;
+    }
+
+    std::vector<ggml_tensor*> runtime_param_tensors(ggml_cgraph* gf, const Segment& segment, const char* log_desc) {
+        std::vector<ggml_tensor*> tensors = param_tensors(gf, segment);
+        std::vector<ggml_tensor*> filtered_tensors;
+        filtered_tensors.reserve(tensors.size());
+        for (ggml_tensor* tensor : tensors) {
+            if (tensor_buffer(tensor) == nullptr) {
+                LOG_WARN("%s graph cut skipping param input without buffer: segment=%s tensor=%s",
+                         log_desc == nullptr ? "unknown" : log_desc,
+                         segment.group_name.c_str(),
+                         tensor->name);
+                continue;
+            }
+            filtered_tensors.push_back(tensor);
+        }
+        return filtered_tensors;
+    }
+
+    std::unordered_set<std::string> collect_future_input_names(ggml_cgraph* gf,
+                                                               const Plan& plan,
+                                                               size_t current_segment_index) {
+        GGML_ASSERT(gf != nullptr);
+        std::unordered_set<std::string> future_input_names;
+        for (size_t seg_idx = current_segment_index + 1; seg_idx < plan.segments.size(); ++seg_idx) {
+            const auto& segment = plan.segments[seg_idx];
+            for (const auto& input_ref : segment.input_refs) {
+                if (input_ref.type != Segment::INPUT_PREVIOUS_CUT) {
+                    continue;
+                }
+                ggml_tensor* current_input = input_tensor(gf, input_ref);
+                if (current_input != nullptr && current_input->name[0] != '\0') {
+                    future_input_names.insert(current_input->name);
+                }
+            }
+        }
+        return future_input_names;
+    }
+
+    ggml_cgraph* build_segment_graph(ggml_cgraph* gf,
+                                     const Segment& segment,
+                                     ggml_context** graph_ctx_out) {
+        GGML_ASSERT(gf != nullptr);
+        GGML_ASSERT(graph_ctx_out != nullptr);
+
+        const size_t graph_size = segment.internal_node_indices.size() + segment.input_refs.size() + 8;
+        ggml_init_params params = {
+            /*.mem_size   =*/ggml_graph_overhead_custom(graph_size, false) + 1024,
+            /*.mem_buffer =*/nullptr,
+            /*.no_alloc   =*/true,
+        };
+        ggml_context* graph_ctx = ggml_init(params);
+        GGML_ASSERT(graph_ctx != nullptr);
+        ggml_cgraph* segment_graph = ggml_new_graph_custom(graph_ctx, graph_size, false);
+        GGML_ASSERT(segment_graph != nullptr);
+
+        for (const auto& input : segment.input_refs) {
+            ggml_tensor* current_input = input_tensor(gf, input);
+            if (current_input == nullptr) {
+                continue;
+            }
+            GGML_ASSERT(segment_graph->n_leafs < segment_graph->size);
+            segment_graph->leafs[segment_graph->n_leafs++] = current_input;
+        }
+
+        for (int output_node_index : segment.output_node_indices) {
+            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
+            if (output == nullptr) {
+                continue;
+            }
+            ggml_set_output(output);
+        }
+        for (int node_idx : segment.internal_node_indices) {
+            ggml_graph_add_node(segment_graph, ggml_graph_node(gf, node_idx));
+        }
+        *graph_ctx_out = graph_ctx;
+        return segment_graph;
+    }
+
+    size_t measure_segment_compute_buffer(ggml_backend_t backend,
+                                          ggml_cgraph* gf,
+                                          const Segment& segment,
+                                          const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        if (segment.internal_node_indices.empty()) {
+            return 0;
+        }
+
+        ggml_context* graph_ctx    = nullptr;
+        ggml_cgraph* segment_graph = build_segment_graph(gf, segment, &graph_ctx);
+        ggml_gallocr_t allocr      = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
+
+        size_t sizes[1] = {0};
+        ggml_gallocr_reserve_n_size(
+            allocr,
+            segment_graph,
+            nullptr,
+            nullptr,
+            sizes);
+        size_t buffer_size = sizes[0];
+
+        ggml_gallocr_free(allocr);
+        ggml_free(graph_ctx);
+        return buffer_size;
+    }
+
+    Plan build_plan(ggml_backend_t backend,
+                    ggml_cgraph* gf,
+                    const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                    const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        Plan plan;
+        plan.available    = true;
+        const int n_nodes = ggml_graph_n_nodes(gf);
+        if (n_nodes <= 0) {
+            return plan;
+        }
+        plan.n_nodes = n_nodes;
+        plan.n_leafs = gf->n_leafs;
+        for (int i = 0; i < gf->n_leafs; ++i) {
+            ggml_tensor* leaf = gf->leafs[i];
+            if (is_params_tensor(params_tensor_set, leaf)) {
+                continue;
+            }
+            auto shape       = input_shape(leaf);
+            shape.leaf_index = i;
+            plan.input_shapes.push_back(shape);
+        }
+
+        std::unordered_map<const ggml_tensor*, int> producer_index;
+        producer_index.reserve(static_cast<size_t>(n_nodes));
+        for (int i = 0; i < n_nodes; ++i) {
+            producer_index[ggml_graph_node(gf, i)] = i;
+        }
+
+        std::vector<Segment> grouped_segments;
+        std::unordered_map<std::string, size_t> group_to_segment;
+        for (int i = 0; i < n_nodes; ++i) {
+            ggml_tensor* node = ggml_graph_node(gf, i);
+            if (!is_graph_cut_tensor(node)) {
+                continue;
+            }
+
+            plan.has_cuts = true;
+            std::string full_name(node->name);
+            std::string payload = full_name.substr(std::strlen(GGML_RUNNER_CUT_PREFIX));
+            size_t sep          = payload.find('|');
+            std::string group   = sep == std::string::npos ? payload : payload.substr(0, sep);
+
+            auto it = group_to_segment.find(group);
+            if (it == group_to_segment.end()) {
+                Segment segment;
+                segment.group_name = group;
+                segment.output_node_indices.push_back(i);
+                group_to_segment[group] = grouped_segments.size();
+                grouped_segments.push_back(std::move(segment));
+            } else {
+                auto& segment = grouped_segments[it->second];
+                segment.output_node_indices.push_back(i);
+            }
+        }
+
+        if (!plan.has_cuts) {
+            return plan;
+        }
+
+        std::unordered_set<int> available_cut_output_node_indices;
+        available_cut_output_node_indices.reserve(static_cast<size_t>(n_nodes));
+        for (auto& segment : grouped_segments) {
+            build_segment(gf,
+                          plan,
+                          segment,
+                          producer_index,
+                          available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          log_desc);
+        }
+
+        ggml_tensor* final_output = ggml_graph_node(gf, -1);
+        if (final_output != nullptr && available_cut_output_node_indices.find(n_nodes - 1) == available_cut_output_node_indices.end()) {
+            Segment final_segment;
+            final_segment.group_name = "ggml_runner.final";
+            final_segment.output_node_indices.push_back(n_nodes - 1);
+            build_segment(gf,
+                          plan,
+                          final_segment,
+                          producer_index,
+                          available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          log_desc);
+        }
+
+        return plan;
+    }
+
+    Plan apply_max_vram_budget(ggml_cgraph* gf,
+                               const Plan& base_plan,
+                               size_t max_graph_vram_bytes,
+                               ggml_backend_t backend,
+                               const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                               const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        int64_t t_budget_begin = ggml_time_ms();
+        if (max_graph_vram_bytes == 0 || !base_plan.has_cuts || base_plan.segments.size() <= 1) {
+            return base_plan;
+        }
+
+        const int n_nodes = ggml_graph_n_nodes(gf);
+        std::unordered_map<const ggml_tensor*, int> producer_index;
+        producer_index.reserve(static_cast<size_t>(n_nodes));
+        for (int i = 0; i < n_nodes; ++i) {
+            producer_index[ggml_graph_node(gf, i)] = i;
+        }
+
+        Plan merged_plan;
+        merged_plan.available = true;
+        merged_plan.has_cuts  = base_plan.has_cuts;
+        merged_plan.valid     = base_plan.valid;
+        merged_plan.n_nodes   = base_plan.n_nodes;
+        merged_plan.n_leafs   = base_plan.n_leafs;
+
+        std::unordered_set<int> available_cut_output_node_indices;
+        available_cut_output_node_indices.reserve(static_cast<size_t>(n_nodes));
+
+        size_t start_segment_index = 0;
+        while (start_segment_index < base_plan.segments.size()) {
+            Plan single_plan;
+            auto single_available_cut_output_node_indices = available_cut_output_node_indices;
+            auto single_seed                              = make_segment_seed(base_plan,
+                                                                              start_segment_index,
+                                                                              start_segment_index);
+            build_segment(gf,
+                          single_plan,
+                          single_seed,
+                          producer_index,
+                          single_available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          log_desc);
+            GGML_ASSERT(!single_plan.segments.empty());
+
+            size_t best_end_segment_index = start_segment_index;
+            bool can_merge_next_segment   = graph_cut_segment_vram_bytes(single_plan.segments.back()) <= max_graph_vram_bytes;
+
+            while (can_merge_next_segment && best_end_segment_index + 1 < base_plan.segments.size()) {
+                const size_t next_end_segment_index = best_end_segment_index + 1;
+                Plan candidate_plan;
+                auto candidate_available_cut_output_node_indices = available_cut_output_node_indices;
+                auto candidate_seed                              = make_segment_seed(base_plan,
+                                                                                     start_segment_index,
+                                                                                     next_end_segment_index);
+                build_segment(gf,
+                              candidate_plan,
+                              candidate_seed,
+                              producer_index,
+                              candidate_available_cut_output_node_indices,
+                              backend,
+                              params_tensor_set,
+                              log_desc);
+                GGML_ASSERT(!candidate_plan.segments.empty());
+
+                const auto& candidate_segment = candidate_plan.segments.back();
+                if (graph_cut_segment_vram_bytes(candidate_segment) > max_graph_vram_bytes) {
+                    break;
+                }
+
+                best_end_segment_index = next_end_segment_index;
+            }
+
+            auto best_seed = make_segment_seed(base_plan,
+                                               start_segment_index,
+                                               best_end_segment_index);
+            build_segment(gf,
+                          merged_plan,
+                          best_seed,
+                          producer_index,
+                          available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          log_desc);
+            start_segment_index = best_end_segment_index + 1;
+        }
+
+        if (log_desc != nullptr && merged_plan.segments.size() != base_plan.segments.size()) {
+            LOG_INFO("%s graph cut max_vram=%.2f MB merged %zu segments -> %zu segments",
+                     log_desc,
+                     max_graph_vram_bytes / 1024.0 / 1024.0,
+                     base_plan.segments.size(),
+                     merged_plan.segments.size());
+        }
+
+        if (log_desc != nullptr) {
+            LOG_INFO("%s graph cut max_vram budget merge took %lld ms",
+                     log_desc,
+                     ggml_time_ms() - t_budget_begin);
+        }
+
+        return merged_plan;
+    }
+
+    Plan resolve_plan(ggml_backend_t backend,
+                      ggml_cgraph* gf,
+                      PlanCache* cache,
+                      size_t max_graph_vram_bytes,
+                      const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                      const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        GGML_ASSERT(cache != nullptr);
+
+        int64_t t_prepare_begin = ggml_time_ms();
+        Plan base_plan;
+        int64_t t_plan_begin = ggml_time_ms();
+        if (cache->graph_cut_plan.available && plan_matches_graph(gf, cache->graph_cut_plan)) {
+            base_plan = cache->graph_cut_plan;
+        } else {
+            base_plan                                = build_plan(backend, gf, params_tensor_set, log_desc);
+            cache->graph_cut_plan                    = base_plan;
+            cache->graph_cut_plan.available          = true;
+            cache->budgeted_graph_cut_plan.available = false;
+            if (log_desc != nullptr) {
+                LOG_INFO("%s build cached graph cut plan done (taking %lld ms)", log_desc, ggml_time_ms() - t_plan_begin);
+            }
+        }
+
+        Plan resolved_plan = base_plan;
+        if (max_graph_vram_bytes > 0 && base_plan.has_cuts) {
+            if (cache->budgeted_graph_cut_plan.available &&
+                cache->budgeted_graph_cut_plan_max_vram_bytes == max_graph_vram_bytes &&
+                plan_matches_graph(gf, cache->budgeted_graph_cut_plan)) {
+                resolved_plan = cache->budgeted_graph_cut_plan;
+            } else {
+                resolved_plan                                 = apply_max_vram_budget(gf,
+                                                                                      base_plan,
+                                                                                      max_graph_vram_bytes,
+                                                                                      backend,
+                                                                                      params_tensor_set,
+                                                                                      log_desc);
+                cache->budgeted_graph_cut_plan                = resolved_plan;
+                cache->budgeted_graph_cut_plan.available      = true;
+                cache->budgeted_graph_cut_plan_max_vram_bytes = max_graph_vram_bytes;
+            }
+        }
+        return resolved_plan;
+    }
+
+}  // namespace sd::ggml_graph_cut

src/ggml_graph_cut.h

@@ -0,0 +1,104 @@
+#ifndef __SD_GGML_GRAPH_CUT_H__
+#define __SD_GGML_GRAPH_CUT_H__
+
+#include <array>
+#include <string>
+#include <unordered_set>
+#include <vector>
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+namespace sd::ggml_graph_cut {
+
+    struct Segment {
+        enum InputType {
+            INPUT_EXTERNAL = 0,
+            INPUT_PREVIOUS_CUT,
+            INPUT_PARAM,
+        };
+
+        struct InputRef {
+            InputType type = INPUT_EXTERNAL;
+            std::string display_name;
+            int leaf_index = -1;
+            int node_index = -1;
+        };
+
+        size_t compute_buffer_size      = 0;
+        size_t output_bytes             = 0;
+        size_t input_external_bytes     = 0;
+        size_t input_previous_cut_bytes = 0;
+        size_t input_param_bytes        = 0;
+        std::string group_name;
+        std::vector<int> internal_node_indices;
+        std::vector<int> output_node_indices;
+        std::vector<InputRef> input_refs;
+    };
+
+    struct Plan {
+        struct InputShape {
+            int leaf_index                        = -1;
+            ggml_type type                        = GGML_TYPE_COUNT;
+            std::array<int64_t, GGML_MAX_DIMS> ne = {0, 0, 0, 0};
+        };
+
+        bool available = false;
+        bool has_cuts  = false;
+        bool valid     = true;
+        int n_nodes    = 0;
+        int n_leafs    = 0;
+        std::vector<InputShape> input_shapes;
+        std::vector<Segment> segments;
+    };
+
+    struct PlanCache {
+        Plan graph_cut_plan;
+        Plan budgeted_graph_cut_plan;
+        size_t budgeted_graph_cut_plan_max_vram_bytes = 0;
+    };
+
+    static constexpr const char* GGML_RUNNER_CUT_PREFIX = "ggml_runner_cut:";
+
+    bool is_graph_cut_tensor(const ggml_tensor* tensor);
+    std::string make_graph_cut_name(const std::string& group, const std::string& output);
+    void mark_graph_cut(ggml_tensor* tensor, const std::string& group, const std::string& output);
+    int leaf_count(ggml_cgraph* gf);
+    ggml_tensor* leaf_tensor(ggml_cgraph* gf, int leaf_index);
+    ggml_backend_buffer_t tensor_buffer(const ggml_tensor* tensor);
+    ggml_tensor* cache_source_tensor(ggml_tensor* tensor);
+    size_t cache_tensor_bytes(const ggml_tensor* tensor);
+    bool plan_matches_graph(ggml_cgraph* gf, const Plan& plan);
+    ggml_tensor* output_tensor(ggml_cgraph* gf, const Segment& segment, size_t output_index);
+    ggml_tensor* input_tensor(ggml_cgraph* gf, const Segment::InputRef& input_ref);
+    std::vector<ggml_tensor*> param_tensors(ggml_cgraph* gf, const Segment& segment);
+    std::vector<ggml_tensor*> runtime_param_tensors(ggml_cgraph* gf, const Segment& segment, const char* log_desc);
+    std::unordered_set<std::string> collect_future_input_names(ggml_cgraph* gf,
+                                                               const Plan& plan,
+                                                               size_t current_segment_index);
+    ggml_cgraph* build_segment_graph(ggml_cgraph* gf,
+                                     const Segment& segment,
+                                     ggml_context** graph_ctx_out);
+    size_t measure_segment_compute_buffer(ggml_backend_t backend,
+                                          ggml_cgraph* gf,
+                                          const Segment& segment,
+                                          const char* log_desc);
+    Plan build_plan(ggml_backend_t backend,
+                    ggml_cgraph* gf,
+                    const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                    const char* log_desc);
+    Plan apply_max_vram_budget(ggml_cgraph* gf,
+                               const Plan& base_plan,
+                               size_t max_graph_vram_bytes,
+                               ggml_backend_t backend,
+                               const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                               const char* log_desc);
+    Plan resolve_plan(ggml_backend_t backend,
+                      ggml_cgraph* gf,
+                      PlanCache* cache,
+                      size_t max_graph_vram_bytes,
+                      const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                      const char* log_desc);
+}  // namespace sd::ggml_graph_cut
+
+#endif

src/latent-preview.h

@@ -1,6 +1,8 @@
+#include <algorithm>
 #include <cstddef>
 #include <cstdint>
 #include "ggml.h"
+#include "tensor.hpp"
 
 const float wan_21_latent_rgb_proj[16][3] = {
     {0.015123f, -0.148418f, 0.479828f},
@@ -232,3 +234,67 @@ void preview_latent_video(uint8_t* buffer, ggml_tensor* latents, const float (*l
         }
     }
 }
+
+static inline bool preview_latent_tensor_is_video(const sd::Tensor<float>& latents) {
+    return latents.dim() == 5;
+}
+
+void preview_latent_video(uint8_t* buffer, const sd::Tensor<float>& latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int patch_size) {
+    uint32_t latent_width  = static_cast<uint32_t>(latents.shape()[0]);
+    uint32_t latent_height = static_cast<uint32_t>(latents.shape()[1]);
+    bool is_video          = preview_latent_tensor_is_video(latents);
+    uint32_t frames        = is_video ? static_cast<uint32_t>(latents.shape()[2]) : 1;
+    uint32_t dim           = is_video ? static_cast<uint32_t>(latents.shape()[3]) : static_cast<uint32_t>(latents.shape()[2]);
+
+    uint32_t rgb_width     = latent_width * patch_size;
+    uint32_t rgb_height    = latent_height * patch_size;
+    uint32_t unpatched_dim = dim / (patch_size * patch_size);
+
+    for (uint32_t k = 0; k < frames; k++) {
+        for (uint32_t rgb_x = 0; rgb_x < rgb_width; rgb_x++) {
+            for (uint32_t rgb_y = 0; rgb_y < rgb_height; rgb_y++) {
+                uint32_t latent_x = rgb_x / patch_size;
+                uint32_t latent_y = rgb_y / patch_size;
+
+                uint32_t channel_offset = 0;
+                if (patch_size > 1) {
+                    channel_offset = ((rgb_y % patch_size) * patch_size + (rgb_x % patch_size));
+                }
+
+                size_t pixel_id   = k * rgb_width * rgb_height + rgb_y * rgb_width + rgb_x;
+                auto latent_value = [&](uint32_t latent_channel) -> float {
+                    return is_video
+                               ? latents.values()[latent_x + latent_width * (latent_y + latent_height * (k + frames * latent_channel))]
+                               : latents.values()[latent_x + latent_width * (latent_y + latent_height * latent_channel)];
+                };
+
+                float r = 0.f, g = 0.f, b = 0.f;
+                if (latent_rgb_proj != nullptr) {
+                    for (uint32_t d = 0; d < unpatched_dim; d++) {
+                        uint32_t latent_channel = d * patch_size * patch_size + channel_offset;
+                        float value             = latent_value(latent_channel);
+                        r += value * latent_rgb_proj[d][0];
+                        g += value * latent_rgb_proj[d][1];
+                        b += value * latent_rgb_proj[d][2];
+                    }
+                } else {
+                    r = latent_value(0);
+                    g = latent_value(1);
+                    b = latent_value(2);
+                }
+                if (latent_rgb_bias != nullptr) {
+                    r += latent_rgb_bias[0];
+                    g += latent_rgb_bias[1];
+                    b += latent_rgb_bias[2];
+                }
+                r = std::min(1.0f, std::max(0.0f, r * .5f + .5f));
+                g = std::min(1.0f, std::max(0.0f, g * .5f + .5f));
+                b = std::min(1.0f, std::max(0.0f, b * .5f + .5f));
+
+                buffer[pixel_id * 3 + 0] = (uint8_t)(r * 255);
+                buffer[pixel_id * 3 + 1] = (uint8_t)(g * 255);
+                buffer[pixel_id * 3 + 2] = (uint8_t)(b * 255);
+            }
+        }
+    }
+}

src/llm.hpp

@@ -14,468 +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);
-            }
-            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,
     };
 
