refactor: simplify DiscreteFlowDenoiser (#1405 )

feat: add flux2 small decoder support (#1402 )
fix: correct double increment on flow denoisers sigma calculations (#1372 )
2026-10-02 10:24:37 +08:00 · 2026-04-11 17:18:23 +08:00 · 2026-04-08 23:13:25 +08:00 · 2026-04-08 23:13:05 +08:00 · 2026-04-06 21:25:34 +08:00 · 2026-04-06 21:11:57 +08:00
107 changed files with 21686 additions and 14792 deletions
@@ -21,11 +21,13 @@ on:
        "**/*.c",
        "**/*.cpp",
        "**/*.cu",
+        "examples/server/frontend/**",
      ]
  pull_request:
    types: [opened, synchronize, reopened]
    paths:
      [
+        ".github/workflows/**",
        "**/CMakeLists.txt",
        "**/Makefile",
        "**/*.h",
@@ -33,6 +35,7 @@ on:
        "**/*.c",
        "**/*.cpp",
        "**/*.cu",
+        "examples/server/frontend/**",
      ]

 env:
@@ -53,6 +56,16 @@ jobs:
        with:
          submodules: recursive

+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
      - name: Dependencies
        id: depends
        run: |
@@ -70,7 +83,7 @@ jobs:
      - name: Get commit hash
        id: commit
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2

      - name: Fetch system info
        id: system-info
@@ -106,6 +119,16 @@ jobs:
        with:
          submodules: recursive

+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
      - name: Dependencies
        id: depends
        run: |
@@ -123,7 +146,7 @@ jobs:
      - name: Get commit hash
        id: commit
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2

      - name: Fetch system info
        id: system-info
@@ -162,7 +185,7 @@ jobs:

    strategy:
      matrix:
-        variant: [musa, sycl, vulkan]
+        variant: [musa, sycl, vulkan, cuda]

    env:
      REGISTRY: ghcr.io
@@ -174,10 +197,20 @@ jobs:
        with:
          submodules: recursive

+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
      - name: Get commit hash
        id: commit
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2

      - name: Set up Docker Buildx
        uses: docker/setup-buildx-action@v3
@@ -206,6 +239,7 @@ 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 }}
@@ -223,6 +257,16 @@ jobs:
        with:
          submodules: recursive

+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
      - name: Dependencies
        id: depends
        run: |
@@ -240,7 +284,7 @@ jobs:
      - name: Get commit hash
        id: commit
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2

      - name: Fetch system info
        id: system-info
@@ -294,6 +338,16 @@ jobs:
        with:
          submodules: recursive

+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
      - name: Install cuda-toolkit
        id: cuda-toolkit
        if: ${{ matrix.build == 'cuda12' }}
@@ -340,7 +394,7 @@ jobs:
      - name: Get commit hash
        id: commit
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2

      - name: Pack artifacts
        id: pack_artifacts
@@ -399,6 +453,16 @@ jobs:
        with:
          submodules: recursive

+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
      - name: Cache ROCm Installation
        id: cache-rocm
        uses: actions/cache@v4
@@ -463,7 +527,7 @@ jobs:
      - name: Get commit hash
        id: commit
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2

      - name: Pack artifacts
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -485,6 +549,156 @@ jobs:
          path: |
            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-x64.zip

+  ubuntu-latest-rocm:
+    runs-on: ubuntu-latest
+    container: rocm/dev-ubuntu-24.04:7.2
+
+    env:
+      ROCM_VERSION: "7.2"
+      UBUNTU_VERSION: "24.04"
+      GPU_TARGETS: "gfx1151;gfx1150;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
+
+    steps:
+      - run: apt-get update && apt-get install -y git
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          submodules: recursive
+
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
+      - name: Free disk space
+        run: |
+          # Remove preinstalled SDKs and caches not needed for this job
+          sudo rm -rf /usr/share/dotnet || true
+          sudo rm -rf /usr/local/lib/android || true
+          sudo rm -rf /opt/ghc || true
+          sudo rm -rf /usr/local/.ghcup || true
+          sudo rm -rf /opt/hostedtoolcache || true
+
+          # Remove old package lists and caches
+          sudo rm -rf /var/lib/apt/lists/* || true
+          sudo apt clean
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt install -y \
+            cmake \
+            hip-dev \
+            hipblas-dev \
+            ninja-build \
+            rocm-dev \
+            zip
+          # Clean apt caches to recover disk space
+          sudo apt clean
+          sudo rm -rf /var/lib/apt/lists/* || true
+
+      - name: Setup ROCm Environment
+        run: |
+          # Add ROCm to PATH for current session
+          echo "/opt/rocm/bin" >> $GITHUB_PATH
+
+          # Build regex pattern from ${{ env.GPU_TARGETS }} (match target as substring)
+          TARGET_REGEX="($(printf '%s' "${{ env.GPU_TARGETS }}" | sed 's/;/|/g'))"
+
+          # Remove library files for architectures we're not building for to save disk space
+          echo "Cleaning up unneeded architecture files..."
+          cd /opt/rocm/lib/rocblas/library
+          # Keep only our target architectures
+          for file in *; do
+            if printf '%s' "$file" | grep -q 'gfx'; then
+              if ! printf '%s' "$file" | grep -Eq "$TARGET_REGEX"; then
+                echo "Removing $file" &&
+                sudo rm -f "$file";
+              fi
+            fi
+          done
+
+          cd /opt/rocm/lib/hipblaslt/library
+          for file in *; do
+            if printf '%s' "$file" | grep -q 'gfx'; then
+              if ! printf '%s' "$file" | grep -Eq "$TARGET_REGEX"; then
+                echo "Removing $file" &&
+                sudo rm -f "$file";
+              fi
+            fi
+          done
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -G Ninja \
+            -DCMAKE_CXX_COMPILER=amdclang++ \
+            -DCMAKE_C_COMPILER=amdclang \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DSD_HIPBLAS=ON \
+            -DGPU_TARGETS="${{ env.GPU_TARGETS }}" \
+            -DAMDGPU_TARGETS="${{ env.GPU_TARGETS }}" \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+            -DSD_BUILD_SHARED_LIBS=ON
+          cmake --build . --config Release
+
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: prompt/actions-commit-hash@v2
+
+      - name: Prepare artifacts
+        id: prepare_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          # Copy licenses
+          cp ggml/LICENSE ./build/bin/ggml.txt
+          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
+
+          # Move ROCm runtime libraries (to avoid double space consumption)
+          sudo mv /opt/rocm/lib/librocsparse.so* ./build/bin/
+          sudo mv /opt/rocm/lib/libhsa-runtime64.so* ./build/bin/
+          sudo mv /opt/rocm/lib/libamdhip64.so* ./build/bin/
+          sudo mv /opt/rocm/lib/libhipblas.so* ./build/bin/
+          sudo mv /opt/rocm/lib/libhipblaslt.so* ./build/bin/
+          sudo mv /opt/rocm/lib/librocblas.so* ./build/bin/
+          sudo mv /opt/rocm/lib/rocblas/ ./build/bin/
+          sudo mv /opt/rocm/lib/hipblaslt/ ./build/bin/
+
+      - name: Fetch system info
+        id: system-info
+        run: |
+          echo "CPU_ARCH=`uname -m`" >> "$GITHUB_OUTPUT"
+          echo "OS_NAME=`lsb_release -s -i`" >> "$GITHUB_OUTPUT"
+          echo "OS_VERSION=`lsb_release -s -r`" >> "$GITHUB_OUTPUT"
+          echo "OS_TYPE=`uname -s`" >> "$GITHUB_OUTPUT"
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp ggml/LICENSE ./build/bin/ggml.txt
+          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
+          zip -y -r sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm.zip ./build/bin
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm.zip
+          path: |
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm.zip
+
  release:
    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}

@@ -493,6 +707,7 @@ jobs:
    needs:
      - ubuntu-latest-cmake
      - ubuntu-latest-cmake-vulkan
+      - ubuntu-latest-rocm
      - build-and-push-docker-images
      - macOS-latest-cmake
      - windows-latest-cmake
@@ -519,7 +734,7 @@ jobs:

      - name: Get commit hash
        id: commit
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2

      - name: Create release
        id: create_release
@@ -1,3 +1,12 @@
 [submodule "ggml"]
    path = ggml
 	url = https://github.com/ggml-org/ggml.git
+[submodule "examples/server/frontend"]
+	path = examples/server/frontend
+	url = https://github.com/leejet/stable-ui.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
@@ -22,6 +22,26 @@ 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 +49,10 @@ 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)
@@ -36,7 +60,6 @@ option(SD_VULKAN                     "sd: vulkan backend" OFF)
 option(SD_OPENCL                     "sd: opencl backend" OFF)
 option(SD_SYCL                       "sd: sycl backend" OFF)
 option(SD_MUSA                       "sd: musa backend" OFF)
-option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
 option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
 option(SD_BUILD_SHARED_GGML_LIB      "sd: build ggml as a separate shared lib" OFF)
 option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
@@ -70,26 +93,71 @@ if (SD_HIPBLAS)
    message("-- Use HIPBLAS as backend stable-diffusion")
    set(GGML_HIP ON)
    add_definitions(-DSD_USE_CUDA)
-    if(SD_FAST_SOFTMAX)
-        set(GGML_CUDA_FAST_SOFTMAX ON)
-    endif()
 endif ()

 if(SD_MUSA)
    message("-- Use MUSA as backend stable-diffusion")
    set(GGML_MUSA ON)
    add_definitions(-DSD_USE_CUDA)
-    if(SD_FAST_SOFTMAX)
-        set(GGML_CUDA_FAST_SOFTMAX ON)
+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
-    "*.h"
-    "*.cpp"
-    "*.hpp"
+    "src/*.h"
+    "src/*.cpp"
+    "src/*.hpp"
+    "src/vocab/*.h"
+    "src/vocab/*.cpp"
 )

 find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
@@ -119,7 +187,7 @@ endif()
 message(STATUS "stable-diffusion.cpp commit ${SDCPP_BUILD_COMMIT}")

 set_property(
-  SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/version.cpp
+  SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/src/version.cpp
  APPEND PROPERTY COMPILE_DEFINITIONS
  SDCPP_BUILD_COMMIT=${SDCPP_BUILD_COMMIT} SDCPP_BUILD_VERSION=${SDCPP_BUILD_VERSION}
 )
@@ -182,6 +250,7 @@ endif()
 add_subdirectory(thirdparty)

 target_link_libraries(${SD_LIB} PUBLIC ggml zip)
+target_include_directories(${SD_LIB} PUBLIC . include)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
 target_compile_features(${SD_LIB} PUBLIC c_std_11 cxx_std_17)

@@ -190,7 +259,7 @@ if (SD_BUILD_EXAMPLES)
    add_subdirectory(examples)
 endif()

-set(SD_PUBLIC_HEADERS stable-diffusion.h)
+set(SD_PUBLIC_HEADERS include/stable-diffusion.h)
 set_target_properties(${SD_LIB} PROPERTIES PUBLIC_HEADER "${SD_PUBLIC_HEADERS}")

 install(TARGETS ${SD_LIB} LIBRARY PUBLIC_HEADER)
@@ -0,0 +1,25 @@
+ARG CUDA_VERSION=12.6.3
+ARG UBUNTU_VERSION=24.04
+
+FROM nvidia/cuda:${CUDA_VERSION}-cudnn-devel-ubuntu${UBUNTU_VERSION} AS build
+
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git ccache cmake
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+ARG CUDACXX=/usr/local/cuda/bin/nvcc
+RUN cmake . -B ./build -DSD_CUDA=ON
+RUN cmake --build ./build --config Release -j$(nproc)
+
+FROM nvidia/cuda:${CUDA_VERSION}-cudnn-runtime-ubuntu${UBUNTU_VERSION} AS runtime
+
+RUN apt-get update && \
+    apt-get install --yes --no-install-recommends libgomp1 && \
+    apt-get clean
+
+COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
+COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
+
+ENTRYPOINT [ "/sd-cli" ]
@@ -15,6 +15,9 @@ API and command-line option may change frequently.***

 ## 🔥Important News

+* **2026/01/18** 🚀 stable-diffusion.cpp now supports **FLUX.2-klein**  
+  👉 Details: [PR #1193](https://github.com/leejet/stable-diffusion.cpp/pull/1193)
+
 * **2025/12/01** 🚀 stable-diffusion.cpp now supports **Z-Image**  
  👉 Details: [PR #1020](https://github.com/leejet/stable-diffusion.cpp/pull/1020)

@@ -43,13 +46,14 @@ API and command-line option may change frequently.***
    - SDXL, [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo)
    - [Some SD1.x and SDXL distilled models](./docs/distilled_sd.md)
    - [SD3/SD3.5](./docs/sd3.md)
-    - [FlUX.1-dev/FlUX.1-schnell](./docs/flux.md)
-    - [FLUX.2-dev](./docs/flux2.md)
+    - [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
+    - [FLUX.2-dev/FLUX.2-klein](./docs/flux2.md)
    - [Chroma](./docs/chroma.md)
    - [Chroma1-Radiance](./docs/chroma_radiance.md)
    - [Qwen Image](./docs/qwen_image.md)
    - [Z-Image](./docs/z_image.md)
    - [Ovis-Image](./docs/ovis_image.md)
+    - [Anima](./docs/anima.md)
  - Image Edit Models
    - [FLUX.1-Kontext-dev](./docs/kontext.md)
    - [Qwen Image Edit series](./docs/qwen_image_edit.md)
@@ -127,8 +131,8 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe

 - [SD1.x/SD2.x/SDXL](./docs/sd.md)
 - [SD3/SD3.5](./docs/sd3.md)
- [FlUX.1-dev/FlUX.1-schnell](./docs/flux.md)
- [FLUX.2-dev](./docs/flux2.md)
+- [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
+- [FLUX.2-dev/FLUX.2-klein](./docs/flux2.md)
 - [FLUX.1-Kontext-dev](./docs/kontext.md)
 - [Chroma](./docs/chroma.md)
 - [🔥Qwen Image](./docs/qwen_image.md)
@@ -136,6 +140,7 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 - [🔥Wan2.1/Wan2.2](./docs/wan.md)
 - [🔥Z-Image](./docs/z_image.md)
 - [Ovis-Image](./docs/ovis_image.md)
+- [Anima](./docs/anima.md)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
 - [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)
@@ -0,0 +1,21 @@
+# How to Use
+
+## Download weights
+
+- Download Anima
+    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/Bedovyy/Anima-GGUF/tree/main
+    - gguf Anima2: https://huggingface.co/JusteLeo/Anima2-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae
+- Download Qwen3-0.6B-Base
+    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/mradermacher/Qwen3-0.6B-Base-GGUF/tree/main
+
+## Examples
+
+```sh
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\anima-preview.safetensors --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_06b_base.safetensors  -p "a lovely cat holding a sign says 'anima.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="anima image example" src="../assets/anima/example.png" />
@@ -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.
@@ -11,6 +11,7 @@ Caching methods accelerate diffusion inference by reusing intermediate computati
 | `dbcache` | DiT models | Block-level L1 residual threshold |
 | `taylorseer` | DiT models | Taylor series approximation |
 | `cache-dit` | DiT models | Combined DBCache + TaylorSeer |
+| `spectrum` | UNET and DiT models | Chebyshev + Taylor output forecasting |

 ### UCache (UNET Models)

@@ -79,7 +80,7 @@ Uses Taylor series approximation to predict block outputs:
 Combines DBCache and TaylorSeer:

 ```bash
--cache-mode cache-dit --cache-preset fast
+--cache-mode cache-dit
 ```

 #### Parameters
@@ -91,14 +92,6 @@ Combines DBCache and TaylorSeer:
 | `threshold` | L1 residual difference threshold | 0.08 |
 | `warmup` | Steps before caching starts | 8 |

-#### Presets
-
-Available presets: `slow`, `medium`, `fast`, `ultra` (or `s`, `m`, `f`, `u`).
-
-```bash
--cache-mode cache-dit --cache-preset fast
-```
-
 #### SCM Options

 Steps Computation Mask controls which steps can be cached:
@@ -118,6 +111,28 @@ Mask values: `1` = compute, `0` = can cache.
 --scm-policy dynamic
 ```

+### Spectrum (UNET and DiT Models)
+
+Spectrum uses Chebyshev polynomial fitting blended with Taylor extrapolation to predict denoised outputs, skipping entire forward passes. Based on the paper [Spectrum: Adaptive Spectral Feature Forecasting for Efficient Diffusion Sampling](https://github.com/tingyu215/Spectrum).
+
+```bash
+sd-cli -m model.safetensors -p "a cat" --cache-mode spectrum
+```
+
+#### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `w` | Chebyshev vs Taylor blend weight (0=Taylor, 1=Chebyshev) | 0.40 |
+| `m` | Chebyshev polynomial degree | 3 |
+| `lam` | Ridge regression regularization | 1.0 |
+| `window` | Initial window size (compute every N steps) | 2 |
+| `flex` | Window growth per computed step after warmup | 0.50 |
+| `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
@@ -1,8 +1,8 @@
-# Running distilled models: SSD1B and SDx.x with tiny U-Nets
+# Running distilled models: SSD1B, Vega and SDx.x with tiny U-Nets

 ## Preface 

-These models feature a reduced U-Net architecture. Unlike standard SDXL models, the SSD-1B U-Net contains only one middle block and fewer attention layers in its up- and down-blocks, resulting in significantly smaller file sizes. Using these models can reduce inference time by more than 33%. For more details, refer to Segmind's paper: https://arxiv.org/abs/2401.02677v1.
+These models feature a reduced U-Net architecture. Unlike standard SDXL models, the SSD-1B and Vega U-Net contains only one middle block and fewer attention layers in its up- and down-blocks, resulting in significantly smaller file sizes. Using these models can reduce inference time by more than 33%. For more details, refer to Segmind's paper: https://arxiv.org/abs/2401.02677v1.
 Similarly, SD1.x- and SD2.x-style models with a tiny U-Net consist of only 6 U-Net blocks, leading to very small files and time savings of up to 50%. For more information, see the paper: https://arxiv.org/pdf/2305.15798.pdf.

 ## SSD1B
@@ -17,7 +17,17 @@ Useful LoRAs are also available:
 * https://huggingface.co/seungminh/lora-swarovski-SSD-1B/resolve/main/pytorch_lora_weights.safetensors
 * https://huggingface.co/kylielee505/mylcmlorassd/resolve/main/pytorch_lora_weights.safetensors

-These files can be used out-of-the-box, unlike the models described in the next section.
+## Vega
+
+Segmind's Vega model is available online here:
+
+ * https://huggingface.co/segmind/Segmind-Vega/resolve/main/segmind-vega.safetensors
+ 
+VegaRT is an example for an LCM-LoRA:
+
+ * https://huggingface.co/segmind/Segmind-VegaRT/resolve/main/pytorch_lora_weights.safetensors
+
+Both files can be used out-of-the-box, unlike the models described in next sections.


 ## SD1.x, SD2.x with tiny U-Nets
@@ -83,7 +93,7 @@ python convert_diffusers_to_original_stable_diffusion.py \
 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.


-### Another available .ckpt file:
+##### Another available .ckpt file:

 * https://huggingface.co/ClashSAN/small-sd/resolve/main/tinySDdistilled.ckpt

@@ -97,3 +107,31 @@ for key, value in ckpt['state_dict'].items():
        ckpt['state_dict'][key] = value.contiguous()
 torch.save(ckpt, "tinySDdistilled_fixed.ckpt")
 ```
+
+
+### SDXS-512
+
+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.
+
+##### 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:
+
+```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.                                                 
@@ -1,6 +1,6 @@
 ## Using ESRGAN to upscale results

-You can use ESRGAN to upscale the generated images. At the moment, only the [RealESRGAN_x4plus_anime_6B.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth) model is supported. Support for more models of this architecture will be added soon.
+You can use ESRGAN—such as the model [RealESRGAN_x4plus_anime_6B.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth)—to upscale the generated images and improve their overall resolution and clarity.

 - Specify the model path using the `--upscale-model PATH` parameter. example:

@@ -1,15 +1,19 @@
 # How to Use

-## Download weights
+## Flux.2-dev
+
+### Download weights

 - Download FLUX.2-dev
    - 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

-## Examples
+### Examples

 ```
 .\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux2-dev-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf -r .\kontext_input.png -p "change 'flux.cpp' to 'flux2-dev.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu
@@ -17,5 +21,76 @@

 <img alt="flux2 example" src="../assets/flux2/example.png" />

+## Flux.2 klein 4B / Flux.2 klein base 4B

+### Download weights

+- Download FLUX.2-klein-4B
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-4B
+    - gguf: https://huggingface.co/leejet/FLUX.2-klein-4B-GGUF/tree/main
+- Download FLUX.2-klein-base-4B
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B
+    - 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
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="flux2-klein-4b" src="../assets/flux2/flux2-klein-4b.png" />
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
+```
+
+<img alt="flux2-klein-4b-edit" src="../assets/flux2/flux2-klein-4b-edit.png" />
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="flux2-klein-base-4b" src="../assets/flux2/flux2-klein-base-4b.png" />
+
+## Flux.2 klein 9B / Flux.2 klein base 9B
+
+### Download weights
+
+- Download FLUX.2-klein-9B
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-9B
+    - gguf: https://huggingface.co/leejet/FLUX.2-klein-9B-GGUF/tree/main
+- Download FLUX.2-klein-base-9B
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B
+    - gguf: https://huggingface.co/leejet/FLUX.2-klein-base-9B-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download Qwen3 8B
+    - safetensors: https://huggingface.co/Comfy-Org/flux2-klein-9B/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/unsloth/Qwen3-8B-GGUF/tree/main
+
+### Examples
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 1.0 --steps 4 -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="flux2-klein-9b" src="../assets/flux2/flux2-klein-9b.png" />
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -r .\kontext_input.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu --steps 4
+```
+
+<img alt="flux2-klein-9b-edit" src="../assets/flux2/flux2-klein-9b-edit.png" />
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-9b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_8b.safetensors -p "a lovely cat" --cfg-scale 4.0 --steps 20 -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="flux2-klein-base-9b" src="../assets/flux2/flux2-klein-base-9b.png" />
@@ -7,6 +7,9 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev
 - Download Z-Image-Turbo
    - safetensors: https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/diffusion_models
    - gguf: https://huggingface.co/leejet/Z-Image-Turbo-GGUF/tree/main
+- Download Z-Image
+    - safetensors: https://huggingface.co/Comfy-Org/z_image/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/unsloth/Z-Image-GGUF/tree/main
 - Download vae
    - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main
 - Download Qwen3 4b
@@ -15,12 +18,22 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev

 ## Examples

+### Z-Image-Turbo
+
 ```
 .\bin\Release\sd-cli.exe --diffusion-model  z_image_turbo-Q3_K.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
 ```

 <img width="256" alt="z-image example" src="../assets/z_image/q3_K.png" />

+### Z-Image-Base
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\z_image_bf16.safetensors --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
+```
+
+<img width="256" alt="z-image example" src="../assets/z_image/base_bf16.png" />
+
 ## Comparison of Different Quantization Types

 | bf16 | q8_0 | q6_K | q5_0 | q4_K | q4_0 | q3_K | q2_K|
@@ -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)
@@ -4,18 +4,24 @@
 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)
-  --output-begin-idx <int>    starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)
-  --preview-path <string>     path to write preview image to (default: ./preview.png)
+  -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). For video generation, single-file outputs support .avi, .webm, and animated .webp
+  --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)
+  --image <string>            path to the image to inspect (for metadata mode)
+  --metadata-format <string>  metadata output format, one of [text, json] (default: text)
  --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
+  --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

@@ -44,7 +50,6 @@ Context Options:
                                           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)
-  --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
  --vae-tiling                             process vae in tiles to reduce memory usage
  --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
@@ -52,13 +57,15 @@ Context Options:
  --control-net-cpu                        keep controlnet in cpu (for low vram)
  --clip-on-cpu                            keep clip in cpu (for low vram)
  --vae-on-cpu                             keep vae in cpu (for low vram)
-  --diffusion-fa                           use flash attention in the diffusion model
+  --fa                                     use flash attention
+  --diffusion-fa                           use flash attention in the diffusion model only
  --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
  --vae-conv-direct                        use ggml_conv2d_direct in the vae model
  --circular                               enable circular padding for convolutions
  --circularx                              enable circular RoPE wrapping on x-axis (width) only
  --circulary                              enable circular RoPE wrapping on y-axis (height) only
  --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
@@ -107,14 +114,15 @@ Generation Options:
                                           medium
  --skip-layer-start <float>               SLG enabling point (default: 0.01)
  --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            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 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-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 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
@@ -122,22 +130,35 @@ Generation Options:
  --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).
  --disable-auto-resize-ref-image          disable auto resize of ref images
+  --disable-image-metadata                 do not embed generation metadata on image files
  -s, --seed                               RNG seed (default: 42, use random seed for < 0)
  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           tcd] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+                                           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] default: euler for Flux/SD3/Wan, euler_a otherwise
+                                           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], default: discrete
+                                           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)
+  --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"
-  --cache-preset                           cache-dit preset: 'slow'/'s', 'medium'/'m', 'fast'/'f', 'ultra'/'u'
+                                           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
  --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
 ```
+
+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
+```
@@ -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__
@@ -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);
@@ -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",
@@ -165,7 +193,7 @@ struct SDCliParams {
    };

    bool process_and_check() {
-        if (output_path.length() == 0) {
+        if (mode != METADATA && output_path.length() == 0) {
            LOG_ERROR("error: the following arguments are required: output_path");
            return false;
        }
@@ -174,6 +202,16 @@ struct SDCliParams {
            if (output_path == "output.png") {
                output_path = "output.gguf";
            }
+        } else if (mode == METADATA) {
+            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;
    }
@@ -183,6 +221,8 @@ struct SDCliParams {
        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 +232,10 @@ 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 +260,25 @@ 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.process_and_check();
+    if (valid && cli_params.mode != METADATA) {
+        valid = ctx_params.process_and_check(cli_params.mode) &&
+                gen_params.process_and_check(cli_params.mode, ctx_params.lora_model_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.width) + "x" + std::to_string(gen_params.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,
+                          SDImageVec& images,
                          int expected_width  = 0,
                          int expected_height = 0,
                          int max_image_num   = 0,
@@ -315,7 +305,7 @@ 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;
@@ -330,7 +320,7 @@ bool load_images_from_dir(const std::string dir,
                              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 +335,17 @@ 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());
+        }
    }
 }