@@ -483,6 +36,7 @@ namespace LLM {
         "qwen2.5vl",
         "qwen3",
         "mistral_small3.2",
+        "ministral3.3b",
     };
 
     struct LLMVisionParams {
@@ -792,6 +346,7 @@ namespace LLM {
             auto merger      = std::dynamic_pointer_cast<PatchMerger>(blocks["merger"]);
 
             auto x = patch_embed->forward(ctx, pixel_values);
+            sd::ggml_graph_cut::mark_graph_cut(x, "llm.vision.prelude", "x");
 
             x = ggml_reshape_4d(ctx->ggml_ctx, x, x->ne[0] * spatial_merge_size * spatial_merge_size, x->ne[1] / spatial_merge_size / spatial_merge_size, x->ne[2], x->ne[3]);
             x = ggml_get_rows(ctx->ggml_ctx, x, window_index);
@@ -805,9 +360,11 @@ namespace LLM {
                     mask = nullptr;
                 }
                 x = block->forward(ctx, x, pe, mask);
+                sd::ggml_graph_cut::mark_graph_cut(x, "llm.vision.blocks." + std::to_string(i), "x");
             }
 
             x = merger->forward(ctx, x);
+            sd::ggml_graph_cut::mark_graph_cut(x, "llm.vision.final", "x");
 
             x = ggml_get_rows(ctx->ggml_ctx, x, window_inverse_index);
 
@@ -867,6 +424,9 @@ 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);
@@ -949,6 +509,7 @@ namespace LLM {
             auto norm         = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
 
             auto x = embed_tokens->forward(ctx, input_ids);
+            sd::ggml_graph_cut::mark_graph_cut(x, "llm.text.prelude", "x");
 
             std::vector<ggml_tensor*> intermediate_outputs;
 
@@ -995,6 +556,10 @@ namespace LLM {
                 auto block = std::dynamic_pointer_cast<TransformerBlock>(blocks["layers." + std::to_string(i)]);
 
                 x = block->forward(ctx, x, input_pos, attention_mask);
+                if (out_layers.size() > 1) {
+                    x = ggml_cont(ctx->ggml_ctx, x);
+                }
+                sd::ggml_graph_cut::mark_graph_cut(x, "llm.text.layers." + std::to_string(i), "x");
                 if (out_layers.find(i + 1) != out_layers.end()) {
                     intermediate_outputs.push_back(x);
                 }
@@ -1082,7 +647,7 @@ 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;
@@ -1180,20 +745,21 @@ namespace LLM {
             return hidden_states;
         }
 
-        ggml_cgraph* build_graph(ggml_tensor* input_ids,
-                                 ggml_tensor* attention_mask,
-                                 std::vector<std::pair<int, ggml_tensor*>> image_embeds,
+        ggml_cgraph* build_graph(const sd::Tensor<int32_t>& input_ids_tensor,
+                                 const sd::Tensor<float>& attention_mask_tensor,
+                                 const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds_tensor,
                                  std::set<int> out_layers) {
-            ggml_cgraph* gf = ggml_new_graph(compute_ctx);
-
-            input_ids = to_backend(input_ids);
-
-            for (auto& image_embed : image_embeds) {
-                image_embed.second = to_backend(image_embed.second);
+            ggml_cgraph* gf        = ggml_new_graph(compute_ctx);
+            ggml_tensor* input_ids = make_input(input_ids_tensor);
+            std::vector<std::pair<int, ggml_tensor*>> image_embeds;
+            image_embeds.reserve(image_embeds_tensor.size());
+            for (const auto& [idx, embed_tensor] : image_embeds_tensor) {
+                ggml_tensor* embed = make_input(embed_tensor);
+                image_embeds.emplace_back(idx, embed);
             }
 
             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;
@@ -1213,8 +779,9 @@ namespace LLM {
                                                 input_pos_vec.size());
             set_backend_tensor_data(input_pos, input_pos_vec.data());
 
-            if (attention_mask != nullptr) {
-                attention_mask = to_backend(attention_mask);
+            ggml_tensor* attention_mask = nullptr;
+            if (!attention_mask_tensor.empty()) {
+                attention_mask = make_input(attention_mask_tensor);
             } else {
                 attention_mask_vec.resize(n_tokens * n_tokens);
                 for (int i0 = 0; i0 < n_tokens; i0++) {
@@ -1239,17 +806,15 @@ namespace LLM {
             return gf;
         }
 
-        bool compute(const int n_threads,
-                     ggml_tensor* input_ids,
-                     ggml_tensor* attention_mask,
-                     std::vector<std::pair<int, ggml_tensor*>> image_embeds,
-                     std::set<int> out_layers,
-                     ggml_tensor** output,
-                     ggml_context* output_ctx = nullptr) {
+        sd::Tensor<float> compute(const int n_threads,
+                                  const sd::Tensor<int32_t>& input_ids,
+                                  const sd::Tensor<float>& attention_mask,
+                                  const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds,
+                                  std::set<int> out_layers) {
             auto get_graph = [&]() -> ggml_cgraph* {
                 return build_graph(input_ids, attention_mask, image_embeds, out_layers);
             };
-            return GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
+            return take_or_empty(GGMLRunner::compute<float>(get_graph, n_threads, true));
         }
 
         int64_t get_num_image_tokens(int64_t t, int64_t h, int64_t w) {
@@ -1288,8 +853,9 @@ namespace LLM {
             return image;
         }
 
-        ggml_cgraph* build_encode_image_graph(ggml_tensor* image) {
-            ggml_cgraph* gf = new_graph_custom(LLM_GRAPH_SIZE);
+        ggml_cgraph* build_encode_image_graph(const sd::Tensor<float>& image_tensor) {
+            ggml_cgraph* gf    = new_graph_custom(LLM_GRAPH_SIZE);
+            ggml_tensor* image = make_input(image_tensor);
 
             GGML_ASSERT(image->ne[1] % (params.vision.patch_size * params.vision.spatial_merge_size) == 0);
             GGML_ASSERT(image->ne[0] % (params.vision.patch_size * params.vision.spatial_merge_size) == 0);
@@ -1301,8 +867,6 @@ namespace LLM {
             int llm_grid_w             = grid_w / params.vision.spatial_merge_size;
             int vit_merger_window_size = params.vision.window_size / params.vision.patch_size / params.vision.spatial_merge_size;
 
-            image = to_backend(image);
-
             auto pixel_values = process_image(compute_ctx, image);
 
             // window index
@@ -1411,14 +975,12 @@ namespace LLM {
             return gf;
         }
 
-        void encode_image(const int n_threads,
-                          ggml_tensor* image,
-                          ggml_tensor** output,
-                          ggml_context* output_ctx = nullptr) {
+        sd::Tensor<float> encode_image(const int n_threads,
+                                       const sd::Tensor<float>& image) {
             auto get_graph = [&]() -> ggml_cgraph* {
                 return build_encode_image_graph(image);
             };
-            GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+            return take_or_empty(GGMLRunner::compute<float>(get_graph, n_threads, false));
         }
     };
 
@@ -1433,7 +995,7 @@ 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>();
@@ -1481,7 +1043,7 @@ 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] << ", ";
@@ -1497,39 +1059,41 @@ namespace LLM {
             params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
-            ggml_context* work_ctx = ggml_init(params);
-            GGML_ASSERT(work_ctx != nullptr);
+            ggml_context* ctx = ggml_init(params);
+            GGML_ASSERT(ctx != nullptr);
             bool test_mistral          = false;
             bool test_qwen3            = true;
             bool test_vit              = false;
             bool test_decoder_with_vit = false;
 
             if (test_decoder_with_vit) {
-                ggml_tensor* image_embed = nullptr;
+                sd::Tensor<float> image_embed;
                 {
-                    auto image = load_tensor_from_file(work_ctx, "qwen2vl_normalized.bin");
-                    print_ggml_tensor(image, false, "image");
-                    ggml_tensor* out = nullptr;
+                    auto image = sd::load_tensor_from_file_as_tensor<float>("qwen2vl_normalized.bin");
+                    print_sd_tensor(image, false, "image");
+                    sd::Tensor<float> out;
 
-                    int64_t t0 = ggml_time_ms();
-                    model.encode_image(8, image, &out, work_ctx);
-                    int64_t t1 = ggml_time_ms();
+                    int64_t t0   = ggml_time_ms();
+                    auto out_opt = model.encode_image(8, image);
+                    int64_t t1   = ggml_time_ms();
 
-                    print_ggml_tensor(out, false, "image_embed");
+                    GGML_ASSERT(!out_opt.empty());
+                    out = std::move(out_opt);
+                    print_sd_tensor(out, false, "image_embed");
                     image_embed = out;
                     LOG_DEBUG("llm encode_image test done in %lldms", t1 - t0);
                 }
 
                 std::string placeholder  = "<|image_pad|>";
                 std::string img_prompt   = "Picture 1: <|vision_start|>";  // [24669, 220, 16, 25, 220, 151652]
-                int64_t num_image_tokens = image_embed->ne[1];
+                int64_t num_image_tokens = image_embed.shape()[1];
                 img_prompt.reserve(num_image_tokens * placeholder.size());
                 for (int i = 0; i < num_image_tokens; i++) {
                     img_prompt += placeholder;
                 }
                 img_prompt += "<|vision_end|>";
 
-                std::vector<std::pair<int, ggml_tensor*>> image_embeds;
+                std::vector<std::pair<int, sd::Tensor<float>>> image_embeds;
                 image_embeds.emplace_back(64, image_embed);
 
                 std::pair<int, int> prompt_attn_range;
@@ -1547,29 +1111,33 @@ namespace LLM {
                     printf("%d ", token);
                 }
                 printf("\n");
-                auto input_ids   = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                ggml_tensor* out = nullptr;
+                auto input_ids = sd::Tensor<int32_t>::from_vector(tokens);
+                sd::Tensor<float> out;
 
-                int64_t t0 = ggml_time_ms();
-                model.compute(8, input_ids, nullptr, image_embeds, {}, &out, work_ctx);
-                int64_t t1 = ggml_time_ms();
+                int64_t t0   = ggml_time_ms();
+                auto out_opt = model.compute(8, input_ids, sd::Tensor<float>(), image_embeds, {});
+                int64_t t1   = ggml_time_ms();
 
-                print_ggml_tensor(out);
+                GGML_ASSERT(!out_opt.empty());
+                out = std::move(out_opt);
+                print_sd_tensor(out);
                 LOG_DEBUG("llm test done in %lldms", t1 - t0);
             } else if (test_vit) {
-                // auto image = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 280, 280, 3);
+                // auto image = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 280, 280, 3);
                 // ggml_set_f32(image, 0.f);
-                auto image = load_tensor_from_file(work_ctx, "qwen2vl_normalized.bin");
-                print_ggml_tensor(image, false, "image");
-                ggml_tensor* out = nullptr;
+                auto image = sd::load_tensor_from_file_as_tensor<float>("qwen2vl_normalized.bin");
+                print_sd_tensor(image, false, "image");
+                sd::Tensor<float> out;
 
-                int64_t t0 = ggml_time_ms();
-                model.encode_image(8, image, &out, work_ctx);
-                int64_t t1 = ggml_time_ms();
+                int64_t t0   = ggml_time_ms();
+                auto out_opt = model.encode_image(8, image);
+                int64_t t1   = ggml_time_ms();
 
-                print_ggml_tensor(out, false, "out");
+                GGML_ASSERT(!out_opt.empty());
+                out = std::move(out_opt);
+                print_sd_tensor(out, false, "out");
 
-                // auto ref_out = load_tensor_from_file(work_ctx, "qwen2vl.bin");
+                // auto ref_out = load_tensor_from_file(ctx, "qwen2vl.bin");
                 // ggml_ext_tensor_diff(ref_out, out, 0.01f);
 
                 LOG_DEBUG("llm test done in %lldms", t1 - t0);
@@ -1587,14 +1155,16 @@ namespace LLM {
                     printf("%d ", token);
                 }
                 printf("\n");
-                auto input_ids   = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                ggml_tensor* out = nullptr;
+                auto input_ids = sd::Tensor<int32_t>::from_vector(tokens);
+                sd::Tensor<float> out;
 
-                int64_t t0 = ggml_time_ms();
-                model.compute(8, input_ids, nullptr, {}, {10, 20, 30}, &out, work_ctx);
-                int64_t t1 = ggml_time_ms();
+                int64_t t0   = ggml_time_ms();
+                auto out_opt = model.compute(8, input_ids, sd::Tensor<float>(), {}, {10, 20, 30});
+                int64_t t1   = ggml_time_ms();
 
-                print_ggml_tensor(out);
+                GGML_ASSERT(!out_opt.empty());
+                out = std::move(out_opt);
+                print_sd_tensor(out);
                 LOG_DEBUG("llm test done in %lldms", t1 - t0);
             } else if (test_qwen3) {
                 std::pair<int, int> prompt_attn_range;
@@ -1610,14 +1180,16 @@ namespace LLM {
                     printf("%d ", token);
                 }
                 printf("\n");
-                auto input_ids   = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                ggml_tensor* out = nullptr;
+                auto input_ids = sd::Tensor<int32_t>::from_vector(tokens);
+                sd::Tensor<float> out;
 
-                int64_t t0 = ggml_time_ms();
-                model.compute(8, input_ids, nullptr, {}, {35}, &out, work_ctx);
-                int64_t t1 = ggml_time_ms();
+                int64_t t0   = ggml_time_ms();
+                auto out_opt = model.compute(8, input_ids, sd::Tensor<float>(), {}, {35});
+                int64_t t1   = ggml_time_ms();
 
-                print_ggml_tensor(out);
+                GGML_ASSERT(!out_opt.empty());
+                out = std::move(out_opt);
+                print_sd_tensor(out);
                 LOG_DEBUG("llm test done in %lldms", t1 - t0);
             } else {
                 std::pair<int, int> prompt_attn_range;
@@ -1633,14 +1205,16 @@ namespace LLM {
                     printf("%d ", token);
                 }
                 printf("\n");
-                auto input_ids   = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                ggml_tensor* out = nullptr;
+                auto input_ids = sd::Tensor<int32_t>::from_vector(tokens);
+                sd::Tensor<float> out;
 
-                int64_t t0 = ggml_time_ms();
-                model.compute(8, input_ids, nullptr, {}, {}, &out, work_ctx);
-                int64_t t1 = ggml_time_ms();
+                int64_t t0   = ggml_time_ms();
+                auto out_opt = model.compute(8, input_ids, sd::Tensor<float>(), {}, {});
+                int64_t t1   = ggml_time_ms();
 
-                print_ggml_tensor(out);
+                GGML_ASSERT(!out_opt.empty());
+                out = std::move(out_opt);
+                print_sd_tensor(out);
                 LOG_DEBUG("llm test done in %lldms", t1 - t0);
             }
         }

src/lora.hpp

@@ -129,7 +129,7 @@ struct LoraModel : public GGMLRunner {
         }
     }
 
-    ggml_tensor* get_lora_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_lora_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -152,17 +152,17 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(lora_up_name);
             if (iter != lora_tensors.end()) {
-                lora_up = ggml_ext_cast_f32(ctx, iter->second);
+                lora_up = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(lora_mid_name);
             if (iter != lora_tensors.end()) {
-                lora_mid = ggml_ext_cast_f32(ctx, iter->second);
+                lora_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(lora_down_name);
             if (iter != lora_tensors.end()) {
-                lora_down = ggml_ext_cast_f32(ctx, iter->second);
+                lora_down = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             if (lora_up == nullptr || lora_down == nullptr) {
@@ -208,7 +208,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_raw_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_raw_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -225,7 +225,7 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(diff_name);
             if (iter != lora_tensors.end()) {
-                curr_updown = ggml_ext_cast_f32(ctx, iter->second);
+                curr_updown = ggml_ext_cast_f32(ctx, backend, iter->second);
             } else {
                 break;
             }
@@ -248,7 +248,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_loha_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_loha_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -276,33 +276,33 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(hada_1_down_name);
             if (iter != lora_tensors.end()) {
-                hada_1_down = ggml_ext_cast_f32(ctx, iter->second);
+                hada_1_down = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_1_up_name);
             if (iter != lora_tensors.end()) {
-                hada_1_up = ggml_ext_cast_f32(ctx, iter->second);
+                hada_1_up = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_1_mid_name);
             if (iter != lora_tensors.end()) {
-                hada_1_mid = ggml_ext_cast_f32(ctx, iter->second);
+                hada_1_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
                 hada_1_up  = ggml_cont(ctx, ggml_transpose(ctx, hada_1_up));
             }
 
             iter = lora_tensors.find(hada_2_down_name);
             if (iter != lora_tensors.end()) {
-                hada_2_down = ggml_ext_cast_f32(ctx, iter->second);
+                hada_2_down = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_2_up_name);
             if (iter != lora_tensors.end()) {
-                hada_2_up = ggml_ext_cast_f32(ctx, iter->second);
+                hada_2_up = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_2_mid_name);
             if (iter != lora_tensors.end()) {
-                hada_2_mid = ggml_ext_cast_f32(ctx, iter->second);
+                hada_2_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
                 hada_2_up  = ggml_cont(ctx, ggml_transpose(ctx, hada_2_up));
             }
 
@@ -351,7 +351,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_lokr_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_lokr_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -378,24 +378,24 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(lokr_w1_name);
             if (iter != lora_tensors.end()) {
-                lokr_w1 = ggml_ext_cast_f32(ctx, iter->second);
+                lokr_w1 = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(lokr_w2_name);
             if (iter != lora_tensors.end()) {
-                lokr_w2 = ggml_ext_cast_f32(ctx, iter->second);
+                lokr_w2 = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             int64_t rank = 1;
             if (lokr_w1 == nullptr) {
                 iter = lora_tensors.find(lokr_w1_a_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w1_a = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w1_a = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 iter = lora_tensors.find(lokr_w1_b_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w1_b = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w1_b = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 if (lokr_w1_a == nullptr || lokr_w1_b == nullptr) {
@@ -410,12 +410,12 @@ struct LoraModel : public GGMLRunner {
             if (lokr_w2 == nullptr) {
                 iter = lora_tensors.find(lokr_w2_a_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w2_a = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w2_a = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 iter = lora_tensors.find(lokr_w2_b_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w2_b = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w2_b = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 if (lokr_w2_a == nullptr || lokr_w2_b == nullptr) {
@@ -468,23 +468,23 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_tensor* model_tensor, bool with_lora_and_lokr = true) {
+    ggml_tensor* get_weight_diff(const std::string& model_tensor_name, ggml_backend_t backend, ggml_context* ctx, ggml_tensor* model_tensor, bool with_lora_and_lokr = true) {
         // lora
         ggml_tensor* diff = nullptr;
         if (with_lora_and_lokr) {
-            diff = get_lora_weight_diff(model_tensor_name, ctx);
+            diff = get_lora_weight_diff(model_tensor_name, ctx, backend);
         }
         // diff
         if (diff == nullptr) {
-            diff = get_raw_weight_diff(model_tensor_name, ctx);
+            diff = get_raw_weight_diff(model_tensor_name, ctx, backend);
         }
         // loha
         if (diff == nullptr) {
-            diff = get_loha_weight_diff(model_tensor_name, ctx);
+            diff = get_loha_weight_diff(model_tensor_name, ctx, backend);
         }
         // lokr
         if (diff == nullptr && with_lora_and_lokr) {
-            diff = get_lokr_weight_diff(model_tensor_name, ctx);
+            diff = get_lokr_weight_diff(model_tensor_name, ctx, backend);
         }
         if (diff != nullptr) {
             if (ggml_nelements(diff) < ggml_nelements(model_tensor)) {
@@ -502,6 +502,7 @@ struct LoraModel : public GGMLRunner {
     }
 
     ggml_tensor* get_out_diff(ggml_context* ctx,
+                              ggml_backend_t backend,
                               ggml_tensor* x,
                               WeightAdapter::ForwardParams forward_params,
                               const std::string& model_tensor_name) {
@@ -590,7 +591,7 @@ struct LoraModel : public GGMLRunner {
                 }
                 scale_value *= multiplier;
 
-                auto curr_out_diff = ggml_ext_lokr_forward(ctx, x, lokr_w1, lokr_w1_a, lokr_w1_b, lokr_w2, lokr_w2_a, lokr_w2_b, is_conv2d, forward_params.conv2d, scale_value);
+                auto curr_out_diff = ggml_ext_lokr_forward(ctx, backend, x, lokr_w1, lokr_w1_a, lokr_w1_b, lokr_w2, lokr_w2_a, lokr_w2_b, is_conv2d, forward_params.conv2d, scale_value);
                 if (out_diff == nullptr) {
                     out_diff = curr_out_diff;
                 } else {
@@ -761,7 +762,7 @@ struct LoraModel : public GGMLRunner {
             ggml_tensor* model_tensor     = it.second;
 
             // lora
-            ggml_tensor* diff = get_weight_diff(model_tensor_name, compute_ctx, model_tensor);
+            ggml_tensor* diff = get_weight_diff(model_tensor_name, runtime_backend, compute_ctx, model_tensor);
             if (diff == nullptr) {
                 continue;
             }
@@ -774,7 +775,7 @@ struct LoraModel : public GGMLRunner {
 
             ggml_tensor* final_tensor;
             if (model_tensor->type != GGML_TYPE_F32 && model_tensor->type != GGML_TYPE_F16) {
-                final_tensor = ggml_ext_cast_f32(compute_ctx, model_tensor);
+                final_tensor = ggml_ext_cast_f32(compute_ctx, runtime_backend, model_tensor);
                 final_tensor = ggml_add_inplace(compute_ctx, final_tensor, diff);
                 final_tensor = ggml_cpy(compute_ctx, final_tensor, model_tensor);
             } else {
@@ -792,7 +793,7 @@ struct LoraModel : public GGMLRunner {
         auto get_graph = [&]() -> ggml_cgraph* {
             return build_lora_graph(model_tensors, version);
         };
-        GGMLRunner::compute(get_graph, n_threads, false);
+        GGMLRunner::compute<float>(get_graph, n_threads, false, true);
         stat();
         for (auto item : original_tensor_to_final_tensor) {
             ggml_tensor* original_tensor = item.first;
@@ -841,34 +842,35 @@ public:
         : lora_models(lora_models) {
     }
 
-    ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name, bool with_lora_and_lokr) {
+    ggml_tensor* patch_weight(ggml_context* ctx, ggml_backend_t backend, ggml_tensor* weight, const std::string& weight_name, bool with_lora_and_lokr) {
         for (auto& lora_model : lora_models) {
-            ggml_tensor* diff = lora_model->get_weight_diff(weight_name, ctx, weight, with_lora_and_lokr);
+            ggml_tensor* diff = lora_model->get_weight_diff(weight_name, backend, ctx, weight, with_lora_and_lokr);
             if (diff == nullptr) {
                 continue;
             }
 
             if (weight->type != GGML_TYPE_F32 && weight->type != GGML_TYPE_F16) {
-                weight = ggml_ext_cast_f32(ctx, weight);
+                weight = ggml_ext_cast_f32(ctx, backend, weight);
             }
             weight = ggml_add(ctx, weight, diff);
         }
         return weight;
     }
 
-    ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name) override {
-        return patch_weight(ctx, weight, weight_name, true);
+    ggml_tensor* patch_weight(ggml_context* ctx, ggml_backend_t backend, ggml_tensor* weight, const std::string& weight_name) override {
+        return patch_weight(ctx, backend, weight, weight_name, true);
     }
 
     ggml_tensor* forward_with_lora(ggml_context* ctx,
+                                   ggml_backend_t backend,
                                    ggml_tensor* x,
                                    ggml_tensor* w,
                                    ggml_tensor* b,
                                    const std::string& prefix,
                                    WeightAdapter::ForwardParams forward_params) override {
-        w = patch_weight(ctx, w, prefix + "weight", false);
+        w = patch_weight(ctx, backend, w, prefix + "weight", false);
         if (b) {
-            b = patch_weight(ctx, b, prefix + "bias", false);
+            b = patch_weight(ctx, backend, b, prefix + "bias", false);
         }
         ggml_tensor* out;
         if (forward_params.op_type == ForwardParams::op_type_t::OP_LINEAR) {
@@ -890,7 +892,7 @@ public:
                                    forward_params.conv2d.scale);
         }
         for (auto& lora_model : lora_models) {
-            ggml_tensor* out_diff = lora_model->get_out_diff(ctx, x, forward_params, prefix + "weight");
+            ggml_tensor* out_diff = lora_model->get_out_diff(ctx, backend, x, forward_params, prefix + "weight");
             if (out_diff == nullptr) {
                 continue;
             }

src/mmdit.hpp

@@ -767,6 +767,8 @@ public:
             auto context_x = block->forward(ctx, context, x, c_mod);
             context        = context_x.first;
             x              = context_x.second;
+            sd::ggml_graph_cut::mark_graph_cut(context, "mmdit.joint_blocks." + std::to_string(i), "context");
+            sd::ggml_graph_cut::mark_graph_cut(x, "mmdit.joint_blocks." + std::to_string(i), "x");
         }
 
         x = final_layer->forward(ctx, x, c_mod);  // (N, T, patch_size ** 2 * out_channels)
@@ -809,6 +811,11 @@ public:
 
             context = context_embedder->forward(ctx, context);  // [N, L, D] aka [N, L, 1536]
         }
+        sd::ggml_graph_cut::mark_graph_cut(x, "mmdit.prelude", "x");
+        sd::ggml_graph_cut::mark_graph_cut(c, "mmdit.prelude", "c");
+        if (context != nullptr) {
+            sd::ggml_graph_cut::mark_graph_cut(context, "mmdit.prelude", "context");
+        }
 
         x = forward_core_with_concat(ctx, x, c, context, skip_layers);  // (N, H*W, patch_size ** 2 * out_channels)
 
@@ -836,17 +843,17 @@ struct MMDiTRunner : public GGMLRunner {
         mmdit.get_param_tensors(tensors, prefix);
     }
 
-    ggml_cgraph* build_graph(ggml_tensor* x,
-                             ggml_tensor* timesteps,
-                             ggml_tensor* context,
-                             ggml_tensor* y,
-                             std::vector<int> skip_layers = std::vector<int>()) {
+    ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                             const sd::Tensor<float>& timesteps_tensor,
+                             const sd::Tensor<float>& context_tensor = {},
+                             const sd::Tensor<float>& y_tensor       = {},
+                             std::vector<int> skip_layers            = std::vector<int>()) {
         ggml_cgraph* gf = new_graph_custom(MMDIT_GRAPH_SIZE);
 
-        x         = to_backend(x);
-        context   = to_backend(context);
-        y         = to_backend(y);
-        timesteps = to_backend(timesteps);
+        ggml_tensor* x         = make_input(x_tensor);
+        ggml_tensor* timesteps = make_input(timesteps_tensor);
+        ggml_tensor* context   = make_optional_input(context_tensor);
+        ggml_tensor* y         = make_optional_input(y_tensor);
 
         auto runner_ctx  = get_context();
         ggml_tensor* out = mmdit.forward(&runner_ctx,
@@ -861,14 +868,12 @@ struct MMDiTRunner : public GGMLRunner {
         return gf;
     }
 
-    bool compute(int n_threads,
-                 ggml_tensor* x,
-                 ggml_tensor* timesteps,
-                 ggml_tensor* context,
-                 ggml_tensor* y,
-                 ggml_tensor** output         = nullptr,
-                 ggml_context* output_ctx     = nullptr,
-                 std::vector<int> skip_layers = std::vector<int>()) {
+    sd::Tensor<float> compute(int n_threads,
+                              const sd::Tensor<float>& x,
+                              const sd::Tensor<float>& timesteps,
+                              const sd::Tensor<float>& context = {},
+                              const sd::Tensor<float>& y       = {},
+                              std::vector<int> skip_layers     = std::vector<int>()) {
         // x: [N, in_channels, h, w]
         // timesteps: [N, ]
         // context: [N, max_position, hidden_size]([N, 154, 4096]) or [1, max_position, hidden_size]
@@ -877,7 +882,7 @@ struct MMDiTRunner : public GGMLRunner {
             return build_graph(x, timesteps, context, y, skip_layers);
         };
 
-        return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
     }
 
     void test() {
@@ -886,35 +891,41 @@ struct MMDiTRunner : public GGMLRunner {
         params.mem_buffer = nullptr;
         params.no_alloc   = false;
 
-        ggml_context* work_ctx = ggml_init(params);
-        GGML_ASSERT(work_ctx != nullptr);
+        ggml_context* ctx = ggml_init(params);
+        GGML_ASSERT(ctx != nullptr);
 
         {
             // cpu f16: pass
             // cpu f32: pass
             // cuda f16: pass
             // cuda f32: pass
-            auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 128, 128, 16, 1);
+            sd::Tensor<float> x({128, 128, 16, 1});
             std::vector<float> timesteps_vec(1, 999.f);
-            auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
-            ggml_set_f32(x, 0.01f);
+            auto timesteps = sd::Tensor<float>::from_vector(timesteps_vec);
+            x.fill_(0.01f);
             // print_ggml_tensor(x);
 
-            auto context = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 4096, 154, 1);
-            ggml_set_f32(context, 0.01f);
+            sd::Tensor<float> context({4096, 154, 1});
+            context.fill_(0.01f);
             // print_ggml_tensor(context);
 
-            auto y = ggml_new_tensor_2d(work_ctx, GGML_TYPE_F32, 2048, 1);
-            ggml_set_f32(y, 0.01f);
+            sd::Tensor<float> y({2048, 1});
+            y.fill_(0.01f);
             // print_ggml_tensor(y);
 
-            ggml_tensor* out = nullptr;
+            sd::Tensor<float> out;
 
-            int64_t t0 = ggml_time_ms();
-            compute(8, x, timesteps, context, y, &out, work_ctx);
-            int64_t t1 = ggml_time_ms();
+            int64_t t0   = ggml_time_ms();
+            auto out_opt = compute(8,
+                                   x,
+                                   timesteps,
+                                   context,
+                                   y);
+            int64_t t1   = ggml_time_ms();
 
-            print_ggml_tensor(out);
+            GGML_ASSERT(!out_opt.empty());
+            out = std::move(out_opt);
+            print_sd_tensor(out);
             LOG_DEBUG("mmdit test done in %lldms", t1 - t0);
         }
     }

src/model.h

@@ -5,20 +5,13 @@
 #include <map>
 #include <memory>
 #include <set>
-#include <sstream>
 #include <string>
-#include <tuple>
-#include <utility>
 #include <vector>
 
 #include "ggml-backend.h"
 #include "ggml.h"
-#include "gguf.h"
-#include "json.hpp"
+#include "model_io/tensor_storage.h"
 #include "ordered_map.hpp"
-#include "zip.h"
-
-#define SD_MAX_DIMS 5
 
 enum SDVersion {
     VERSION_SD1,
@@ -28,7 +21,8 @@ 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 +44,19 @@ 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 +132,20 @@ 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 +164,8 @@ 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;
@@ -178,116 +188,10 @@ enum PMVersion {
     PM_VERSION_2,
 };
 
-struct TensorStorage {
-    std::string name;
-    ggml_type type          = GGML_TYPE_F32;
-    ggml_type expected_type = GGML_TYPE_COUNT;
-    bool is_f8_e4m3         = false;
-    bool is_f8_e5m2         = false;
-    bool is_f64             = false;
-    bool is_i64             = false;
-    int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
-    int n_dims              = 0;
-
-    size_t file_index = 0;
-    int index_in_zip  = -1;  // >= means stored in a zip file
-    uint64_t offset   = 0;   // offset in file
-
-    TensorStorage() = default;
-
-    TensorStorage(std::string name, ggml_type type, const int64_t* ne, int n_dims, size_t file_index, size_t offset = 0)
-        : name(std::move(name)), type(type), n_dims(n_dims), file_index(file_index), offset(offset) {
-        for (int i = 0; i < n_dims; i++) {
-            this->ne[i] = ne[i];
-        }
-    }
-
-    int64_t nelements() const {
-        int64_t n = 1;
-        for (int i = 0; i < SD_MAX_DIMS; i++) {
-            n *= ne[i];
-        }
-        return n;
-    }
-
-    int64_t nbytes() const {
-        return nelements() * ggml_type_size(type) / ggml_blck_size(type);
-    }
-
-    int64_t nbytes_to_read() const {
-        if (is_f8_e4m3 || is_f8_e5m2) {
-            return nbytes() / 2;
-        } else if (is_f64 || is_i64) {
-            return nbytes() * 2;
-        } else {
-            return nbytes();
-        }
-    }
-
-    void unsqueeze() {
-        if (n_dims == 2) {
-            n_dims = 4;
-            ne[3]  = ne[1];
-            ne[2]  = ne[0];
-            ne[1]  = 1;
-            ne[0]  = 1;
-        }
-    }
-
-    std::vector<TensorStorage> chunk(size_t n) {
-        std::vector<TensorStorage> chunks;
-        uint64_t chunk_size = nbytes_to_read() / n;
-        // printf("%d/%d\n", chunk_size, nbytes_to_read());
-        reverse_ne();
-        for (size_t i = 0; i < n; i++) {
-            TensorStorage chunk_i = *this;
-            chunk_i.ne[0]         = ne[0] / n;
-            chunk_i.offset        = offset + i * chunk_size;
-            chunk_i.reverse_ne();
-            chunks.push_back(chunk_i);
-        }
-        reverse_ne();
-        return chunks;
-    }
-
-    void reverse_ne() {
-        int64_t new_ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
-        for (int i = 0; i < n_dims; i++) {
-            new_ne[i] = ne[n_dims - 1 - i];
-        }
-        for (int i = 0; i < n_dims; i++) {
-            ne[i] = new_ne[i];
-        }
-    }
-
-    std::string to_string() const {
-        std::stringstream ss;
-        const char* type_name = ggml_type_name(type);
-        if (is_f8_e4m3) {
-            type_name = "f8_e4m3";
-        } else if (is_f8_e5m2) {
-            type_name = "f8_e5m2";
-        } else if (is_f64) {
-            type_name = "f64";
-        } else if (is_i64) {
-            type_name = "i64";
-        }
-        ss << name << " | " << type_name << " | ";
-        ss << n_dims << " [";
-        for (int i = 0; i < SD_MAX_DIMS; i++) {
-            ss << ne[i];
-            if (i != SD_MAX_DIMS - 1) {
-                ss << ", ";
-            }
-        }
-        ss << "]";
-        return ss.str();
-    }
-};
-
-typedef std::function<bool(const TensorStorage&, ggml_tensor**)> on_new_tensor_cb_t;
-
 typedef OrderedMap<std::string, TensorStorage> String2TensorStorage;
+using TensorTypeRules = std::vector<std::pair<std::string, ggml_type>>;
+
+TensorTypeRules parse_tensor_type_rules(const std::string& tensor_type_rules);
 
 class ModelLoader {
 protected:
@@ -297,16 +201,10 @@ protected:
 
     void add_tensor_storage(const TensorStorage& tensor_storage);
 
-    bool parse_data_pkl(uint8_t* buffer,
-                        size_t buffer_size,
-                        zip_t* zip,
-                        std::string dir,
-                        size_t file_index,
-                        const std::string prefix);
-
     bool init_from_gguf_file(const std::string& file_path, const std::string& prefix = "");
     bool init_from_safetensors_file(const std::string& file_path, const std::string& prefix = "");
-    bool init_from_ckpt_file(const std::string& file_path, const std::string& prefix = "");
+    bool init_from_torch_zip_file(const std::string& file_path, const std::string& prefix = "");
+    bool init_from_torch_legacy_file(const std::string& file_path, const std::string& prefix = "");
     bool init_from_diffusers_file(const std::string& file_path, const std::string& prefix = "");
 
 public:
@@ -336,7 +234,6 @@ public:
         return names;
     }
 
-    bool save_to_gguf_file(const std::string& file_path, ggml_type type, const std::string& tensor_type_rules);
     bool tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type);
     int64_t get_params_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);
     ~ModelLoader() = default;

src/model_io/binary_io.h

@@ -0,0 +1,57 @@
+#ifndef __SD_MODEL_IO_BINARY_IO_H__
+#define __SD_MODEL_IO_BINARY_IO_H__
+
+#include <cstdint>
+#include <ostream>
+
+namespace model_io {
+
+    inline int32_t read_int(const uint8_t* buffer) {
+        uint32_t value = 0;
+        value |= static_cast<uint32_t>(buffer[3]) << 24;
+        value |= static_cast<uint32_t>(buffer[2]) << 16;
+        value |= static_cast<uint32_t>(buffer[1]) << 8;
+        value |= static_cast<uint32_t>(buffer[0]);
+        return static_cast<int32_t>(value);
+    }
+
+    inline uint16_t read_short(const uint8_t* buffer) {
+        uint16_t value = 0;
+        value |= static_cast<uint16_t>(buffer[1]) << 8;
+        value |= static_cast<uint16_t>(buffer[0]);
+        return value;
+    }
+
+    inline uint64_t read_u64(const uint8_t* buffer) {
+        uint64_t value = 0;
+        value |= static_cast<uint64_t>(buffer[7]) << 56;
+        value |= static_cast<uint64_t>(buffer[6]) << 48;
+        value |= static_cast<uint64_t>(buffer[5]) << 40;
+        value |= static_cast<uint64_t>(buffer[4]) << 32;
+        value |= static_cast<uint64_t>(buffer[3]) << 24;
+        value |= static_cast<uint64_t>(buffer[2]) << 16;
+        value |= static_cast<uint64_t>(buffer[1]) << 8;
+        value |= static_cast<uint64_t>(buffer[0]);
+        return value;
+    }
+
+    inline void write_u64(std::ostream& stream, uint64_t value) {
+        uint8_t buffer[8];
+        for (int i = 0; i < 8; ++i) {
+            buffer[i] = static_cast<uint8_t>((value >> (8 * i)) & 0xFF);
+        }
+        stream.write((const char*)buffer, sizeof(buffer));
+    }
+
+    inline int find_char(const uint8_t* buffer, int len, char c) {
+        for (int pos = 0; pos < len; pos++) {
+            if (buffer[pos] == (uint8_t)c) {
+                return pos;
+            }
+        }
+        return -1;
+    }
+
+}  // namespace model_io
+
+#endif  // __SD_MODEL_IO_BINARY_IO_H__

src/model_io/gguf_io.cpp

@@ -0,0 +1,123 @@
+#include "gguf_io.h"
+
+#include <cstdint>
+#include <fstream>
+#include <string>
+#include <vector>
+
+#include "gguf.h"
+#include "gguf_reader_ext.h"
+#include "util.h"
+
+static void set_error(std::string* error, const std::string& message) {
+    if (error != nullptr) {
+        *error = message;
+    }
+}
+
+bool is_gguf_file(const std::string& file_path) {
+    std::ifstream file(file_path, std::ios::binary);
+    if (!file.is_open()) {
+        return false;
+    }
+
+    char magic[4];
+
+    file.read(magic, sizeof(magic));
+    if (!file) {
+        return false;
+    }
+    for (uint32_t i = 0; i < sizeof(magic); i++) {
+        if (magic[i] != GGUF_MAGIC[i]) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+bool read_gguf_file(const std::string& file_path,
+                    std::vector<TensorStorage>& tensor_storages,
+                    std::string* error) {
+    tensor_storages.clear();
+
+    gguf_context* ctx_gguf_ = nullptr;
+    ggml_context* ctx_meta_ = nullptr;
+
+    ctx_gguf_ = gguf_init_from_file(file_path.c_str(), {true, &ctx_meta_});
+    if (!ctx_gguf_) {
+        GGUFReader gguf_reader;
+        if (!gguf_reader.load(file_path)) {
+            set_error(error, "failed to open '" + file_path + "' with GGUFReader");
+            return false;
+        }
+
+        size_t data_offset = gguf_reader.data_offset();
+        for (const auto& gguf_tensor_info : gguf_reader.tensors()) {
+            TensorStorage tensor_storage(
+                gguf_tensor_info.name,
+                gguf_tensor_info.type,
+                gguf_tensor_info.shape.data(),
+                static_cast<int>(gguf_tensor_info.shape.size()),
+                0,
+                data_offset + gguf_tensor_info.offset);
+
+            tensor_storages.push_back(tensor_storage);
+        }
+
+        return true;
+    }
+
+    int n_tensors = static_cast<int>(gguf_get_n_tensors(ctx_gguf_));
+
+    size_t data_offset = gguf_get_data_offset(ctx_gguf_);
+    for (int i = 0; i < n_tensors; i++) {
+        std::string name   = gguf_get_tensor_name(ctx_gguf_, i);
+        ggml_tensor* dummy = ggml_get_tensor(ctx_meta_, name.c_str());
+        size_t offset      = data_offset + gguf_get_tensor_offset(ctx_gguf_, i);
+
+        TensorStorage tensor_storage(name, dummy->type, dummy->ne, ggml_n_dims(dummy), 0, offset);
+
+        if (ggml_nbytes(dummy) != tensor_storage.nbytes()) {
+            gguf_free(ctx_gguf_);
+            ggml_free(ctx_meta_);
+            set_error(error, "size mismatch for tensor '" + name + "'");
+            return false;
+        }
+
+        tensor_storages.push_back(tensor_storage);
+    }
+
+    gguf_free(ctx_gguf_);
+    ggml_free(ctx_meta_);
+
+    return true;
+}
+
+bool write_gguf_file(const std::string& file_path,
+                     const std::vector<TensorWriteInfo>& tensors,
+                     std::string* error) {
+    gguf_context* gguf_ctx = gguf_init_empty();
+    if (gguf_ctx == nullptr) {
+        set_error(error, "gguf_init_empty failed");
+        return false;
+    }
+
+    for (const TensorWriteInfo& write_tensor : tensors) {
+        ggml_tensor* tensor = write_tensor.tensor;
+        if (tensor == nullptr) {
+            set_error(error, "null tensor cannot be written to GGUF");
+            gguf_free(gguf_ctx);
+            return false;
+        }
+        gguf_add_tensor(gguf_ctx, tensor);
+    }
+
+    LOG_INFO("trying to save tensors to %s", file_path.c_str());
+    bool success = gguf_write_to_file(gguf_ctx, file_path.c_str(), false);
+    if (!success) {
+        set_error(error, "failed to write GGUF file '" + file_path + "'");
+    }
+    gguf_free(gguf_ctx);
+    return success;
+}

src/model_io/gguf_io.h

@@ -0,0 +1,17 @@
+#ifndef __SD_MODEL_IO_GGUF_IO_H__
+#define __SD_MODEL_IO_GGUF_IO_H__
+
+#include <string>
+#include <vector>
+
+#include "tensor_storage.h"
+
+bool is_gguf_file(const std::string& file_path);
+bool read_gguf_file(const std::string& file_path,
+                    std::vector<TensorStorage>& tensor_storages,
+                    std::string* error = nullptr);
+bool write_gguf_file(const std::string& file_path,
+                     const std::vector<TensorWriteInfo>& tensors,
+                     std::string* error = nullptr);
+
+#endif  // __SD_MODEL_IO_GGUF_IO_H__

src/model_io/gguf_reader_ext.h

@@ -1,5 +1,5 @@
-#ifndef __GGUF_READER_HPP__
-#define __GGUF_READER_HPP__
+#ifndef __SD_MODEL_IO_GGUF_READER_EXT_H__
+#define __SD_MODEL_IO_GGUF_READER_EXT_H__
 