@@ -394,12 +392,19 @@ bool save_results(const SDCliParams& cli_params,

    fs::path base_path = out_path;
    fs::path ext       = out_path.has_extension() ? out_path.extension() : fs::path{};
-    if (!ext.empty())
-        base_path.replace_extension();

    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 (output_format == EncodedImageFormat::JPEG ||
+            output_format == EncodedImageFormat::PNG ||
+            output_format == EncodedImageFormat::WEBP ||
+            ext_lower == ".avi" ||
+            ext_lower == ".webm") {
+            base_path.replace_extension();
+        }
+    }

    int output_begin_idx = cli_params.output_begin_idx;
    if (output_begin_idx < 0) {
@@ -409,42 +414,49 @@ bool save_results(const SDCliParams& cli_params,
    auto write_image = [&](const fs::path& path, int idx) {
        const sd_image_t& img = results[idx];
        if (!img.data)
-            return;
+            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());
-        }
+        std::string params = gen_params.embed_image_metadata
+                                 ? get_image_params(ctx_params, gen_params, gen_params.seed + idx)
+                                 : "";
+        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;
    };

+    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;

        for (int i = 0; i < num_results; ++i) {
            fs::path img_path = format_frame_idx(pattern.string(), output_begin_idx + i);
-            write_image(img_path, i);
+            if (write_image(img_path, i)) {
+                sucessful_reults++;
+            }
        }
-        return true;
+        LOG_INFO("%d/%d images saved", sucessful_reults, num_results);
+        return sucessful_reults != 0;
    }

    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;
-        create_mjpg_avi_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps);
-        LOG_INFO("save result MJPG AVI video to '%s'", video_path.string().c_str());
-        return true;
+        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 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) {
@@ -453,10 +465,12 @@ bool save_results(const SDCliParams& cli_params,
            img_path += "_" + std::to_string(output_begin_idx + i);
        }
        img_path += ext;
-        write_image(img_path, i);
+        if (write_image(img_path, i)) {
+            sucessful_reults++;
+        }
    }
-
-    return true;
+    LOG_INFO("%d/%d images saved", sucessful_reults, num_results);
+    return sucessful_reults != 0;
 }

 int main(int argc, const char* argv[]) {
@@ -470,6 +484,27 @@ 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;
@@ -487,9 +522,6 @@ 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,
@@ -525,94 +557,92 @@ int main(int argc, const char* argv[]) {
        }
    }

-    bool vae_decode_only     = true;
-    sd_image_t init_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
-    sd_image_t end_image     = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
-    sd_image_t control_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
-    sd_image_t mask_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 1, nullptr};
-    std::vector<sd_image_t> ref_images;
-    std::vector<sd_image_t> pmid_images;
-    std::vector<sd_image_t> control_frames;
+    bool vae_decode_only = true;
+    SDImageOwner init_image({0, 0, 3, nullptr});
+    SDImageOwner end_image({0, 0, 3, nullptr});
+    SDImageOwner control_image({0, 0, 3, nullptr});
+    SDImageOwner mask_image({0, 0, 1, nullptr});
+    SDImageVec ref_images;
+    SDImageVec pmid_images;
+    SDImageVec 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;
+    auto load_image_and_update_size = [&](const std::string& path,
+                                          SDImageOwner& image,
+                                          bool resize_image    = true,
+                                          int expected_channel = 3) -> bool {
+        int expected_width  = 0;
+        int expected_height = 0;
+        if (resize_image && gen_params.width_and_height_are_set()) {
+            expected_width  = gen_params.width;
+            expected_height = gen_params.height;
        }
-        ref_images.clear();
-        for (auto image : pmid_images) {
-            free(image.data);
-            image.data = nullptr;
+
+        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());
+            return false;
        }
-        pmid_images.clear();
-        for (auto image : control_frames) {
-            free(image.data);
-            image.data = nullptr;
-        }
-        control_frames.clear();
+
+        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;
-
-        int width       = 0;
-        int height      = 0;
-        init_image.data = load_image_from_file(gen_params.init_image_path.c_str(), width, height, gen_params.width, gen_params.height);
-        if (init_image.data == nullptr) {
-            LOG_ERROR("load image from '%s' failed", gen_params.init_image_path.c_str());
-            release_all_resources();
+        if (!load_image_and_update_size(gen_params.init_image_path, init_image)) {
            return 1;
        }
    }

    if (gen_params.end_image_path.size() > 0) {
        vae_decode_only = false;
-
-        int width      = 0;
-        int height     = 0;
-        end_image.data = load_image_from_file(gen_params.end_image_path.c_str(), width, height, gen_params.width, gen_params.height);
-        if (end_image.data == nullptr) {
-            LOG_ERROR("load image from '%s' failed", gen_params.end_image_path.c_str());
-            release_all_resources();
+        if (!load_image_and_update_size(gen_params.end_image_path, end_image)) {
            return 1;
        }
    }

+    if (gen_params.ref_image_paths.size() > 0) {
+        vae_decode_only = false;
+        for (auto& path : gen_params.ref_image_paths) {
+            SDImageOwner ref_image({0, 0, 3, nullptr});
+            if (!load_image_and_update_size(path, ref_image, false)) {
+                return 1;
+            }
+            ref_images.push_back(std::move(ref_image));
+        }
+    }
+
    if (gen_params.mask_image_path.size() > 0) {
-        int c           = 0;
-        int width       = 0;
-        int height      = 0;
-        mask_image.data = load_image_from_file(gen_params.mask_image_path.c_str(), width, height, gen_params.width, gen_params.height, 1);
-        if (mask_image.data == nullptr) {
+        if (!load_sd_image_from_file(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.width * gen_params.height);
-        memset(mask_image.data, 255, gen_params.width * gen_params.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;
        }
+        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());
+        mask_image.reset(generated_mask);
    }

    if (gen_params.control_image_path.size() > 0) {
-        int width          = 0;
-        int height         = 0;
-        control_image.data = load_image_from_file(gen_params.control_image_path.c_str(), width, height, gen_params.width, gen_params.height);
-        if (control_image.data == nullptr) {
+        if (!load_sd_image_from_file(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(control_image.get(),
                             0.08f,
                             0.08f,
                             0.8f,
@@ -621,32 +651,13 @@ int main(int argc, const char* argv[]) {
        }
    }

-    if (gen_params.ref_image_paths.size() > 0) {
-        vae_decode_only = false;
-        for (auto& path : gen_params.ref_image_paths) {
-            int width             = 0;
-            int height            = 0;
-            uint8_t* image_buffer = load_image_from_file(path.c_str(), width, height);
-            if (image_buffer == nullptr) {
-                LOG_ERROR("load image from '%s' failed", path.c_str());
-                release_all_resources();
-                return 1;
-            }
-            ref_images.push_back({(uint32_t)width,
-                                  (uint32_t)height,
-                                  3,
-                                  image_buffer});
-        }
-    }
-
    if (!gen_params.control_video_path.empty()) {
        if (!load_images_from_dir(gen_params.control_video_path,
                                  control_frames,
-                                  gen_params.width,
-                                  gen_params.height,
+                                  gen_params.get_resolved_width(),
+                                  gen_params.get_resolved_height(),
                                  gen_params.video_frames,
                                  cli_params.verbose)) {
-            release_all_resources();
            return 1;
        }
    }
@@ -658,7 +669,6 @@ int main(int argc, const char* argv[]) {
                                  0,
                                  0,
                                  cli_params.verbose)) {
-            release_all_resources();
            return 1;
        }
    }
@@ -669,39 +679,30 @@ int main(int argc, const char* argv[]) {

    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(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) {
@@ -711,19 +712,19 @@ int main(int argc, const char* argv[]) {
                gen_params.prompt.c_str(),
                gen_params.negative_prompt.c_str(),
                gen_params.clip_skip,
-                init_image,
+                init_image.get(),
                ref_images.data(),
                (int)ref_images.size(),
                gen_params.auto_resize_ref_image,
                gen_params.increase_ref_index,
-                mask_image,
-                gen_params.width,
-                gen_params.height,
+                mask_image.get(),
+                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,
+                control_image.get(),
                gen_params.control_strength,
                {
                    pmid_images.data(),
@@ -731,12 +732,12 @@ int main(int argc, const char* argv[]) {
                    gen_params.pm_id_embed_path.c_str(),
                    gen_params.pm_style_strength,
                },  // pm_params
-                ctx_params.vae_tiling_params,
+                gen_params.vae_tiling_params,
                gen_params.cache_params,
            };

-            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(),
@@ -744,12 +745,12 @@ int main(int argc, const char* argv[]) {
                gen_params.prompt.c_str(),
                gen_params.negative_prompt.c_str(),
                gen_params.clip_skip,
-                init_image,
-                end_image,
+                init_image.get(),
+                end_image.get(),
                control_frames.data(),
                (int)control_frames.size(),
-                gen_params.width,
-                gen_params.height,
+                gen_params.get_resolved_width(),
+                gen_params.get_resolved_height(),
                gen_params.sample_params,
                gen_params.high_noise_sample_params,
                gen_params.moe_boundary,
@@ -757,29 +758,27 @@ int main(int argc, const char* argv[]) {
                gen_params.seed,
                gen_params.video_frames,
                gen_params.vace_strength,
-                ctx_params.vae_tiling_params,
+                gen_params.vae_tiling_params,
                gen_params.cache_params,
            };

-            results = generate_video(sd_ctx, &vid_gen_params, &num_results);
+            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");
@@ -788,32 +787,24 @@ 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;
 }
@@ -0,0 +1,207 @@
+#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 "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);
+
+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::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;
+    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 process_and_check(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 {
+    std::string prompt;
+    std::string prompt_with_lora;  // for metadata record only
+    std::string negative_prompt;
+    int clip_skip   = -1;  // <= 0 represents unspecified
+    int width       = -1;
+    int height      = -1;
+    int batch_count = 1;
+    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;
+    bool auto_resize_ref_image = true;
+    bool increase_ref_index    = false;
+    bool embed_image_metadata  = true;
+
+    std::vector<int> skip_layers = {7, 8, 9};
+    sd_sample_params_t sample_params;
+
+    std::vector<int> high_noise_skip_layers = {7, 8, 9};
+    sd_sample_params_t high_noise_sample_params;
+
+    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;
+
+    float strength         = 0.75f;
+    float control_strength = 0.9f;
+
+    int64_t seed = 42;
+
+    sd_tiling_params_t vae_tiling_params = {false, 0, 0, 0.5f, 0.0f, 0.0f};
+
+    // Photo Maker
+    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;
+
+    std::map<std::string, float> lora_map;
+    std::map<std::string, float> high_noise_lora_map;
+    std::vector<sd_lora_t> lora_vec;
+    SDGenerationParams();
+    ArgOptions get_options();
+    bool from_json_str(const std::string& json_str);
+    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 process_and_check(SDMode mode, const std::string& lora_model_dir);
+    std::string to_string() const;
+};
+
+std::string version_string();
+std::string get_image_params(const SDContextParams& ctx_params, const SDGenerationParams& gen_params, int64_t seed);
+
+#endif  // __EXAMPLES_COMMON_COMMON_H__
@@ -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);
+}
@@ -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__
@@ -0,0 +1,84 @@
+#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);
+
+#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);
+#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);
+#endif
+
+int create_video_from_sd_images(const char* filename,
+                                sd_image_t* images,
+                                int num_images,
+                                int fps,
+                                int quality = 90);
+
+#endif  // __MEDIA_IO_H__
@@ -0,0 +1,207 @@
+#ifndef __EXAMPLE_RESOURCE_OWNERS_H__
+#define __EXAMPLE_RESOURCE_OWNERS_H__
+
+#include <cstdio>
+#include <cstdlib>
+#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 {
+public:
+    SDImageOwner() = default;
+    explicit SDImageOwner(sd_image_t image)
+        : image_(image) {
+    }
+
+    SDImageOwner(const SDImageOwner&)            = delete;
+    SDImageOwner& operator=(const SDImageOwner&) = delete;
+
+    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;
+        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;
+    }
+
+private:
+    sd_image_t image_ = {0, 0, 0, nullptr};
+};
+
+class SDImageVec {
+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();
+    }
+
+    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();
+    }
+
+private:
+    std::vector<sd_image_t> images_;
+};
+
+#endif  // __EXAMPLE_RESOURCE_OWNERS_H__
@@ -1,6 +1,92 @@
 set(TARGET sd-server)

-add_executable(${TARGET} main.cpp)
+option(SD_SERVER_BUILD_FRONTEND "Build server frontend with pnpm" ON)
+
+set(FRONTEND_DIR "${CMAKE_CURRENT_SOURCE_DIR}/frontend")
+set(GENERATED_HTML_HEADER "${FRONTEND_DIR}/dist/gen_index_html.h")
+
+set(HAVE_FRONTEND_BUILD OFF)
+
+if(SD_SERVER_BUILD_FRONTEND AND EXISTS "${FRONTEND_DIR}")
+    if(WIN32)
+        find_program(PNPM_EXECUTABLE NAMES pnpm.cmd pnpm)
+    else()
+        find_program(PNPM_EXECUTABLE NAMES pnpm)
+    endif()
+
+    if(PNPM_EXECUTABLE)
+        message(STATUS "Frontend dir found: ${FRONTEND_DIR}")
+        message(STATUS "pnpm found: ${PNPM_EXECUTABLE}")
+
+        set(HAVE_FRONTEND_BUILD ON)
+
+        add_custom_target(${TARGET}_frontend_install
+            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" install
+            WORKING_DIRECTORY "${FRONTEND_DIR}"
+            COMMENT "Installing frontend dependencies"
+            VERBATIM
+        )
+
+        add_custom_target(${TARGET}_frontend_build
+            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" run build
+            WORKING_DIRECTORY "${FRONTEND_DIR}"
+            COMMENT "Building frontend"
+            VERBATIM
+        )
+
+        add_custom_target(${TARGET}_frontend_header
+            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" run build:header
+            WORKING_DIRECTORY "${FRONTEND_DIR}"
+            COMMENT "Generating gen_index_html.h"
+            VERBATIM
+        )
+
+        add_dependencies(${TARGET}_frontend_build ${TARGET}_frontend_install)
+        add_dependencies(${TARGET}_frontend_header ${TARGET}_frontend_build)
+
+        add_custom_target(${TARGET}_frontend
+            DEPENDS ${TARGET}_frontend_header
+        )
+
+        set_source_files_properties("${GENERATED_HTML_HEADER}" PROPERTIES GENERATED TRUE)
+    else()
+        message(WARNING "pnpm not found, frontend build disabled")
+    endif()
+else()
+    message(STATUS "Frontend disabled or directory not found: ${FRONTEND_DIR}")
+endif()
+
+add_executable(${TARGET}
+    ../common/common.cpp
+    ../common/log.cpp
+    ../common/media_io.cpp
+    main.cpp
+)
+
+if(HAVE_FRONTEND_BUILD)
+    add_dependencies(${TARGET} ${TARGET}_frontend)
+    target_sources(${TARGET} PRIVATE "${GENERATED_HTML_HEADER}")
+    target_include_directories(${TARGET} PRIVATE "${FRONTEND_DIR}/dist")
+    target_compile_definitions(${TARGET} PRIVATE HAVE_INDEX_HTML)
+    message(STATUS "HAVE_INDEX_HTML enabled")
+else()
+    message(STATUS "HAVE_INDEX_HTML disabled")
+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)
@@ -1,15 +1,104 @@
+# Frontend
+
+## Build with Frontend
+
+The server can optionally build the web frontend and embed it into the binary as `gen_index_html.h`.
+
+### Requirements
+
+Install the following tools:
+
+* **Node.js** ≥ 22.18
+  https://nodejs.org/
+
+* **pnpm** ≥ 10
+  Install via npm:
+
+```bash
+npm install -g pnpm
+```
+
+Verify installation:
+
+```bash
+node -v
+pnpm -v
+```
+
+### Install frontend dependencies
+
+Go to the frontend directory and install dependencies:
+
+```bash
+cd examples/server/frontend
+pnpm install
+```
+
+### Build the server with CMake
+
+Enable the frontend build option when configuring CMake:
+
+```bash
+cmake -B build -DSD_SERVER_BUILD_FRONTEND=ON
+cmake --build build --config Release
+```
+
+If `pnpm` is available, the build system will automatically run:
+
+```
+pnpm run build
+pnpm run build:header
+```
+
+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.
+
+If you want to modify the UI or frontend logic, please submit pull requests to the **frontend repository**.
+
+This repository (`stable-diffusion.cpp`) only vendors the frontend periodically. Changes from the frontend repo are synchronized:
+
+* approximately **every 1–2 weeks**, or
+* when there are **major frontend updates**
+
+Because of this, frontend changes will **not appear here immediately** after being merged upstream.
+
+## Using an external frontend
+
+By default, the server uses the **embedded frontend** generated during the build (`gen_index_html.h`).
+
+You can also serve a custom frontend file instead of the embedded one by using:
+
+```bash
+--serve-html-path <path-to-index.html>
+```
+
+For example:
+
+```bash
+sd-server --serve-html-path ./index.html
+```
+
+In this case, the server will load and serve the specified `index.html` file instead of the embedded frontend. This is useful when:
+
+* developing or testing frontend changes
+* using a custom UI
+* avoiding rebuilding the binary after frontend modifications
+
 # Run

 ```
 usage: ./bin/sd-server  [options]

 Svr Options:
-  -l, --listen-ip <string>    server listen ip (default: 127.0.0.1)
-  --listen-port <int>         server listen port (default: 1234)
-  --serve-html-path <string>  path to HTML file to serve at root (optional)
-  -v, --verbose               print extra info
-  --color                     colors the logging tags according to level
-  -h, --help                  show this help message and exit
+  -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   
+  -h, --help                    show this help message and exit

 Context Options:
  -m, --model <string>                     path to full model
@@ -36,21 +125,22 @@ Context Options:
                                           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)
-  --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
  --vae-tiling                             process vae in tiles to reduce memory usage
  --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
+  --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)
  --vae-on-cpu                             keep vae in cpu (for low vram)
-  --mmap                                   whether to memory-map model
-  --diffusion-fa                           use flash attention in the diffusion model
+  --fa                                     use flash attention
+  --diffusion-fa                           use flash attention in the diffusion model only
  --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
  --vae-conv-direct                        use ggml_conv2d_direct in the vae model
  --circular                               enable circular padding for convolutions
  --circularx                              enable circular RoPE wrapping on x-axis (width) only
  --circulary                              enable circular RoPE wrapping on y-axis (height) only
  --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
@@ -99,14 +189,15 @@ Default Generation Options:
                                           medium
  --skip-layer-start <float>               SLG enabling point (default: 0.01)
  --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            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 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-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 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
@@ -114,22 +205,24 @@ Default Generation Options:
  --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).
  --disable-auto-resize-ref-image          disable auto resize of ref images
+  --disable-image-metadata                 do not embed generation metadata on image files
  -s, --seed                               RNG seed (default: 42, use random seed for < 0)
  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           tcd] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+                                           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] default: euler for Flux/SD3/Wan, euler_a otherwise
+                                           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], default: discrete
+                                           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)
+  --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"
-  --cache-preset                           cache-dit preset: 'slow'/'s', 'medium'/'m', 'fast'/'f', 'ultra'/'u'
  --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'
 ```
@@ -1,4 +1,6 @@
-for f in *.cpp *.h *.hpp examples/cli/*.cpp examples/common/*.hpp examples/cli/*.h examples/server/*.cpp; do
+for f in src/*.cpp src/*.h src/*.hpp src/vocab/*.h src/vocab/*.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
@@ -48,6 +48,8 @@ enum sample_method_t {
    LCM_SAMPLE_METHOD,
    DDIM_TRAILING_SAMPLE_METHOD,
    TCD_SAMPLE_METHOD,
+    RES_MULTISTEP_SAMPLE_METHOD,
+    RES_2S_SAMPLE_METHOD,
    SAMPLE_METHOD_COUNT
 };

@@ -62,6 +64,7 @@ enum scheduler_t {
    SMOOTHSTEP_SCHEDULER,
    KL_OPTIMAL_SCHEDULER,
    LCM_SCHEDULER,
+    BONG_TANGENT_SCHEDULER,
    SCHEDULER_COUNT
 };

@@ -117,7 +120,8 @@ 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 {
@@ -186,6 +190,7 @@ typedef struct {
    bool keep_clip_on_cpu;
    bool keep_control_net_on_cpu;
    bool keep_vae_on_cpu;
+    bool flash_attn;
    bool diffusion_flash_attn;
    bool tae_preview_only;
    bool diffusion_conv_direct;
@@ -197,7 +202,6 @@ typedef struct {
    bool chroma_use_t5_mask;
    int chroma_t5_mask_pad;
    bool qwen_image_zero_cond_t;
-    float flow_shift;
 } sd_ctx_params_t;

 typedef struct {
@@ -231,6 +235,7 @@ typedef struct {
    int shifted_timestep;
    float* custom_sigmas;
    int custom_sigmas_count;
+    float flow_shift;
 } sd_sample_params_t;

 typedef struct {
@@ -247,6 +252,7 @@ enum sd_cache_mode_t {
    SD_CACHE_DBCACHE,
    SD_CACHE_TAYLORSEER,
    SD_CACHE_CACHE_DIT,
+    SD_CACHE_SPECTRUM,
 };

 typedef struct {
@@ -267,6 +273,13 @@ typedef struct {
    int taylorseer_skip_interval;
    const char* scm_mask;
    bool scm_policy_dynamic;
+    float spectrum_w;
+    int spectrum_m;
+    float spectrum_lam;
+    int spectrum_window_size;
+    float spectrum_flex_window;
+    int spectrum_warmup_steps;
+    float spectrum_stop_percent;
 } sd_cache_params_t;