 #include <cstdint>
 #include <fstream>
@@ -59,6 +59,9 @@ private:
         if (!safe_read(fin, key_len))
             return false;
 
+        if (key_len > 4096)
+            return false;
+
         std::string key(key_len, '\0');
         if (!safe_read(fin, (char*)key.data(), key_len))
             return false;
@@ -228,4 +231,4 @@ public:
     size_t data_offset() const { return data_offset_; }
 };
 
-#endif  // __GGUF_READER_HPP__
+#endif  // __SD_MODEL_IO_GGUF_READER_EXT_H__

src/model_io/pickle_io.h

@@ -0,0 +1,21 @@
+#ifndef __SD_MODEL_IO_PICKLE_IO_H__
+#define __SD_MODEL_IO_PICKLE_IO_H__
+
+#include <cstddef>
+#include <cstdint>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "tensor_storage.h"
+
+bool skip_pickle_object(const uint8_t* buffer, size_t buffer_size, size_t* object_size);
+bool pickle_object_is_torch_magic_number(const uint8_t* buffer, size_t buffer_size);
+bool parse_pickle_uint32_object(const uint8_t* buffer, size_t buffer_size, uint32_t* value);
+bool parse_torch_state_dict_pickle(const uint8_t* buffer,
+                                   size_t buffer_size,
+                                   std::vector<TensorStorage>& tensor_storages,
+                                   std::unordered_map<std::string, uint64_t>& storage_nbytes,
+                                   std::string* error = nullptr);
+
+#endif  // __SD_MODEL_IO_PICKLE_IO_H__

src/model_io/safetensors_io.cpp

@@ -0,0 +1,316 @@
+#include "safetensors_io.h"
+
+#include <cstdint>
+#include <exception>
+#include <fstream>
+#include <string>
+#include <vector>
+
+#include "binary_io.h"
+#include "json.hpp"
+#include "util.h"
+
+static constexpr size_t ST_HEADER_SIZE_LEN = 8;
+
+static void set_error(std::string* error, const std::string& message) {
+    if (error != nullptr) {
+        *error = message;
+    }
+}
+
+bool is_safetensors_file(const std::string& file_path) {
+    std::ifstream file(file_path, std::ios::binary);
+    if (!file.is_open()) {
+        return false;
+    }
+
+    // get file size
+    file.seekg(0, file.end);
+    size_t file_size_ = file.tellg();
+    file.seekg(0, file.beg);
+
+    // read header size
+    if (file_size_ <= ST_HEADER_SIZE_LEN) {
+        return false;
+    }
+
+    uint8_t header_size_buf[ST_HEADER_SIZE_LEN];
+    file.read((char*)header_size_buf, ST_HEADER_SIZE_LEN);
+    if (!file) {
+        return false;
+    }
+
+    size_t header_size_ = model_io::read_u64(header_size_buf);
+    if (header_size_ >= file_size_ || header_size_ <= 2) {
+        return false;
+    }
+
+    // read header
+    std::vector<char> header_buf;
+    header_buf.resize(header_size_ + 1);
+    header_buf[header_size_] = '\0';
+    file.read(header_buf.data(), header_size_);
+    if (!file) {
+        return false;
+    }
+    try {
+        nlohmann::json header_ = nlohmann::json::parse(header_buf.data());
+    } catch (const std::exception&) {
+        return false;
+    }
+    return true;
+}
+
+static ggml_type safetensors_dtype_to_ggml_type(const std::string& dtype) {
+    ggml_type ttype = GGML_TYPE_COUNT;
+    if (dtype == "F16") {
+        ttype = GGML_TYPE_F16;
+    } else if (dtype == "BF16") {
+        ttype = GGML_TYPE_BF16;
+    } else if (dtype == "F32") {
+        ttype = GGML_TYPE_F32;
+    } else if (dtype == "F64") {
+        ttype = GGML_TYPE_F32;
+    } else if (dtype == "F8_E4M3") {
+        ttype = GGML_TYPE_F16;
+    } else if (dtype == "F8_E5M2") {
+        ttype = GGML_TYPE_F16;
+    } else if (dtype == "I32") {
+        ttype = GGML_TYPE_I32;
+    } else if (dtype == "I64") {
+        ttype = GGML_TYPE_I32;
+    }
+    return ttype;
+}
+
+// https://huggingface.co/docs/safetensors/index
+bool read_safetensors_file(const std::string& file_path,
+                           std::vector<TensorStorage>& tensor_storages,
+                           std::string* error) {
+    std::ifstream file(file_path, std::ios::binary);
+    if (!file.is_open()) {
+        set_error(error, "failed to open '" + file_path + "'");
+        return false;
+    }
+
+    // get file size
+    file.seekg(0, file.end);
+    size_t file_size_ = file.tellg();
+    file.seekg(0, file.beg);
+
+    // read header size
+    if (file_size_ <= ST_HEADER_SIZE_LEN) {
+        set_error(error, "invalid safetensor file '" + file_path + "'");
+        return false;
+    }
+
+    uint8_t header_size_buf[ST_HEADER_SIZE_LEN];
+    file.read((char*)header_size_buf, ST_HEADER_SIZE_LEN);
+    if (!file) {
+        set_error(error, "read safetensors header size failed: '" + file_path + "'");
+        return false;
+    }
+
+    size_t header_size_ = model_io::read_u64(header_size_buf);
+    if (header_size_ >= file_size_) {
+        set_error(error, "invalid safetensor file '" + file_path + "'");
+        return false;
+    }
+
+    // read header
+    std::vector<char> header_buf;
+    header_buf.resize(header_size_ + 1);
+    header_buf[header_size_] = '\0';
+    file.read(header_buf.data(), header_size_);
+    if (!file) {
+        set_error(error, "read safetensors header failed: '" + file_path + "'");
+        return false;
+    }
+
+    nlohmann::json header_;
+    try {
+        header_ = nlohmann::json::parse(header_buf.data());
+    } catch (const std::exception&) {
+        set_error(error, "parsing safetensors header failed: '" + file_path + "'");
+        return false;
+    }
+
+    tensor_storages.clear();
+    for (auto& item : header_.items()) {
+        std::string name           = item.key();
+        nlohmann::json tensor_info = item.value();
+        // LOG_DEBUG("%s %s\n", name.c_str(), tensor_info.dump().c_str());
+
+        if (name == "__metadata__") {
+            continue;
+        }
+
+        std::string dtype    = tensor_info["dtype"];
+        nlohmann::json shape = tensor_info["shape"];
+
+        if (dtype == "U8") {
+            continue;
+        }
+
+        size_t begin = tensor_info["data_offsets"][0].get<size_t>();
+        size_t end   = tensor_info["data_offsets"][1].get<size_t>();
+
+        ggml_type type = safetensors_dtype_to_ggml_type(dtype);
+        if (type == GGML_TYPE_COUNT) {
+            set_error(error, "unsupported dtype '" + dtype + "' (tensor '" + name + "')");
+            return false;
+        }
+
+        if (shape.size() > SD_MAX_DIMS) {
+            set_error(error, "invalid tensor '" + name + "'");
+            return false;
+        }
+
+        int n_dims              = (int)shape.size();
+        int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
+        for (int i = 0; i < n_dims; i++) {
+            ne[i] = shape[i].get<int64_t>();
+        }
+
+        if (n_dims == 5) {
+            n_dims = 4;
+            ne[0]  = ne[0] * ne[1];
+            ne[1]  = ne[2];
+            ne[2]  = ne[3];
+            ne[3]  = ne[4];
+        }
+
+        // ggml_n_dims returns 1 for scalars
+        if (n_dims == 0) {
+            n_dims = 1;
+        }
+
+        TensorStorage tensor_storage(name, type, ne, n_dims, 0, ST_HEADER_SIZE_LEN + header_size_ + begin);
+        tensor_storage.reverse_ne();
+
+        size_t tensor_data_size = end - begin;
+
+        bool tensor_size_ok;
+        if (dtype == "F8_E4M3") {
+            tensor_storage.is_f8_e4m3 = true;
+            // f8 -> f16
+            tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size * 2);
+        } else if (dtype == "F8_E5M2") {
+            tensor_storage.is_f8_e5m2 = true;
+            // f8 -> f16
+            tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size * 2);
+        } else if (dtype == "F64") {
+            tensor_storage.is_f64 = true;
+            // f64 -> f32
+            tensor_size_ok = (tensor_storage.nbytes() * 2 == tensor_data_size);
+        } else if (dtype == "I64") {
+            tensor_storage.is_i64 = true;
+            // i64 -> i32
+            tensor_size_ok = (tensor_storage.nbytes() * 2 == tensor_data_size);
+        } else {
+            tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size);
+        }
+        if (!tensor_size_ok) {
+            set_error(error, "size mismatch for tensor '" + name + "' (" + dtype + ")");
+            return false;
+        }
+
+        tensor_storages.push_back(tensor_storage);
+
+        // LOG_DEBUG("%s %s", tensor_storage.to_string().c_str(), dtype.c_str());
+    }
+
+    return true;
+}
+
+static bool ggml_type_to_safetensors_dtype(ggml_type type, std::string* dtype) {
+    switch (type) {
+        case GGML_TYPE_F16:
+            *dtype = "F16";
+            return true;
+        case GGML_TYPE_BF16:
+            *dtype = "BF16";
+            return true;
+        case GGML_TYPE_F32:
+            *dtype = "F32";
+            return true;
+        case GGML_TYPE_I32:
+            *dtype = "I32";
+            return true;
+        default:
+            return false;
+    }
+}
+
+bool write_safetensors_file(const std::string& file_path,
+                            const std::vector<TensorWriteInfo>& tensors,
+                            std::string* error) {
+    nlohmann::ordered_json header = nlohmann::ordered_json::object();
+
+    uint64_t data_offset = 0;
+    for (const TensorWriteInfo& write_tensor : tensors) {
+        ggml_tensor* tensor = write_tensor.tensor;
+        if (tensor == nullptr) {
+            set_error(error, "null tensor cannot be written to safetensors");
+            return false;
+        }
+
+        const std::string name = ggml_get_name(tensor);
+        std::string dtype;
+        if (!ggml_type_to_safetensors_dtype(tensor->type, &dtype)) {
+            set_error(error,
+                      "unsupported safetensors dtype '" + std::string(ggml_type_name(tensor->type)) +
+                          "' for tensor '" + name + "'");
+            return false;
+        }
+
+        const uint64_t tensor_nbytes = ggml_nbytes(tensor);
+
+        nlohmann::ordered_json json_tensor_info = nlohmann::ordered_json::object();
+        json_tensor_info["dtype"]               = dtype;
+
+        nlohmann::ordered_json shape = nlohmann::ordered_json::array();
+        for (int i = 0; i < write_tensor.n_dims; ++i) {
+            shape.push_back(write_tensor.ne[write_tensor.n_dims - 1 - i]);
+        }
+        json_tensor_info["shape"] = shape;
+
+        nlohmann::ordered_json data_offsets = nlohmann::ordered_json::array();
+        data_offsets.push_back(data_offset);
+        data_offsets.push_back(data_offset + tensor_nbytes);
+        json_tensor_info["data_offsets"] = data_offsets;
+
+        header[name] = json_tensor_info;
+        data_offset += tensor_nbytes;
+    }
+
+    const std::string header_str = header.dump();
+
+    std::ofstream file(file_path, std::ios::binary);
+    if (!file.is_open()) {
+        set_error(error, "failed to open '" + file_path + "' for writing");
+        return false;
+    }
+
+    LOG_INFO("trying to save tensors to %s", file_path.c_str());
+    model_io::write_u64(file, header_str.size());
+    file.write(header_str.data(), header_str.size());
+    if (!file) {
+        set_error(error, "failed to write safetensors header to '" + file_path + "'");
+        return false;
+    }
+
+    for (const TensorWriteInfo& write_tensor : tensors) {
+        ggml_tensor* tensor        = write_tensor.tensor;
+        const std::string name     = ggml_get_name(tensor);
+        const size_t tensor_nbytes = ggml_nbytes(tensor);
+        file.write((const char*)tensor->data, tensor_nbytes);
+        if (!file) {
+            set_error(error,
+                      "failed to write tensor '" + name + "' to '" + file_path + "'");
+            return false;
+        }
+    }
+
+    return true;
+}

src/model_io/safetensors_io.h

@@ -0,0 +1,17 @@
+#ifndef __SD_MODEL_IO_SAFETENSORS_IO_H__
+#define __SD_MODEL_IO_SAFETENSORS_IO_H__
+
+#include <string>
+#include <vector>
+
+#include "tensor_storage.h"
+
+bool is_safetensors_file(const std::string& file_path);
+bool read_safetensors_file(const std::string& file_path,
+                           std::vector<TensorStorage>& tensor_storages,
+                           std::string* error = nullptr);
+bool write_safetensors_file(const std::string& file_path,
+                            const std::vector<TensorWriteInfo>& tensors,
+                            std::string* error = nullptr);
+
+#endif  // __SD_MODEL_IO_SAFETENSORS_IO_H__

src/model_io/tensor_storage.h

@@ -0,0 +1,132 @@
+#ifndef __SD_TENSOR_STORAGE_H__
+#define __SD_TENSOR_STORAGE_H__
+
+#include <cstddef>
+#include <cstdint>
+#include <functional>
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "ggml.h"
+
+#define SD_MAX_DIMS 5
+
+struct TensorStorage {
+    std::string name;
+    ggml_type type          = GGML_TYPE_F32;
+    ggml_type expected_type = GGML_TYPE_COUNT;
+    bool is_f8_e4m3         = false;
+    bool is_f8_e5m2         = false;
+    bool is_f64             = false;
+    bool is_i64             = false;
+    int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
+    int n_dims              = 0;
+
+    std::string storage_key;
+    size_t file_index = 0;
+    int index_in_zip  = -1;  // >= means stored in a zip file
+    uint64_t offset   = 0;   // offset in file
+
+    TensorStorage() = default;
+
+    TensorStorage(std::string name, ggml_type type, const int64_t* ne, int n_dims, size_t file_index, size_t offset = 0)
+        : name(std::move(name)), type(type), n_dims(n_dims), file_index(file_index), offset(offset) {
+        for (int i = 0; i < n_dims; i++) {
+            this->ne[i] = ne[i];
+        }
+    }
+
+    int64_t nelements() const {
+        int64_t n = 1;
+        for (int i = 0; i < SD_MAX_DIMS; i++) {
+            n *= ne[i];
+        }
+        return n;
+    }
+
+    int64_t nbytes() const {
+        return nelements() * ggml_type_size(type) / ggml_blck_size(type);
+    }
+
+    int64_t nbytes_to_read() const {
+        if (is_f8_e4m3 || is_f8_e5m2) {
+            return nbytes() / 2;
+        } else if (is_f64 || is_i64) {
+            return nbytes() * 2;
+        } else {
+            return nbytes();
+        }
+    }
+
+    void unsqueeze() {
+        if (n_dims == 2) {
+            n_dims = 4;
+            ne[3]  = ne[1];
+            ne[2]  = ne[0];
+            ne[1]  = 1;
+            ne[0]  = 1;
+        }
+    }
+
+    std::vector<TensorStorage> chunk(size_t n) {
+        std::vector<TensorStorage> chunks;
+        uint64_t chunk_size = nbytes_to_read() / n;
+        // printf("%d/%d\n", chunk_size, nbytes_to_read());
+        reverse_ne();
+        for (size_t i = 0; i < n; i++) {
+            TensorStorage chunk_i = *this;
+            chunk_i.ne[0]         = ne[0] / n;
+            chunk_i.offset        = offset + i * chunk_size;
+            chunk_i.reverse_ne();
+            chunks.push_back(chunk_i);
+        }
+        reverse_ne();
+        return chunks;
+    }
+
+    void reverse_ne() {
+        int64_t new_ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
+        for (int i = 0; i < n_dims; i++) {
+            new_ne[i] = ne[n_dims - 1 - i];
+        }
+        for (int i = 0; i < n_dims; i++) {
+            ne[i] = new_ne[i];
+        }
+    }
+
+    std::string to_string() const {
+        std::stringstream ss;
+        const char* type_name = ggml_type_name(type);
+        if (is_f8_e4m3) {
+            type_name = "f8_e4m3";
+        } else if (is_f8_e5m2) {
+            type_name = "f8_e5m2";
+        } else if (is_f64) {
+            type_name = "f64";
+        } else if (is_i64) {
+            type_name = "i64";
+        }
+        ss << name << " | " << type_name << " | ";
+        ss << n_dims << " [";
+        for (int i = 0; i < SD_MAX_DIMS; i++) {
+            ss << ne[i];
+            if (i != SD_MAX_DIMS - 1) {
+                ss << ", ";
+            }
+        }
+        ss << "]";
+        return ss.str();
+    }
+};
+
+struct TensorWriteInfo {
+    int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
+    int n_dims              = 0;
+    ggml_tensor* tensor     = nullptr;
+};
+
+typedef std::function<bool(const TensorStorage&, ggml_tensor**)> on_new_tensor_cb_t;
+
+#endif  // __SD_TENSOR_STORAGE_H__