 typedef struct {
@@ -1,226 +0,0 @@
-#ifndef __PREPROCESSING_HPP__
-#define __PREPROCESSING_HPP__
-
-#include "ggml_extend.hpp"
-#define M_PI_ 3.14159265358979323846f
-
-void convolve(struct ggml_tensor* input, struct ggml_tensor* output, struct ggml_tensor* kernel, int padding) {
-    struct 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;
-    struct ggml_context* ctx0       = ggml_init(params);
-    struct 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);
-}
-
-void gaussian_kernel(struct ggml_tensor* kernel) {
-    int ks_mid   = static_cast<int>(kernel->ne[0] / 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 gx = static_cast<float>(-ks_mid + y);
-        for (int x = 0; x < kernel->ne[1]; 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);
-        }
-    }
-}
-
-void grayscale(struct ggml_tensor* rgb_img, struct 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);
-            float gray = 0.2989f * r + 0.5870f * g + 0.1140f * b;
-            ggml_ext_tensor_set_f32(grayscale, gray, ix, iy);
-        }
-    }
-}
-
-void prop_hypot(struct ggml_tensor* x, struct ggml_tensor* y, struct 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]);
-    }
-}
-
-void prop_arctan2(struct ggml_tensor* x, struct ggml_tensor* y, struct 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]);
-    }
-}
-
-void normalize_tensor(struct 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;
-    }
-    max = 1.0f / max;
-    for (int i = 0; i < n_elements; i++) {
-        dg[i] *= max;
-    }
-}
-
-void non_max_supression(struct ggml_tensor* result, struct ggml_tensor* G, struct 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;
-            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);
-            }
-
-            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);
-            }
-        }
-    }
-}
-
-void threshold_hystersis(struct 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;
-    }
-    float ht = max * 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 (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);
-            }
-        }
-    }
-
-    // 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);
-            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 preprocess_canny(sd_image_t img, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
-    struct 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;
-    struct 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,
-        -1, 0, 1};
-
-    float kY[9] = {
-        1, 2, 1,
-        0, 0, 0,
-        -1, -2, -1};
-
-    // generate kernel
-    int kernel_size             = 5;
-    struct ggml_tensor* gkernel = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, kernel_size, kernel_size, 1, 1);
-    struct 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));
-    struct 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);
-    struct ggml_tensor* image      = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 3, 1);
-    struct ggml_tensor* image_gray = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 1, 1);
-    struct ggml_tensor* iX         = ggml_dup_tensor(work_ctx, image_gray);
-    struct ggml_tensor* iY         = ggml_dup_tensor(work_ctx, image_gray);
-    struct ggml_tensor* G          = ggml_dup_tensor(work_ctx, image_gray);
-    struct 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);
-            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);
-        }
-    }
-    ggml_tensor_to_sd_image(image, img.data);
-    ggml_free(work_ctx);
-    return true;
-}
-
-#endif  // __PREPROCESSING_HPP__
@@ -1,88 +1,88 @@
-import os
-import sys
-
-import numpy as np
-import torch
-from diffusers.utils import load_image
-# pip install insightface==0.7.3
-from insightface.app import FaceAnalysis
-from insightface.data import get_image as ins_get_image
-from safetensors.torch import save_file
-
-### 
-# https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/165#issue-2055829543
-###
-class FaceAnalysis2(FaceAnalysis):
-    # NOTE: allows setting det_size for each detection call.
-    # the model allows it but the wrapping code from insightface
-    # doesn't show it, and people end up loading duplicate models
-    # for different sizes where there is absolutely no need to
-    def get(self, img, max_num=0, det_size=(640, 640)):
-        if det_size is not None:
-            self.det_model.input_size = det_size
-
-        return super().get(img, max_num)
-
-def analyze_faces(face_analysis: FaceAnalysis, img_data: np.ndarray, det_size=(640, 640)):
-    # NOTE: try detect faces, if no faces detected, lower det_size until it does
-    detection_sizes = [None] + [(size, size) for size in range(640, 256, -64)] + [(256, 256)]
-
-    for size in detection_sizes:
-        faces = face_analysis.get(img_data, det_size=size)
-        if len(faces) > 0:
-            return faces
-
-    return []
-
-if __name__ == "__main__":
-    #face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
-    face_detector = FaceAnalysis2(providers=['CPUExecutionProvider'], allowed_modules=['detection', 'recognition'])
-    face_detector.prepare(ctx_id=0, det_size=(640, 640))
-    #input_folder_name = './scarletthead_woman'
-    input_folder_name = sys.argv[1]
-    image_basename_list = os.listdir(input_folder_name)
-    image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
-
-    input_id_images = []
-    for image_path in image_path_list:
-        input_id_images.append(load_image(image_path))
-    
-    id_embed_list = []
-    
-    for img in input_id_images:
-        img = np.array(img)
-        img = img[:, :, ::-1]
-        faces = analyze_faces(face_detector, img)
-        if len(faces) > 0:
-            id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
-    
-    if len(id_embed_list) == 0:
-        raise ValueError(f"No face detected in input image pool")
-    
-    id_embeds = torch.stack(id_embed_list)    
-    
-    # for r in id_embeds:
-    #     print(r)
-    # #torch.save(id_embeds, input_folder_name+'/id_embeds.pt');
-    # weights = dict()
-    # weights["id_embeds"] = id_embeds
-    # save_file(weights, input_folder_name+'/id_embeds.safetensors')
-
-    binary_data = id_embeds.numpy().tobytes()
-    two = 4
-    zero = 0
-    one = 1
-    tensor_name = "id_embeds"
-# Write binary data to a file
-    with open(input_folder_name+'/id_embeds.bin', "wb") as f:
-        f.write(two.to_bytes(4, byteorder='little'))
-        f.write((len(tensor_name)).to_bytes(4, byteorder='little'))
-        f.write(zero.to_bytes(4, byteorder='little'))
-        f.write((id_embeds.shape[1]).to_bytes(4, byteorder='little'))
-        f.write((id_embeds.shape[0]).to_bytes(4, byteorder='little'))
-        f.write(one.to_bytes(4, byteorder='little'))
-        f.write(one.to_bytes(4, byteorder='little'))
-        f.write(tensor_name.encode('ascii'))
-        f.write(binary_data)
-
+import os
+import sys
+
+import numpy as np
+import torch
+from diffusers.utils import load_image
+# pip install insightface==0.7.3
+from insightface.app import FaceAnalysis
+from insightface.data import get_image as ins_get_image
+from safetensors.torch import save_file
+
+### 
+# https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/165#issue-2055829543
+###
+class FaceAnalysis2(FaceAnalysis):
+    # NOTE: allows setting det_size for each detection call.
+    # the model allows it but the wrapping code from insightface
+    # doesn't show it, and people end up loading duplicate models
+    # for different sizes where there is absolutely no need to
+    def get(self, img, max_num=0, det_size=(640, 640)):
+        if det_size is not None:
+            self.det_model.input_size = det_size
+
+        return super().get(img, max_num)
+
+def analyze_faces(face_analysis: FaceAnalysis, img_data: np.ndarray, det_size=(640, 640)):
+    # NOTE: try detect faces, if no faces detected, lower det_size until it does
+    detection_sizes = [None] + [(size, size) for size in range(640, 256, -64)] + [(256, 256)]
+
+    for size in detection_sizes:
+        faces = face_analysis.get(img_data, det_size=size)
+        if len(faces) > 0:
+            return faces
+
+    return []
+
+if __name__ == "__main__":
+    #face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
+    face_detector = FaceAnalysis2(providers=['CPUExecutionProvider'], allowed_modules=['detection', 'recognition'])
+    face_detector.prepare(ctx_id=0, det_size=(640, 640))
+    #input_folder_name = './scarletthead_woman'
+    input_folder_name = sys.argv[1]
+    image_basename_list = os.listdir(input_folder_name)
+    image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
+
+    input_id_images = []
+    for image_path in image_path_list:
+        input_id_images.append(load_image(image_path))
+    
+    id_embed_list = []
+    
+    for img in input_id_images:
+        img = np.array(img)
+        img = img[:, :, ::-1]
+        faces = analyze_faces(face_detector, img)
+        if len(faces) > 0:
+            id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
+    
+    if len(id_embed_list) == 0:
+        raise ValueError(f"No face detected in input image pool")
+    
+    id_embeds = torch.stack(id_embed_list)    
+    
+    # for r in id_embeds:
+    #     print(r)
+    # #torch.save(id_embeds, input_folder_name+'/id_embeds.pt');
+    # weights = dict()
+    # weights["id_embeds"] = id_embeds
+    # save_file(weights, input_folder_name+'/id_embeds.safetensors')
+
+    binary_data = id_embeds.numpy().tobytes()
+    two = 4
+    zero = 0
+    one = 1
+    tensor_name = "id_embeds"
+# Write binary data to a file
+    with open(input_folder_name+'/id_embeds.bin', "wb") as f:
+        f.write(two.to_bytes(4, byteorder='little'))
+        f.write((len(tensor_name)).to_bytes(4, byteorder='little'))
+        f.write(zero.to_bytes(4, byteorder='little'))
+        f.write((id_embeds.shape[1]).to_bytes(4, byteorder='little'))
+        f.write((id_embeds.shape[0]).to_bytes(4, byteorder='little'))
+        f.write(one.to_bytes(4, byteorder='little'))
+        f.write(one.to_bytes(4, byteorder='little'))
+        f.write(tensor_name.encode('ascii'))
+        f.write(binary_data)
+
    
@@ -0,0 +1,683 @@
+#ifndef __ANIMA_HPP__
+#define __ANIMA_HPP__
+
+#include <cmath>
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include "common_block.hpp"
+#include "flux.hpp"
+#include "rope.hpp"
+
+namespace Anima {
+    constexpr int ANIMA_GRAPH_SIZE = 65536;
+
+    __STATIC_INLINE__ ggml_tensor* apply_gate(ggml_context* ctx,
+                                              ggml_tensor* x,
+                                              ggml_tensor* gate) {
+        gate = ggml_reshape_3d(ctx, gate, gate->ne[0], 1, gate->ne[1]);  // [N, 1, C]
+        return ggml_mul(ctx, x, gate);
+    }
+
+    struct XEmbedder : public GGMLBlock {
+    public:
+        XEmbedder(int64_t in_dim, int64_t out_dim) {
+            blocks["proj.1"] = std::make_shared<Linear>(in_dim, out_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj.1"]);
+            return proj->forward(ctx, x);
+        }
+    };
+
+    struct TimestepEmbedder : public GGMLBlock {
+    public:
+        TimestepEmbedder(int64_t in_dim, int64_t out_dim) {
+            blocks["1.linear_1"] = std::make_shared<Linear>(in_dim, in_dim, false);
+            blocks["1.linear_2"] = std::make_shared<Linear>(in_dim, out_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1.linear_1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["1.linear_2"]);
+
+            x = linear_1->forward(ctx, x);
+            x = ggml_silu_inplace(ctx->ggml_ctx, x);
+            x = linear_2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct AdaLayerNormZero : public GGMLBlock {
+    protected:
+        int64_t in_features;
+
+    public:
+        AdaLayerNormZero(int64_t in_features, int64_t hidden_features = 256)
+            : in_features(in_features) {
+            blocks["norm"] = std::make_shared<LayerNorm>(in_features, 1e-6f, false, false);
+            blocks["1"]    = std::make_shared<Linear>(in_features, hidden_features, false);
+            blocks["2"]    = std::make_shared<Linear>(hidden_features, 3 * in_features, false);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* hidden_states,
+                                                      ggml_tensor* embedded_timestep,
+                                                      ggml_tensor* temb = nullptr) {
+            auto norm     = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
+
+            auto emb = ggml_silu(ctx->ggml_ctx, embedded_timestep);
+            emb      = linear_1->forward(ctx, emb);
+            emb      = linear_2->forward(ctx, emb);  // [N, 3*C]
+
+            if (temb != nullptr) {
+                emb = ggml_add(ctx->ggml_ctx, emb, temb);
+            }
+
+            auto emb_chunks = ggml_ext_chunk(ctx->ggml_ctx, emb, 3, 0);
+            auto shift      = emb_chunks[0];
+            auto scale      = emb_chunks[1];
+            auto gate       = emb_chunks[2];
+
+            auto x = norm->forward(ctx, hidden_states);
+            x      = Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+
+            return {x, gate};
+        }
+    };
+
+    struct AdaLayerNorm : public GGMLBlock {
+    protected:
+        int64_t embedding_dim;
+
+    public:
+        AdaLayerNorm(int64_t in_features, int64_t hidden_features = 256)
+            : embedding_dim(in_features) {
+            blocks["norm"] = std::make_shared<LayerNorm>(in_features, 1e-6f, false, false);
+            blocks["1"]    = std::make_shared<Linear>(in_features, hidden_features, false);
+            blocks["2"]    = std::make_shared<Linear>(hidden_features, 2 * in_features, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* hidden_states,
+                             ggml_tensor* embedded_timestep,
+                             ggml_tensor* temb = nullptr) {
+            auto norm     = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
+
+            auto emb = ggml_silu(ctx->ggml_ctx, embedded_timestep);
+            emb      = linear_1->forward(ctx, emb);
+            emb      = linear_2->forward(ctx, emb);  // [N, 2*C]
+
+            if (temb != nullptr) {
+                auto temb_2c = ggml_view_2d(ctx->ggml_ctx, temb, 2 * embedding_dim, temb->ne[1], temb->nb[1], 0);
+                emb          = ggml_add(ctx->ggml_ctx, emb, temb_2c);
+            }
+
+            auto emb_chunks = ggml_ext_chunk(ctx->ggml_ctx, emb, 2, 0);
+            auto shift      = emb_chunks[0];
+            auto scale      = emb_chunks[1];
+
+            auto x = norm->forward(ctx, hidden_states);
+            x      = Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+            return x;
+        }
+    };
+
+    struct AnimaAttention : public GGMLBlock {
+    protected:
+        int64_t num_heads;
+        int64_t head_dim;
+        std::string out_proj_name;
+
+    public:
+        AnimaAttention(int64_t query_dim,
+                       int64_t context_dim,
+                       int64_t num_heads,
+                       int64_t head_dim,
+                       const std::string& out_proj_name = "output_proj")
+            : num_heads(num_heads), head_dim(head_dim), out_proj_name(out_proj_name) {
+            int64_t inner_dim = num_heads * head_dim;
+
+            blocks["q_proj"]            = std::make_shared<Linear>(query_dim, inner_dim, false);
+            blocks["k_proj"]            = std::make_shared<Linear>(context_dim, inner_dim, false);
+            blocks["v_proj"]            = std::make_shared<Linear>(context_dim, inner_dim, false);
+            blocks["q_norm"]            = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm"]            = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks[this->out_proj_name] = std::make_shared<Linear>(inner_dim, query_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* hidden_states,
+                             ggml_tensor* encoder_hidden_states = nullptr,
+                             ggml_tensor* pe_q                  = nullptr,
+                             ggml_tensor* pe_k                  = nullptr) {
+            if (encoder_hidden_states == nullptr) {
+                encoder_hidden_states = hidden_states;
+            }
+
+            auto q_proj   = std::dynamic_pointer_cast<Linear>(blocks["q_proj"]);
+            auto k_proj   = std::dynamic_pointer_cast<Linear>(blocks["k_proj"]);
+            auto v_proj   = std::dynamic_pointer_cast<Linear>(blocks["v_proj"]);
+            auto q_norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"]);
+            auto k_norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"]);
+            auto out_proj = std::dynamic_pointer_cast<Linear>(blocks[out_proj_name]);
+
+            auto q = q_proj->forward(ctx, hidden_states);
+            auto k = k_proj->forward(ctx, encoder_hidden_states);
+            auto v = v_proj->forward(ctx, encoder_hidden_states);
+
+            int64_t N   = q->ne[2];
+            int64_t L_q = q->ne[1];
+            int64_t L_k = k->ne[1];
+
+            auto q4 = ggml_reshape_4d(ctx->ggml_ctx, q, head_dim, num_heads, L_q, N);  // [N, L_q, H, D]
+            auto k4 = ggml_reshape_4d(ctx->ggml_ctx, k, head_dim, num_heads, L_k, N);  // [N, L_k, H, D]
+            auto v4 = ggml_reshape_4d(ctx->ggml_ctx, v, head_dim, num_heads, L_k, N);  // [N, L_k, H, D]
+
+            q4 = q_norm->forward(ctx, q4);
+            k4 = k_norm->forward(ctx, k4);
+
+            ggml_tensor* attn_out = nullptr;
+            if (pe_q != nullptr || pe_k != nullptr) {
+                if (pe_q == nullptr) {
+                    pe_q = pe_k;
+                }
+                if (pe_k == nullptr) {
+                    pe_k = pe_q;
+                }
+                auto q_rope = Rope::apply_rope(ctx->ggml_ctx, q4, pe_q, false);
+                auto k_rope = Rope::apply_rope(ctx->ggml_ctx, k4, pe_k, false);
+                attn_out    = ggml_ext_attention_ext(ctx->ggml_ctx,
+                                                     ctx->backend,
+                                                     q_rope,
+                                                     k_rope,
+                                                     v4,
+                                                     num_heads,
+                                                     nullptr,
+                                                     true,
+                                                     ctx->flash_attn_enabled);
+            } else {
+                auto q_flat = ggml_reshape_3d(ctx->ggml_ctx, q4, head_dim * num_heads, L_q, N);
+                auto k_flat = ggml_reshape_3d(ctx->ggml_ctx, k4, head_dim * num_heads, L_k, N);
+                attn_out    = ggml_ext_attention_ext(ctx->ggml_ctx,
+                                                     ctx->backend,
+                                                     q_flat,
+                                                     k_flat,
+                                                     v,
+                                                     num_heads,
+                                                     nullptr,
+                                                     false,
+                                                     ctx->flash_attn_enabled);
+            }
+
+            return out_proj->forward(ctx, attn_out);
+        }
+    };
+
+    struct AnimaMLP : public GGMLBlock {
+    public:
+        AnimaMLP(int64_t dim, int64_t hidden_dim) {
+            blocks["layer1"] = std::make_shared<Linear>(dim, hidden_dim, false);
+            blocks["layer2"] = std::make_shared<Linear>(hidden_dim, dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto layer1 = std::dynamic_pointer_cast<Linear>(blocks["layer1"]);
+            auto layer2 = std::dynamic_pointer_cast<Linear>(blocks["layer2"]);
+
+            x = layer1->forward(ctx, x);
+            x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
+            x = layer2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct AdapterMLP : public GGMLBlock {
+    public:
+        AdapterMLP(int64_t dim, int64_t hidden_dim) {
+            blocks["0"] = std::make_shared<Linear>(dim, hidden_dim, true);
+            blocks["2"] = std::make_shared<Linear>(hidden_dim, dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto layer0 = std::dynamic_pointer_cast<Linear>(blocks["0"]);
+            auto layer2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
+
+            x = layer0->forward(ctx, x);
+            x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
+            x = layer2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct LLMAdapterBlock : public GGMLBlock {
+    public:
+        LLMAdapterBlock(int64_t model_dim = 1024, int64_t source_dim = 1024, int64_t num_heads = 16, int64_t head_dim = 64) {
+            blocks["norm_self_attn"]  = std::make_shared<RMSNorm>(model_dim, 1e-6f);
+            blocks["self_attn"]       = std::make_shared<AnimaAttention>(model_dim, model_dim, num_heads, head_dim, "o_proj");
+            blocks["norm_cross_attn"] = std::make_shared<RMSNorm>(model_dim, 1e-6f);
+            blocks["cross_attn"]      = std::make_shared<AnimaAttention>(model_dim, source_dim, num_heads, head_dim, "o_proj");
+            blocks["norm_mlp"]        = std::make_shared<RMSNorm>(model_dim, 1e-6f);
+            blocks["mlp"]             = std::make_shared<AdapterMLP>(model_dim, model_dim * 4);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* context,
+                             ggml_tensor* target_pe,
+                             ggml_tensor* context_pe) {
+            auto norm_self_attn  = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_self_attn"]);
+            auto self_attn       = std::dynamic_pointer_cast<AnimaAttention>(blocks["self_attn"]);
+            auto norm_cross_attn = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_cross_attn"]);
+            auto cross_attn      = std::dynamic_pointer_cast<AnimaAttention>(blocks["cross_attn"]);
+            auto norm_mlp        = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_mlp"]);
+            auto mlp             = std::dynamic_pointer_cast<AdapterMLP>(blocks["mlp"]);
+
+            auto h = norm_self_attn->forward(ctx, x);
+            h      = self_attn->forward(ctx, h, nullptr, target_pe, target_pe);
+            x      = ggml_add(ctx->ggml_ctx, x, h);
+
+            h = norm_cross_attn->forward(ctx, x);
+            h = cross_attn->forward(ctx, h, context, target_pe, context_pe);
+            x = ggml_add(ctx->ggml_ctx, x, h);
+
+            h = norm_mlp->forward(ctx, x);
+            h = mlp->forward(ctx, h);
+            x = ggml_add(ctx->ggml_ctx, x, h);
+
+            return x;
+        }
+    };
+
+    struct LLMAdapter : public GGMLBlock {
+    protected:
+        int num_layers;
+
+    public:
+        LLMAdapter(int64_t source_dim = 1024,
+                   int64_t target_dim = 1024,
+                   int64_t model_dim  = 1024,
+                   int num_layers     = 6,
+                   int num_heads      = 16)
+            : num_layers(num_layers) {
+            int64_t head_dim = model_dim / num_heads;
+
+            blocks["embed"] = std::make_shared<Embedding>(32128, target_dim);
+            for (int i = 0; i < num_layers; i++) {
+                blocks["blocks." + std::to_string(i)] =
+                    std::make_shared<LLMAdapterBlock>(model_dim, source_dim, num_heads, head_dim);
+            }
+            blocks["out_proj"] = std::make_shared<Linear>(model_dim, target_dim, true);
+            blocks["norm"]     = std::make_shared<RMSNorm>(target_dim, 1e-6f);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* source_hidden_states,
+                             ggml_tensor* target_input_ids,
+                             ggml_tensor* target_pe,
+                             ggml_tensor* source_pe) {
+            GGML_ASSERT(target_input_ids != nullptr);
+            if (ggml_n_dims(target_input_ids) == 1) {
+                target_input_ids = ggml_reshape_2d(ctx->ggml_ctx, target_input_ids, target_input_ids->ne[0], 1);
+            }
+
+            auto embed    = std::dynamic_pointer_cast<Embedding>(blocks["embed"]);
+            auto out_proj = std::dynamic_pointer_cast<Linear>(blocks["out_proj"]);
+            auto norm     = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
+
+            auto x = embed->forward(ctx, target_input_ids);  // [N, target_len, target_dim]
+
+            for (int i = 0; i < num_layers; i++) {
+                auto block = std::dynamic_pointer_cast<LLMAdapterBlock>(blocks["blocks." + std::to_string(i)]);
+                x          = block->forward(ctx, x, source_hidden_states, target_pe, source_pe);
+            }
+
+            x = out_proj->forward(ctx, x);
+            x = norm->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct TransformerBlock : public GGMLBlock {
+    public:
+        TransformerBlock(int64_t hidden_size,
+                         int64_t text_embed_dim,
+                         int64_t num_heads,
+                         int64_t head_dim,
+                         int64_t mlp_ratio      = 4,
+                         int64_t adaln_lora_dim = 256) {
+            blocks["adaln_modulation_self_attn"]  = std::make_shared<AdaLayerNormZero>(hidden_size, adaln_lora_dim);
+            blocks["self_attn"]                   = std::make_shared<AnimaAttention>(hidden_size, hidden_size, num_heads, head_dim);
+            blocks["adaln_modulation_cross_attn"] = std::make_shared<AdaLayerNormZero>(hidden_size, adaln_lora_dim);
+            blocks["cross_attn"]                  = std::make_shared<AnimaAttention>(hidden_size, text_embed_dim, num_heads, head_dim);
+            blocks["adaln_modulation_mlp"]        = std::make_shared<AdaLayerNormZero>(hidden_size, adaln_lora_dim);
+            blocks["mlp"]                         = std::make_shared<AnimaMLP>(hidden_size, hidden_size * mlp_ratio);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* hidden_states,
+                             ggml_tensor* encoder_hidden_states,
+                             ggml_tensor* embedded_timestep,
+                             ggml_tensor* temb,
+                             ggml_tensor* image_pe) {
+            auto norm1 = std::dynamic_pointer_cast<AdaLayerNormZero>(blocks["adaln_modulation_self_attn"]);
+            auto attn1 = std::dynamic_pointer_cast<AnimaAttention>(blocks["self_attn"]);
+            auto norm2 = std::dynamic_pointer_cast<AdaLayerNormZero>(blocks["adaln_modulation_cross_attn"]);
+            auto attn2 = std::dynamic_pointer_cast<AnimaAttention>(blocks["cross_attn"]);
+            auto norm3 = std::dynamic_pointer_cast<AdaLayerNormZero>(blocks["adaln_modulation_mlp"]);
+            auto mlp   = std::dynamic_pointer_cast<AnimaMLP>(blocks["mlp"]);
+
+            auto [normed1, gate1] = norm1->forward(ctx, hidden_states, embedded_timestep, temb);
+            auto h                = attn1->forward(ctx, normed1, nullptr, image_pe, image_pe);
+            hidden_states         = ggml_add(ctx->ggml_ctx, hidden_states, apply_gate(ctx->ggml_ctx, h, gate1));
+
+            auto [normed2, gate2] = norm2->forward(ctx, hidden_states, embedded_timestep, temb);
+            h                     = attn2->forward(ctx, normed2, encoder_hidden_states, nullptr, nullptr);
+            hidden_states         = ggml_add(ctx->ggml_ctx, hidden_states, apply_gate(ctx->ggml_ctx, h, gate2));
+
+            auto [normed3, gate3] = norm3->forward(ctx, hidden_states, embedded_timestep, temb);
+            h                     = mlp->forward(ctx, normed3);
+            hidden_states         = ggml_add(ctx->ggml_ctx, hidden_states, apply_gate(ctx->ggml_ctx, h, gate3));
+
+            return hidden_states;
+        }
+    };
+
+    struct FinalLayer : public GGMLBlock {
+    protected:
+        int64_t hidden_size;
+        int64_t patch_size;
+        int64_t out_channels;
+
+    public:
+        FinalLayer(int64_t hidden_size, int64_t patch_size, int64_t out_channels)
+            : hidden_size(hidden_size), patch_size(patch_size), out_channels(out_channels) {
+            blocks["adaln_modulation"] = std::make_shared<AdaLayerNorm>(hidden_size, 256);
+            blocks["linear"]           = std::make_shared<Linear>(hidden_size, patch_size * patch_size * out_channels, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* hidden_states,
+                             ggml_tensor* embedded_timestep,
+                             ggml_tensor* temb) {
+            auto adaln  = std::dynamic_pointer_cast<AdaLayerNorm>(blocks["adaln_modulation"]);
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+
+            hidden_states = adaln->forward(ctx, hidden_states, embedded_timestep, temb);
+            hidden_states = linear->forward(ctx, hidden_states);
+            return hidden_states;
+        }
+    };
+
+    struct AnimaNet : public GGMLBlock {
+    public:
+        int64_t in_channels       = 16;
+        int64_t out_channels      = 16;
+        int64_t hidden_size       = 2048;
+        int64_t text_embed_dim    = 1024;
+        int64_t num_heads         = 16;
+        int64_t head_dim          = 128;
+        int patch_size            = 2;
+        int64_t num_layers        = 28;
+        std::vector<int> axes_dim = {44, 42, 42};
+        int theta                 = 10000;
+
+    public:
+        AnimaNet() = default;
+        explicit AnimaNet(int64_t num_layers)
+            : num_layers(num_layers) {
+            blocks["x_embedder"]       = std::make_shared<XEmbedder>((in_channels + 1) * patch_size * patch_size, hidden_size);
+            blocks["t_embedder"]       = std::make_shared<TimestepEmbedder>(hidden_size, hidden_size * 3);
+            blocks["t_embedding_norm"] = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            for (int i = 0; i < num_layers; i++) {
+                blocks["blocks." + std::to_string(i)] = std::make_shared<TransformerBlock>(hidden_size,
+                                                                                           text_embed_dim,
+                                                                                           num_heads,
+                                                                                           head_dim);
+            }
+            blocks["final_layer"] = std::make_shared<FinalLayer>(hidden_size, patch_size, out_channels);
+            blocks["llm_adapter"] = std::make_shared<LLMAdapter>(1024, 1024, 1024, 6, 16);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timestep,
+                             ggml_tensor* encoder_hidden_states,
+                             ggml_tensor* image_pe,
+                             ggml_tensor* t5_ids       = nullptr,
+                             ggml_tensor* t5_weights   = nullptr,
+                             ggml_tensor* adapter_q_pe = nullptr,
+                             ggml_tensor* adapter_k_pe = nullptr) {
+            GGML_ASSERT(x->ne[3] == 1);
+
+            auto x_embedder       = std::dynamic_pointer_cast<XEmbedder>(blocks["x_embedder"]);
+            auto t_embedder       = std::dynamic_pointer_cast<TimestepEmbedder>(blocks["t_embedder"]);
+            auto t_embedding_norm = std::dynamic_pointer_cast<RMSNorm>(blocks["t_embedding_norm"]);
+            auto final_layer      = std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer"]);
+            auto llm_adapter      = std::dynamic_pointer_cast<LLMAdapter>(blocks["llm_adapter"]);
+
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+
+            auto padding_mask = ggml_ext_zeros(ctx->ggml_ctx, x->ne[0], x->ne[1], 1, x->ne[3]);
+            x                 = ggml_concat(ctx->ggml_ctx, x, padding_mask, 2);  // [N, C + 1, H, W]
+
+            x = DiT::pad_and_patchify(ctx, x, patch_size, patch_size);  // [N, h*w, (C+1)*ph*pw]
+
+            x = x_embedder->forward(ctx, x);
+
+            auto timestep_proj     = ggml_ext_timestep_embedding(ctx->ggml_ctx, timestep, static_cast<int>(hidden_size));
+            auto temb              = t_embedder->forward(ctx, timestep_proj);
+            auto embedded_timestep = t_embedding_norm->forward(ctx, timestep_proj);
+
+            if (t5_ids != nullptr) {
+                auto adapted_context = llm_adapter->forward(ctx, encoder_hidden_states, t5_ids, adapter_q_pe, adapter_k_pe);
+                if (t5_weights != nullptr) {
+                    auto w = t5_weights;
+                    if (ggml_n_dims(w) == 1) {
+                        w = ggml_reshape_3d(ctx->ggml_ctx, w, 1, w->ne[0], 1);
+                    }
+                    w               = ggml_repeat_4d(ctx->ggml_ctx, w, adapted_context->ne[0], adapted_context->ne[1], adapted_context->ne[2], 1);
+                    adapted_context = ggml_mul(ctx->ggml_ctx, adapted_context, w);
+                }
+                if (adapted_context->ne[1] < 512) {
+                    auto pad_ctx    = ggml_ext_zeros(ctx->ggml_ctx,
+                                                     adapted_context->ne[0],
+                                                     512 - adapted_context->ne[1],
+                                                     adapted_context->ne[2],
+                                                     1);
+                    adapted_context = ggml_concat(ctx->ggml_ctx, adapted_context, pad_ctx, 1);
+                } else if (adapted_context->ne[1] > 512) {
+                    adapted_context = ggml_ext_slice(ctx->ggml_ctx, adapted_context, 1, 0, 512);
+                }
+                encoder_hidden_states = adapted_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);
+            }
+
+            x = final_layer->forward(ctx, x, embedded_timestep, temb);  // [N, h*w, ph*pw*C]
+
+            x = DiT::unpatchify_and_crop(ctx->ggml_ctx, x, H, W, patch_size, patch_size, false);  // [N, C, H, W]
+
+            return x;
+        }
+    };
+
+    struct AnimaRunner : public GGMLRunner {
+    public:
+        std::vector<float> image_pe_vec;
+        std::vector<float> adapter_q_pe_vec;
+        std::vector<float> adapter_k_pe_vec;
+        AnimaNet net;
+
+        AnimaRunner(ggml_backend_t backend,
+                    bool offload_params_to_cpu,
+                    const String2TensorStorage& tensor_storage_map = {},
+                    const std::string prefix                       = "model.diffusion_model")
+            : GGMLRunner(backend, offload_params_to_cpu) {
+            int64_t num_layers    = 0;
+            std::string layer_tag = prefix + ".net.blocks.";
+            for (const auto& kv : tensor_storage_map) {
+                const std::string& tensor_name = kv.first;
+                size_t pos                     = tensor_name.find(layer_tag);
+                if (pos == std::string::npos) {
+                    continue;
+                }
+                size_t start = pos + layer_tag.size();
+                size_t end   = tensor_name.find('.', start);
+                if (end == std::string::npos) {
+                    continue;
+                }
+                int64_t layer_id = atoll(tensor_name.substr(start, end - start).c_str());
+                num_layers       = std::max(num_layers, layer_id + 1);
+            }
+            if (num_layers <= 0) {
+                num_layers = 28;
+            }
+            LOG_INFO("anima net layers: %" PRId64, num_layers);
+
+            net = AnimaNet(num_layers);
+            net.init(params_ctx, tensor_storage_map, prefix + ".net");
+        }
+
+        std::string get_desc() override {
+            return "anima";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+            net.get_param_tensors(tensors, prefix + ".net");
+        }
+
+        static std::vector<float> gen_1d_rope_pe_vec(int64_t seq_len, int dim, float theta = 10000.f) {
+            std::vector<float> pos(seq_len);
+            for (int64_t i = 0; i < seq_len; i++) {
+                pos[i] = static_cast<float>(i);
+            }
+            auto rope_emb = Rope::rope(pos, dim, theta);
+            return Rope::flatten(rope_emb);
+        }
+
+        static float calc_ntk_factor(float extrapolation_ratio, int axis_dim) {
+            if (extrapolation_ratio == 1.0f || axis_dim <= 2) {
+                return 1.0f;
+            }
+            return std::pow(extrapolation_ratio, static_cast<float>(axis_dim) / static_cast<float>(axis_dim - 2));
+        }
+
+        static std::vector<float> gen_anima_image_pe_vec(int bs,
+                                                         int h,
+                                                         int w,
+                                                         int patch_size,
+                                                         int theta,
+                                                         const std::vector<int>& axes_dim,
+                                                         float h_extrapolation_ratio,
+                                                         float w_extrapolation_ratio,
+                                                         float t_extrapolation_ratio) {
+            static const std::vector<ggml_tensor*> empty_ref_latents;
+            auto ids = Rope::gen_flux_ids(h,
+                                          w,
+                                          patch_size,
+                                          bs,
+                                          static_cast<int>(axes_dim.size()),
+                                          0,
+                                          {},
+                                          empty_ref_latents,
+                                          false,
+                                          1.0f);
+
+            std::vector<float> axis_thetas = {
+                static_cast<float>(theta) * calc_ntk_factor(t_extrapolation_ratio, axes_dim[0]),
+                static_cast<float>(theta) * calc_ntk_factor(h_extrapolation_ratio, axes_dim[1]),
+                static_cast<float>(theta) * calc_ntk_factor(w_extrapolation_ratio, axes_dim[2]),
+            };
+            return Rope::embed_nd(ids, bs, axis_thetas, axes_dim);
+        }
+
+        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);
+
+            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;
+            int64_t w_pad = x->ne[0] + pad_w;
+
+            image_pe_vec          = gen_anima_image_pe_vec(1,
+                                                           static_cast<int>(h_pad),
+                                                           static_cast<int>(w_pad),
+                                                           static_cast<int>(net.patch_size),
+                                                           net.theta,
+                                                           net.axes_dim,
+                                                           4.0f,
+                                                           4.0f,
+                                                           1.0f);
+            int64_t image_pos_len = static_cast<int64_t>(image_pe_vec.size()) / (2 * 2 * (net.head_dim / 2));
+            auto image_pe         = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, net.head_dim / 2, image_pos_len);
+            set_backend_tensor_data(image_pe, image_pe_vec.data());
+
+            ggml_tensor* adapter_q_pe = nullptr;
+            ggml_tensor* adapter_k_pe = nullptr;
+            if (t5_ids != nullptr) {
+                int64_t target_len = t5_ids->ne[0];
+                int64_t source_len = context->ne[1];
+
+                adapter_q_pe_vec = gen_1d_rope_pe_vec(target_len, 64, 10000.f);
+                adapter_k_pe_vec = gen_1d_rope_pe_vec(source_len, 64, 10000.f);
+
+                int64_t target_pos_len = static_cast<int64_t>(adapter_q_pe_vec.size()) / (2 * 2 * 32);
+                int64_t source_pos_len = static_cast<int64_t>(adapter_k_pe_vec.size()) / (2 * 2 * 32);
+
+                adapter_q_pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, 32, target_pos_len);
+                adapter_k_pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, 32, source_pos_len);
+                set_backend_tensor_data(adapter_q_pe, adapter_q_pe_vec.data());
+                set_backend_tensor_data(adapter_k_pe, adapter_k_pe_vec.data());
+            }
+
+            auto runner_ctx = get_context();
+            auto out        = net.forward(&runner_ctx,
+                                          x,
+                                          timesteps,
+                                          context,
+                                          image_pe,
+                                          t5_ids,
+                                          t5_weights,
+                                          adapter_q_pe,
+                                          adapter_k_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    = {},
+                                  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 restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+        }
+    };
+}  // namespace Anima
+
+#endif  // __ANIMA_HPP__
@@ -1,8 +1,7 @@
-#ifndef __VAE_HPP__
-#define __VAE_HPP__
+#ifndef __AUTO_ENCODER_KL_HPP__
+#define __AUTO_ENCODER_KL_HPP__