src/model_io/torch_legacy_io.cpp

@@ -0,0 +1,252 @@
+#include "torch_legacy_io.h"
+
+#include <algorithm>
+#include <cstdint>
+#include <fstream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "pickle_io.h"
+#include "util.h"
+
+// torch.save format background:
+//
+//   - Before PyTorch 1.6.0, torch.save used this legacy non-zip format by
+//     default.
+//   - Since PyTorch 1.6.0, torch.save defaults to an uncompressed ZIP64 archive
+//     containing data.pkl, data/, version, and, since PyTorch 2.1.0, byteorder.
+//   - The old format can still be produced explicitly with:
+//       torch.save(obj, path, _use_new_zipfile_serialization=False)
+//
+// Whether obj is a state_dict or a whole nn.Module does not change the outer
+// container format selected by torch.save. It changes the pickled object inside:
+//
+//   - state_dict: usually an OrderedDict[str, Tensor]. pickle_io.cpp supports a
+//     restricted subset of this layout because tensor metadata and raw storages
+//     can be recovered without executing pickle callables.
+//   - whole module/checkpoint object: arbitrary Python object graph. This may
+//     require importing user classes and executing pickle GLOBAL/REDUCE rebuild
+//     logic, so it is intentionally not supported here.
+//
+// Legacy non-zip PyTorch files are not a single pickle object:
+//
+//   1. pickle object: PyTorch legacy magic number
+//   2. pickle object: legacy protocol version, expected to be 1001
+//   3. pickle object: sys_info metadata, ignored by this reader
+//   4. pickle object: state_dict metadata, parsed by pickle_io.cpp
+//   5. pickle object: serialized storage key list, skipped here
+//   6. raw storage data payloads
+//      - PyTorch writes storages after the pickles, ordered by storage key
+//      - each storage has an 8-byte legacy storage header followed by raw bytes
+static constexpr size_t LEGACY_STORAGE_HEADER_SIZE = 8;
+
+static void set_error(std::string* error, const std::string& message) {
+    if (error != nullptr) {
+        *error = message;
+    }
+}
+
+static std::string bytes_to_hex(const std::vector<uint8_t>& bytes) {
+    static const char* hex = "0123456789ABCDEF";
+    std::string result;
+    result.reserve(bytes.size() * 3);
+    for (size_t i = 0; i < bytes.size(); ++i) {
+        if (i > 0) {
+            result.push_back('-');
+        }
+        result.push_back(hex[(bytes[i] >> 4) & 0x0F]);
+        result.push_back(hex[bytes[i] & 0x0F]);
+    }
+    return result;
+}
+
+static bool is_probably_tar_file(const std::vector<uint8_t>& header) {
+    return header.size() >= 262 &&
+           header[257] == 'u' &&
+           header[258] == 's' &&
+           header[259] == 't' &&
+           header[260] == 'a' &&
+           header[261] == 'r';
+}
+
+static std::string torch_legacy_diagnostics(const std::string& file_path, const std::vector<uint8_t>& buffer) {
+    if (!ends_with(file_path, ".pt") && !ends_with(file_path, ".pth")) {
+        return "";
+    }
+    if (buffer.empty()) {
+        return "unsupported PyTorch file '" + file_path + "': empty file";
+    }
+
+    size_t short_len = std::min<size_t>(buffer.size(), 32);
+    std::vector<uint8_t> short_header(buffer.begin(), buffer.begin() + short_len);
+    const bool raw_pickle = buffer[0] == 0x80;
+    const bool tar_file   = is_probably_tar_file(buffer);
+
+    std::string message = "unsupported PyTorch file '" + file_path + "': first bytes " +
+                          bytes_to_hex(short_header) +
+                          ", raw_pickle=" + (raw_pickle ? "true" : "false") +
+                          ", tar=" + (tar_file ? "true" : "false");
+    if (raw_pickle) {
+        message += "; raw pickle did not match the restricted state_dict layouts currently supported";
+    } else if (tar_file) {
+        message += "; legacy tar PyTorch checkpoints are not supported yet";
+    }
+    return message;
+}
+
+bool read_torch_legacy_file(const std::string& file_path,
+                            std::vector<TensorStorage>& tensor_storages,
+                            std::string* error) {
+    std::ifstream file(file_path, std::ios::binary);
+    if (!file.is_open()) {
+        set_error(error, "failed to open '" + file_path + "'");
+        return false;
+    }
+
+    file.seekg(0, file.end);
+    size_t file_size = (size_t)file.tellg();
+    file.seekg(0, file.beg);
+    if (file_size == 0) {
+        set_error(error, "empty file '" + file_path + "'");
+        return false;
+    }
+
+    std::vector<uint8_t> buffer(file_size);
+    file.read((char*)buffer.data(), file_size);
+    if (!file) {
+        set_error(error, "failed to read '" + file_path + "'");
+        return false;
+    }
+
+    auto finalize_tensor_offsets = [&](size_t storage_data_offset,
+                                       const std::unordered_map<std::string, uint64_t>& legacy_storage_map) -> bool {
+        if (storage_data_offset > file_size) {
+            return false;
+        }
+
+        std::vector<std::string> storage_keys;
+        storage_keys.reserve(legacy_storage_map.size());
+        for (const auto& [storage_key, _] : legacy_storage_map) {
+            storage_keys.push_back(storage_key);
+        }
+        std::sort(storage_keys.begin(), storage_keys.end());
+
+        std::unordered_map<std::string, uint64_t> storage_offsets;
+        uint64_t current_offset = storage_data_offset;
+        for (const auto& storage_key : storage_keys) {
+            auto it = legacy_storage_map.find(storage_key);
+            if (it == legacy_storage_map.end()) {
+                return false;
+            }
+            if (current_offset + LEGACY_STORAGE_HEADER_SIZE + it->second > file_size) {
+                return false;
+            }
+            storage_offsets[storage_key] = current_offset + LEGACY_STORAGE_HEADER_SIZE;
+            current_offset += LEGACY_STORAGE_HEADER_SIZE + it->second;
+        }
+
+        for (auto& tensor_storage : tensor_storages) {
+            if (tensor_storage.storage_key.empty()) {
+                continue;
+            }
+
+            auto it_offset = storage_offsets.find(tensor_storage.storage_key);
+            auto it_size   = legacy_storage_map.find(tensor_storage.storage_key);
+            if (it_offset == storage_offsets.end() || it_size == legacy_storage_map.end()) {
+                return false;
+            }
+
+            uint64_t base_offset    = it_offset->second;
+            uint64_t storage_nbytes = it_size->second;
+            uint64_t tensor_nbytes  = tensor_storage.nbytes_to_read();
+            if (tensor_storage.offset + tensor_nbytes > storage_nbytes) {
+                return false;
+            }
+
+            tensor_storage.offset = base_offset + tensor_storage.offset;
+            tensor_storage.storage_key.clear();
+        }
+
+        return true;
+    };
+
+    auto parse_state_dict_at = [&](size_t state_dict_offset, size_t state_dict_size, size_t* storage_data_offset) -> bool {
+        tensor_storages.clear();
+        std::unordered_map<std::string, uint64_t> legacy_storage_map;
+        if (!parse_torch_state_dict_pickle(buffer.data() + state_dict_offset,
+                                           state_dict_size,
+                                           tensor_storages,
+                                           legacy_storage_map,
+                                           error)) {
+            return false;
+        }
+
+        size_t offset_after_state_dict = state_dict_offset + state_dict_size;
+        size_t storage_keys_size       = 0;
+        if (!skip_pickle_object(buffer.data() + offset_after_state_dict,
+                                buffer.size() - offset_after_state_dict,
+                                &storage_keys_size)) {
+            return false;
+        }
+
+        *storage_data_offset = offset_after_state_dict + storage_keys_size;
+        return finalize_tensor_offsets(*storage_data_offset, legacy_storage_map);
+    };
+
+    size_t object_size_1 = 0;
+    size_t offset        = 0;
+
+    if (skip_pickle_object(buffer.data(), buffer.size(), &object_size_1) &&
+        pickle_object_is_torch_magic_number(buffer.data(), object_size_1)) {
+        offset += object_size_1;
+
+        size_t object_size_2 = 0;
+        if (!skip_pickle_object(buffer.data() + offset, buffer.size() - offset, &object_size_2)) {
+            set_error(error, torch_legacy_diagnostics(file_path, buffer));
+            return false;
+        }
+        uint32_t protocol_version = 0;
+        if (!parse_pickle_uint32_object(buffer.data() + offset, object_size_2, &protocol_version) || protocol_version != 1001) {
+            set_error(error, torch_legacy_diagnostics(file_path, buffer));
+            return false;
+        }
+        offset += object_size_2;
+
+        size_t object_size_3 = 0;
+        if (!skip_pickle_object(buffer.data() + offset, buffer.size() - offset, &object_size_3)) {
+            set_error(error, torch_legacy_diagnostics(file_path, buffer));
+            return false;
+        }
+        offset += object_size_3;
+
+        size_t state_dict_size = 0;
+        if (!skip_pickle_object(buffer.data() + offset, buffer.size() - offset, &state_dict_size)) {
+            set_error(error, torch_legacy_diagnostics(file_path, buffer));
+            return false;
+        }
+
+        size_t storage_data_offset = 0;
+        if (parse_state_dict_at(offset, state_dict_size, &storage_data_offset)) {
+            return true;
+        }
+
+        if (error != nullptr && error->empty()) {
+            set_error(error, torch_legacy_diagnostics(file_path, buffer));
+        }
+        return false;
+    }
+
+    size_t state_dict_size = 0;
+    if (skip_pickle_object(buffer.data(), buffer.size(), &state_dict_size)) {
+        size_t storage_data_offset = 0;
+        if (parse_state_dict_at(0, state_dict_size, &storage_data_offset)) {
+            return true;
+        }
+    }
+
+    if (error != nullptr && error->empty()) {
+        set_error(error, torch_legacy_diagnostics(file_path, buffer));
+    }
+    return false;
+}

src/model_io/torch_legacy_io.h

@@ -0,0 +1,13 @@
+#ifndef __SD_MODEL_IO_TORCH_LEGACY_IO_H__
+#define __SD_MODEL_IO_TORCH_LEGACY_IO_H__
+
+#include <string>
+#include <vector>
+
+#include "tensor_storage.h"
+
+bool read_torch_legacy_file(const std::string& file_path,
+                            std::vector<TensorStorage>& tensor_storages,
+                            std::string* error = nullptr);
+
+#endif  // __SD_MODEL_IO_TORCH_LEGACY_IO_H__

src/model_io/torch_zip_io.cpp

@@ -0,0 +1,140 @@
+#include "torch_zip_io.h"
+
+#include <cstdint>
+#include <cstdlib>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "pickle_io.h"
+
+#include "zip.h"
+
+static void set_error(std::string* error, const std::string& message) {
+    if (error != nullptr) {
+        *error = message;
+    }
+}
+
+bool is_torch_zip_file(const std::string& file_path) {
+    zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
+    if (zip == nullptr) {
+        return false;
+    }
+    zip_close(zip);
+    return true;
+}
+
+static bool find_zip_entry(zip_t* zip, const std::string& entry_name, int* index, uint64_t* size) {
+    size_t n = zip_entries_total(zip);
+    for (size_t i = 0; i < n; ++i) {
+        zip_entry_openbyindex(zip, i);
+        std::string name = zip_entry_name(zip);
+        if (name == entry_name) {
+            *index = (int)i;
+            *size  = zip_entry_size(zip);
+            zip_entry_close(zip);
+            return true;
+        }
+        zip_entry_close(zip);
+    }
+    return false;
+}
+
+static bool parse_zip_data_pkl(const uint8_t* buffer,
+                               size_t buffer_size,
+                               zip_t* zip,
+                               const std::string& dir,
+                               std::vector<TensorStorage>& tensor_storages,
+                               std::string* error) {
+    std::vector<TensorStorage> parsed_tensors;
+    std::unordered_map<std::string, uint64_t> storage_nbytes;
+    if (!parse_torch_state_dict_pickle(buffer, buffer_size, parsed_tensors, storage_nbytes, error)) {
+        if (error != nullptr && error->empty()) {
+            *error = "failed to parse torch zip pickle metadata";
+        }
+        return false;
+    }
+
+    for (auto& tensor_storage : parsed_tensors) {
+        if (tensor_storage.storage_key.empty()) {
+            set_error(error, "tensor '" + tensor_storage.name + "' has no storage key");
+            return false;
+        }
+
+        const std::string entry_name = dir + "data/" + tensor_storage.storage_key;
+        int zip_index                = -1;
+        uint64_t entry_size          = 0;
+        if (!find_zip_entry(zip, entry_name, &zip_index, &entry_size)) {
+            set_error(error, "storage entry '" + entry_name + "' was not found");
+            return false;
+        }
+
+        auto it_storage_size = storage_nbytes.find(tensor_storage.storage_key);
+        if (it_storage_size != storage_nbytes.end() && entry_size < it_storage_size->second) {
+            set_error(error, "storage entry '" + entry_name + "' is smaller than pickle metadata");
+            return false;
+        }
+
+        uint64_t tensor_nbytes = tensor_storage.nbytes_to_read();
+        if (tensor_storage.offset + tensor_nbytes > entry_size) {
+            set_error(error, "tensor '" + tensor_storage.name + "' exceeds storage entry '" + entry_name + "'");
+            return false;
+        }
+
+        tensor_storage.index_in_zip = zip_index;
+        tensor_storage.storage_key.clear();
+        tensor_storages.push_back(tensor_storage);
+    }
+
+    return true;
+}
+
+bool read_torch_zip_file(const std::string& file_path,
+                         std::vector<TensorStorage>& tensor_storages,
+                         std::string* error) {
+    zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
+    if (zip == nullptr) {
+        set_error(error, "failed to open '" + file_path + "'");
+        return false;
+    }
+
+    tensor_storages.clear();
+    bool success        = true;
+    bool found_data_pkl = false;
+    int n               = (int)zip_entries_total(zip);
+    for (int i = 0; i < n; ++i) {
+        zip_entry_openbyindex(zip, i);
+        std::string name = zip_entry_name(zip);
+        size_t pos       = name.find("data.pkl");
+        if (pos != std::string::npos) {
+            found_data_pkl  = true;
+            std::string dir = name.substr(0, pos);
+            void* pkl_data  = nullptr;
+            size_t pkl_size = 0;
+            zip_entry_read(zip, &pkl_data, &pkl_size);
+
+            if (pkl_data == nullptr || pkl_size == 0) {
+                set_error(error, "failed to read '" + name + "' from '" + file_path + "'");
+                success = false;
+            } else if (!parse_zip_data_pkl((const uint8_t*)pkl_data, pkl_size, zip, dir, tensor_storages, error)) {
+                success = false;
+            }
+
+            free(pkl_data);
+        }
+        zip_entry_close(zip);
+
+        if (!success) {
+            break;
+        }
+    }
+
+    if (success && !found_data_pkl) {
+        set_error(error, "data.pkl was not found in '" + file_path + "'");
+        success = false;
+    }
+
+    zip_close(zip);
+    return success;
+}

src/model_io/torch_zip_io.h

@@ -0,0 +1,14 @@
+#ifndef __SD_MODEL_IO_TORCH_ZIP_IO_H__
+#define __SD_MODEL_IO_TORCH_ZIP_IO_H__
+
+#include <string>
+#include <vector>
+
+#include "tensor_storage.h"
+
+bool is_torch_zip_file(const std::string& file_path);
+bool read_torch_zip_file(const std::string& file_path,
+                         std::vector<TensorStorage>& tensor_storages,
+                         std::string* error = nullptr);
+
+#endif  // __SD_MODEL_IO_TORCH_ZIP_IO_H__

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/pmid.hpp

@@ -443,11 +443,10 @@ public:
             id_encoder2.get_param_tensors(tensors, prefix);
     }
 
-    ggml_cgraph* build_graph(  // ggml_allocr* allocr,
-        ggml_tensor* id_pixel_values,
-        ggml_tensor* prompt_embeds,
-        std::vector<bool>& class_tokens_mask,
-        ggml_tensor* id_embeds) {
+    ggml_cgraph* build_graph(const sd::Tensor<float>& id_pixel_values_tensor,
+                             const sd::Tensor<float>& prompt_embeds_tensor,
+                             std::vector<bool>& class_tokens_mask,
+                             const sd::Tensor<float>& id_embeds_tensor = {}) {
         ctm.clear();
         ctmf16.clear();
         ctmpos.clear();
@@ -460,16 +459,16 @@ public:
 
         ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
+        ggml_tensor* id_pixel_values = make_input(id_pixel_values_tensor);
+        ggml_tensor* prompt_embeds   = make_input(prompt_embeds_tensor);
+        ggml_tensor* id_embeds       = make_optional_input(id_embeds_tensor);
+
         int64_t hidden_size = prompt_embeds->ne[0];
         int64_t seq_length  = prompt_embeds->ne[1];
         ggml_type type      = GGML_TYPE_F32;
 
         ggml_tensor* class_tokens_mask_d = ggml_new_tensor_1d(runner_ctx.ggml_ctx, type, class_tokens_mask.size());
 
-        ggml_tensor* id_pixel_values_d = to_backend(id_pixel_values);
-        ggml_tensor* prompt_embeds_d   = to_backend(prompt_embeds);
-        ggml_tensor* id_embeds_d       = to_backend(id_embeds);
-
         ggml_tensor* left  = nullptr;
         ggml_tensor* right = nullptr;
         for (int i = 0; i < class_tokens_mask.size(); i++) {
@@ -529,18 +528,18 @@ public:
         ggml_tensor* updated_prompt_embeds = nullptr;
         if (pm_version == PM_VERSION_1)
             updated_prompt_embeds = id_encoder.forward(&runner_ctx,
-                                                       id_pixel_values_d,
-                                                       prompt_embeds_d,
+                                                       id_pixel_values,
+                                                       prompt_embeds,
                                                        class_tokens_mask_d,
                                                        class_tokens_mask_pos,
                                                        left, right);
         else if (pm_version == PM_VERSION_2)
             updated_prompt_embeds = id_encoder2.forward(&runner_ctx,
-                                                        id_pixel_values_d,
-                                                        prompt_embeds_d,
+                                                        id_pixel_values,
+                                                        prompt_embeds,
                                                         class_tokens_mask_d,
                                                         class_tokens_mask_pos,
-                                                        id_embeds_d,
+                                                        id_embeds,
                                                         left, right);
 
         ggml_build_forward_expand(gf, updated_prompt_embeds);
@@ -548,20 +547,16 @@ public:
         return gf;
     }
 
-    bool compute(const int n_threads,
-                 ggml_tensor* id_pixel_values,
-                 ggml_tensor* prompt_embeds,
-                 ggml_tensor* id_embeds,
-                 std::vector<bool>& class_tokens_mask,
-                 ggml_tensor** updated_prompt_embeds,
-                 ggml_context* output_ctx) {
+    sd::Tensor<float> compute(const int n_threads,
+                              const sd::Tensor<float>& id_pixel_values,
+                              const sd::Tensor<float>& prompt_embeds,
+                              const sd::Tensor<float>& id_embeds,
+                              std::vector<bool>& class_tokens_mask) {
         auto get_graph = [&]() -> ggml_cgraph* {
-            // return build_graph(compute_allocr, id_pixel_values, prompt_embeds, class_tokens_mask);
             return build_graph(id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds);
         };
 
-        // GGMLRunner::compute(get_graph, n_threads, updated_prompt_embeds);
-        return GGMLRunner::compute(get_graph, n_threads, true, updated_prompt_embeds, output_ctx);
+        return take_or_empty(GGMLRunner::compute<float>(get_graph, n_threads, true));
     }
 };
 

src/preprocessing.hpp

@@ -1,179 +1,297 @@
 #ifndef __PREPROCESSING_HPP__
 #define __PREPROCESSING_HPP__
 
+#include <cmath>
+#include <limits>
+
 #include "ggml_extend.hpp"
+
 #define M_PI_ 3.14159265358979323846f
 
-void convolve(ggml_tensor* input, ggml_tensor* output, ggml_tensor* kernel, int padding) {
-    ggml_init_params params;
-    params.mem_size          = 80 * input->ne[0] * input->ne[1];  // 20M for 512x512
-    params.mem_buffer        = nullptr;
-    params.no_alloc          = false;
-    ggml_context* ctx0       = ggml_init(params);
-    ggml_tensor* kernel_fp16 = ggml_new_tensor_4d(ctx0, GGML_TYPE_F16, kernel->ne[0], kernel->ne[1], 1, 1);
-    ggml_fp32_to_fp16_row((float*)kernel->data, (ggml_fp16_t*)kernel_fp16->data, ggml_nelements(kernel));
-    ggml_tensor* h  = ggml_conv_2d(ctx0, kernel_fp16, input, 1, 1, padding, padding, 1, 1);
-    ggml_cgraph* gf = ggml_new_graph(ctx0);
-    ggml_build_forward_expand(gf, ggml_cpy(ctx0, h, output));
-    ggml_graph_compute_with_ctx(ctx0, gf, 1);
-    ggml_free(ctx0);
+static inline int64_t preprocessing_offset_4d(const sd::Tensor<float>& tensor, int64_t i0, int64_t i1 = 0, int64_t i2 = 0, int64_t i3 = 0) {
+    const auto& shape = tensor.shape();
+    int64_t n0        = shape.size() > 0 ? shape[0] : 1;
+    int64_t n1        = shape.size() > 1 ? shape[1] : 1;
+    int64_t n2        = shape.size() > 2 ? shape[2] : 1;
+    return ((i3 * n2 + i2) * n1 + i1) * n0 + i0;
 }
 