-#include "common.hpp"
-#include "ggml_extend.hpp"
+#include "vae.hpp"

 /*================================================== AutoEncoderKL ===================================================*/

@@ -30,7 +29,7 @@ public:
        }
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
        // x: [N, in_channels, h, w]
        // t_emb is always None
        auto norm1 = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm1"]);
@@ -66,7 +65,7 @@ protected:
    int64_t in_channels;
    bool use_linear;

-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") {
        auto iter = tensor_storage_map.find(prefix + "proj_out.weight");
        if (iter != tensor_storage_map.end()) {
            if (iter->second.n_dims == 4 && use_linear) {
@@ -102,7 +101,7 @@ public:
        }
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
        // x: [N, in_channels, h, w]
        auto norm     = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm"]);
        auto q_proj   = std::dynamic_pointer_cast<UnaryBlock>(blocks["q"]);
@@ -141,7 +140,7 @@ public:
            v = ggml_reshape_3d(ctx->ggml_ctx, v, c, h * w, n);                        // [N, h * w, in_channels]
        }

-        h_ = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, false, true, false);
+        h_ = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);

        if (use_linear) {
            h_ = proj_out->forward(ctx, h_);  // [N, h * w, in_channels]
@@ -179,8 +178,8 @@ public:
                                                                        {kernel_padding, 0, 0}));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x) override {
        // timesteps always None
        // skip_video always False
        // x: [N, IC, IH, IW]
@@ -209,7 +208,7 @@ public:

 class VideoResnetBlock : public ResnetBlock {
 protected:
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
        enum ggml_type wtype = get_type(prefix + "mix_factor", tensor_storage_map, GGML_TYPE_F32);
        params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
    }
@@ -228,7 +227,7 @@ public:
        blocks["time_stack"] = std::shared_ptr<GGMLBlock>(new ResBlock(out_channels, 0, out_channels, {video_kernel_size, 1}, 3, false, true));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
        // x: [N, in_channels, h, w] aka [b*t, in_channels, h, w]
        // return: [N, out_channels, h, w] aka [b*t, out_channels, h, w]
        // t_emb is always None
@@ -253,8 +252,8 @@ public:

        float alpha = get_alpha();
        x           = ggml_add(ctx->ggml_ctx,
-                               ggml_scale(ctx->ggml_ctx, x, alpha),
-                               ggml_scale(ctx->ggml_ctx, x_mix, 1.0f - alpha));
+                               ggml_ext_scale(ctx->ggml_ctx, x, alpha),
+                               ggml_ext_scale(ctx->ggml_ctx, x_mix, 1.0f - alpha));

        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
        x = ggml_reshape_4d(ctx->ggml_ctx, x, W, H, C, T * B);                     // b t c (h w) -> (b t) c h w
@@ -318,7 +317,7 @@ public:
        blocks["conv_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(block_in, double_z ? z_channels * 2 : z_channels, {3, 3}, {1, 1}, {1, 1}));
    }

-    virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    virtual ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [N, in_channels, h, w]

        auto conv_in     = std::dynamic_pointer_cast<Conv2d>(blocks["conv_in"]);
@@ -436,7 +435,7 @@ public:
        blocks["conv_out"] = get_conv_out(block_in, out_ch, {3, 3}, {1, 1}, {1, 1});
    }

-    virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* z) {
+    virtual ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* z) {
        // z: [N, z_channels, h, w]
        // alpha is always 0
        // merge_strategy is always learned
@@ -484,7 +483,7 @@ public:
 };

 // ldm.models.autoencoder.AutoencoderKL
-class AutoencodingEngine : public GGMLBlock {
+class AutoEncoderKLModel : public GGMLBlock {
 protected:
    SDVersion version;
    bool decode_only       = true;
@@ -502,11 +501,36 @@ 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:
-    AutoencodingEngine(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)) {
@@ -520,7 +544,9 @@ 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,
@@ -550,7 +576,7 @@ public:
        }
    }

-    struct ggml_tensor* decode(GGMLRunnerContext* ctx, struct ggml_tensor* z) {
+    ggml_tensor* decode(GGMLRunnerContext* ctx, ggml_tensor* z) {
        // z: [N, z_channels, h, w]
        if (sd_version_is_flux2(version)) {
            // [N, C*p*p, h, w] -> [N, C, h*p, w*p]
@@ -582,7 +608,7 @@ public:
        return h;
    }

-    struct ggml_tensor* encode(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* encode(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [N, in_channels, h, w]
        auto encoder = std::dynamic_pointer_cast<Encoder>(blocks["encoder"]);

@@ -611,48 +637,21 @@ public:
        }
        return z;
    }
-};

-struct VAE : public GGMLRunner {
-    VAE(ggml_backend_t backend, bool offload_params_to_cpu)
-        : GGMLRunner(backend, offload_params_to_cpu) {}
-    virtual bool compute(const int n_threads,
-                         struct ggml_tensor* z,
-                         bool decode_graph,
-                         struct ggml_tensor** output,
-                         struct ggml_context* output_ctx)                                                         = 0;
-    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) = 0;
-    virtual void set_conv2d_scale(float scale) { SD_UNUSED(scale); };
-};
-
-struct FakeVAE : public VAE {
-    FakeVAE(ggml_backend_t backend, bool offload_params_to_cpu)
-        : VAE(backend, offload_params_to_cpu) {}
-    bool compute(const int n_threads,
-                 struct ggml_tensor* z,
-                 bool decode_graph,
-                 struct ggml_tensor** output,
-                 struct ggml_context* output_ctx) override {
-        if (*output == nullptr && output_ctx != nullptr) {
-            *output = ggml_dup_tensor(output_ctx, z);
+    int get_encoder_output_channels() {
+        int factor = dd_config.double_z ? 2 : 1;
+        if (sd_version_is_flux2(version)) {
+            return dd_config.z_channels * 4;
        }
-        ggml_ext_tensor_iter(z, [&](ggml_tensor* z, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
-            float value = ggml_ext_tensor_get_f32(z, i0, i1, i2, i3);
-            ggml_ext_tensor_set_f32(*output, value, i0, i1, i2, i3);
-        });
-        return true;
-    }
-
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) override {}
-
-    std::string get_desc() override {
-        return "fake_vae";
+        return dd_config.z_channels * factor;
    }
 };

 struct AutoEncoderKL : public VAE {
-    bool decode_only = true;
-    AutoencodingEngine ae;
+    float scale_factor = 1.f;
+    float shift_factor = 0.f;
+    bool decode_only   = true;
+    AutoEncoderKLModel ae;

    AutoEncoderKL(ggml_backend_t backend,
                  bool offload_params_to_cpu,
@@ -661,7 +660,23 @@ struct AutoEncoderKL : public VAE {
                  bool decode_only       = false,
                  bool use_video_decoder = false,
                  SDVersion version      = VERSION_SD1)
-        : decode_only(decode_only), VAE(backend, offload_params_to_cpu) {
+        : decode_only(decode_only), VAE(version, backend, offload_params_to_cpu) {
+        if (sd_version_is_sd1(version) || sd_version_is_sd2(version)) {
+            scale_factor = 0.18215f;
+            shift_factor = 0.f;
+        } else if (sd_version_is_sdxl(version)) {
+            scale_factor = 0.13025f;
+            shift_factor = 0.f;
+        } else if (sd_version_is_sd3(version)) {
+            scale_factor = 1.5305f;
+            shift_factor = 0.0609f;
+        } 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)) {
+            scale_factor = 1.0f;
+            shift_factor = 0.f;
+        }
        bool use_linear_projection = false;
        for (const auto& [name, tensor_storage] : tensor_storage_map) {
            if (!starts_with(name, prefix)) {
@@ -674,7 +689,7 @@ struct AutoEncoderKL : public VAE {
                break;
            }
        }
-        ae = AutoencodingEngine(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);
    }

@@ -693,62 +708,149 @@ struct AutoEncoderKL : public VAE {
        return "vae";
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) override {
        ae.get_param_tensors(tensors, prefix);
    }

-    struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
-        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
-
-        z = to_backend(z);
+    ggml_cgraph* build_graph(const sd::Tensor<float>& z_tensor, bool decode_graph) {
+        ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+        ggml_tensor* z  = make_input(z_tensor);

        auto runner_ctx = get_context();

-        struct ggml_tensor* out = decode_graph ? ae.decode(&runner_ctx, z) : ae.encode(&runner_ctx, z);
+        ggml_tensor* out = decode_graph ? ae.decode(&runner_ctx, z) : ae.encode(&runner_ctx, z);

        ggml_build_forward_expand(gf, out);

        return gf;
    }

-    bool compute(const int n_threads,
-                 struct ggml_tensor* z,
-                 bool decode_graph,
-                 struct ggml_tensor** output,
-                 struct 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 = [&]() -> struct ggml_cgraph* {
+        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());
+    }
+
+    sd::Tensor<float> gaussian_latent_sample(const sd::Tensor<float>& moments, std::shared_ptr<RNG> rng) {
+        // ldm.modules.distributions.distributions.DiagonalGaussianDistribution.sample
+        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;
+    }
+
+    sd::Tensor<float> vae_output_to_latents(const sd::Tensor<float>& vae_output, std::shared_ptr<RNG> rng) override {
+        if (sd_version_is_flux2(version)) {
+            return vae_output;
+        } else if (version == VERSION_SD1_PIX2PIX) {
+            return sd::ops::chunk(vae_output, 2, 2)[0];
+        } else {
+            return gaussian_latent_sample(vae_output, rng);
+        }
+    }
+
+    std::pair<sd::Tensor<float>, sd::Tensor<float>> get_latents_mean_std(const sd::Tensor<float>& latents, int channel_dim) {
+        GGML_ASSERT(channel_dim >= 0 && static_cast<size_t>(channel_dim) < static_cast<size_t>(latents.dim()));
+        if (sd_version_is_flux2(version)) {
+            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);
+        }
+    }
+
+    sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
+        if (sd_version_is_flux2(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 (latents / scale_factor) + shift_factor;
+    }
+
+    sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents) override {
+        if (sd_version_is_flux2(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 (latents - shift_factor) * scale_factor;
+    }
+
+    int get_encoder_output_channels(int input_channels) {
+        return ae.get_encoder_output_channels();
    }

    void test() {
-        struct ggml_init_params params;
+        ggml_init_params params;
        params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
        params.mem_buffer = nullptr;
        params.no_alloc   = false;

-        struct 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);
-            struct 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);
        }

@@ -757,19 +859,21 @@ 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);
-            struct 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);
        }
    };
 };

-#endif
+#endif  // __AUTO_ENCODER_KL_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;
@@ -603,87 +605,6 @@ inline std::vector<int> generate_scm_mask(
    return mask;
 }

-inline std::vector<int> get_scm_preset(const std::string& preset, int total_steps) {
-    struct Preset {
-        std::vector<int> compute_bins;
-        std::vector<int> cache_bins;
-    };
-
-    Preset slow   = {{8, 3, 3, 2, 1, 1}, {1, 2, 2, 2, 3}};
-    Preset medium = {{6, 2, 2, 2, 2, 1}, {1, 3, 3, 3, 3}};
-    Preset fast   = {{6, 1, 1, 1, 1, 1}, {1, 3, 4, 5, 4}};
-    Preset ultra  = {{4, 1, 1, 1, 1}, {2, 5, 6, 7}};
-
-    Preset* p = nullptr;
-    if (preset == "slow" || preset == "s" || preset == "S")
-        p = &slow;
-    else if (preset == "medium" || preset == "m" || preset == "M")
-        p = &medium;
-    else if (preset == "fast" || preset == "f" || preset == "F")
-        p = &fast;
-    else if (preset == "ultra" || preset == "u" || preset == "U")
-        p = &ultra;
-    else
-        return {};
-
-    if (total_steps != 28 && total_steps > 0) {
-        float scale = static_cast<float>(total_steps) / 28.0f;
-        std::vector<int> scaled_compute, scaled_cache;
-
-        for (int v : p->compute_bins) {
-            scaled_compute.push_back(std::max(1, static_cast<int>(v * scale + 0.5f)));
-        }
-        for (int v : p->cache_bins) {
-            scaled_cache.push_back(std::max(1, static_cast<int>(v * scale + 0.5f)));
-        }
-
-        return generate_scm_mask(scaled_compute, scaled_cache, total_steps);
-    }
-
-    return generate_scm_mask(p->compute_bins, p->cache_bins, total_steps);
-}
-
-inline float get_preset_threshold(const std::string& preset) {
-    if (preset == "slow" || preset == "s" || preset == "S")
-        return 0.20f;
-    if (preset == "medium" || preset == "m" || preset == "M")
-        return 0.25f;
-    if (preset == "fast" || preset == "f" || preset == "F")
-        return 0.30f;
-    if (preset == "ultra" || preset == "u" || preset == "U")
-        return 0.34f;
-    return 0.08f;
-}
-
-inline int get_preset_warmup(const std::string& preset) {
-    if (preset == "slow" || preset == "s" || preset == "S")
-        return 8;
-    if (preset == "medium" || preset == "m" || preset == "M")
-        return 6;
-    if (preset == "fast" || preset == "f" || preset == "F")
-        return 6;
-    if (preset == "ultra" || preset == "u" || preset == "U")
-        return 4;
-    return 8;
-}
-
-inline int get_preset_Fn(const std::string& preset) {
-    if (preset == "slow" || preset == "s" || preset == "S")
-        return 8;
-    if (preset == "medium" || preset == "m" || preset == "M")
-        return 8;
-    if (preset == "fast" || preset == "f" || preset == "F")
-        return 6;
-    if (preset == "ultra" || preset == "u" || preset == "U")
-        return 4;
-    return 8;
-}
-
-inline int get_preset_Bn(const std::string& preset) {
-    (void)preset;
-    return 0;
-}
-
 inline void parse_dbcache_options(const std::string& opts, DBCacheConfig& cfg) {
    if (opts.empty())
        return;
@@ -852,35 +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, struct ggml_tensor* input, struct 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;

@@ -900,8 +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;
@@ -914,13 +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) {
@@ -940,7 +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;
        }

@@ -948,15 +870,14 @@ struct CacheDitConditionState {
        return false;
    }

-    void after_condition(const void* cond, struct ggml_tensor* input, struct 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);
        }
    }

@@ -4,6 +4,7 @@
 #include "ggml_extend.hpp"
 #include "model.h"
 #include "tokenize_util.h"
+#include "vocab/vocab.h"

 /*================================================== CLIPTokenizer ===================================================*/

@@ -110,7 +111,7 @@ public:
        if (merges_utf8_str.size() > 0) {
            load_from_merges(merges_utf8_str);
        } else {
-            load_from_merges(ModelLoader::load_merges());
+            load_from_merges(load_clip_merges());
        }
        add_special_token("<|startoftext|>");
        add_special_token("<|endoftext|>");
@@ -472,16 +473,16 @@ public:
        }
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [N, n_token, d_model]
        auto fc1 = std::dynamic_pointer_cast<Linear>(blocks["fc1"]);
        auto fc2 = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);

        x = fc1->forward(ctx, x);
        if (use_gelu) {
-            x = ggml_gelu_inplace(ctx->ggml_ctx, x);
+            x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
        } else {
-            x = ggml_gelu_quick_inplace(ctx->ggml_ctx, x);
+            x = ggml_ext_gelu_quick(ctx->ggml_ctx, x, true);
        }
        x = fc2->forward(ctx, x);
        return x;
@@ -510,7 +511,7 @@ public:
        blocks["mlp"] = std::shared_ptr<GGMLBlock>(new CLIPMLP(d_model, intermediate_size));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x, bool mask = true) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* mask = nullptr) {
        // x: [N, n_token, d_model]
        auto self_attn   = std::dynamic_pointer_cast<MultiheadAttention>(blocks["self_attn"]);
        auto layer_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm1"]);
@@ -540,10 +541,10 @@ public:
        }
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
-                                int clip_skip = -1,
-                                bool mask     = true) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x,
+                         ggml_tensor* mask = nullptr,
+                         int clip_skip     = -1) {
        // x: [N, n_token, d_model]
        int layer_idx = n_layer - 1;
        // LOG_DEBUG("clip_skip %d", clip_skip);
@@ -572,7 +573,7 @@ protected:
    int64_t num_positions;
    bool force_clip_f32;

-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
        enum ggml_type token_wtype = GGML_TYPE_F32;
        if (!force_clip_f32) {
            token_wtype = get_type(prefix + "token_embedding.weight", tensor_storage_map, GGML_TYPE_F32);
@@ -596,13 +597,13 @@ public:
          force_clip_f32(force_clip_f32) {
    }

-    struct ggml_tensor* get_token_embed_weight() {
+    ggml_tensor* get_token_embed_weight() {
        return params["token_embedding.weight"];
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* input_ids,
-                                struct ggml_tensor* custom_embed_weight) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* input_ids,
+                         ggml_tensor* custom_embed_weight) {
        // input_ids: [N, n_token]
        auto token_embed_weight    = params["token_embedding.weight"];
        auto position_embed_weight = params["position_embedding.weight"];
@@ -629,7 +630,7 @@ protected:
    int num_patches;
    int64_t num_positions;

-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
        enum ggml_type patch_wtype    = GGML_TYPE_F16;
        enum ggml_type class_wtype    = GGML_TYPE_F32;
        enum ggml_type position_wtype = GGML_TYPE_F32;
@@ -652,7 +653,7 @@ public:
        num_positions = num_patches + 1;
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* pixel_values) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* pixel_values) {
        // pixel_values: [N, num_channels, image_size, image_size]
        // return: [N, num_positions, embed_dim]
        GGML_ASSERT(pixel_values->ne[0] == image_size && pixel_values->ne[1] == image_size && pixel_values->ne[2] == num_channels);
@@ -662,20 +663,20 @@ public:
        auto position_embed_weight = params["position_embedding.weight"];

        // concat(patch_embedding, class_embedding) + position_embedding
-        struct ggml_tensor* patch_embedding;
+        ggml_tensor* patch_embedding;
        int64_t N       = pixel_values->ne[3];
        patch_embedding = ggml_ext_conv_2d(ctx->ggml_ctx, pixel_values, patch_embed_weight, nullptr, patch_size, patch_size);  // [N, embed_dim, image_size // pacht_size, image_size // pacht_size]
        patch_embedding = ggml_reshape_3d(ctx->ggml_ctx, patch_embedding, num_patches, embed_dim, N);                          // [N, embed_dim, num_patches]
        patch_embedding = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, patch_embedding, 1, 0, 2, 3));                  // [N, num_patches, embed_dim]
        patch_embedding = ggml_reshape_4d(ctx->ggml_ctx, patch_embedding, 1, embed_dim, num_patches, N);                       // [N, num_patches, embed_dim, 1]

-        struct ggml_tensor* class_embedding = ggml_new_tensor_2d(ctx->ggml_ctx, GGML_TYPE_F32, embed_dim, N);
-        class_embedding                     = ggml_repeat(ctx->ggml_ctx, class_embed_weight, class_embedding);      // [N, embed_dim]
-        class_embedding                     = ggml_reshape_4d(ctx->ggml_ctx, class_embedding, 1, embed_dim, 1, N);  // [N, 1, embed_dim, 1]
+        ggml_tensor* class_embedding = ggml_new_tensor_2d(ctx->ggml_ctx, GGML_TYPE_F32, embed_dim, N);
+        class_embedding              = ggml_repeat(ctx->ggml_ctx, class_embed_weight, class_embedding);      // [N, embed_dim]
+        class_embedding              = ggml_reshape_4d(ctx->ggml_ctx, class_embedding, 1, embed_dim, 1, N);  // [N, 1, embed_dim, 1]

-        struct ggml_tensor* x = ggml_concat(ctx->ggml_ctx, class_embedding, patch_embedding, 2);  // [N, num_positions, embed_dim, 1]
-        x                     = ggml_reshape_3d(ctx->ggml_ctx, x, embed_dim, num_positions, N);   // [N, num_positions, embed_dim]
-        x                     = ggml_add(ctx->ggml_ctx, x, position_embed_weight);
+        ggml_tensor* x = ggml_concat(ctx->ggml_ctx, class_embedding, patch_embedding, 2);  // [N, num_positions, embed_dim, 1]
+        x              = ggml_reshape_3d(ctx->ggml_ctx, x, embed_dim, num_positions, N);   // [N, num_positions, embed_dim]
+        x              = ggml_add(ctx->ggml_ctx, x, position_embed_weight);
        return x;  // [N, num_positions, embed_dim]
    }
 };
@@ -692,7 +693,7 @@ enum CLIPVersion {

 class CLIPTextModel : public GGMLBlock {
 protected:
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
        if (version == OPEN_CLIP_VIT_BIGG_14) {
            enum ggml_type wtype      = GGML_TYPE_F32;
            params["text_projection"] = ggml_new_tensor_2d(ctx, wtype, projection_dim, hidden_size);
@@ -733,24 +734,25 @@ public:
        blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
    }

-    struct ggml_tensor* get_token_embed_weight() {
+    ggml_tensor* get_token_embed_weight() {
        auto embeddings = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
        return embeddings->get_token_embed_weight();
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* input_ids,
-                                struct ggml_tensor* tkn_embeddings,
-                                size_t max_token_idx = 0,
-                                bool return_pooled   = false,
-                                int clip_skip        = -1) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* input_ids,
+                         ggml_tensor* tkn_embeddings,
+                         ggml_tensor* mask    = nullptr,
+                         size_t max_token_idx = 0,
+                         bool return_pooled   = false,
+                         int clip_skip        = -1) {
        // input_ids: [N, n_token]
        auto embeddings       = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
        auto encoder          = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
        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, return_pooled ? -1 : clip_skip, true);
+        x      = encoder->forward(ctx, x, mask, return_pooled ? -1 : clip_skip);
        if (return_pooled || with_final_ln) {
            x = final_layer_norm->forward(ctx, x);
        }
@@ -802,10 +804,10 @@ public:
        blocks["post_layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* pixel_values,
-                                bool return_pooled = true,
-                                int clip_skip      = -1) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* pixel_values,
+                         bool return_pooled = true,
+                         int clip_skip      = -1) {
        // pixel_values: [N, num_channels, image_size, image_size]
        auto embeddings     = std::dynamic_pointer_cast<CLIPVisionEmbeddings>(blocks["embeddings"]);
        auto pre_layernorm  = std::dynamic_pointer_cast<LayerNorm>(blocks["pre_layernorm"]);
@@ -814,10 +816,11 @@ public:

        auto x = embeddings->forward(ctx, pixel_values);  // [N, num_positions, embed_dim]
        x      = pre_layernorm->forward(ctx, x);
-        x      = encoder->forward(ctx, x, clip_skip, false);
-        // print_ggml_tensor(x, true, "ClipVisionModel x: ");
+        x      = encoder->forward(ctx, x, nullptr, clip_skip);
+
        auto last_hidden_state = x;
-        x                      = post_layernorm->forward(ctx, x);  // [N, n_token, hidden_size]
+
+        x = post_layernorm->forward(ctx, x);  // [N, n_token, hidden_size]

        GGML_ASSERT(x->ne[3] == 1);
        if (return_pooled) {
@@ -836,7 +839,7 @@ protected:
    int64_t out_features;
    bool transpose_weight;

-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
        enum ggml_type wtype = get_type(prefix + "weight", tensor_storage_map, GGML_TYPE_F32);
        if (transpose_weight) {
            params["weight"] = ggml_new_tensor_2d(ctx, wtype, out_features, in_features);
@@ -853,8 +856,8 @@ public:
          out_features(out_features),
          transpose_weight(transpose_weight) {}

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
-        struct ggml_tensor* w = params["weight"];
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+        ggml_tensor* w = params["weight"];
        if (transpose_weight) {
            w = ggml_cont(ctx->ggml_ctx, ggml_transpose(ctx->ggml_ctx, w));
        }
@@ -883,10 +886,10 @@ public:
        blocks["visual_projection"] = std::shared_ptr<GGMLBlock>(new CLIPProjection(hidden_size, projection_dim, transpose_proj_w));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* pixel_values,
-                                bool return_pooled = true,
-                                int clip_skip      = -1) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* pixel_values,
+                         bool return_pooled = true,
+                         int clip_skip      = -1) {
        // pixel_values: [N, num_channels, image_size, image_size]
        // return: [N, projection_dim] if return_pooled else [N, n_token, hidden_size]
        auto vision_model      = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
@@ -905,6 +908,8 @@ public:
 struct CLIPTextModelRunner : public GGMLRunner {
    CLIPTextModel model;

+    std::vector<float> attention_mask_vec;
+
    CLIPTextModelRunner(ggml_backend_t backend,
                        bool offload_params_to_cpu,
                        const String2TensorStorage& tensor_storage_map,
@@ -931,16 +936,17 @@ struct CLIPTextModelRunner : public GGMLRunner {
        return "clip";
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
        model.get_param_tensors(tensors, prefix);
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* input_ids,
-                                struct ggml_tensor* embeddings,
-                                size_t max_token_idx = 0,
-                                bool return_pooled   = false,
-                                int clip_skip        = -1) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* input_ids,
+                         ggml_tensor* embeddings,
+                         ggml_tensor* mask,
+                         size_t max_token_idx = 0,
+                         bool return_pooled   = false,
+                         int clip_skip        = -1) {
        size_t N       = input_ids->ne[1];
        size_t n_token = input_ids->ne[0];
        if (input_ids->ne[0] > model.n_token) {
@@ -948,20 +954,19 @@ struct CLIPTextModelRunner : public GGMLRunner {
            input_ids = ggml_reshape_2d(ctx->ggml_ctx, input_ids, model.n_token, input_ids->ne[0] / model.n_token);
        }

-        return model.forward(ctx, input_ids, embeddings, max_token_idx, return_pooled, clip_skip);
+        return model.forward(ctx, input_ids, embeddings, mask, max_token_idx, return_pooled, clip_skip);
    }

-    struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
-                                    int num_custom_embeddings    = 0,
-                                    void* custom_embeddings_data = nullptr,
-                                    size_t max_token_idx         = 0,
-                                    bool return_pooled           = false,
-                                    int clip_skip                = -1) {
-        struct ggml_cgraph* gf = new_graph_custom(2048);
+    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);
+        ggml_tensor* input_ids = make_input(input_ids_tensor);

-        input_ids = to_backend(input_ids);
-
-        struct ggml_tensor* embeddings = nullptr;
+        ggml_tensor* embeddings = nullptr;

        if (num_custom_embeddings > 0 && custom_embeddings_data != nullptr) {
            auto token_embed_weight = model.get_token_embed_weight();
@@ -975,28 +980,44 @@ struct CLIPTextModelRunner : public GGMLRunner {
            embeddings = ggml_concat(compute_ctx, token_embed_weight, custom_embeddings, 1);
        }

+        int n_tokens = static_cast<int>(input_ids->ne[0]);
+        attention_mask_vec.resize(n_tokens * n_tokens);
+        for (int i0 = 0; i0 < n_tokens; i0++) {
+            for (int i1 = 0; i1 < n_tokens; i1++) {
+                float value = 0.f;
+                if (i0 > i1) {
+                    value = -INFINITY;
+                }
+                attention_mask_vec[i1 * n_tokens + i0] = value;
+            }
+        }
+        auto attention_mask = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, n_tokens, n_tokens);
+        set_backend_tensor_data(attention_mask, attention_mask_vec.data());
+
        auto runner_ctx = get_context();

-        struct ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, embeddings, max_token_idx, return_pooled, clip_skip);
+        ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, embeddings, attention_mask, max_token_idx, return_pooled, clip_skip);

        ggml_build_forward_expand(gf, hidden_states);

        return gf;
    }

-    bool compute(const int n_threads,
-                 struct 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) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+    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);
    }
 };

@@ -1,5 +1,5 @@
-#ifndef __COMMON_HPP__
-#define __COMMON_HPP__
+#ifndef __COMMON_BLOCK_HPP__
+#define __COMMON_BLOCK_HPP__

 #include "ggml_extend.hpp"

@@ -23,7 +23,7 @@ public:
        }
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [N, channels, h, w]
        if (vae_downsample) {
            auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
@@ -52,7 +52,7 @@ public:
        blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [N, channels, h, w]
        auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);

@@ -121,7 +121,7 @@ public:
        }
    }

-    virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x, struct ggml_tensor* emb = nullptr) {
+    virtual ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* emb = nullptr) {
        // For dims==3, we reduce dimension from 5d to 4d by merging h and w, in order not to change ggml
        // [N, c, t, h, w] => [N, c, t, h * w]
        // x: [N, channels, h, w] if dims == 2 else [N, channels, t, h, w]
@@ -188,7 +188,7 @@ public:
        blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim_in, dim_out * 2));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
        // x: [ne3, ne2, ne1, dim_in]
        // return: [ne3, ne2, ne1, dim_out]
        auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
@@ -200,7 +200,7 @@ public:

        gate = ggml_cont(ctx->ggml_ctx, gate);

-        gate = ggml_gelu_inplace(ctx->ggml_ctx, gate);
+        gate = ggml_ext_gelu(ctx->ggml_ctx, gate, true);

        x = ggml_mul(ctx->ggml_ctx, x, gate);  // [ne3, ne2, ne1, dim_out]

@@ -214,13 +214,13 @@ public:
        blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim_in, dim_out, bias));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
        // x: [ne3, ne2, ne1, dim_in]
        // return: [ne3, ne2, ne1, dim_out]
        auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);

        x = proj->forward(ctx, x);
-        x = ggml_gelu_inplace(ctx->ggml_ctx, x);
+        x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
        return x;
    }
 };
@@ -258,7 +258,7 @@ public:
        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out, true, false, force_prec_f32, scale));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [ne3, ne2, ne1, dim]
        // return: [ne3, ne2, ne1, dim_out]

@@ -297,9 +297,9 @@ public:
        // to_out_1 is nn.Dropout(), skip for inference
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
-                                struct ggml_tensor* context) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x,
+                         ggml_tensor* context) {
        // x: [N, n_token, query_dim]
        // context: [N, n_context, context_dim]
        // return: [N, n_token, query_dim]
@@ -317,7 +317,7 @@ public:
        auto k = to_k->forward(ctx, context);  // [N, n_context, inner_dim]
        auto v = to_v->forward(ctx, context);  // [N, n_context, inner_dim]

-        x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, nullptr, false, false, ctx->flash_attn_enabled);  // [N, n_token, inner_dim]
+        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]
        return x;
@@ -355,9 +355,9 @@ public:
        }
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
-                                struct ggml_tensor* context) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x,
+                         ggml_tensor* context) {
        // x: [N, n_token, query_dim]
        // context: [N, n_context, context_dim]
        // return: [N, n_token, query_dim]
@@ -406,7 +406,7 @@ protected:
    int64_t context_dim = 768;  // hidden_size, 1024 for VERSION_SD2
    bool use_linear     = false;

-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") {
        auto iter = tensor_storage_map.find(prefix + "proj_out.weight");
        if (iter != tensor_storage_map.end()) {
            int64_t inner_dim = n_head * d_head;
@@ -456,9 +456,9 @@ public:
        }
    }

-    virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                        struct ggml_tensor* x,
-                                        struct ggml_tensor* context) {
+    virtual ggml_tensor* forward(GGMLRunnerContext* ctx,
+                                 ggml_tensor* x,
+                                 ggml_tensor* context) {
        // x: [N, in_channels, h, w]
        // context: [N, max_position(aka n_token), hidden_size(aka context_dim)]
        auto norm     = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm"]);
@@ -510,7 +510,7 @@ public:

 class AlphaBlender : public GGMLBlock {
 protected:
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") override {
        // Get the type of the "mix_factor" tensor from the input tensors map with the specified prefix
        enum ggml_type wtype = GGML_TYPE_F32;
        params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
@@ -530,14 +530,14 @@ public:
        // since mix_factor.shape is [1,], we don't need rearrange using  rearrange_pattern
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x_spatial,
-                                struct ggml_tensor* x_temporal) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x_spatial,
+                         ggml_tensor* x_temporal) {
        // image_only_indicator is always tensor([0.])
        float alpha = get_alpha();
        auto x      = ggml_add(ctx->ggml_ctx,
-                               ggml_scale(ctx->ggml_ctx, x_spatial, alpha),
-                               ggml_scale(ctx->ggml_ctx, x_temporal, 1.0f - alpha));
+                               ggml_ext_scale(ctx->ggml_ctx, x_spatial, alpha),
+                               ggml_ext_scale(ctx->ggml_ctx, x_temporal, 1.0f - alpha));
        return x;
    }
 };
@@ -555,10 +555,10 @@ public:
        blocks["time_mixer"] = std::shared_ptr<GGMLBlock>(new AlphaBlender());
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
-                                struct ggml_tensor* emb,
-                                int num_video_frames) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x,
+                         ggml_tensor* emb,
+                         int num_video_frames) {
        // x: [N, channels, h, w] aka [b*t, channels, h, w]
        // emb: [N, emb_channels] aka [b*t, emb_channels]
        // image_only_indicator is always tensor([0.])
@@ -590,4 +590,4 @@ public:
    }
 };

-#endif  // __COMMON_HPP__
+#endif  // __COMMON_BLOCK_HPP__
@@ -0,0 +1,108 @@
+#ifndef __COMMON_DIT_HPP__
+#define __COMMON_DIT_HPP__
+
+#include "ggml_extend.hpp"
+
+namespace DiT {
+    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];
+        int64_t C = x->ne[2];
+        int64_t H = x->ne[1];
+        int64_t W = x->ne[0];
+        int64_t h = H / ph;
+        int64_t w = W / pw;
+
+        GGML_ASSERT(h * ph == H && w * pw == W);
+
+        x = ggml_reshape_4d(ctx, x, pw, w, ph, h * C * N);     // [N*C*h, ph, w, pw]
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, w, ph, pw]
+        x = ggml_reshape_4d(ctx, x, pw * ph, w * h, C, N);     // [N, C, h*w, ph*pw]
+        if (patch_last) {
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, h*w, C, ph*pw]
+            x = ggml_reshape_3d(ctx, x, pw * ph * C, w * h, N);    // [N, h*w, C*ph*pw]
+        } else {
+            x = ggml_cont(ctx, ggml_ext_torch_permute(ctx, x, 2, 0, 1, 3));  // [N, h*w, C, ph*pw]
+            x = ggml_reshape_3d(ctx, x, C * pw * ph, w * h, N);              // [N, h*w, ph*pw*C]
+        }
+        return x;
+    }
+
+    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];
+        int64_t C = x->ne[0] / ph / pw;
+        int64_t H = h * ph;
+        int64_t W = w * pw;
+
+        GGML_ASSERT(C * ph * pw == x->ne[0]);
+
+        if (patch_last) {
+            x = ggml_reshape_4d(ctx, x, pw * ph, C, w * h, N);     // [N, h*w, C, ph*pw]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, C, h*w, ph*pw]
+        } else {
+            x = ggml_reshape_4d(ctx, x, C, pw * ph, w * h, N);     // [N, h*w, ph*pw, C]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 2, 0, 1, 3));  // [N, C, h*w, ph*pw]
+        }
+
+        x = ggml_reshape_4d(ctx, x, pw, ph, w, h * C * N);     // [N*C*h, w, ph, pw]
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, ph, w, pw]
+        x = ggml_reshape_4d(ctx, x, W, H, C, N);               // [N, C, h*ph, w*pw]
+
+        return x;
+    }
+
+    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];
+
+        int pad_h = (ph - H % ph) % ph;
+        int pad_w = (pw - W % pw) % pw;
+        x         = ggml_ext_pad(ctx->ggml_ctx, x, pad_w, pad_h, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
+        return x;
+    }
+
+    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;
+    }
+
+    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);
+        int64_t w = ((W + pad_w) / pw);
+        x         = unpatchify(ctx, x, h, w, ph, pw, patch_last);  // [N, C, H + pad_h, W + pad_w]
+        x         = ggml_ext_slice(ctx, x, 1, 0, H);               // [N, C, H, W + pad_w]
+        x         = ggml_ext_slice(ctx, x, 0, 0, W);               // [N, C, H, W]
+        return x;
+    }
+}  // namespace DiT
+
+#endif  // __COMMON_DIT_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__
@@ -1,8 +1,7 @@
 #ifndef __CONTROL_HPP__
 #define __CONTROL_HPP__

-#include "common.hpp"
-#include "ggml_extend.hpp"
+#include "common_block.hpp"
 #include "model.h"

 #define CONTROL_NET_GRAPH_SIZE 1536
@@ -165,26 +164,26 @@ public:
        blocks["middle_block_out.0"] = std::shared_ptr<GGMLBlock>(make_zero_conv(ch));
    }

-    struct ggml_tensor* resblock_forward(std::string name,
-                                         GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* x,
-                                         struct ggml_tensor* emb) {
+    ggml_tensor* resblock_forward(std::string name,
+                                  GGMLRunnerContext* ctx,
+                                  ggml_tensor* x,
+                                  ggml_tensor* emb) {
        auto block = std::dynamic_pointer_cast<ResBlock>(blocks[name]);
        return block->forward(ctx, x, emb);
    }

-    struct ggml_tensor* attention_layer_forward(std::string name,
-                                                GGMLRunnerContext* ctx,
-                                                struct ggml_tensor* x,
-                                                struct ggml_tensor* context) {
+    ggml_tensor* attention_layer_forward(std::string name,
+                                         GGMLRunnerContext* ctx,
+                                         ggml_tensor* x,
+                                         ggml_tensor* context) {
        auto block = std::dynamic_pointer_cast<SpatialTransformer>(blocks[name]);
        return block->forward(ctx, x, context);
    }

-    struct ggml_tensor* input_hint_block_forward(GGMLRunnerContext* ctx,
-                                                 struct ggml_tensor* hint,
-                                                 struct ggml_tensor* emb,
-                                                 struct ggml_tensor* context) {
+    ggml_tensor* input_hint_block_forward(GGMLRunnerContext* ctx,
+                                          ggml_tensor* hint,
+                                          ggml_tensor* emb,
+                                          ggml_tensor* context) {
        int num_input_blocks = 15;
        auto h               = hint;
        for (int i = 0; i < num_input_blocks; i++) {
@@ -199,13 +198,13 @@ public:
        return h;
    }

-    std::vector<struct ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                             struct ggml_tensor* x,
-                                             struct ggml_tensor* hint,
-                                             struct ggml_tensor* guided_hint,
-                                             struct ggml_tensor* timesteps,
-                                             struct ggml_tensor* context,
-                                             struct ggml_tensor* y = nullptr) {
+    std::vector<ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                      ggml_tensor* x,
+                                      ggml_tensor* hint,
+                                      ggml_tensor* guided_hint,
+                                      ggml_tensor* timesteps,
+                                      ggml_tensor* context,
+                                      ggml_tensor* y = nullptr) {
        // x: [N, in_channels, h, w] or [N, in_channels/2, h, w]
        // timesteps: [N,]
        // context: [N, max_position, hidden_size] or [1, max_position, hidden_size]. for example, [N, 77, 768]
@@ -247,7 +246,7 @@ public:
            emb = ggml_add(ctx->ggml_ctx, emb, label_emb);  // [N, time_embed_dim]
        }

-        std::vector<struct ggml_tensor*> outs;
+        std::vector<ggml_tensor*> outs;

        if (guided_hint == nullptr) {
            guided_hint = input_hint_block_forward(ctx, hint, emb, context);
@@ -311,11 +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<struct ggml_tensor*> controls;  // (12 input block outputs, 1 middle block output) SD 1.5
-    struct 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,
@@ -329,23 +330,23 @@ struct ControlNet : public GGMLRunner {
        free_control_ctx();
    }

-    void alloc_control_ctx(std::vector<struct ggml_tensor*> outs) {
-        struct ggml_init_params params;
+    void alloc_control_ctx(std::vector<ggml_tensor*> outs) {
+        ggml_init_params params;
        params.mem_size   = static_cast<size_t>(outs.size() * ggml_tensor_overhead()) + 1024 * 1024;
        params.mem_buffer = nullptr;
        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);
@@ -362,8 +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();
    }

@@ -371,33 +374,37 @@ struct ControlNet : public GGMLRunner {
        return "control_net";
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
        control_net.get_param_tensors(tensors, prefix);
    }

-    struct ggml_cgraph* build_graph(struct ggml_tensor* x,
-                                    struct ggml_tensor* hint,
-                                    struct ggml_tensor* timesteps,
-                                    struct ggml_tensor* context,
-                                    struct ggml_tensor* y = nullptr) {
-        struct ggml_cgraph* gf = new_graph_custom(CONTROL_NET_GRAPH_SIZE);
+    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);
@@ -406,36 +413,46 @@ 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,
-                 struct ggml_tensor* x,
-                 struct ggml_tensor* hint,
-                 struct ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
-                 struct ggml_tensor* y,
-                 struct ggml_tensor** output     = nullptr,
-                 struct 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]
        // y: [N, adm_in_channels] or [1, adm_in_channels]
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
            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) {
@@ -463,4 +480,4 @@ struct ControlNet : public GGMLRunner {
    }
 };

-#endif  // __CONTROL_HPP__
+#endif  // __CONTROL_HPP__
@@ -1,44 +1,53 @@
 #ifndef __DIFFUSION_MODEL_H__
 #define __DIFFUSION_MODEL_H__

+#include <optional>
+#include "anima.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 {
-    struct ggml_tensor* x                     = nullptr;
-    struct ggml_tensor* timesteps             = nullptr;
-    struct ggml_tensor* context               = nullptr;
-    struct ggml_tensor* c_concat              = nullptr;
-    struct ggml_tensor* y                     = nullptr;
-    struct ggml_tensor* guidance              = nullptr;
-    std::vector<ggml_tensor*> ref_latents     = {};
-    bool increase_ref_index                   = false;
-    int num_video_frames                      = -1;
-    std::vector<struct ggml_tensor*> controls = {};
-    float control_strength                    = 0.f;
-    struct 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,
-                         struct ggml_tensor** output     = nullptr,
-                         struct ggml_context* output_ctx = nullptr)                     = 0;
-    virtual void alloc_params_buffer()                                                  = 0;
-    virtual void free_params_buffer()                                                   = 0;
-    virtual void free_compute_buffer()                                                  = 0;
-    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) = 0;
-    virtual size_t get_params_buffer_size()                                             = 0;
+    virtual std::string get_desc()                                               = 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;
+    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) = 0;
+    virtual size_t get_params_buffer_size()                                      = 0;
    virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter){};
    virtual int64_t get_adm_in_channels()                            = 0;
-    virtual void set_flash_attn_enabled(bool enabled)                = 0;
+    virtual void set_flash_attention_enabled(bool enabled)           = 0;
    virtual void set_circular_axes(bool circular_x, bool circular_y) = 0;
 };

@@ -68,7 +77,7 @@ struct UNetModel : public DiffusionModel {
        unet.free_compute_buffer();
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
        unet.get_param_tensors(tensors, "model.diffusion_model");
    }

@@ -84,7 +93,7 @@ struct UNetModel : public DiffusionModel {
        return unet.unet.adm_in_channels;
    }

-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
        unet.set_flash_attention_enabled(enabled);
    }

@@ -92,19 +101,20 @@ struct UNetModel : public DiffusionModel {
        unet.set_circular_axes(circular_x, circular_y);
    }

-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
-                 struct 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);
    }
 };