-void gaussian_kernel(ggml_tensor* kernel) {
-    int ks_mid   = static_cast<int>(kernel->ne[0] / 2);
+static inline float preprocessing_get_4d(const sd::Tensor<float>& tensor, int64_t i0, int64_t i1 = 0, int64_t i2 = 0, int64_t i3 = 0) {
+    return tensor.values()[static_cast<size_t>(preprocessing_offset_4d(tensor, i0, i1, i2, i3))];
+}
+
+static inline void preprocessing_set_4d(sd::Tensor<float>& tensor, float value, int64_t i0, int64_t i1 = 0, int64_t i2 = 0, int64_t i3 = 0) {
+    tensor.values()[static_cast<size_t>(preprocessing_offset_4d(tensor, i0, i1, i2, i3))] = value;
+}
+
+static inline uint8_t preprocessing_float_to_u8(float value) {
+    if (value <= 0.0f) {
+        return 0;
+    }
+    if (value >= 1.0f) {
+        return 255;
+    }
+    return static_cast<uint8_t>(value * 255.0f + 0.5f);
+}
+
+static inline void preprocessing_tensor_frame_to_sd_image(const sd::Tensor<float>& tensor, int frame_index, uint8_t* image_data) {
+    const auto& shape = tensor.shape();
+    GGML_ASSERT(shape.size() == 4 || shape.size() == 5);
+    GGML_ASSERT(image_data != nullptr);
+
+    const int width     = static_cast<int>(shape[0]);
+    const int height    = static_cast<int>(shape[1]);
+    const int channel   = static_cast<int>(shape[shape.size() == 5 ? 3 : 2]);
+    const size_t pixels = static_cast<size_t>(width) * static_cast<size_t>(height);
+    const float* src    = tensor.data();
+
+    if (shape.size() == 4) {
+        GGML_ASSERT(frame_index >= 0 && frame_index < shape[3]);
+        const size_t frame_stride = pixels * static_cast<size_t>(channel);
+        const float* frame_ptr    = src + static_cast<size_t>(frame_index) * frame_stride;
+        if (channel == 3) {
+            const float* c0 = frame_ptr;
+            const float* c1 = frame_ptr + pixels;
+            const float* c2 = frame_ptr + pixels * 2;
+            for (size_t i = 0; i < pixels; ++i) {
+                image_data[i * 3 + 0] = preprocessing_float_to_u8(c0[i]);
+                image_data[i * 3 + 1] = preprocessing_float_to_u8(c1[i]);
+                image_data[i * 3 + 2] = preprocessing_float_to_u8(c2[i]);
+            }
+            return;
+        }
+
+        for (size_t i = 0; i < pixels; ++i) {
+            for (int c = 0; c < channel; ++c) {
+                image_data[i * static_cast<size_t>(channel) + static_cast<size_t>(c)] =
+                    preprocessing_float_to_u8(frame_ptr[i + pixels * static_cast<size_t>(c)]);
+            }
+        }
+        return;
+    }
+
+    GGML_ASSERT(frame_index >= 0 && frame_index < shape[2]);
+    const size_t channel_stride = pixels * static_cast<size_t>(shape[2]);
+    const float* frame_ptr      = src + static_cast<size_t>(frame_index) * pixels;
+    if (channel == 3) {
+        const float* c0 = frame_ptr;
+        const float* c1 = frame_ptr + channel_stride;
+        const float* c2 = frame_ptr + channel_stride * 2;
+        for (size_t i = 0; i < pixels; ++i) {
+            image_data[i * 3 + 0] = preprocessing_float_to_u8(c0[i]);
+            image_data[i * 3 + 1] = preprocessing_float_to_u8(c1[i]);
+            image_data[i * 3 + 2] = preprocessing_float_to_u8(c2[i]);
+        }
+        return;
+    }
+
+    for (size_t i = 0; i < pixels; ++i) {
+        for (int c = 0; c < channel; ++c) {
+            image_data[i * static_cast<size_t>(channel) + static_cast<size_t>(c)] =
+                preprocessing_float_to_u8(frame_ptr[i + channel_stride * static_cast<size_t>(c)]);
+        }
+    }
+}
+
+static inline sd::Tensor<float> sd_image_to_preprocessing_tensor(sd_image_t image) {
+    sd::Tensor<float> tensor({static_cast<int64_t>(image.width), static_cast<int64_t>(image.height), static_cast<int64_t>(image.channel), 1});
+    for (uint32_t y = 0; y < image.height; ++y) {
+        for (uint32_t x = 0; x < image.width; ++x) {
+            for (uint32_t c = 0; c < image.channel; ++c) {
+                preprocessing_set_4d(tensor, sd_image_get_f32(image, x, y, c), x, y, c, 0);
+            }
+        }
+    }
+    return tensor;
+}
+
+static inline void preprocessing_tensor_to_sd_image(const sd::Tensor<float>& tensor, uint8_t* image_data) {
+    GGML_ASSERT(tensor.dim() == 4);
+    GGML_ASSERT(tensor.shape()[3] == 1);
+    preprocessing_tensor_frame_to_sd_image(tensor, 0, image_data);
+}
+
+static inline sd::Tensor<float> gaussian_kernel_tensor(int kernel_size) {
+    sd::Tensor<float> kernel({kernel_size, kernel_size, 1, 1});
+    int ks_mid   = kernel_size / 2;
     float sigma  = 1.4f;
-    float normal = 1.f / (2.0f * M_PI_ * powf(sigma, 2.0f));
-    for (int y = 0; y < kernel->ne[0]; y++) {
+    float normal = 1.f / (2.0f * M_PI_ * std::pow(sigma, 2.0f));
+    for (int y = 0; y < kernel_size; ++y) {
         float gx = static_cast<float>(-ks_mid + y);
-        for (int x = 0; x < kernel->ne[1]; x++) {
+        for (int x = 0; x < kernel_size; ++x) {
             float gy = static_cast<float>(-ks_mid + x);
-            float k_ = expf(-((gx * gx + gy * gy) / (2.0f * powf(sigma, 2.0f)))) * normal;
-            ggml_ext_tensor_set_f32(kernel, k_, x, y);
+            float k  = std::exp(-((gx * gx + gy * gy) / (2.0f * std::pow(sigma, 2.0f)))) * normal;
+            preprocessing_set_4d(kernel, k, x, y, 0, 0);
         }
     }
+    return kernel;
 }
 
-void grayscale(ggml_tensor* rgb_img, ggml_tensor* grayscale) {
-    for (int iy = 0; iy < rgb_img->ne[1]; iy++) {
-        for (int ix = 0; ix < rgb_img->ne[0]; ix++) {
-            float r    = ggml_ext_tensor_get_f32(rgb_img, ix, iy);
-            float g    = ggml_ext_tensor_get_f32(rgb_img, ix, iy, 1);
-            float b    = ggml_ext_tensor_get_f32(rgb_img, ix, iy, 2);
+static inline sd::Tensor<float> convolve_tensor(const sd::Tensor<float>& input, const sd::Tensor<float>& kernel, int padding) {
+    GGML_ASSERT(input.dim() == 4);
+    GGML_ASSERT(kernel.dim() == 4);
+    GGML_ASSERT(input.shape()[3] == 1);
+    GGML_ASSERT(kernel.shape()[2] == 1);
+    GGML_ASSERT(kernel.shape()[3] == 1);
+
+    sd::Tensor<float> output(input.shape());
+    int64_t width    = input.shape()[0];
+    int64_t height   = input.shape()[1];
+    int64_t channels = input.shape()[2];
+    int64_t kernel_w = kernel.shape()[0];
+    int64_t kernel_h = kernel.shape()[1];
+
+    for (int64_t c = 0; c < channels; ++c) {
+        for (int64_t y = 0; y < height; ++y) {
+            for (int64_t x = 0; x < width; ++x) {
+                float sum = 0.0f;
+                for (int64_t ky = 0; ky < kernel_h; ++ky) {
+                    int64_t iy = y + ky - padding;
+                    if (iy < 0 || iy >= height) {
+                        continue;
+                    }
+                    for (int64_t kx = 0; kx < kernel_w; ++kx) {
+                        int64_t ix = x + kx - padding;
+                        if (ix < 0 || ix >= width) {
+                            continue;
+                        }
+                        sum += preprocessing_get_4d(input, ix, iy, c, 0) * preprocessing_get_4d(kernel, kx, ky, 0, 0);
+                    }
+                }
+                preprocessing_set_4d(output, sum, x, y, c, 0);
+            }
+        }
+    }
+    return output;
+}
+
+static inline sd::Tensor<float> grayscale_tensor(const sd::Tensor<float>& rgb_img) {
+    GGML_ASSERT(rgb_img.dim() == 4);
+    GGML_ASSERT(rgb_img.shape()[2] >= 3);
+    sd::Tensor<float> grayscale({rgb_img.shape()[0], rgb_img.shape()[1], 1, rgb_img.shape()[3]});
+    for (int64_t iy = 0; iy < rgb_img.shape()[1]; ++iy) {
+        for (int64_t ix = 0; ix < rgb_img.shape()[0]; ++ix) {
+            float r    = preprocessing_get_4d(rgb_img, ix, iy, 0, 0);
+            float g    = preprocessing_get_4d(rgb_img, ix, iy, 1, 0);
+            float b    = preprocessing_get_4d(rgb_img, ix, iy, 2, 0);
             float gray = 0.2989f * r + 0.5870f * g + 0.1140f * b;
-            ggml_ext_tensor_set_f32(grayscale, gray, ix, iy);
+            preprocessing_set_4d(grayscale, gray, ix, iy, 0, 0);
         }
     }
+    return grayscale;
 }
 
-void prop_hypot(ggml_tensor* x, ggml_tensor* y, ggml_tensor* h) {
-    int n_elements = static_cast<int>(ggml_nelements(h));
-    float* dx      = (float*)x->data;
-    float* dy      = (float*)y->data;
-    float* dh      = (float*)h->data;
-    for (int i = 0; i < n_elements; i++) {
-        dh[i] = sqrtf(dx[i] * dx[i] + dy[i] * dy[i]);
+static inline sd::Tensor<float> tensor_hypot(const sd::Tensor<float>& x, const sd::Tensor<float>& y) {
+    sd::tensor_check_same_shape(x, y);
+    sd::Tensor<float> out(x.shape());
+    for (int64_t i = 0; i < out.numel(); ++i) {
+        out[i] = std::sqrt(x[i] * x[i] + y[i] * y[i]);
     }
+    return out;
 }
 
-void prop_arctan2(ggml_tensor* x, ggml_tensor* y, ggml_tensor* h) {
-    int n_elements = static_cast<int>(ggml_nelements(h));
-    float* dx      = (float*)x->data;
-    float* dy      = (float*)y->data;
-    float* dh      = (float*)h->data;
-    for (int i = 0; i < n_elements; i++) {
-        dh[i] = atan2f(dy[i], dx[i]);
+static inline sd::Tensor<float> tensor_arctan2(const sd::Tensor<float>& x, const sd::Tensor<float>& y) {
+    sd::tensor_check_same_shape(x, y);
+    sd::Tensor<float> out(x.shape());
+    for (int64_t i = 0; i < out.numel(); ++i) {
+        out[i] = std::atan2(y[i], x[i]);
     }
+    return out;
 }
 
-void normalize_tensor(ggml_tensor* g) {
-    int n_elements = static_cast<int>(ggml_nelements(g));
-    float* dg      = (float*)g->data;
-    float max      = -INFINITY;
-    for (int i = 0; i < n_elements; i++) {
-        max = dg[i] > max ? dg[i] : max;
+static inline void normalize_tensor(sd::Tensor<float>* g) {
+    GGML_ASSERT(g != nullptr);
+    if (g->empty()) {
+        return;
     }
-    max = 1.0f / max;
-    for (int i = 0; i < n_elements; i++) {
-        dg[i] *= max;
+    float max_value = -std::numeric_limits<float>::infinity();
+    for (int64_t i = 0; i < g->numel(); ++i) {
+        max_value = std::max(max_value, (*g)[i]);
     }
+    if (max_value == 0.0f || !std::isfinite(max_value)) {
+        return;
+    }
+    *g *= (1.0f / max_value);
 }
 
-void non_max_supression(ggml_tensor* result, ggml_tensor* G, ggml_tensor* D) {
-    for (int iy = 1; iy < result->ne[1] - 1; iy++) {
-        for (int ix = 1; ix < result->ne[0] - 1; ix++) {
-            float angle = ggml_ext_tensor_get_f32(D, ix, iy) * 180.0f / M_PI_;
-            angle       = angle < 0.0f ? angle += 180.0f : angle;
+static inline sd::Tensor<float> non_max_supression(const sd::Tensor<float>& G, const sd::Tensor<float>& D) {
+    GGML_ASSERT(G.shape() == D.shape());
+    sd::Tensor<float> result = sd::Tensor<float>::zeros(G.shape());
+    for (int64_t iy = 1; iy < result.shape()[1] - 1; ++iy) {
+        for (int64_t ix = 1; ix < result.shape()[0] - 1; ++ix) {
+            float angle = preprocessing_get_4d(D, ix, iy, 0, 0) * 180.0f / M_PI_;
+            angle       = angle < 0.0f ? angle + 180.0f : angle;
             float q     = 1.0f;
             float r     = 1.0f;
 
-            // angle 0
-            if ((0 >= angle && angle < 22.5f) || (157.5f >= angle && angle <= 180)) {
-                q = ggml_ext_tensor_get_f32(G, ix, iy + 1);
-                r = ggml_ext_tensor_get_f32(G, ix, iy - 1);
-            }
-            // angle 45
-            else if (22.5f >= angle && angle < 67.5f) {
-                q = ggml_ext_tensor_get_f32(G, ix + 1, iy - 1);
-                r = ggml_ext_tensor_get_f32(G, ix - 1, iy + 1);
-            }
-            // angle 90
-            else if (67.5f >= angle && angle < 112.5) {
-                q = ggml_ext_tensor_get_f32(G, ix + 1, iy);
-                r = ggml_ext_tensor_get_f32(G, ix - 1, iy);
-            }
-            // angle 135
-            else if (112.5 >= angle && angle < 157.5f) {
-                q = ggml_ext_tensor_get_f32(G, ix - 1, iy - 1);
-                r = ggml_ext_tensor_get_f32(G, ix + 1, iy + 1);
+            if ((0 >= angle && angle < 22.5f) || (157.5f >= angle && angle <= 180.0f)) {
+                q = preprocessing_get_4d(G, ix, iy + 1, 0, 0);
+                r = preprocessing_get_4d(G, ix, iy - 1, 0, 0);
+            } else if (22.5f >= angle && angle < 67.5f) {
+                q = preprocessing_get_4d(G, ix + 1, iy - 1, 0, 0);
+                r = preprocessing_get_4d(G, ix - 1, iy + 1, 0, 0);
+            } else if (67.5f >= angle && angle < 112.5f) {
+                q = preprocessing_get_4d(G, ix + 1, iy, 0, 0);
+                r = preprocessing_get_4d(G, ix - 1, iy, 0, 0);
+            } else if (112.5f >= angle && angle < 157.5f) {
+                q = preprocessing_get_4d(G, ix - 1, iy - 1, 0, 0);
+                r = preprocessing_get_4d(G, ix + 1, iy + 1, 0, 0);
             }
 
-            float cur = ggml_ext_tensor_get_f32(G, ix, iy);
-            if ((cur >= q) && (cur >= r)) {
-                ggml_ext_tensor_set_f32(result, cur, ix, iy);
-            } else {
-                ggml_ext_tensor_set_f32(result, 0.0f, ix, iy);
-            }
+            float cur = preprocessing_get_4d(G, ix, iy, 0, 0);
+            preprocessing_set_4d(result, (cur >= q && cur >= r) ? cur : 0.0f, ix, iy, 0, 0);
         }
     }
+    return result;
 }
 
-void threshold_hystersis(ggml_tensor* img, float high_threshold, float low_threshold, float weak, float strong) {
-    int n_elements = static_cast<int>(ggml_nelements(img));
-    float* imd     = (float*)img->data;
-    float max      = -INFINITY;
-    for (int i = 0; i < n_elements; i++) {
-        max = imd[i] > max ? imd[i] : max;
+static inline void threshold_hystersis(sd::Tensor<float>* img, float high_threshold, float low_threshold, float weak, float strong) {
+    GGML_ASSERT(img != nullptr);
+    if (img->empty()) {
+        return;
     }
-    float ht = max * high_threshold;
+    float max_value = -std::numeric_limits<float>::infinity();
+    for (int64_t i = 0; i < img->numel(); ++i) {
+        max_value = std::max(max_value, (*img)[i]);
+    }
+
+    float ht = max_value * high_threshold;
     float lt = ht * low_threshold;
-    for (int i = 0; i < n_elements; i++) {
-        float img_v = imd[i];
-        if (img_v >= ht) {  // strong pixel
-            imd[i] = strong;
-        } else if (img_v <= ht && img_v >= lt) {  // strong pixel
-            imd[i] = weak;
+    for (int64_t i = 0; i < img->numel(); ++i) {
+        float img_v = (*img)[i];
+        if (img_v >= ht) {
+            (*img)[i] = strong;
+        } else if (img_v <= ht && img_v >= lt) {
+            (*img)[i] = weak;
         }
     }
 
-    for (int iy = 0; iy < img->ne[1]; iy++) {
-        for (int ix = 0; ix < img->ne[0]; ix++) {
-            if (ix >= 3 && ix <= img->ne[0] - 3 && iy >= 3 && iy <= img->ne[1] - 3) {
-                ggml_ext_tensor_set_f32(img, ggml_ext_tensor_get_f32(img, ix, iy), ix, iy);
-            } else {
-                ggml_ext_tensor_set_f32(img, 0.0f, ix, iy);
+    for (int64_t iy = 0; iy < img->shape()[1]; ++iy) {
+        for (int64_t ix = 0; ix < img->shape()[0]; ++ix) {
+            if (!(ix >= 3 && ix <= img->shape()[0] - 3 && iy >= 3 && iy <= img->shape()[1] - 3)) {
+                preprocessing_set_4d(*img, 0.0f, ix, iy, 0, 0);
             }
         }
     }
 
-    // hysteresis
-    for (int iy = 1; iy < img->ne[1] - 1; iy++) {
-        for (int ix = 1; ix < img->ne[0] - 1; ix++) {
-            float imd_v = ggml_ext_tensor_get_f32(img, ix, iy);
+    for (int64_t iy = 1; iy < img->shape()[1] - 1; ++iy) {
+        for (int64_t ix = 1; ix < img->shape()[0] - 1; ++ix) {
+            float imd_v = preprocessing_get_4d(*img, ix, iy, 0, 0);
             if (imd_v == weak) {
-                if (ggml_ext_tensor_get_f32(img, ix + 1, iy - 1) == strong || ggml_ext_tensor_get_f32(img, ix + 1, iy) == strong ||
-                    ggml_ext_tensor_get_f32(img, ix, iy - 1) == strong || ggml_ext_tensor_get_f32(img, ix, iy + 1) == strong ||
-                    ggml_ext_tensor_get_f32(img, ix - 1, iy - 1) == strong || ggml_ext_tensor_get_f32(img, ix - 1, iy) == strong) {
-                    ggml_ext_tensor_set_f32(img, strong, ix, iy);
-                } else {
-                    ggml_ext_tensor_set_f32(img, 0.0f, ix, iy);
-                }
+                bool has_strong_neighbor =
+                    preprocessing_get_4d(*img, ix + 1, iy - 1, 0, 0) == strong ||
+                    preprocessing_get_4d(*img, ix + 1, iy, 0, 0) == strong ||
+                    preprocessing_get_4d(*img, ix, iy - 1, 0, 0) == strong ||
+                    preprocessing_get_4d(*img, ix, iy + 1, 0, 0) == strong ||
+                    preprocessing_get_4d(*img, ix - 1, iy - 1, 0, 0) == strong ||
+                    preprocessing_get_4d(*img, ix - 1, iy, 0, 0) == strong;
+                preprocessing_set_4d(*img, has_strong_neighbor ? strong : 0.0f, ix, iy, 0, 0);
             }
         }
     }
 }
 
 bool preprocess_canny(sd_image_t img, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
-    ggml_init_params params;
-    params.mem_size        = static_cast<size_t>(40 * img.width * img.height);  // 10MB for 512x512
-    params.mem_buffer      = nullptr;
-    params.no_alloc        = false;
-    ggml_context* work_ctx = ggml_init(params);
-
-    if (!work_ctx) {
-        LOG_ERROR("ggml_init() failed");
-        return false;
-    }
-
     float kX[9] = {
         -1, 0, 1,
         -2, 0, 2,
@@ -184,43 +302,33 @@ bool preprocess_canny(sd_image_t img, float high_threshold, float low_threshold,
         0, 0, 0,
         -1, -2, -1};
 