@@ -133,7 +143,7 @@ struct MMDiTModel : public DiffusionModel {
        mmdit.free_compute_buffer();
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
        mmdit.get_param_tensors(tensors, "model.diffusion_model");
    }

@@ -149,7 +159,7 @@ struct MMDiTModel : public DiffusionModel {
        return 768 + 1280;
    }

-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
        mmdit.set_flash_attention_enabled(enabled);
    }

@@ -157,18 +167,17 @@ struct MMDiTModel : public DiffusionModel {
        mmdit.set_circular_axes(circular_x, circular_y);
    }

-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
-                 struct 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);
    }
 };

@@ -199,7 +208,7 @@ struct FluxModel : public DiffusionModel {
        flux.free_compute_buffer();
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
        flux.get_param_tensors(tensors, "model.diffusion_model");
    }

@@ -215,7 +224,7 @@ struct FluxModel : public DiffusionModel {
        return 768;
    }

-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
        flux.set_flash_attention_enabled(enabled);
    }

@@ -223,22 +232,86 @@ struct FluxModel : public DiffusionModel {
        flux.set_circular_axes(circular_x, circular_y);
    }

-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
-                 struct 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);
+    }
+};
+
+struct AnimaModel : public DiffusionModel {
+    std::string prefix;
+    Anima::AnimaRunner anima;
+
+    AnimaModel(ggml_backend_t backend,
+               bool offload_params_to_cpu,
+               const String2TensorStorage& tensor_storage_map = {},
+               const std::string prefix                       = "model.diffusion_model")
+        : prefix(prefix), anima(backend, offload_params_to_cpu, tensor_storage_map, prefix) {
+    }
+
+    std::string get_desc() override {
+        return anima.get_desc();
+    }
+
+    void alloc_params_buffer() override {
+        anima.alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        anima.free_params_buffer();
+    }
+
+    void free_compute_buffer() override {
+        anima.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        anima.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() override {
+        return anima.get_params_buffer_size();
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        anima.set_weight_adapter(adapter);
+    }
+
+    int64_t get_adm_in_channels() override {
+        return 768;
+    }
+
+    void set_flash_attention_enabled(bool enabled) {
+        anima.set_flash_attention_enabled(enabled);
+    }
+
+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        anima.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 anima.compute(n_threads,
+                             *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));
    }
 };

@@ -270,7 +343,7 @@ struct WanModel : public DiffusionModel {
        wan.free_compute_buffer();
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
        wan.get_param_tensors(tensors, prefix);
    }

@@ -286,7 +359,7 @@ struct WanModel : public DiffusionModel {
        return 768;
    }

-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
        wan.set_flash_attention_enabled(enabled);
    }

@@ -294,21 +367,19 @@ struct WanModel : public DiffusionModel {
        wan.set_circular_axes(circular_x, circular_y);
    }

-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
-                 struct 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);
    }
 };

@@ -341,7 +412,7 @@ struct QwenImageModel : public DiffusionModel {
        qwen_image.free_compute_buffer();
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
        qwen_image.get_param_tensors(tensors, prefix);
    }

@@ -357,7 +428,7 @@ struct QwenImageModel : public DiffusionModel {
        return 768;
    }

-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
        qwen_image.set_flash_attention_enabled(enabled);
    }

@@ -365,18 +436,17 @@ struct QwenImageModel : public DiffusionModel {
        qwen_image.set_circular_axes(circular_x, circular_y);
    }

-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
-                 struct 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);
    }
 };

@@ -408,7 +478,7 @@ struct ZImageModel : public DiffusionModel {
        z_image.free_compute_buffer();
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
        z_image.get_param_tensors(tensors, prefix);
    }

@@ -424,7 +494,7 @@ struct ZImageModel : public DiffusionModel {
        return 768;
    }

-    void set_flash_attn_enabled(bool enabled) {
+    void set_flash_attention_enabled(bool enabled) {
        z_image.set_flash_attention_enabled(enabled);
    }

@@ -432,18 +502,17 @@ struct ZImageModel : public DiffusionModel {
        z_image.set_circular_axes(circular_x, circular_y);
    }

-    bool compute(int n_threads,
-                 DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
-                 struct 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);
    }
 };

@@ -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
@@ -27,11 +27,11 @@ public:
        blocks["conv5"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 4 * num_grow_ch, num_feat, {3, 3}, {1, 1}, {1, 1}));
    }

-    struct ggml_tensor* lrelu(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* lrelu(GGMLRunnerContext* ctx, ggml_tensor* x) {
        return ggml_leaky_relu(ctx->ggml_ctx, x, 0.2f, true);
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [n, num_feat, h, w]
        // return: [n, num_feat, h, w]

@@ -51,7 +51,7 @@ public:
        x_cat      = ggml_concat(ctx->ggml_ctx, x_cat, x4, 2);
        auto x5    = conv5->forward(ctx, x_cat);

-        x5 = ggml_add(ctx->ggml_ctx, ggml_scale(ctx->ggml_ctx, x5, 0.2f), x);
+        x5 = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx->ggml_ctx, x5, 0.2f), x);
        return x5;
    }
 };
@@ -64,7 +64,7 @@ public:
        blocks["rdb3"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [n, num_feat, h, w]
        // return: [n, num_feat, h, w]

@@ -76,7 +76,7 @@ public:
        out      = rdb2->forward(ctx, out);
        out      = rdb3->forward(ctx, out);

-        out = ggml_add(ctx->ggml_ctx, ggml_scale(ctx->ggml_ctx, out, 0.2f), x);
+        out = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx->ggml_ctx, out, 0.2f), x);
        return out;
    }
 };
@@ -112,11 +112,11 @@ public:
    int get_scale() { return scale; }
    int get_num_block() { return num_block; }

-    struct ggml_tensor* lrelu(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* lrelu(GGMLRunnerContext* ctx, ggml_tensor* x) {
        return ggml_leaky_relu(ctx->ggml_ctx, x, 0.2f, true);
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [n, num_in_ch, h, w]
        // return: [n, num_out_ch, h*scale, w*scale]
        auto conv_first = std::dynamic_pointer_cast<Conv2d>(blocks["conv_first"]);
@@ -341,28 +341,25 @@ struct ESRGAN : public GGMLRunner {
        return success;
    }

-    struct ggml_cgraph* build_graph(struct ggml_tensor* x) {
+    ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor) {
        if (!rrdb_net)
            return nullptr;
        constexpr int kGraphNodes = 1 << 16;  // 65k
-        struct ggml_cgraph* gf    = new_graph_custom(kGraphNodes);
-        x                         = to_backend(x);
+        ggml_cgraph* gf           = new_graph_custom(kGraphNodes);
+        ggml_tensor* x            = make_input(x_tensor);

-        auto runner_ctx         = get_context();
-        struct ggml_tensor* out = rrdb_net->forward(&runner_ctx, x);
+        auto runner_ctx  = get_context();
+        ggml_tensor* out = rrdb_net->forward(&runner_ctx, x);
        ggml_build_forward_expand(gf, out);
        return gf;
    }

-    bool compute(const int n_threads,
-                 struct ggml_tensor* x,
-                 ggml_tensor** output,
-                 ggml_context* output_ctx = nullptr) {
-        auto get_graph = [&]() -> struct 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__
@@ -1,234 +1,300 @@
-#include <cstddef>
-#include <cstdint>
-#include "ggml.h"
-
-const float wan_21_latent_rgb_proj[16][3] = {
-    {0.015123f, -0.148418f, 0.479828f},
-    {0.003652f, -0.010680f, -0.037142f},
-    {0.212264f, 0.063033f, 0.016779f},
-    {0.232999f, 0.406476f, 0.220125f},
-    {-0.051864f, -0.082384f, -0.069396f},
-    {0.085005f, -0.161492f, 0.010689f},
-    {-0.245369f, -0.506846f, -0.117010f},
-    {-0.151145f, 0.017721f, 0.007207f},
-    {-0.293239f, -0.207936f, -0.421135f},
-    {-0.187721f, 0.050783f, 0.177649f},
-    {-0.013067f, 0.265964f, 0.166578f},
-    {0.028327f, 0.109329f, 0.108642f},
-    {-0.205343f, 0.043991f, 0.148914f},
-    {0.014307f, -0.048647f, -0.007219f},
-    {0.217150f, 0.053074f, 0.319923f},
-    {0.155357f, 0.083156f, 0.064780f}};
-float wan_21_latent_rgb_bias[3] = {-0.270270f, -0.234976f, -0.456853f};
-
-const float wan_22_latent_rgb_proj[48][3] = {
-    {0.017126f, -0.027230f, -0.019257f},
-    {-0.113739f, -0.028715f, -0.022885f},
-    {-0.000106f, 0.021494f, 0.004629f},
-    {-0.013273f, -0.107137f, -0.033638f},
-    {-0.000381f, 0.000279f, 0.025877f},
-    {-0.014216f, -0.003975f, 0.040528f},
-    {0.001638f, -0.000748f, 0.011022f},
-    {0.029238f, -0.006697f, 0.035933f},
-    {0.021641f, -0.015874f, 0.040531f},
-    {-0.101984f, -0.070160f, -0.028855f},
-    {0.033207f, -0.021068f, 0.002663f},
-    {-0.104711f, 0.121673f, 0.102981f},
-    {0.082647f, -0.004991f, 0.057237f},
-    {-0.027375f, 0.031581f, 0.006868f},
-    {-0.045434f, 0.029444f, 0.019287f},
-    {-0.046572f, -0.012537f, 0.006675f},
-    {0.074709f, 0.033690f, 0.025289f},
-    {-0.008251f, -0.002745f, -0.006999f},
-    {0.012685f, -0.061856f, -0.048658f},
-    {0.042304f, -0.007039f, 0.000295f},
-    {-0.007644f, -0.060843f, -0.033142f},
-    {0.159909f, 0.045628f, 0.367541f},
-    {0.095171f, 0.086438f, 0.010271f},
-    {0.006812f, 0.019643f, 0.029637f},
-    {0.003467f, -0.010705f, 0.014252f},
-    {-0.099681f, -0.066272f, -0.006243f},
-    {0.047357f, 0.037040f, 0.000185f},
-    {-0.041797f, -0.089225f, -0.032257f},
-    {0.008928f, 0.017028f, 0.018684f},
-    {-0.042255f, 0.016045f, 0.006849f},
-    {0.011268f, 0.036462f, 0.037387f},
-    {0.011553f, -0.016375f, -0.048589f},
-    {0.046266f, -0.027189f, 0.056979f},
-    {0.009640f, -0.017576f, 0.030324f},
-    {-0.045794f, -0.036083f, -0.010616f},
-    {0.022418f, 0.039783f, -0.032939f},
-    {-0.052714f, -0.015525f, 0.007438f},
-    {0.193004f, 0.223541f, 0.264175f},
-    {-0.059406f, -0.008188f, 0.022867f},
-    {-0.156742f, -0.263791f, -0.007385f},
-    {-0.015717f, 0.016570f, 0.033969f},
-    {0.037969f, 0.109835f, 0.200449f},
-    {-0.000782f, -0.009566f, -0.008058f},
-    {0.010709f, 0.052960f, -0.044195f},
-    {0.017271f, 0.045839f, 0.034569f},
-    {0.009424f, 0.013088f, -0.001714f},
-    {-0.024805f, -0.059378f, -0.033756f},
-    {-0.078293f, 0.029070f, 0.026129f}};
-float wan_22_latent_rgb_bias[3] = {0.013160f, -0.096492f, -0.071323f};
-
-const float flux_latent_rgb_proj[16][3] = {
-    {-0.041168f, 0.019917f, 0.097253f},
-    {0.028096f, 0.026730f, 0.129576f},
-    {0.065618f, -0.067950f, -0.014651f},
-    {-0.012998f, -0.014762f, 0.081251f},
-    {0.078567f, 0.059296f, -0.024687f},
-    {-0.015987f, -0.003697f, 0.005012f},
-    {0.033605f, 0.138999f, 0.068517f},
-    {-0.024450f, -0.063567f, -0.030101f},
-    {-0.040194f, -0.016710f, 0.127185f},
-    {0.112681f, 0.088764f, -0.041940f},
-    {-0.023498f, 0.093664f, 0.025543f},
-    {0.082899f, 0.048320f, 0.007491f},
-    {0.075712f, 0.074139f, 0.081965f},
-    {-0.143501f, 0.018263f, -0.136138f},
-    {-0.025767f, -0.082035f, -0.040023f},
-    {-0.111849f, -0.055589f, -0.032361f}};
-float flux_latent_rgb_bias[3] = {0.024600f, -0.006937f, -0.008089f};
-
-const float flux2_latent_rgb_proj[32][3] = {
-    {0.000736f, -0.008385f, -0.019710f},
-    {-0.001352f, -0.016392f, 0.020693f},
-    {-0.006376f, 0.002428f, 0.036736f},
-    {0.039384f, 0.074167f, 0.119789f},
-    {0.007464f, -0.005705f, -0.004734f},
-    {-0.004086f, 0.005287f, -0.000409f},
-    {-0.032835f, 0.050802f, -0.028120f},
-    {-0.003158f, -0.000835f, 0.000406f},
-    {-0.112840f, -0.084337f, -0.023083f},
-    {0.001462f, -0.006656f, 0.000549f},
-    {-0.009980f, -0.007480f, 0.009702f},
-    {0.032540f, 0.000214f, -0.061388f},
-    {0.011023f, 0.000694f, 0.007143f},
-    {-0.001468f, -0.006723f, -0.001678f},
-    {-0.005921f, -0.010320f, -0.003907f},
-    {-0.028434f, 0.027584f, 0.018457f},
-    {0.014349f, 0.011523f, 0.000441f},
-    {0.009874f, 0.003081f, 0.001507f},
-    {0.002218f, 0.005712f, 0.001563f},
-    {0.053010f, -0.019844f, 0.008683f},
-    {-0.002507f, 0.005384f, 0.000938f},
-    {-0.002177f, -0.011366f, 0.003559f},
-    {-0.000261f, 0.015121f, -0.003240f},
-    {-0.003944f, -0.002083f, 0.005043f},
-    {-0.009138f, 0.011336f, 0.003781f},
-    {0.011429f, 0.003985f, -0.003855f},
-    {0.010518f, -0.005586f, 0.010131f},
-    {0.007883f, 0.002912f, -0.001473f},
-    {-0.003318f, -0.003160f, 0.003684f},
-    {-0.034560f, -0.008740f, 0.012996f},
-    {0.000166f, 0.001079f, -0.012153f},
-    {0.017772f, 0.000937f, -0.011953f}};
-float flux2_latent_rgb_bias[3] = {-0.028738f, -0.098463f, -0.107619f};
-
-// This one was taken straight from
-// https://github.com/Stability-AI/sd3.5/blob/8565799a3b41eb0c7ba976d18375f0f753f56402/sd3_impls.py#L288-L303
-// (MiT Licence)
-const float sd3_latent_rgb_proj[16][3] = {
-    {-0.0645f, 0.0177f, 0.1052f},
-    {0.0028f, 0.0312f, 0.0650f},
-    {0.1848f, 0.0762f, 0.0360f},
-    {0.0944f, 0.0360f, 0.0889f},
-    {0.0897f, 0.0506f, -0.0364f},
-    {-0.0020f, 0.1203f, 0.0284f},
-    {0.0855f, 0.0118f, 0.0283f},
-    {-0.0539f, 0.0658f, 0.1047f},
-    {-0.0057f, 0.0116f, 0.0700f},
-    {-0.0412f, 0.0281f, -0.0039f},
-    {0.1106f, 0.1171f, 0.1220f},
-    {-0.0248f, 0.0682f, -0.0481f},
-    {0.0815f, 0.0846f, 0.1207f},
-    {-0.0120f, -0.0055f, -0.0867f},
-    {-0.0749f, -0.0634f, -0.0456f},
-    {-0.1418f, -0.1457f, -0.1259f},
-};
-float sd3_latent_rgb_bias[3] = {0, 0, 0};
-
-const float sdxl_latent_rgb_proj[4][3] = {
-    {0.258303f, 0.277640f, 0.329699f},
-    {-0.299701f, 0.105446f, 0.014194f},
-    {0.050522f, 0.186163f, -0.143257f},
-    {-0.211938f, -0.149892f, -0.080036f}};
-float sdxl_latent_rgb_bias[3] = {0.144381f, -0.033313f, 0.007061f};
-
-const float sd_latent_rgb_proj[4][3] = {
-    {0.337366f, 0.216344f, 0.257386f},
-    {0.165636f, 0.386828f, 0.046994f},
-    {-0.267803f, 0.237036f, 0.223517f},
-    {-0.178022f, -0.200862f, -0.678514f}};
-float sd_latent_rgb_bias[3] = {-0.017478f, -0.055834f, -0.105825f};
-
-void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int patch_size) {
-    size_t buffer_head = 0;
-
-    uint32_t latent_width  = static_cast<uint32_t>(latents->ne[0]);
-    uint32_t latent_height = static_cast<uint32_t>(latents->ne[1]);
-    uint32_t dim           = static_cast<uint32_t>(latents->ne[ggml_n_dims(latents) - 1]);
-    uint32_t frames        = 1;
-    if (ggml_n_dims(latents) == 4) {
-        frames = static_cast<uint32_t>(latents->ne[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++) {
-                int latent_x = rgb_x / patch_size;
-                int latent_y = rgb_y / patch_size;
-
-                int channel_offset = 0;
-                if (patch_size > 1) {
-                    channel_offset = ((rgb_y % patch_size) * patch_size + (rgb_x % patch_size));
-                }
-
-                size_t latent_id = (latent_x * latents->nb[0] + latent_y * latents->nb[1] + k * latents->nb[2]);
-
-                // should be incremented by 1 for each pixel
-                size_t pixel_id = k * rgb_width * rgb_height + rgb_y * rgb_width + rgb_x;
-
-                float r = 0, g = 0, b = 0;
-                if (latent_rgb_proj != nullptr) {
-                    for (uint32_t d = 0; d < unpatched_dim; d++) {
-                        float value = *(float*)((char*)latents->data + latent_id + (d * patch_size * patch_size + channel_offset) * latents->nb[ggml_n_dims(latents) - 1]);
-                        r += value * latent_rgb_proj[d][0];
-                        g += value * latent_rgb_proj[d][1];
-                        b += value * latent_rgb_proj[d][2];
-                    }
-                } else {
-                    // interpret first 3 channels as RGB
-                    r = *(float*)((char*)latents->data + latent_id + 0 * latents->nb[ggml_n_dims(latents) - 1]);
-                    g = *(float*)((char*)latents->data + latent_id + 1 * latents->nb[ggml_n_dims(latents) - 1]);
-                    b = *(float*)((char*)latents->data + latent_id + 2 * latents->nb[ggml_n_dims(latents) - 1]);
-                }
-                if (latent_rgb_bias != nullptr) {
-                    // bias
-                    r += latent_rgb_bias[0];
-                    g += latent_rgb_bias[1];
-                    b += latent_rgb_bias[2];
-                }
-                // change range
-                r = r * .5f + .5f;
-                g = g * .5f + .5f;
-                b = b * .5f + .5f;
-
-                // clamp rgb values to [0,1] range
-                r = r >= 0 ? r <= 1 ? r : 1 : 0;
-                g = g >= 0 ? g <= 1 ? g : 1 : 0;
-                b = b >= 0 ? b <= 1 ? b : 1 : 0;
-
-                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);
-            }
-        }
-    }
-}
+#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},
+    {0.003652f, -0.010680f, -0.037142f},
+    {0.212264f, 0.063033f, 0.016779f},
+    {0.232999f, 0.406476f, 0.220125f},
+    {-0.051864f, -0.082384f, -0.069396f},
+    {0.085005f, -0.161492f, 0.010689f},
+    {-0.245369f, -0.506846f, -0.117010f},
+    {-0.151145f, 0.017721f, 0.007207f},
+    {-0.293239f, -0.207936f, -0.421135f},
+    {-0.187721f, 0.050783f, 0.177649f},
+    {-0.013067f, 0.265964f, 0.166578f},
+    {0.028327f, 0.109329f, 0.108642f},
+    {-0.205343f, 0.043991f, 0.148914f},
+    {0.014307f, -0.048647f, -0.007219f},
+    {0.217150f, 0.053074f, 0.319923f},
+    {0.155357f, 0.083156f, 0.064780f}};
+float wan_21_latent_rgb_bias[3] = {-0.270270f, -0.234976f, -0.456853f};
+
+const float wan_22_latent_rgb_proj[48][3] = {
+    {0.017126f, -0.027230f, -0.019257f},
+    {-0.113739f, -0.028715f, -0.022885f},
+    {-0.000106f, 0.021494f, 0.004629f},
+    {-0.013273f, -0.107137f, -0.033638f},
+    {-0.000381f, 0.000279f, 0.025877f},
+    {-0.014216f, -0.003975f, 0.040528f},
+    {0.001638f, -0.000748f, 0.011022f},
+    {0.029238f, -0.006697f, 0.035933f},
+    {0.021641f, -0.015874f, 0.040531f},
+    {-0.101984f, -0.070160f, -0.028855f},
+    {0.033207f, -0.021068f, 0.002663f},
+    {-0.104711f, 0.121673f, 0.102981f},
+    {0.082647f, -0.004991f, 0.057237f},
+    {-0.027375f, 0.031581f, 0.006868f},
+    {-0.045434f, 0.029444f, 0.019287f},
+    {-0.046572f, -0.012537f, 0.006675f},
+    {0.074709f, 0.033690f, 0.025289f},
+    {-0.008251f, -0.002745f, -0.006999f},
+    {0.012685f, -0.061856f, -0.048658f},
+    {0.042304f, -0.007039f, 0.000295f},
+    {-0.007644f, -0.060843f, -0.033142f},
+    {0.159909f, 0.045628f, 0.367541f},
+    {0.095171f, 0.086438f, 0.010271f},
+    {0.006812f, 0.019643f, 0.029637f},
+    {0.003467f, -0.010705f, 0.014252f},
+    {-0.099681f, -0.066272f, -0.006243f},
+    {0.047357f, 0.037040f, 0.000185f},
+    {-0.041797f, -0.089225f, -0.032257f},
+    {0.008928f, 0.017028f, 0.018684f},
+    {-0.042255f, 0.016045f, 0.006849f},
+    {0.011268f, 0.036462f, 0.037387f},
+    {0.011553f, -0.016375f, -0.048589f},
+    {0.046266f, -0.027189f, 0.056979f},
+    {0.009640f, -0.017576f, 0.030324f},
+    {-0.045794f, -0.036083f, -0.010616f},
+    {0.022418f, 0.039783f, -0.032939f},
+    {-0.052714f, -0.015525f, 0.007438f},
+    {0.193004f, 0.223541f, 0.264175f},
+    {-0.059406f, -0.008188f, 0.022867f},
+    {-0.156742f, -0.263791f, -0.007385f},
+    {-0.015717f, 0.016570f, 0.033969f},
+    {0.037969f, 0.109835f, 0.200449f},
+    {-0.000782f, -0.009566f, -0.008058f},
+    {0.010709f, 0.052960f, -0.044195f},
+    {0.017271f, 0.045839f, 0.034569f},
+    {0.009424f, 0.013088f, -0.001714f},
+    {-0.024805f, -0.059378f, -0.033756f},
+    {-0.078293f, 0.029070f, 0.026129f}};
+float wan_22_latent_rgb_bias[3] = {0.013160f, -0.096492f, -0.071323f};
+
+const float flux_latent_rgb_proj[16][3] = {
+    {-0.041168f, 0.019917f, 0.097253f},
+    {0.028096f, 0.026730f, 0.129576f},
+    {0.065618f, -0.067950f, -0.014651f},
+    {-0.012998f, -0.014762f, 0.081251f},
+    {0.078567f, 0.059296f, -0.024687f},
+    {-0.015987f, -0.003697f, 0.005012f},
+    {0.033605f, 0.138999f, 0.068517f},
+    {-0.024450f, -0.063567f, -0.030101f},
+    {-0.040194f, -0.016710f, 0.127185f},
+    {0.112681f, 0.088764f, -0.041940f},
+    {-0.023498f, 0.093664f, 0.025543f},
+    {0.082899f, 0.048320f, 0.007491f},
+    {0.075712f, 0.074139f, 0.081965f},
+    {-0.143501f, 0.018263f, -0.136138f},
+    {-0.025767f, -0.082035f, -0.040023f},
+    {-0.111849f, -0.055589f, -0.032361f}};
+float flux_latent_rgb_bias[3] = {0.024600f, -0.006937f, -0.008089f};
+
+const float flux2_latent_rgb_proj[32][3] = {
+    {0.000736f, -0.008385f, -0.019710f},
+    {-0.001352f, -0.016392f, 0.020693f},
+    {-0.006376f, 0.002428f, 0.036736f},
+    {0.039384f, 0.074167f, 0.119789f},
+    {0.007464f, -0.005705f, -0.004734f},
+    {-0.004086f, 0.005287f, -0.000409f},
+    {-0.032835f, 0.050802f, -0.028120f},
+    {-0.003158f, -0.000835f, 0.000406f},
+    {-0.112840f, -0.084337f, -0.023083f},
+    {0.001462f, -0.006656f, 0.000549f},
+    {-0.009980f, -0.007480f, 0.009702f},
+    {0.032540f, 0.000214f, -0.061388f},
+    {0.011023f, 0.000694f, 0.007143f},
+    {-0.001468f, -0.006723f, -0.001678f},
+    {-0.005921f, -0.010320f, -0.003907f},
+    {-0.028434f, 0.027584f, 0.018457f},
+    {0.014349f, 0.011523f, 0.000441f},
+    {0.009874f, 0.003081f, 0.001507f},
+    {0.002218f, 0.005712f, 0.001563f},
+    {0.053010f, -0.019844f, 0.008683f},
+    {-0.002507f, 0.005384f, 0.000938f},
+    {-0.002177f, -0.011366f, 0.003559f},
+    {-0.000261f, 0.015121f, -0.003240f},
+    {-0.003944f, -0.002083f, 0.005043f},
+    {-0.009138f, 0.011336f, 0.003781f},
+    {0.011429f, 0.003985f, -0.003855f},
+    {0.010518f, -0.005586f, 0.010131f},
+    {0.007883f, 0.002912f, -0.001473f},
+    {-0.003318f, -0.003160f, 0.003684f},
+    {-0.034560f, -0.008740f, 0.012996f},
+    {0.000166f, 0.001079f, -0.012153f},
+    {0.017772f, 0.000937f, -0.011953f}};
+float flux2_latent_rgb_bias[3] = {-0.028738f, -0.098463f, -0.107619f};
+
+// This one was taken straight from
+// https://github.com/Stability-AI/sd3.5/blob/8565799a3b41eb0c7ba976d18375f0f753f56402/sd3_impls.py#L288-L303
+// (MiT Licence)
+const float sd3_latent_rgb_proj[16][3] = {
+    {-0.0645f, 0.0177f, 0.1052f},
+    {0.0028f, 0.0312f, 0.0650f},
+    {0.1848f, 0.0762f, 0.0360f},
+    {0.0944f, 0.0360f, 0.0889f},
+    {0.0897f, 0.0506f, -0.0364f},
+    {-0.0020f, 0.1203f, 0.0284f},
+    {0.0855f, 0.0118f, 0.0283f},
+    {-0.0539f, 0.0658f, 0.1047f},
+    {-0.0057f, 0.0116f, 0.0700f},
+    {-0.0412f, 0.0281f, -0.0039f},
+    {0.1106f, 0.1171f, 0.1220f},
+    {-0.0248f, 0.0682f, -0.0481f},
+    {0.0815f, 0.0846f, 0.1207f},
+    {-0.0120f, -0.0055f, -0.0867f},
+    {-0.0749f, -0.0634f, -0.0456f},
+    {-0.1418f, -0.1457f, -0.1259f},
+};
+float sd3_latent_rgb_bias[3] = {0, 0, 0};
+
+const float sdxl_latent_rgb_proj[4][3] = {
+    {0.258303f, 0.277640f, 0.329699f},
+    {-0.299701f, 0.105446f, 0.014194f},
+    {0.050522f, 0.186163f, -0.143257f},
+    {-0.211938f, -0.149892f, -0.080036f}};
+float sdxl_latent_rgb_bias[3] = {0.144381f, -0.033313f, 0.007061f};
+
+const float sd_latent_rgb_proj[4][3] = {
+    {0.337366f, 0.216344f, 0.257386f},
+    {0.165636f, 0.386828f, 0.046994f},
+    {-0.267803f, 0.237036f, 0.223517f},
+    {-0.178022f, -0.200862f, -0.678514f}};
+float sd_latent_rgb_bias[3] = {-0.017478f, -0.055834f, -0.105825f};
+
+void preview_latent_video(uint8_t* buffer, ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int patch_size) {
+    size_t buffer_head = 0;
+
+    uint32_t latent_width  = static_cast<uint32_t>(latents->ne[0]);
+    uint32_t latent_height = static_cast<uint32_t>(latents->ne[1]);
+    uint32_t dim           = static_cast<uint32_t>(latents->ne[ggml_n_dims(latents) - 1]);
+    uint32_t frames        = 1;
+    if (ggml_n_dims(latents) == 4) {
+        frames = static_cast<uint32_t>(latents->ne[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++) {
+                int latent_x = rgb_x / patch_size;
+                int latent_y = rgb_y / patch_size;
+
+                int channel_offset = 0;
+                if (patch_size > 1) {
+                    channel_offset = ((rgb_y % patch_size) * patch_size + (rgb_x % patch_size));
+                }
+
+                size_t latent_id = (latent_x * latents->nb[0] + latent_y * latents->nb[1] + k * latents->nb[2]);
+
+                // should be incremented by 1 for each pixel
+                size_t pixel_id = k * rgb_width * rgb_height + rgb_y * rgb_width + rgb_x;
+
+                float r = 0, g = 0, b = 0;
+                if (latent_rgb_proj != nullptr) {
+                    for (uint32_t d = 0; d < unpatched_dim; d++) {
+                        float value = *(float*)((char*)latents->data + latent_id + (d * patch_size * patch_size + channel_offset) * latents->nb[ggml_n_dims(latents) - 1]);
+                        r += value * latent_rgb_proj[d][0];
+                        g += value * latent_rgb_proj[d][1];
+                        b += value * latent_rgb_proj[d][2];
+                    }
+                } else {
+                    // interpret first 3 channels as RGB
+                    r = *(float*)((char*)latents->data + latent_id + 0 * latents->nb[ggml_n_dims(latents) - 1]);
+                    g = *(float*)((char*)latents->data + latent_id + 1 * latents->nb[ggml_n_dims(latents) - 1]);
+                    b = *(float*)((char*)latents->data + latent_id + 2 * latents->nb[ggml_n_dims(latents) - 1]);
+                }
+                if (latent_rgb_bias != nullptr) {
+                    // bias
+                    r += latent_rgb_bias[0];
+                    g += latent_rgb_bias[1];
+                    b += latent_rgb_bias[2];
+                }
+                // change range
+                r = r * .5f + .5f;
+                g = g * .5f + .5f;
+                b = b * .5f + .5f;
+
+                // clamp rgb values to [0,1] range
+                r = r >= 0 ? r <= 1 ? r : 1 : 0;
+                g = g >= 0 ? g <= 1 ? g : 1 : 0;
+                b = b >= 0 ? b <= 1 ? b : 1 : 0;
+
+                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);
+            }
+        }
+    }
+}
+
+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);
+            }
+        }
+    }
+}
@@ -19,6 +19,7 @@
 #include "json.hpp"
 #include "rope.hpp"
 #include "tokenize_util.h"
+#include "vocab/vocab.h"