-    // generate kernel
-    int kernel_size      = 5;
-    ggml_tensor* gkernel = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, kernel_size, kernel_size, 1, 1);
-    ggml_tensor* sf_kx   = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 3, 3, 1, 1);
-    memcpy(sf_kx->data, kX, ggml_nbytes(sf_kx));
-    ggml_tensor* sf_ky = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 3, 3, 1, 1);
-    memcpy(sf_ky->data, kY, ggml_nbytes(sf_ky));
-    gaussian_kernel(gkernel);
-    ggml_tensor* image      = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 3, 1);
-    ggml_tensor* image_gray = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 1, 1);
-    ggml_tensor* iX         = ggml_dup_tensor(work_ctx, image_gray);
-    ggml_tensor* iY         = ggml_dup_tensor(work_ctx, image_gray);
-    ggml_tensor* G          = ggml_dup_tensor(work_ctx, image_gray);
-    ggml_tensor* tetha      = ggml_dup_tensor(work_ctx, image_gray);
-    sd_image_to_ggml_tensor(img, image);
-    grayscale(image, image_gray);
-    convolve(image_gray, image_gray, gkernel, 2);
-    convolve(image_gray, iX, sf_kx, 1);
-    convolve(image_gray, iY, sf_ky, 1);
-    prop_hypot(iX, iY, G);
-    normalize_tensor(G);
-    prop_arctan2(iX, iY, tetha);
-    non_max_supression(image_gray, G, tetha);
-    threshold_hystersis(image_gray, high_threshold, low_threshold, weak, strong);
-    // to RGB channels
-    for (uint32_t iy = 0; iy < img.height; iy++) {
-        for (uint32_t ix = 0; ix < img.width; ix++) {
-            float gray = ggml_ext_tensor_get_f32(image_gray, ix, iy);
+    sd::Tensor<float> gkernel = gaussian_kernel_tensor(5);
+    sd::Tensor<float> sf_kx({3, 3, 1, 1}, std::vector<float>(kX, kX + 9));
+    sd::Tensor<float> sf_ky({3, 3, 1, 1}, std::vector<float>(kY, kY + 9));
+
+    sd::Tensor<float> image      = sd_image_to_preprocessing_tensor(img);
+    sd::Tensor<float> image_gray = grayscale_tensor(image);
+    image_gray                   = convolve_tensor(image_gray, gkernel, 2);
+    sd::Tensor<float> iX         = convolve_tensor(image_gray, sf_kx, 1);
+    sd::Tensor<float> iY         = convolve_tensor(image_gray, sf_ky, 1);
+    sd::Tensor<float> G          = tensor_hypot(iX, iY);
+    normalize_tensor(&G);
+    sd::Tensor<float> theta = tensor_arctan2(iX, iY);
+    image_gray              = non_max_supression(G, theta);
+    threshold_hystersis(&image_gray, high_threshold, low_threshold, weak, strong);
+
+    for (uint32_t iy = 0; iy < img.height; ++iy) {
+        for (uint32_t ix = 0; ix < img.width; ++ix) {
+            float gray = preprocessing_get_4d(image_gray, ix, iy, 0, 0);
             gray       = inverse ? 1.0f - gray : gray;
-            ggml_ext_tensor_set_f32(image, gray, ix, iy);
-            ggml_ext_tensor_set_f32(image, gray, ix, iy, 1);
-            ggml_ext_tensor_set_f32(image, gray, ix, iy, 2);
+            for (uint32_t c = 0; c < img.channel; ++c) {
+                preprocessing_set_4d(image, gray, ix, iy, c, 0);
+            }
         }
     }
-    ggml_tensor_to_sd_image(image, img.data);
-    ggml_free(work_ctx);
+
+    preprocessing_tensor_to_sd_image(image, img.data);
     return true;
 }
 
-#endif  // __PREPROCESSING_HPP__
+#endif  // __PREPROCESSING_HPP__

src/qwen_image.hpp

@@ -95,9 +95,7 @@ namespace Qwen {
 
             float scale         = 1.f / 32.f;
             bool force_prec_f32 = false;
-#ifdef SD_USE_VULKAN
-            force_prec_f32 = true;
-#endif
+
             // The purpose of the scale here is to prevent NaN issues in certain situations.
             // For example when using CUDA but the weights are k-quants (not all prompts).
             blocks["to_out.0"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_dim, out_bias, false, force_prec_f32, scale));
@@ -124,6 +122,10 @@ namespace Qwen {
             auto to_v     = std::dynamic_pointer_cast<Linear>(blocks["to_v"]);
             auto to_out_0 = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
 
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                to_out_0->set_force_prec_f32(true);
+            }
+
             auto norm_added_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_q"]);
             auto norm_added_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_k"]);
 
@@ -410,6 +412,9 @@ namespace Qwen {
             auto img = img_in->forward(ctx, x);
             auto txt = txt_norm->forward(ctx, context);
             txt      = txt_in->forward(ctx, txt);
+            sd::ggml_graph_cut::mark_graph_cut(img, "qwen_image.prelude", "img");
+            sd::ggml_graph_cut::mark_graph_cut(txt, "qwen_image.prelude", "txt");
+            // sd::ggml_graph_cut::mark_graph_cut(t_emb, "qwen_image.prelude", "t_emb");
 
             for (int i = 0; i < params.num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<QwenImageTransformerBlock>(blocks["transformer_blocks." + std::to_string(i)]);
@@ -417,6 +422,8 @@ namespace Qwen {
                 auto result = block->forward(ctx, img, txt, t_emb, pe, modulate_index);
                 img         = result.first;
                 txt         = result.second;
+                sd::ggml_graph_cut::mark_graph_cut(img, "qwen_image.transformer_blocks." + std::to_string(i), "img");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "qwen_image.transformer_blocks." + std::to_string(i), "txt");
             }
 
             if (params.zero_cond_t) {
@@ -525,20 +532,21 @@ namespace Qwen {
             qwen_image.get_param_tensors(tensors, prefix);
         }
 
-        ggml_cgraph* build_graph(ggml_tensor* x,
-                                 ggml_tensor* timesteps,
-                                 ggml_tensor* context,
-                                 std::vector<ggml_tensor*> ref_latents = {},
-                                 bool increase_ref_index               = false) {
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor,
+                                 const std::vector<sd::Tensor<float>>& ref_latents_tensor = {},
+                                 bool increase_ref_index                                  = false) {
+            ggml_cgraph* gf        = new_graph_custom(QWEN_IMAGE_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timesteps = make_input(timesteps_tensor);
             GGML_ASSERT(x->ne[3] == 1);
-            ggml_cgraph* gf = new_graph_custom(QWEN_IMAGE_GRAPH_SIZE);
-
-            x         = to_backend(x);
-            context   = to_backend(context);
-            timesteps = to_backend(timesteps);
-
-            for (int i = 0; i < ref_latents.size(); i++) {
-                ref_latents[i] = to_backend(ref_latents[i]);
+            GGML_ASSERT(!context_tensor.empty());
+            ggml_tensor* context = make_input(context_tensor);
+            std::vector<ggml_tensor*> ref_latents;
+            ref_latents.reserve(ref_latents_tensor.size());
+            for (const auto& ref_latent_tensor : ref_latents_tensor) {
+                ref_latents.push_back(make_input(ref_latent_tensor));
             }
 
             pe_vec      = Rope::gen_qwen_image_pe(static_cast<int>(x->ne[1]),
@@ -600,14 +608,12 @@ namespace Qwen {
             return gf;
         }
 
-        bool compute(int n_threads,
-                     ggml_tensor* x,
-                     ggml_tensor* timesteps,
-                     ggml_tensor* context,
-                     std::vector<ggml_tensor*> ref_latents = {},
-                     bool increase_ref_index               = false,
-                     ggml_tensor** output                  = nullptr,
-                     ggml_context* output_ctx              = nullptr) {
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context,
+                                  const std::vector<sd::Tensor<float>>& ref_latents = {},
+                                  bool increase_ref_index                           = false) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
@@ -615,7 +621,7 @@ namespace Qwen {
                 return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
             };
 
-            return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
         }
 
         void test() {
@@ -624,30 +630,37 @@ namespace Qwen {
             params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
-            ggml_context* work_ctx = ggml_init(params);
-            GGML_ASSERT(work_ctx != nullptr);
+            ggml_context* ctx = ggml_init(params);
+            GGML_ASSERT(ctx != nullptr);
 
             {
-                // auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 16, 16, 16, 1);
+                // auto x = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, 16, 16, 16, 1);
                 // ggml_set_f32(x, 0.01f);
-                auto x = load_tensor_from_file(work_ctx, "./qwen_image_x.bin");
-                print_ggml_tensor(x);
+                auto x = sd::load_tensor_from_file_as_tensor<float>("./qwen_image_x.bin");
+                print_sd_tensor(x);
 
                 std::vector<float> timesteps_vec(1, 1000.f);
-                auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
+                auto timesteps = sd::Tensor<float>::from_vector(timesteps_vec);
 
-                // auto context = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 3584, 256, 1);
+                // auto context = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 3584, 256, 1);
                 // ggml_set_f32(context, 0.01f);
-                auto context = load_tensor_from_file(work_ctx, "./qwen_image_context.bin");
-                print_ggml_tensor(context);
+                auto context = sd::load_tensor_from_file_as_tensor<float>("./qwen_image_context.bin");
+                print_sd_tensor(context);
 
-                ggml_tensor* out = nullptr;
+                sd::Tensor<float> out;
 
-                int64_t t0 = ggml_time_ms();
-                compute(8, x, timesteps, context, {}, false, &out, work_ctx);
-                int64_t t1 = ggml_time_ms();
+                int64_t t0   = ggml_time_ms();
+                auto out_opt = compute(8,
+                                       x,
+                                       timesteps,
+                                       context,
+                                       {},
+                                       false);
+                int64_t t1   = ggml_time_ms();
 
-                print_ggml_tensor(out);
+                GGML_ASSERT(!out_opt.empty());
+                out = std::move(out_opt);
+                print_sd_tensor(out);
                 LOG_DEBUG("qwen_image test done in %lldms", t1 - t0);
             }
         }

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/sample-cache.cpp

@@ -0,0 +1,361 @@
+#include "sample-cache.h"
+
+namespace sd_sample {
+
+    static float get_cache_reuse_threshold(const sd_cache_params_t& params) {
+        float reuse_threshold = params.reuse_threshold;
+        if (reuse_threshold == INFINITY) {
+            if (params.mode == SD_CACHE_EASYCACHE) {
+                reuse_threshold = 0.2f;
+            } else if (params.mode == SD_CACHE_UCACHE) {
+                reuse_threshold = 1.0f;
+            }
+        }
+        return std::max(0.0f, reuse_threshold);
+    }
+
+    bool SampleCacheRuntime::easycache_enabled() const {
+        return mode == SampleCacheMode::EASYCACHE;
+    }
+
+    bool SampleCacheRuntime::ucache_enabled() const {
+        return mode == SampleCacheMode::UCACHE;
+    }
+
+    bool SampleCacheRuntime::cachedit_enabled() const {
+        return mode == SampleCacheMode::CACHEDIT;
+    }
+
+    static bool has_valid_cache_percent_range(const sd_cache_params_t& cache_params) {
+        if (cache_params.mode != SD_CACHE_EASYCACHE && cache_params.mode != SD_CACHE_UCACHE) {
+            return true;
+        }
+
+        return cache_params.start_percent >= 0.0f &&
+               cache_params.start_percent < 1.0f &&
+               cache_params.end_percent > 0.0f &&
+               cache_params.end_percent <= 1.0f &&
+               cache_params.start_percent < cache_params.end_percent;
+    }
+
+    static void init_easycache_runtime(SampleCacheRuntime& runtime,
+                                       SDVersion version,
+                                       const sd_cache_params_t& cache_params,
+                                       Denoiser* denoiser) {
+        if (!sd_version_is_dit(version)) {
+            LOG_WARN("EasyCache requested but not supported for this model type");
+            return;
+        }
+
+        EasyCacheConfig config;
+        config.enabled         = true;
+        config.reuse_threshold = get_cache_reuse_threshold(cache_params);
+        config.start_percent   = cache_params.start_percent;
+        config.end_percent     = cache_params.end_percent;
+
+        runtime.easycache.init(config, denoiser);
+        if (!runtime.easycache.enabled()) {
+            LOG_WARN("EasyCache requested but could not be initialized for this run");
+            return;
+        }
+
+        runtime.mode = SampleCacheMode::EASYCACHE;
+        LOG_INFO("EasyCache enabled - threshold: %.3f, start: %.2f, end: %.2f",
+                 config.reuse_threshold,
+                 config.start_percent,
+                 config.end_percent);
+    }
+
+    static void init_ucache_runtime(SampleCacheRuntime& runtime,
+                                    SDVersion version,
+                                    const sd_cache_params_t& cache_params,
+                                    Denoiser* denoiser,
+                                    const std::vector<float>& sigmas) {
+        if (!sd_version_is_unet(version)) {
+            LOG_WARN("UCache requested but not supported for this model type (only UNET models)");
+            return;
+        }
+
+        UCacheConfig config;
+        config.enabled                = true;
+        config.reuse_threshold        = get_cache_reuse_threshold(cache_params);
+        config.start_percent          = cache_params.start_percent;
+        config.end_percent            = cache_params.end_percent;
+        config.error_decay_rate       = std::max(0.0f, std::min(1.0f, cache_params.error_decay_rate));
+        config.use_relative_threshold = cache_params.use_relative_threshold;
+        config.reset_error_on_compute = cache_params.reset_error_on_compute;
+
+        runtime.ucache.init(config, denoiser);
+        if (!runtime.ucache.enabled()) {
+            LOG_WARN("UCache requested but could not be initialized for this run");
+            return;
+        }
+
+        runtime.ucache.set_sigmas(sigmas);
+        runtime.mode = SampleCacheMode::UCACHE;
+        LOG_INFO("UCache enabled - threshold: %.3f, start: %.2f, end: %.2f, decay: %.2f, relative: %s, reset: %s",
+                 config.reuse_threshold,
+                 config.start_percent,
+                 config.end_percent,
+                 config.error_decay_rate,
+                 config.use_relative_threshold ? "true" : "false",
+                 config.reset_error_on_compute ? "true" : "false");
+    }
+
+    static void init_cachedit_runtime(SampleCacheRuntime& runtime,
+                                      SDVersion version,
+                                      const sd_cache_params_t& cache_params,
+                                      const std::vector<float>& sigmas) {
+        if (!sd_version_is_dit(version)) {
+            LOG_WARN("CacheDIT requested but not supported for this model type (only DiT models)");
+            return;
+        }
+
+        DBCacheConfig dbcfg;
+        dbcfg.enabled                     = (cache_params.mode == SD_CACHE_DBCACHE || cache_params.mode == SD_CACHE_CACHE_DIT);
+        dbcfg.Fn_compute_blocks           = cache_params.Fn_compute_blocks;
+        dbcfg.Bn_compute_blocks           = cache_params.Bn_compute_blocks;
+        dbcfg.residual_diff_threshold     = cache_params.residual_diff_threshold;
+        dbcfg.max_warmup_steps            = cache_params.max_warmup_steps;
+        dbcfg.max_cached_steps            = cache_params.max_cached_steps;
+        dbcfg.max_continuous_cached_steps = cache_params.max_continuous_cached_steps;
+        if (cache_params.scm_mask != nullptr && strlen(cache_params.scm_mask) > 0) {
+            dbcfg.steps_computation_mask = parse_scm_mask(cache_params.scm_mask);
+        }
+        dbcfg.scm_policy_dynamic = cache_params.scm_policy_dynamic;
+
+        TaylorSeerConfig tcfg;
+        tcfg.enabled             = (cache_params.mode == SD_CACHE_TAYLORSEER || cache_params.mode == SD_CACHE_CACHE_DIT);
+        tcfg.n_derivatives       = cache_params.taylorseer_n_derivatives;
+        tcfg.skip_interval_steps = cache_params.taylorseer_skip_interval;
+
+        runtime.cachedit.init(dbcfg, tcfg);
+        if (!runtime.cachedit.enabled()) {
+            LOG_WARN("CacheDIT requested but could not be initialized for this run");
+            return;
+        }
+
+        runtime.cachedit.set_sigmas(sigmas);
+        runtime.mode = SampleCacheMode::CACHEDIT;
+        LOG_INFO("CacheDIT enabled - mode: %s, Fn: %d, Bn: %d, threshold: %.3f, warmup: %d",
+                 cache_params.mode == SD_CACHE_CACHE_DIT ? "DBCache+TaylorSeer" : (cache_params.mode == SD_CACHE_DBCACHE ? "DBCache" : "TaylorSeer"),
+                 dbcfg.Fn_compute_blocks,
+                 dbcfg.Bn_compute_blocks,
+                 dbcfg.residual_diff_threshold,
+                 dbcfg.max_warmup_steps);
+    }
+
+    static void init_spectrum_runtime(SampleCacheRuntime& runtime,
+                                      SDVersion version,
+                                      const sd_cache_params_t& cache_params,
+                                      const std::vector<float>& sigmas) {
+        if (!sd_version_is_unet(version) && !sd_version_is_dit(version)) {
+            LOG_WARN("Spectrum requested but not supported for this model type (only UNET and DiT models)");
+            return;
+        }
+
+        SpectrumConfig config;
+        config.w            = cache_params.spectrum_w;
+        config.m            = cache_params.spectrum_m;
+        config.lam          = cache_params.spectrum_lam;
+        config.window_size  = cache_params.spectrum_window_size;
+        config.flex_window  = cache_params.spectrum_flex_window;
+        config.warmup_steps = cache_params.spectrum_warmup_steps;
+        config.stop_percent = cache_params.spectrum_stop_percent;
+
+        size_t total_steps = sigmas.size() > 0 ? sigmas.size() - 1 : 0;
+        runtime.spectrum.init(config, total_steps);
+        runtime.spectrum_enabled = true;
+
+        LOG_INFO("Spectrum enabled - w: %.2f, m: %d, lam: %.2f, window: %d, flex: %.2f, warmup: %d, stop: %.0f%%",
+                 config.w, config.m, config.lam,
+                 config.window_size, config.flex_window,
+                 config.warmup_steps, config.stop_percent * 100.0f);
+    }
+
+    SampleCacheRuntime init_sample_cache_runtime(SDVersion version,
+                                                 const sd_cache_params_t* cache_params,
+                                                 Denoiser* denoiser,
+                                                 const std::vector<float>& sigmas) {
+        SampleCacheRuntime runtime;
+        if (cache_params == nullptr || cache_params->mode == SD_CACHE_DISABLED) {
+            return runtime;
+        }
+
+        if (!has_valid_cache_percent_range(*cache_params)) {
+            LOG_WARN("Cache disabled due to invalid percent range (start=%.3f, end=%.3f)",
+                     cache_params->start_percent,
+                     cache_params->end_percent);
+            return runtime;
+        }
+
+        switch (cache_params->mode) {
+            case SD_CACHE_EASYCACHE:
+                init_easycache_runtime(runtime, version, *cache_params, denoiser);
+                break;
+            case SD_CACHE_UCACHE:
+                init_ucache_runtime(runtime, version, *cache_params, denoiser, sigmas);
+                break;
+            case SD_CACHE_DBCACHE:
+            case SD_CACHE_TAYLORSEER:
+            case SD_CACHE_CACHE_DIT:
+                init_cachedit_runtime(runtime, version, *cache_params, sigmas);
+                break;
+            case SD_CACHE_SPECTRUM:
+                init_spectrum_runtime(runtime, version, *cache_params, sigmas);
+                break;
+            default:
+                break;
+        }
+
+        return runtime;
+    }
+
+    SampleStepCacheDispatcher::SampleStepCacheDispatcher(SampleCacheRuntime& runtime, int step, float sigma)
+        : runtime(runtime), step(step), sigma(sigma), step_index(step > 0 ? (step - 1) : -1) {
+        if (step_index < 0) {
+            return;
+        }
+
+        switch (runtime.mode) {
+            case SampleCacheMode::EASYCACHE:
+                runtime.easycache.begin_step(step_index, sigma);
+                break;
+            case SampleCacheMode::UCACHE:
+                runtime.ucache.begin_step(step_index, sigma);
+                break;
+            case SampleCacheMode::CACHEDIT:
+                runtime.cachedit.begin_step(step_index, sigma);
+                break;
+            case SampleCacheMode::NONE:
+                break;
+        }
+    }
+
+    bool SampleStepCacheDispatcher::before_condition(const void* condition,
+                                                     const sd::Tensor<float>& input,
+                                                     sd::Tensor<float>* output) {
+        if (step_index < 0 || condition == nullptr || output == nullptr) {
+            return false;
+        }
+
+        switch (runtime.mode) {
+            case SampleCacheMode::EASYCACHE:
+                return runtime.easycache.before_condition(condition, input, output, sigma, step_index);
+            case SampleCacheMode::UCACHE:
+                return runtime.ucache.before_condition(condition, input, output, sigma, step_index);
+            case SampleCacheMode::CACHEDIT:
+                return runtime.cachedit.before_condition(condition, input, output, sigma, step_index);
+            case SampleCacheMode::NONE:
+                return false;
+        }
+
+        return false;
+    }
+
+    void SampleStepCacheDispatcher::after_condition(const void* condition,
+                                                    const sd::Tensor<float>& input,
+                                                    const sd::Tensor<float>& output) {
+        if (step_index < 0 || condition == nullptr) {
+            return;
+        }
+
+        switch (runtime.mode) {
+            case SampleCacheMode::EASYCACHE:
+                runtime.easycache.after_condition(condition, input, output);
+                break;
+            case SampleCacheMode::UCACHE:
+                runtime.ucache.after_condition(condition, input, output);
+                break;
+            case SampleCacheMode::CACHEDIT:
+                runtime.cachedit.after_condition(condition, input, output);
+                break;
+            case SampleCacheMode::NONE:
+                break;
+        }
+    }
+
+    bool SampleStepCacheDispatcher::is_step_skipped() const {
+        switch (runtime.mode) {
+            case SampleCacheMode::EASYCACHE:
+                return runtime.easycache.is_step_skipped();
+            case SampleCacheMode::UCACHE:
+                return runtime.ucache.is_step_skipped();
+            case SampleCacheMode::CACHEDIT:
+                return runtime.cachedit.is_step_skipped();
+            case SampleCacheMode::NONE:
+                return false;
+        }
+
+        return false;
+    }
+
+    void log_sample_cache_summary(const SampleCacheRuntime& runtime, size_t total_steps) {
+        if (runtime.easycache_enabled()) {
+            if (runtime.easycache.total_steps_skipped > 0 && total_steps > 0) {
+                if (runtime.easycache.total_steps_skipped < static_cast<int>(total_steps)) {
+                    double speedup = static_cast<double>(total_steps) /
+                                     static_cast<double>(total_steps - runtime.easycache.total_steps_skipped);
+                    LOG_INFO("EasyCache skipped %d/%zu steps (%.2fx estimated speedup)",
+                             runtime.easycache.total_steps_skipped,
+                             total_steps,
+                             speedup);
+                } else {
+                    LOG_INFO("EasyCache skipped %d/%zu steps",
+                             runtime.easycache.total_steps_skipped,
+                             total_steps);
+                }
+            } else if (total_steps > 0) {
+                LOG_INFO("EasyCache completed without skipping steps");
+            }
+        }
+
+        if (runtime.ucache_enabled()) {
+            if (runtime.ucache.total_steps_skipped > 0 && total_steps > 0) {
+                if (runtime.ucache.total_steps_skipped < static_cast<int>(total_steps)) {
+                    double speedup = static_cast<double>(total_steps) /
+                                     static_cast<double>(total_steps - runtime.ucache.total_steps_skipped);
+                    LOG_INFO("UCache skipped %d/%zu steps (%.2fx estimated speedup)",
+                             runtime.ucache.total_steps_skipped,
+                             total_steps,
+                             speedup);
+                } else {
+                    LOG_INFO("UCache skipped %d/%zu steps",
+                             runtime.ucache.total_steps_skipped,
+                             total_steps);
+                }
+            } else if (total_steps > 0) {
+                LOG_INFO("UCache completed without skipping steps");
+            }
+        }
+
+        if (runtime.cachedit_enabled()) {
+            if (runtime.cachedit.total_steps_skipped > 0 && total_steps > 0) {
+                if (runtime.cachedit.total_steps_skipped < static_cast<int>(total_steps)) {
+                    double speedup = static_cast<double>(total_steps) /
+                                     static_cast<double>(total_steps - runtime.cachedit.total_steps_skipped);
+                    LOG_INFO("CacheDIT skipped %d/%zu steps (%.2fx estimated speedup)",
+                             runtime.cachedit.total_steps_skipped,
+                             total_steps,
+                             speedup);
+                } else {
+                    LOG_INFO("CacheDIT skipped %d/%zu steps",
+                             runtime.cachedit.total_steps_skipped,
+                             total_steps);
+                }
+            } else if (total_steps > 0) {
+                LOG_INFO("CacheDIT completed without skipping steps");
+            }
+        }
+
+        if (runtime.spectrum_enabled && runtime.spectrum.total_steps_skipped > 0 && total_steps > 0) {
+            double speedup = static_cast<double>(total_steps) /
+                             static_cast<double>(total_steps - runtime.spectrum.total_steps_skipped);
+            LOG_INFO("Spectrum skipped %d/%zu steps (%.2fx estimated speedup)",
+                     runtime.spectrum.total_steps_skipped,
+                     total_steps,
+                     speedup);
+        }
+    }
+
+}  // namespace sd_sample