 namespace LLM {
    constexpr int LLM_GRAPH_SIZE = 10240;
@@ -193,6 +194,7 @@ namespace LLM {
                        bool padding      = false) {
            if (add_bos_token) {
                tokens.insert(tokens.begin(), BOS_TOKEN_ID);
+                weights.insert(weights.begin(), 1.f);
            }
            if (max_length > 0 && padding) {
                size_t n = static_cast<size_t>(std::ceil(tokens.size() * 1.f / max_length));
@@ -365,7 +367,7 @@ namespace LLM {
            if (merges_utf8_str.size() > 0) {
                load_from_merges(merges_utf8_str);
            } else {
-                load_from_merges(ModelLoader::load_qwen2_merges());
+                load_from_merges(load_qwen2_merges());
            }
        }
    };
@@ -466,7 +468,7 @@ namespace LLM {
            if (merges_utf8_str.size() > 0 && vocab_utf8_str.size() > 0) {
                load_from_merges(merges_utf8_str, vocab_utf8_str);
            } else {
-                load_from_merges(ModelLoader::load_mistral_merges(), ModelLoader::load_mistral_vocab_json());
+                load_from_merges(load_mistral_merges(), load_mistral_vocab_json());
            }
        }
    };
@@ -521,7 +523,7 @@ namespace LLM {
            blocks["down_proj"] = std::shared_ptr<GGMLBlock>(new Linear(intermediate_size, hidden_size, bias));
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
            // x: [N, n_token, hidden_size]
            auto gate_proj = std::dynamic_pointer_cast<Linear>(blocks["gate_proj"]);
            auto up_proj   = std::dynamic_pointer_cast<Linear>(blocks["up_proj"]);
@@ -581,7 +583,7 @@ namespace LLM {
            }
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
            // x: [N*grid_t*grid_h*grid_w, in_channels, temporal_patch_size*patch_size*patch_size]
            // return: [N*grid_t*grid_h*grid_w, embed_dim]
            x = ggml_reshape_4d(ctx->ggml_ctx,
@@ -630,7 +632,7 @@ namespace LLM {
            blocks["mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, dim));
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
            auto ln_q  = std::dynamic_pointer_cast<RMSNorm>(blocks["ln_q"]);
            auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
            auto mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["mlp.2"]);
@@ -638,7 +640,7 @@ namespace LLM {
            x = ln_q->forward(ctx, x);
            x = ggml_reshape_2d(ctx->ggml_ctx, x, hidden_size, ggml_nelements(x) / hidden_size);
            x = mlp_0->forward(ctx, x);
-            x = ggml_gelu(ctx->ggml_ctx, x);
+            x = ggml_ext_gelu(ctx->ggml_ctx, x);
            x = mlp_2->forward(ctx, x);
            return x;
        }
@@ -667,10 +669,10 @@ namespace LLM {
            blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size));
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
-                                    struct ggml_tensor* pe,
-                                    struct ggml_tensor* mask = nullptr) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* pe,
+                             ggml_tensor* mask = nullptr) {
            // x: [N, n_token, hidden_size]
            int64_t n_token = x->ne[1];
            int64_t N       = x->ne[2];
@@ -717,10 +719,10 @@ namespace LLM {
            blocks["norm2"] = std::shared_ptr<GGMLBlock>(new RMSNorm(hidden_size, eps));
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
-                                    struct ggml_tensor* pe,
-                                    struct ggml_tensor* mask = nullptr) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* pe,
+                             ggml_tensor* mask = nullptr) {
            // x: [N, n_token, hidden_size]
            auto attn  = std::dynamic_pointer_cast<VisionAttention>(blocks["attn"]);
            auto mlp   = std::dynamic_pointer_cast<MLP>(blocks["mlp"]);
@@ -777,12 +779,12 @@ namespace LLM {
            blocks["merger"] = std::shared_ptr<GGMLBlock>(new PatchMerger(out_hidden_size, hidden_size, spatial_merge_size));
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* pixel_values,
-                                    struct ggml_tensor* pe,
-                                    struct ggml_tensor* window_index,
-                                    struct ggml_tensor* window_inverse_index,
-                                    struct ggml_tensor* window_mask) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* pixel_values,
+                             ggml_tensor* pe,
+                             ggml_tensor* window_index,
+                             ggml_tensor* window_inverse_index,
+                             ggml_tensor* window_mask) {
            // pixel_values: [grid_t*(H/mh/ph)*(W/mw/pw)*mh*mw, C*pt*ph*pw]
            // window_index: [grid_t*(H/mh/ph)*(W/mw/pw)]
            // window_inverse_index: [grid_t*(H/mh/ph)*(W/mw/pw)]
@@ -835,9 +837,10 @@ namespace LLM {
            }
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
-                                    struct ggml_tensor* input_pos) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* input_pos,
+                             ggml_tensor* attention_mask = nullptr) {
            // x: [N, n_token, hidden_size]
            int64_t n_token = x->ne[1];
            int64_t N       = x->ne[2];
@@ -880,7 +883,7 @@ namespace LLM {
            k = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, k, 0, 2, 1, 3));  // [N, num_kv_heads, n_token, head_dim]
            k = ggml_reshape_3d(ctx->ggml_ctx, k, k->ne[0], k->ne[1], k->ne[2] * k->ne[3]);      // [N*num_kv_heads, n_token, head_dim]

-            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, nullptr, true, true, false);  // [N, n_token, hidden_size]
+            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, attention_mask, true, false);  // [N, n_token, hidden_size]

            x = out_proj->forward(ctx, x);  // [N, n_token, hidden_size]
            return x;
@@ -896,9 +899,10 @@ namespace LLM {
            blocks["post_attention_layernorm"] = std::make_shared<RMSNorm>(params.hidden_size, params.rms_norm_eps);
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
-                                    struct ggml_tensor* input_pos) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* input_pos,
+                             ggml_tensor* attention_mask = nullptr) {
            // x: [N, n_token, hidden_size]
            auto self_attn                = std::dynamic_pointer_cast<Attention>(blocks["self_attn"]);
            auto mlp                      = std::dynamic_pointer_cast<MLP>(blocks["mlp"]);
@@ -907,7 +911,7 @@ namespace LLM {

            auto residual = x;
            x             = input_layernorm->forward(ctx, x);
-            x             = self_attn->forward(ctx, x, input_pos);
+            x             = self_attn->forward(ctx, x, input_pos, attention_mask);
            x             = ggml_add_inplace(ctx->ggml_ctx, x, residual);

            residual = x;
@@ -933,11 +937,12 @@ namespace LLM {
            blocks["norm"] = std::shared_ptr<GGMLBlock>(new RMSNorm(params.hidden_size, params.rms_norm_eps));
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* input_ids,
-                                    struct ggml_tensor* input_pos,
-                                    std::vector<std::pair<int, ggml_tensor*>> image_embeds,
-                                    std::set<int> out_layers) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* input_ids,
+                             ggml_tensor* input_pos,
+                             ggml_tensor* attention_mask,
+                             std::vector<std::pair<int, ggml_tensor*>> image_embeds,
+                             std::set<int> out_layers) {
            // input_ids: [N, n_token]
            // return: [N, n_token, hidden_size]

@@ -990,7 +995,7 @@ namespace LLM {
            for (int i = 0; i < num_layers; i++) {
                auto block = std::dynamic_pointer_cast<TransformerBlock>(blocks["layers." + std::to_string(i)]);

-                x = block->forward(ctx, x, input_pos);
+                x = block->forward(ctx, x, input_pos, attention_mask);
                if (out_layers.find(i + 1) != out_layers.end()) {
                    intermediate_outputs.push_back(x);
                }
@@ -1033,24 +1038,25 @@ namespace LLM {
            }
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* input_ids,
-                                    struct ggml_tensor* input_pos,
-                                    std::vector<std::pair<int, ggml_tensor*>> image_embeds,
-                                    std::set<int> out_layers) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* input_ids,
+                             ggml_tensor* input_pos,
+                             ggml_tensor* attention_mask,
+                             std::vector<std::pair<int, ggml_tensor*>> image_embeds,
+                             std::set<int> out_layers) {
            // input_ids: [N, n_token]
            auto model = std::dynamic_pointer_cast<TextModel>(blocks["model"]);

-            auto x = model->forward(ctx, input_ids, input_pos, image_embeds, out_layers);
+            auto x = model->forward(ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers);
            return x;
        }

-        struct ggml_tensor* vision_forward(GGMLRunnerContext* ctx,
-                                           struct ggml_tensor* pixel_values,
-                                           struct ggml_tensor* pe,
-                                           struct ggml_tensor* window_index,
-                                           struct ggml_tensor* window_inverse_index,
-                                           struct ggml_tensor* window_mask) {
+        ggml_tensor* vision_forward(GGMLRunnerContext* ctx,
+                                    ggml_tensor* pixel_values,
+                                    ggml_tensor* pe,
+                                    ggml_tensor* window_index,
+                                    ggml_tensor* window_inverse_index,
+                                    ggml_tensor* window_mask) {
            GGML_ASSERT(enable_vision);
            auto vision_model = std::dynamic_pointer_cast<VisionModel>(blocks["visual"]);
            return vision_model->forward(ctx, pixel_values, pe, window_index, window_inverse_index, window_mask);
@@ -1063,6 +1069,7 @@ namespace LLM {
        LLM model;

        std::vector<int> input_pos_vec;
+        std::vector<float> attention_mask_vec;
        std::vector<float> window_mask_vec;
        std::vector<int> window_index_vec;
        std::vector<int> window_inverse_index_vec;
@@ -1150,38 +1157,41 @@ namespace LLM {
            return llm_arch_to_str[static_cast<int>(params.arch)];
        }

-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
            model.get_param_tensors(tensors, prefix);
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* input_ids,
-                                    struct ggml_tensor* input_pos,
-                                    std::vector<std::pair<int, ggml_tensor*>> image_embeds,
-                                    std::set<int> out_layers) {
-            auto hidden_states = model.forward(ctx, input_ids, input_pos, image_embeds, out_layers);  // [N, n_token, hidden_size]
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* input_ids,
+                             ggml_tensor* input_pos,
+                             ggml_tensor* attention_mask,
+                             std::vector<std::pair<int, ggml_tensor*>> image_embeds,
+                             std::set<int> out_layers) {
+            auto hidden_states = model.forward(ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers);  // [N, n_token, hidden_size]
            return hidden_states;
        }

-        struct ggml_tensor* vision_forward(GGMLRunnerContext* ctx,
-                                           struct ggml_tensor* pixel_values,
-                                           struct ggml_tensor* input_pos,
-                                           struct ggml_tensor* window_index,
-                                           struct ggml_tensor* window_inverse_index,
-                                           struct ggml_tensor* window_mask) {
+        ggml_tensor* vision_forward(GGMLRunnerContext* ctx,
+                                    ggml_tensor* pixel_values,
+                                    ggml_tensor* input_pos,
+                                    ggml_tensor* window_index,
+                                    ggml_tensor* window_inverse_index,
+                                    ggml_tensor* window_mask) {
            auto hidden_states = model.vision_forward(ctx, pixel_values, input_pos, window_index, window_inverse_index, window_mask);
            return hidden_states;
        }

-        struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
-                                        std::vector<std::pair<int, ggml_tensor*>> image_embeds,
-                                        std::set<int> out_layers) {
-            struct 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* 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);
+            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];
@@ -1205,25 +1215,42 @@ namespace LLM {
                                                input_pos_vec.size());
            set_backend_tensor_data(input_pos, input_pos_vec.data());

+            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++) {
+                    for (int i1 = 0; i1 < n_tokens; i1++) {
+                        float value = 0.f;
+                        if (i0 > i1) {
+                            value = -INFINITY;
+                        }
+                        attention_mask_vec[i1 * n_tokens + i0] = value;
+                    }
+                }
+                attention_mask = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, n_tokens, n_tokens);
+                set_backend_tensor_data(attention_mask, attention_mask_vec.data());
+            }
+
            auto runner_ctx = get_context();

-            struct ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, input_pos, image_embeds, out_layers);
+            ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers);

            ggml_build_forward_expand(gf, hidden_states);

            return gf;
        }

-        bool compute(const int n_threads,
-                     struct ggml_tensor* input_ids,
-                     std::vector<std::pair<int, ggml_tensor*>> image_embeds,
-                     std::set<int> out_layers,
-                     ggml_tensor** output,
-                     ggml_context* output_ctx = nullptr) {
-            auto get_graph = [&]() -> struct ggml_cgraph* {
-                return build_graph(input_ids, image_embeds, out_layers);
+        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) {
@@ -1235,7 +1262,7 @@ namespace LLM {
            return grid_t * grid_h * grid_w;
        }

-        struct ggml_tensor* process_image(struct ggml_context* ctx, struct ggml_tensor* image) {
+        ggml_tensor* process_image(ggml_context* ctx, ggml_tensor* image) {
            // image: [C, H, W]
            // return: [grid_t*(H/mh/ph)*(W/mw/pw)*mh*mw, C*pt*ph*pw], grid_t == 1
            int64_t C  = image->ne[2];
@@ -1262,8 +1289,9 @@ namespace LLM {
            return image;
        }

-        struct ggml_cgraph* build_encode_image_graph(struct ggml_tensor* image) {
-            struct 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);
@@ -1275,8 +1303,6 @@ 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
@@ -1373,26 +1399,24 @@ namespace LLM {
            // pe->data = nullptr;
            set_backend_tensor_data(pe, pe_vec.data());

-            auto runnter_ctx                  = get_context();
-            struct ggml_tensor* hidden_states = vision_forward(&runnter_ctx,
-                                                               pixel_values,
-                                                               pe,
-                                                               window_index,
-                                                               window_inverse_index,
-                                                               window_mask);
+            auto runnter_ctx           = get_context();
+            ggml_tensor* hidden_states = vision_forward(&runnter_ctx,
+                                                        pixel_values,
+                                                        pe,
+                                                        window_index,
+                                                        window_inverse_index,
+                                                        window_mask);
            ggml_build_forward_expand(gf, hidden_states);

            return gf;
        }

-        void encode_image(const int n_threads,
-                          struct ggml_tensor* image,
-                          ggml_tensor** output,
-                          ggml_context* output_ctx = nullptr) {
-            auto get_graph = [&]() -> struct ggml_cgraph* {
+        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));
        }
    };

@@ -1414,7 +1438,7 @@ namespace LLM {
            }
        }

-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
            model.get_param_tensors(tensors, prefix);
        }

@@ -1466,44 +1490,46 @@ namespace LLM {
        }

        void test() {
-            struct ggml_init_params params;
+            ggml_init_params params;
            params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
            params.mem_buffer = nullptr;
            params.no_alloc   = false;

-            struct 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");
-                    struct 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;
@@ -1521,29 +1547,33 @@ namespace LLM {
                    printf("%d ", token);
                }
                printf("\n");
-                auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                struct 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, 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");
-                struct 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);
@@ -1561,14 +1591,16 @@ namespace LLM {
                    printf("%d ", token);
                }
                printf("\n");
-                auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                struct 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, {}, {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;
@@ -1584,14 +1616,16 @@ namespace LLM {
                    printf("%d ", token);
                }
                printf("\n");
-                auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                struct 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, {}, {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;
@@ -1607,14 +1641,16 @@ namespace LLM {
                    printf("%d ", token);
                }
                printf("\n");
-                auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                struct 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, {}, {}, &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);
            }
        }
@@ -9,7 +9,7 @@
 struct LoraModel : public GGMLRunner {
    std::string lora_id;
    float multiplier = 1.0f;
-    std::unordered_map<std::string, struct ggml_tensor*> lora_tensors;
+    std::unordered_map<std::string, ggml_tensor*> lora_tensors;
    std::map<ggml_tensor*, ggml_tensor*> original_tensor_to_final_tensor;
    std::set<std::string> applied_lora_tensors;
    std::string file_path;
@@ -76,13 +76,13 @@ struct LoraModel : public GGMLRunner {
        }

        for (const auto& pair : tensors_to_create) {
-            const auto& name         = pair.first;
-            const auto& ts           = pair.second;
-            struct ggml_tensor* real = ggml_new_tensor(params_ctx,
-                                                       ts.type,
-                                                       ts.n_dims,
-                                                       ts.ne);
-            lora_tensors[name]       = real;
+            const auto& name   = pair.first;
+            const auto& ts     = pair.second;
+            ggml_tensor* real  = ggml_new_tensor(params_ctx,
+                                                 ts.type,
+                                                 ts.n_dims,
+                                                 ts.ne);
+            lora_tensors[name] = real;
        }

        alloc_params_buffer();
@@ -195,7 +195,7 @@ struct LoraModel : public GGMLRunner {
            scale_value *= multiplier;

            auto curr_updown = ggml_ext_merge_lora(ctx, lora_down, lora_up, lora_mid);
-            curr_updown      = ggml_scale_inplace(ctx, curr_updown, scale_value);
+            curr_updown      = ggml_ext_scale(ctx, curr_updown, scale_value, true);

            if (updown == nullptr) {
                updown = curr_updown;
@@ -235,7 +235,7 @@ struct LoraModel : public GGMLRunner {
            float scale_value = 1.0f;
            scale_value *= multiplier;

-            curr_updown = ggml_scale_inplace(ctx, curr_updown, scale_value);
+            curr_updown = ggml_ext_scale(ctx, curr_updown, scale_value, true);

            if (updown == nullptr) {
                updown = curr_updown;
@@ -337,10 +337,10 @@ struct LoraModel : public GGMLRunner {
            }
            scale_value *= multiplier;