src/sample-cache.h

@@ -0,0 +1,61 @@
+#ifndef __SAMPLE_CACHE_H__
+#define __SAMPLE_CACHE_H__
+
+#include <vector>
+
+#include "cache_dit.hpp"
+#include "denoiser.hpp"
+#include "easycache.hpp"
+#include "model.h"
+#include "spectrum.hpp"
+#include "tensor.hpp"
+#include "ucache.hpp"
+#include "util.h"
+
+namespace sd_sample {
+
+    enum class SampleCacheMode {
+        NONE,
+        EASYCACHE,
+        UCACHE,
+        CACHEDIT,
+    };
+
+    struct SampleCacheRuntime {
+        SampleCacheMode mode = SampleCacheMode::NONE;
+
+        EasyCacheState easycache;
+        UCacheState ucache;
+        CacheDitConditionState cachedit;
+        SpectrumState spectrum;
+
+        bool spectrum_enabled = false;
+
+        bool easycache_enabled() const;
+        bool ucache_enabled() const;
+        bool cachedit_enabled() const;
+    };
+
+    struct SampleStepCacheDispatcher {
+        SampleCacheRuntime& runtime;
+        int step;
+        float sigma;
+        int step_index;
+
+        SampleStepCacheDispatcher(SampleCacheRuntime& runtime, int step, float sigma);
+
+        bool before_condition(const void* condition, const sd::Tensor<float>& input, sd::Tensor<float>* output);
+        void after_condition(const void* condition, const sd::Tensor<float>& input, const sd::Tensor<float>& output);
+        bool is_step_skipped() const;
+    };
+
+    SampleCacheRuntime init_sample_cache_runtime(SDVersion version,
+                                                 const sd_cache_params_t* cache_params,
+                                                 Denoiser* denoiser,
+                                                 const std::vector<float>& sigmas);
+
+    void log_sample_cache_summary(const SampleCacheRuntime& runtime, size_t total_steps);
+
+}  // namespace sd_sample
+
+#endif  // __SAMPLE_CACHE_H__

src/spectrum.hpp

@@ -6,6 +6,7 @@
 #include <vector>
 
 #include "ggml_extend.hpp"
+#include "tensor.hpp"
 
 struct SpectrumConfig {
     float w            = 0.40f;
@@ -57,11 +58,8 @@ struct SpectrumState {
         return (num_cached + 1) % ws != 0;
     }
 
-    void update(const ggml_tensor* denoised) {
-        int64_t ne        = ggml_nelements(denoised);
-        const float* data = (const float*)denoised->data;
-
-        H_buf.emplace_back(data, data + ne);
+    void update(const sd::Tensor<float>& denoised) {
+        H_buf.emplace_back(denoised.data(), denoised.data() + denoised.numel());
         T_buf.push_back(taus(cnt));
 
         while ((int)H_buf.size() > K) {
@@ -76,13 +74,13 @@ struct SpectrumState {
         cnt++;
     }
 
-    void predict(ggml_tensor* denoised) {
+    void predict(sd::Tensor<float>* denoised) {
+        GGML_ASSERT(denoised != nullptr);
         int64_t F    = (int64_t)H_buf[0].size();
         int K_curr   = (int)H_buf.size();
         int M1       = config.m + 1;
         float tau_at = taus(cnt);
 
-        // Design matrix X: K_curr x M1 (Chebyshev basis)
         std::vector<float> X(K_curr * M1);
         for (int i = 0; i < K_curr; i++) {
             X[i * M1] = 1.0f;
@@ -92,7 +90,6 @@ struct SpectrumState {
                 X[i * M1 + j] = 2.0f * T_buf[i] * X[i * M1 + j - 1] - X[i * M1 + j - 2];
         }
 
-        // x_star: Chebyshev basis at current tau
         std::vector<float> x_star(M1);
         x_star[0] = 1.0f;
         if (M1 > 1)
@@ -100,7 +97,6 @@ struct SpectrumState {
         for (int j = 2; j < M1; j++)
             x_star[j] = 2.0f * tau_at * x_star[j - 1] - x_star[j - 2];
 
-        // XtX = X^T X + lambda I
         std::vector<float> XtX(M1 * M1, 0.0f);
         for (int i = 0; i < M1; i++) {
             for (int j = 0; j < M1; j++) {
@@ -111,7 +107,6 @@ struct SpectrumState {
             }
         }
 
-        // Cholesky decomposition
         std::vector<float> L(M1 * M1, 0.0f);
         if (!cholesky_decompose(XtX.data(), L.data(), M1)) {
             float trace = 0.0f;
@@ -122,18 +117,15 @@ struct SpectrumState {
             cholesky_decompose(XtX.data(), L.data(), M1);
         }
 
-        // Solve XtX v = x_star
         std::vector<float> v(M1);
         cholesky_solve(L.data(), x_star.data(), v.data(), M1);
 
-        // Prediction weights per history entry
         std::vector<float> weights(K_curr, 0.0f);
         for (int k = 0; k < K_curr; k++)
             for (int j = 0; j < M1; j++)
                 weights[k] += X[k * M1 + j] * v[j];
 
-        // Blend Chebyshev and Taylor predictions
-        float* out          = (float*)denoised->data;
+        float* out          = denoised->data();
         float w_cheb        = config.w;
         float w_taylor      = 1.0f - w_cheb;
         const float* h_last = H_buf.back().data();

src/tae.hpp

@@ -562,41 +562,40 @@ struct TinyImageAutoEncoder : public VAE {
         taesd.get_param_tensors(tensors, prefix);
     }
 
-    ggml_tensor* vae_output_to_latents(ggml_context* work_ctx, ggml_tensor* vae_output, std::shared_ptr<RNG> rng) {
+    sd::Tensor<float> vae_output_to_latents(const sd::Tensor<float>& vae_output, std::shared_ptr<RNG> rng) override {
+        SD_UNUSED(rng);
         return vae_output;
     }
 
-    ggml_tensor* diffusion_to_vae_latents(ggml_context* work_ctx, ggml_tensor* latents) {
-        return ggml_ext_dup_and_cpy_tensor(work_ctx, latents);
+    sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
+        return latents;
     }
 
-    ggml_tensor* vae_to_diffuison_latents(ggml_context* work_ctx, ggml_tensor* latents) {
-        return ggml_ext_dup_and_cpy_tensor(work_ctx, latents);
+    sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents) override {
+        return latents;
     }
 
     int get_encoder_output_channels(int input_channels) {
         return taesd.z_channels;
     }
 
-    ggml_cgraph* build_graph(ggml_tensor* z, bool decode_graph) {
+    ggml_cgraph* build_graph(const sd::Tensor<float>& z_tensor, bool decode_graph) {
         ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
-        z                = to_backend(z);
+        ggml_tensor* z   = make_input(z_tensor);
         auto runner_ctx  = get_context();
         ggml_tensor* out = decode_graph ? taesd.decode(&runner_ctx, z) : taesd.encode(&runner_ctx, z);
         ggml_build_forward_expand(gf, out);
         return gf;
     }
 
-    bool _compute(const int n_threads,
-                  ggml_tensor* z,
-                  bool decode_graph,
-                  ggml_tensor** output,
-                  ggml_context* output_ctx = nullptr) {
+    sd::Tensor<float> _compute(const int n_threads,
+                               const sd::Tensor<float>& z_tensor,
+                               bool decode_graph) override {
         auto get_graph = [&]() -> ggml_cgraph* {
-            return build_graph(z, decode_graph);
+            return build_graph(z_tensor, decode_graph);
         };
 
-        return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), z_tensor.dim());
     }
 };
 
@@ -625,42 +624,41 @@ struct TinyVideoAutoEncoder : public VAE {
         taehv.get_param_tensors(tensors, prefix);
     }
 
-    ggml_tensor* vae_output_to_latents(ggml_context* work_ctx, ggml_tensor* vae_output, std::shared_ptr<RNG> rng) {
+    sd::Tensor<float> vae_output_to_latents(const sd::Tensor<float>& vae_output, std::shared_ptr<RNG> rng) override {
+        SD_UNUSED(rng);
         return vae_output;
     }
 
-    ggml_tensor* diffusion_to_vae_latents(ggml_context* work_ctx, ggml_tensor* latents) {
-        return ggml_ext_dup_and_cpy_tensor(work_ctx, latents);
+    sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
+        return latents;
     }
 
-    ggml_tensor* vae_to_diffuison_latents(ggml_context* work_ctx, ggml_tensor* latents) {
-        return ggml_ext_dup_and_cpy_tensor(work_ctx, latents);
+    sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents) override {
+        return latents;
     }
 
     int get_encoder_output_channels(int input_channels) {
         return taehv.z_channels;
     }
 
-    ggml_cgraph* build_graph(ggml_tensor* z, bool decode_graph) {
+    ggml_cgraph* build_graph(const sd::Tensor<float>& z_tensor, bool decode_graph) {
         ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
-        z                = to_backend(z);
+        ggml_tensor* z   = make_input(z_tensor);
         auto runner_ctx  = get_context();
         ggml_tensor* out = decode_graph ? taehv.decode(&runner_ctx, z) : taehv.encode(&runner_ctx, z);
         ggml_build_forward_expand(gf, out);
         return gf;
     }
 
-    bool _compute(const int n_threads,
-                  ggml_tensor* z,
-                  bool decode_graph,
-                  ggml_tensor** output,
-                  ggml_context* output_ctx = nullptr) {
+    sd::Tensor<float> _compute(const int n_threads,
+                               const sd::Tensor<float>& z_tensor,
+                               bool decode_graph) override {
         auto get_graph = [&]() -> ggml_cgraph* {
-            return build_graph(z, decode_graph);
+            return build_graph(z_tensor, decode_graph);
         };
 
-        return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), z_tensor.dim());
     }
 };
 
-#endif  // __TAE_HPP__
+#endif  // __TAE_HPP__

src/tensor_ggml.hpp

@@ -0,0 +1,127 @@
+#ifndef __SD_TENSOR_GGML_HPP__
+#define __SD_TENSOR_GGML_HPP__
+
+#include <array>
+#include <cstring>
+#include <fstream>
+#include <stdexcept>
+#include <string>
+#include <type_traits>
+
+#include "ggml.h"
+#include "tensor.hpp"
+
+namespace sd {
+
+    template <typename T>
+    struct GGMLTypeTraits;
+
+    template <>
+    struct GGMLTypeTraits<float> {
+        static constexpr ggml_type type = GGML_TYPE_F32;
+    };
+
+    template <>
+    struct GGMLTypeTraits<ggml_fp16_t> {
+        static constexpr ggml_type type = GGML_TYPE_F16;
+    };
+
+    template <>
+    struct GGMLTypeTraits<int32_t> {
+        static constexpr ggml_type type = GGML_TYPE_I32;
+    };
+
+    template <>
+    struct GGMLTypeTraits<int64_t> {
+        static constexpr ggml_type type = GGML_TYPE_I64;
+    };
+
+    inline std::vector<int64_t> shape_from_ggml(const ggml_tensor* tensor) {
+        std::vector<int64_t> shape;
+        shape.reserve(static_cast<size_t>(ggml_n_dims(tensor)));
+        for (int i = 0; i < ggml_n_dims(tensor); ++i) {
+            shape.push_back(tensor->ne[i]);
+        }
+        return shape;
+    }
+
+    template <typename T>
+    inline Tensor<T> make_sd_tensor_from_ggml(const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return {};
+        }
+        if (tensor->type != GGMLTypeTraits<T>::type) {
+            GGML_ABORT("ggml tensor type does not match sd::Tensor type");
+        }
+        Tensor<T> result(shape_from_ggml(tensor));
+        if (tensor->buffer != nullptr) {
+            ggml_backend_tensor_get(tensor, result.data(), 0, ggml_nbytes(tensor));
+        } else {
+            std::memcpy(result.data(), tensor->data, ggml_nbytes(tensor));
+        }
+        return result;
+    }
+
+    template <typename T>
+    inline ggml_tensor* make_ggml_tensor(ggml_context* ctx, const Tensor<T>& tensor, bool copy_data = true) {
+        GGML_ASSERT(tensor.dim() > 0 && tensor.dim() <= 5);
+
+        int n_dims = std::min(static_cast<int>(tensor.dim()), GGML_MAX_DIMS);
+
+        std::array<int64_t, GGML_MAX_DIMS> ne = {1, 1, 1, 1};
+        for (int64_t i = 0; i < n_dims; ++i) {
+            ne[static_cast<size_t>(i)] = tensor.shape()[static_cast<size_t>(i)];
+        }
+
+        if (tensor.dim() == 5) {
+            ne[3] *= tensor.shape()[4];
+        }
+
+        ggml_tensor* result = ggml_new_tensor(ctx, GGMLTypeTraits<T>::type, n_dims, ne.data());
+        if (copy_data && tensor.numel() > 0) {
+            std::memcpy(result->data, tensor.data(), static_cast<size_t>(ggml_nbytes(result)));
+        }
+        return result;
+    }
+
+    template <typename T>
+    inline Tensor<T> load_tensor_from_file_as_tensor(const std::string& file_path) {
+        std::ifstream file(file_path, std::ios::binary);
+        if (!file.is_open()) {
+            throw std::runtime_error("failed to open tensor file: " + file_path);
+        }
+
+        int32_t n_dims = 0;
+        int32_t length = 0;
+        int32_t ttype  = 0;
+        file.read(reinterpret_cast<char*>(&n_dims), sizeof(n_dims));
+        file.read(reinterpret_cast<char*>(&length), sizeof(length));
+        file.read(reinterpret_cast<char*>(&ttype), sizeof(ttype));
+        if (!file.good()) {
+            throw std::runtime_error("incomplete tensor file header: " + file_path);
+        }
+        if (static_cast<ggml_type>(ttype) != GGMLTypeTraits<T>::type) {
+            throw std::invalid_argument("tensor file type does not match requested sd::Tensor type");
+        }
+
+        std::vector<int64_t> shape(4, 1);
+        for (int i = 0; i < n_dims; ++i) {
+            int32_t dim = 1;
+            file.read(reinterpret_cast<char*>(&dim), sizeof(dim));
+            shape[static_cast<size_t>(i)] = dim;
+        }
+        std::string name(static_cast<size_t>(length), '\0');
+        file.read(name.data(), length);
+
+        shape.resize(static_cast<size_t>(n_dims));
+        Tensor<T> tensor(shape);
+        file.read(reinterpret_cast<char*>(tensor.data()), static_cast<std::streamsize>(tensor.numel() * sizeof(T)));
+        if (!file.good()) {
+            throw std::runtime_error("incomplete tensor file data: " + file_path);
+        }
+        return tensor;
+    }
+
+}  // namespace sd
+
+#endif

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: %zu", 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 %zu", 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__