-            struct ggml_tensor* updown_1 = ggml_ext_merge_lora(ctx, hada_1_down, hada_1_up, hada_1_mid);
-            struct ggml_tensor* updown_2 = ggml_ext_merge_lora(ctx, hada_2_down, hada_2_up, hada_2_mid);
-            auto curr_updown             = ggml_mul_inplace(ctx, updown_1, updown_2);
-            curr_updown                  = ggml_scale_inplace(ctx, curr_updown, scale_value);
+            ggml_tensor* updown_1 = ggml_ext_merge_lora(ctx, hada_1_down, hada_1_up, hada_1_mid);
+            ggml_tensor* updown_2 = ggml_ext_merge_lora(ctx, hada_2_down, hada_2_up, hada_2_mid);
+            auto curr_updown      = ggml_mul_inplace(ctx, updown_1, updown_2);
+            curr_updown           = ggml_ext_scale(ctx, curr_updown, scale_value, true);
            if (updown == nullptr) {
                updown = curr_updown;
            } else {
@@ -456,7 +456,7 @@ struct LoraModel : public GGMLRunner {
            scale_value *= multiplier;

            auto curr_updown = ggml_ext_kronecker(ctx, lokr_w1, lokr_w2);
-            curr_updown      = ggml_scale_inplace(ctx, curr_updown, scale_value);
+            curr_updown      = ggml_ext_scale(ctx, curr_updown, scale_value, true);

            if (updown == nullptr) {
                updown = curr_updown;
@@ -468,10 +468,10 @@ 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 = true) {
+    ggml_tensor* get_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_tensor* model_tensor, bool with_lora_and_lokr = true) {
        // lora
        ggml_tensor* diff = nullptr;
-        if (with_lora) {
+        if (with_lora_and_lokr) {
            diff = get_lora_weight_diff(model_tensor_name, ctx);
        }
        // diff
@@ -483,7 +483,7 @@ struct LoraModel : public GGMLRunner {
            diff = get_loha_weight_diff(model_tensor_name, ctx);
        }
        // lokr
-        if (diff == nullptr) {
+        if (diff == nullptr && with_lora_and_lokr) {
            diff = get_lokr_weight_diff(model_tensor_name, ctx);
        }
        if (diff != nullptr) {
@@ -514,6 +514,108 @@ struct LoraModel : public GGMLRunner {
            } else {
                key = model_tensor_name + "." + std::to_string(index);
            }
+            bool is_conv2d = forward_params.op_type == WeightAdapter::ForwardParams::op_type_t::OP_CONV2D;
+
+            std::string lokr_w1_name   = "lora." + key + ".lokr_w1";
+            std::string lokr_w1_a_name = "lora." + key + ".lokr_w1_a";
+            // if either of these is found, then we have a lokr lora
+            auto iter   = lora_tensors.find(lokr_w1_name);
+            auto iter_a = lora_tensors.find(lokr_w1_a_name);
+            if (iter != lora_tensors.end() || iter_a != lora_tensors.end()) {
+                std::string lokr_w1_b_name = "lora." + key + ".lokr_w1_b";
+                std::string lokr_w2_name   = "lora." + key + ".lokr_w2";
+                std::string lokr_w2_a_name = "lora." + key + ".lokr_w2_a";
+                std::string lokr_w2_b_name = "lora." + key + ".lokr_w2_b";
+                std::string alpha_name     = "lora." + key + ".alpha";
+
+                ggml_tensor* lokr_w1   = nullptr;
+                ggml_tensor* lokr_w1_a = nullptr;
+                ggml_tensor* lokr_w1_b = nullptr;
+                ggml_tensor* lokr_w2   = nullptr;
+                ggml_tensor* lokr_w2_a = nullptr;
+                ggml_tensor* lokr_w2_b = nullptr;
+
+                if (iter != lora_tensors.end()) {
+                    lokr_w1 = iter->second;
+                }
+                iter = iter_a;
+                if (iter != lora_tensors.end()) {
+                    lokr_w1_a = iter->second;
+                }
+                iter = lora_tensors.find(lokr_w1_b_name);
+                if (iter != lora_tensors.end()) {
+                    lokr_w1_b = iter->second;
+                }
+
+                iter = lora_tensors.find(lokr_w2_name);
+                if (iter != lora_tensors.end()) {
+                    lokr_w2 = iter->second;
+                    if (is_conv2d && lokr_w2->type != GGML_TYPE_F16) {
+                        lokr_w2 = ggml_cast(ctx, lokr_w2, GGML_TYPE_F16);
+                    }
+                }
+                iter = lora_tensors.find(lokr_w2_a_name);
+                if (iter != lora_tensors.end()) {
+                    lokr_w2_a = iter->second;
+                    if (is_conv2d && lokr_w2_a->type != GGML_TYPE_F16) {
+                        lokr_w2_a = ggml_cast(ctx, lokr_w2_a, GGML_TYPE_F16);
+                    }
+                }
+                iter = lora_tensors.find(lokr_w2_b_name);
+                if (iter != lora_tensors.end()) {
+                    lokr_w2_b = iter->second;
+                    if (is_conv2d && lokr_w2_b->type != GGML_TYPE_F16) {
+                        lokr_w2_b = ggml_cast(ctx, lokr_w2_b, GGML_TYPE_F16);
+                    }
+                }
+
+                int rank = 1;
+                if (lokr_w1_b) {
+                    rank = (int)lokr_w1_b->ne[ggml_n_dims(lokr_w1_b) - 1];
+                }
+                if (lokr_w2_b) {
+                    rank = (int)lokr_w2_b->ne[ggml_n_dims(lokr_w2_b) - 1];
+                }
+
+                float scale_value = 1.0f;
+                iter              = lora_tensors.find(alpha_name);
+                if (iter != lora_tensors.end()) {
+                    float alpha = ggml_ext_backend_tensor_get_f32(iter->second);
+                    scale_value = alpha / rank;
+                    applied_lora_tensors.insert(alpha_name);
+                }
+
+                if (rank == 1) {
+                    scale_value = 1.0f;
+                }
+                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);
+                if (out_diff == nullptr) {
+                    out_diff = curr_out_diff;
+                } else {
+                    out_diff = ggml_concat(ctx, out_diff, curr_out_diff, 0);
+                }
+
+                if (lokr_w1)
+                    applied_lora_tensors.insert(lokr_w1_name);
+                if (lokr_w1_a)
+                    applied_lora_tensors.insert(lokr_w1_a_name);
+                if (lokr_w1_b)
+                    applied_lora_tensors.insert(lokr_w1_b_name);
+                if (lokr_w2)
+                    applied_lora_tensors.insert(lokr_w2_name);
+                if (lokr_w2_a)
+                    applied_lora_tensors.insert(lokr_w2_name);
+                if (lokr_w2_b)
+                    applied_lora_tensors.insert(lokr_w2_b_name);
+                applied_lora_tensors.insert(alpha_name);
+
+                index++;
+                continue;
+            }
+
+            // not a lokr, normal lora path

            std::string lora_down_name = "lora." + key + ".lora_down";
            std::string lora_up_name   = "lora." + key + ".lora_up";
@@ -525,9 +627,7 @@ struct LoraModel : public GGMLRunner {
            ggml_tensor* lora_mid  = nullptr;
            ggml_tensor* lora_down = nullptr;

-            bool is_conv2d = forward_params.op_type == WeightAdapter::ForwardParams::op_type_t::OP_CONV2D;
-
-            auto iter = lora_tensors.find(lora_up_name);
+            iter = lora_tensors.find(lora_up_name);
            if (iter != lora_tensors.end()) {
                lora_up = iter->second;
                if (is_conv2d && lora_up->type != GGML_TYPE_F16) {
@@ -634,7 +734,7 @@ struct LoraModel : public GGMLRunner {
                                      forward_params.conv2d.scale);
            }

-            auto curr_out_diff = ggml_scale_inplace(ctx, lx, scale_value);
+            auto curr_out_diff = ggml_ext_scale(ctx, lx, scale_value, true);

            if (out_diff == nullptr) {
                out_diff = curr_out_diff;
@@ -647,9 +747,9 @@ struct LoraModel : public GGMLRunner {
        return out_diff;
    }

-    struct ggml_cgraph* build_lora_graph(const std::map<std::string, ggml_tensor*>& model_tensors, SDVersion version) {
+    ggml_cgraph* build_lora_graph(const std::map<std::string, ggml_tensor*>& model_tensors, SDVersion version) {
        size_t lora_graph_size = LORA_GRAPH_BASE_SIZE + lora_tensors.size() * 10;
-        struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, lora_graph_size, false);
+        ggml_cgraph* gf        = ggml_new_graph_custom(compute_ctx, lora_graph_size, false);

        preprocess_lora_tensors(model_tensors);

@@ -688,11 +788,11 @@ struct LoraModel : public GGMLRunner {
        return gf;
    }

-    void apply(std::map<std::string, struct ggml_tensor*> model_tensors, SDVersion version, int n_threads) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+    void apply(std::map<std::string, ggml_tensor*> model_tensors, SDVersion version, int n_threads) {
+        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;
@@ -741,9 +841,9 @@ public:
        : lora_models(lora_models) {
    }

-    ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name, bool with_lora) {
+    ggml_tensor* patch_weight(ggml_context* ctx, 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);
+            ggml_tensor* diff = lora_model->get_weight_diff(weight_name, ctx, weight, with_lora_and_lokr);
            if (diff == nullptr) {
                continue;
            }
@@ -1,8 +1,7 @@
 #ifndef __LTXV_HPP__
 #define __LTXV_HPP__

-#include "common.hpp"
-#include "ggml_extend.hpp"
+#include "common_block.hpp"

 namespace LTXV {

@@ -27,9 +26,9 @@ namespace LTXV {
                                                                     bias));
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
-                                    bool causal = true) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             bool causal = true) {
            // x: [N*IC, ID, IH, IW]
            // result: [N*OC, OD, OH, OW]
            auto conv = std::dynamic_pointer_cast<Conv3d>(blocks["conv"]);
@@ -27,13 +27,13 @@ public:
        blocks["fc2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_features, out_features, bias));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [N, n_token, in_features]
        auto fc1 = std::dynamic_pointer_cast<Linear>(blocks["fc1"]);
        auto fc2 = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);

        x = fc1->forward(ctx, x);
-        x = ggml_gelu_inplace(ctx->ggml_ctx, x);
+        x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
        x = fc2->forward(ctx, x);
        return x;
    }
@@ -72,7 +72,7 @@ public:
                                                               bias));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [N, C, H, W]
        // return: [N, H*W, embed_dim]
        auto proj = std::dynamic_pointer_cast<Conv2d>(blocks["proj"]);
@@ -111,7 +111,7 @@ public:
        blocks["mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, out_channels, true, true));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* t) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* t) {
        // t: [N, ]
        // return: [N, hidden_size]
        auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
@@ -135,7 +135,7 @@ public:
        blocks["mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size, true, true));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [N, input_dim]
        // return: [N, hidden_size]
        auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
@@ -175,7 +175,7 @@ public:
        }
    }

-    std::vector<struct ggml_tensor*> pre_attention(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    std::vector<ggml_tensor*> pre_attention(GGMLRunnerContext* ctx, ggml_tensor* x) {
        auto qkv_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);

        auto qkv         = qkv_proj->forward(ctx, x);
@@ -198,7 +198,7 @@ public:
        return {q, k, v};
    }

-    struct ggml_tensor* post_attention(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* post_attention(GGMLRunnerContext* ctx, ggml_tensor* x) {
        GGML_ASSERT(!pre_only);

        auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
@@ -208,19 +208,19 @@ public:
    }

    // x: [N, n_token, dim]
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x) {
        auto qkv = pre_attention(ctx, x);
-        x        = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
-        x        = post_attention(ctx, x);                                                                                                                  // [N, n_token, dim]
+        x        = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
+        x        = post_attention(ctx, x);                                                                                                           // [N, n_token, dim]
        return x;
    }
 };

-__STATIC_INLINE__ struct ggml_tensor* modulate(struct ggml_context* ctx,
-                                               struct ggml_tensor* x,
-                                               struct ggml_tensor* shift,
-                                               struct ggml_tensor* scale) {
+__STATIC_INLINE__ ggml_tensor* modulate(ggml_context* ctx,
+                                        ggml_tensor* x,
+                                        ggml_tensor* shift,
+                                        ggml_tensor* scale) {
    // x: [N, L, C]
    // scale: [N, C]
    // shift: [N, C]
@@ -274,8 +274,8 @@ public:
    }

    std::tuple<std::vector<ggml_tensor*>, std::vector<ggml_tensor*>, std::vector<ggml_tensor*>> pre_attention_x(GGMLRunnerContext* ctx,
-                                                                                                                struct ggml_tensor* x,
-                                                                                                                struct ggml_tensor* c) {
+                                                                                                                ggml_tensor* x,
+                                                                                                                ggml_tensor* c) {
        GGML_ASSERT(self_attn);
        // x: [N, n_token, hidden_size]
        // c: [N, hidden_size]
@@ -284,23 +284,19 @@ public:
        auto attn2              = std::dynamic_pointer_cast<SelfAttention>(blocks["attn2"]);
        auto adaLN_modulation_1 = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation.1"]);

-        int64_t n_mods = 9;
-        auto m         = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c));         // [N, n_mods * hidden_size]
-        m              = ggml_reshape_3d(ctx->ggml_ctx, m, c->ne[0], n_mods, c->ne[1]);         // [N, n_mods, hidden_size]
-        m              = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, m, 0, 2, 1, 3));  // [n_mods, N, hidden_size]
+        int n_mods = 9;
+        auto m     = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c));  // [N, n_mods * hidden_size]
+        auto m_vec = ggml_ext_chunk(ctx->ggml_ctx, m, n_mods, 0);

-        int64_t offset = m->nb[1] * m->ne[1];
-        auto shift_msa = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 0);  // [N, hidden_size]
-        auto scale_msa = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 1);  // [N, hidden_size]
-        auto gate_msa  = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 2);  // [N, hidden_size]
-
-        auto shift_mlp = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 3);  // [N, hidden_size]
-        auto scale_mlp = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 4);  // [N, hidden_size]
-        auto gate_mlp  = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 5);  // [N, hidden_size]
-
-        auto shift_msa2 = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 6);  // [N, hidden_size]
-        auto scale_msa2 = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 7);  // [N, hidden_size]
-        auto gate_msa2  = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 8);  // [N, hidden_size]
+        auto shift_msa  = m_vec[0];  // [N, hidden_size]
+        auto scale_msa  = m_vec[1];  // [N, hidden_size]
+        auto gate_msa   = m_vec[2];  // [N, hidden_size]
+        auto shift_mlp  = m_vec[3];  // [N, hidden_size]
+        auto scale_mlp  = m_vec[4];  // [N, hidden_size]
+        auto gate_mlp   = m_vec[5];  // [N, hidden_size]
+        auto shift_msa2 = m_vec[6];  // [N, hidden_size]
+        auto scale_msa2 = m_vec[7];  // [N, hidden_size]
+        auto gate_msa2  = m_vec[8];  // [N, hidden_size]

        auto x_norm = norm1->forward(ctx, x);

@@ -313,31 +309,29 @@ public:
        return {qkv, qkv2, {x, gate_msa, shift_mlp, scale_mlp, gate_mlp, gate_msa2}};
    }

-    std::pair<std::vector<struct ggml_tensor*>, std::vector<struct ggml_tensor*>> pre_attention(GGMLRunnerContext* ctx,
-                                                                                                struct ggml_tensor* x,
-                                                                                                struct ggml_tensor* c) {
+    std::pair<std::vector<ggml_tensor*>, std::vector<ggml_tensor*>> pre_attention(GGMLRunnerContext* ctx,
+                                                                                  ggml_tensor* x,
+                                                                                  ggml_tensor* c) {
        // x: [N, n_token, hidden_size]
        // c: [N, hidden_size]
        auto norm1              = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
        auto attn               = std::dynamic_pointer_cast<SelfAttention>(blocks["attn"]);
        auto adaLN_modulation_1 = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation.1"]);

-        int64_t n_mods = 6;
+        int n_mods = 6;
        if (pre_only) {
            n_mods = 2;
        }
-        auto m = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c));         // [N, n_mods * hidden_size]
-        m      = ggml_reshape_3d(ctx->ggml_ctx, m, c->ne[0], n_mods, c->ne[1]);         // [N, n_mods, hidden_size]
-        m      = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, m, 0, 2, 1, 3));  // [n_mods, N, hidden_size]
+        auto m     = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c));  // [N, n_mods * hidden_size]
+        auto m_vec = ggml_ext_chunk(ctx->ggml_ctx, m, n_mods, 0);

-        int64_t offset = m->nb[1] * m->ne[1];
-        auto shift_msa = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 0);  // [N, hidden_size]
-        auto scale_msa = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 1);  // [N, hidden_size]
+        auto shift_msa = m_vec[0];  // [N, hidden_size]
+        auto scale_msa = m_vec[1];  // [N, hidden_size]
        if (!pre_only) {
-            auto gate_msa  = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 2);  // [N, hidden_size]
-            auto shift_mlp = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 3);  // [N, hidden_size]
-            auto scale_mlp = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 4);  // [N, hidden_size]
-            auto gate_mlp  = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 5);  // [N, hidden_size]
+            auto gate_msa  = m_vec[2];  // [N, hidden_size]
+            auto shift_mlp = m_vec[3];  // [N, hidden_size]
+            auto scale_mlp = m_vec[4];  // [N, hidden_size]
+            auto gate_mlp  = m_vec[5];  // [N, hidden_size]

            auto attn_in = modulate(ctx->ggml_ctx, norm1->forward(ctx, x), shift_msa, scale_msa);

@@ -352,15 +346,15 @@ public:
        }
    }

-    struct ggml_tensor* post_attention_x(GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* attn_out,
-                                         struct ggml_tensor* attn2_out,
-                                         struct ggml_tensor* x,
-                                         struct ggml_tensor* gate_msa,
-                                         struct ggml_tensor* shift_mlp,
-                                         struct ggml_tensor* scale_mlp,
-                                         struct ggml_tensor* gate_mlp,
-                                         struct ggml_tensor* gate_msa2) {
+    ggml_tensor* post_attention_x(GGMLRunnerContext* ctx,
+                                  ggml_tensor* attn_out,
+                                  ggml_tensor* attn2_out,
+                                  ggml_tensor* x,
+                                  ggml_tensor* gate_msa,
+                                  ggml_tensor* shift_mlp,
+                                  ggml_tensor* scale_mlp,
+                                  ggml_tensor* gate_mlp,
+                                  ggml_tensor* gate_msa2) {
        // attn_out: [N, n_token, hidden_size]
        // x: [N, n_token, hidden_size]
        // gate_msa: [N, hidden_size]
@@ -390,13 +384,13 @@ public:
        return x;
    }

-    struct ggml_tensor* post_attention(GGMLRunnerContext* ctx,
-                                       struct ggml_tensor* attn_out,
-                                       struct ggml_tensor* x,
-                                       struct ggml_tensor* gate_msa,
-                                       struct ggml_tensor* shift_mlp,
-                                       struct ggml_tensor* scale_mlp,
-                                       struct ggml_tensor* gate_mlp) {
+    ggml_tensor* post_attention(GGMLRunnerContext* ctx,
+                                ggml_tensor* attn_out,
+                                ggml_tensor* x,
+                                ggml_tensor* gate_msa,
+                                ggml_tensor* shift_mlp,
+                                ggml_tensor* scale_mlp,
+                                ggml_tensor* gate_mlp) {
        // attn_out: [N, n_token, hidden_size]
        // x: [N, n_token, hidden_size]
        // gate_msa: [N, hidden_size]
@@ -422,9 +416,9 @@ public:
        return x;
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
-                                struct ggml_tensor* c) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x,
+                         ggml_tensor* c) {
        // x: [N, n_token, hidden_size]
        // c: [N, hidden_size]
        // return: [N, n_token, hidden_size]
@@ -439,8 +433,8 @@ public:
            auto qkv2          = std::get<1>(qkv_intermediates);
            auto intermediates = std::get<2>(qkv_intermediates);

-            auto attn_out  = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, false, ctx->flash_attn_enabled);     // [N, n_token, dim]
-            auto attn2_out = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv2[0], qkv2[1], qkv2[2], num_heads, nullptr, false, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
+            auto attn_out  = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, ctx->flash_attn_enabled);     // [N, n_token, dim]
+            auto attn2_out = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv2[0], qkv2[1], qkv2[2], num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
            x              = post_attention_x(ctx,
                                              attn_out,
                                              attn2_out,
@@ -456,7 +450,7 @@ public:
            auto qkv               = qkv_intermediates.first;
            auto intermediates     = qkv_intermediates.second;

-            auto attn_out = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
+            auto attn_out = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
            x             = post_attention(ctx,
                                           attn_out,
                                           intermediates[0],
@@ -469,11 +463,11 @@ public:
    }
 };

-__STATIC_INLINE__ std::pair<struct ggml_tensor*, struct ggml_tensor*>
+__STATIC_INLINE__ std::pair<ggml_tensor*, ggml_tensor*>
 block_mixing(GGMLRunnerContext* ctx,
-             struct ggml_tensor* context,
-             struct ggml_tensor* x,
-             struct ggml_tensor* c,
+             ggml_tensor* context,
+             ggml_tensor* x,
+             ggml_tensor* c,
             std::shared_ptr<DismantledBlock> context_block,
             std::shared_ptr<DismantledBlock> x_block) {
    // context: [N, n_context, hidden_size]
@@ -495,31 +489,29 @@ block_mixing(GGMLRunnerContext* ctx,
        x_qkv                    = x_qkv_intermediates.first;
        x_intermediates          = x_qkv_intermediates.second;
    }
-    std::vector<struct ggml_tensor*> qkv;
+    std::vector<ggml_tensor*> qkv;
    for (int i = 0; i < 3; i++) {
        qkv.push_back(ggml_concat(ctx->ggml_ctx, context_qkv[i], x_qkv[i], 1));
    }

-    auto attn         = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], x_block->num_heads, nullptr, false, false, ctx->flash_attn_enabled);  // [N, n_context + n_token, hidden_size]
-    attn              = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, attn, 0, 2, 1, 3));                                                                          // [n_context + n_token, N, hidden_size]
+    auto attn = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], x_block->num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_context + n_token, hidden_size]
+
    auto context_attn = ggml_view_3d(ctx->ggml_ctx,
                                     attn,
                                     attn->ne[0],
-                                     attn->ne[1],
                                     context->ne[1],
+                                     attn->ne[2],
                                     attn->nb[1],
                                     attn->nb[2],
-                                     0);                                                                  // [n_context, N, hidden_size]
-    context_attn      = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, context_attn, 0, 2, 1, 3));  // [N, n_context, hidden_size]
+                                     0);  // [N, n_context, hidden_size]
    auto x_attn       = ggml_view_3d(ctx->ggml_ctx,
                                     attn,
                                     attn->ne[0],
-                                     attn->ne[1],
                                     x->ne[1],
+                                     attn->ne[2],
                                     attn->nb[1],
                                     attn->nb[2],
-                                     attn->nb[2] * context->ne[1]);                                 // [n_token, N, hidden_size]
-    x_attn            = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x_attn, 0, 2, 1, 3));  // [N, n_token, hidden_size]
+                                     context->ne[1] * attn->nb[1]);  // [N, n_token, hidden_size]

    if (!context_block->pre_only) {
        context = context_block->post_attention(ctx,
@@ -534,7 +526,7 @@ block_mixing(GGMLRunnerContext* ctx,
    }

    if (x_block->self_attn) {
-        auto attn2 = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, x_qkv2[0], x_qkv2[1], x_qkv2[2], x_block->num_heads, nullptr, false, false, ctx->flash_attn_enabled);  // [N, n_token, hidden_size]
+        auto attn2 = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, x_qkv2[0], x_qkv2[1], x_qkv2[2], x_block->num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, hidden_size]

        x = x_block->post_attention_x(ctx,
                                      x_attn,
@@ -571,10 +563,10 @@ public:
        blocks["x_block"]       = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, false, self_attn_x));
    }

-    std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                                struct ggml_tensor* context,
-                                                                struct ggml_tensor* x,
-                                                                struct ggml_tensor* c) {
+    std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                  ggml_tensor* context,
+                                                  ggml_tensor* x,
+                                                  ggml_tensor* c) {
        auto context_block = std::dynamic_pointer_cast<DismantledBlock>(blocks["context_block"]);
        auto x_block       = std::dynamic_pointer_cast<DismantledBlock>(blocks["x_block"]);

@@ -594,9 +586,9 @@ public:
        blocks["adaLN_modulation.1"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, 2 * hidden_size));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
-                                struct ggml_tensor* c) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x,
+                         ggml_tensor* c) {
        // x: [N, n_token, hidden_size]
        // c: [N, hidden_size]
        // return: [N, n_token, patch_size * patch_size * out_channels]
@@ -604,13 +596,10 @@ public:
        auto linear             = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
        auto adaLN_modulation_1 = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation.1"]);

-        auto m = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c));         // [N, 2 * hidden_size]
-        m      = ggml_reshape_3d(ctx->ggml_ctx, m, c->ne[0], 2, c->ne[1]);              // [N, 2, hidden_size]
-        m      = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, m, 0, 2, 1, 3));  // [2, N, hidden_size]
-
-        int64_t offset = m->nb[1] * m->ne[1];
-        auto shift     = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 0);  // [N, hidden_size]
-        auto scale     = ggml_view_2d(ctx->ggml_ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 1);  // [N, hidden_size]
+        auto m     = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c));  // [N, 2 * hidden_size]
+        auto m_vec = ggml_ext_chunk(ctx->ggml_ctx, m, 2, 0);
+        auto shift = m_vec[0];  // [N, hidden_size]
+        auto scale = m_vec[1];  // [N, hidden_size]

        x = modulate(ctx->ggml_ctx, norm_final->forward(ctx, x), shift, scale);
        x = linear->forward(ctx, x);
@@ -637,7 +626,7 @@ protected:
    int64_t hidden_size;
    std::string qk_norm;

-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") override {
        enum ggml_type wtype = GGML_TYPE_F32;
        params["pos_embed"]  = ggml_new_tensor_3d(ctx, wtype, hidden_size, num_patchs, 1);
    }
@@ -716,8 +705,8 @@ public:
        blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new FinalLayer(hidden_size, patch_size, out_channels));
    }

-    struct ggml_tensor*
-    cropped_pos_embed(struct ggml_context* ctx,
+    ggml_tensor*
+    cropped_pos_embed(ggml_context* ctx,
                      int64_t h,
                      int64_t w) {
        auto pos_embed = params["pos_embed"];
@@ -756,33 +745,11 @@ public:
        return spatial_pos_embed;
    }

-    struct ggml_tensor* unpatchify(struct ggml_context* ctx,
-                                   struct ggml_tensor* x,
-                                   int64_t h,
-                                   int64_t w) {
-        // x: [N, H*W, patch_size * patch_size * C]
-        // return: [N, C, H, W]
-        int64_t n = x->ne[2];
-        int64_t c = out_channels;
-        int64_t p = patch_size;
-        h         = (h + 1) / p;
-        w         = (w + 1) / p;
-
-        GGML_ASSERT(h * w == x->ne[1]);
-
-        x = ggml_reshape_4d(ctx, x, c, p * p, w * h, n);       // [N, H*W, P*P, C]
-        x = ggml_cont(ctx, ggml_permute(ctx, x, 2, 0, 1, 3));  // [N, C, H*W, P*P]
-        x = ggml_reshape_4d(ctx, x, p, p, w, h * c * n);       // [N*C*H, W, P, P]
-        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*H, P, W, P]
-        x = ggml_reshape_4d(ctx, x, p * w, p * h, c, n);       // [N, C, H*P, W*P]
-        return x;
-    }
-
-    struct ggml_tensor* forward_core_with_concat(GGMLRunnerContext* ctx,
-                                                 struct ggml_tensor* x,
-                                                 struct ggml_tensor* c_mod,
-                                                 struct ggml_tensor* context,
-                                                 std::vector<int> skip_layers = std::vector<int>()) {
+    ggml_tensor* forward_core_with_concat(GGMLRunnerContext* ctx,
+                                          ggml_tensor* x,
+                                          ggml_tensor* c_mod,
+                                          ggml_tensor* context,
+                                          std::vector<int> skip_layers = std::vector<int>()) {
        // x: [N, H*W, hidden_size]
        // context: [N, n_context, d_context]
        // c: [N, hidden_size]
@@ -807,12 +774,12 @@ public:
        return x;
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
-                                struct ggml_tensor* t,
-                                struct ggml_tensor* y        = nullptr,
-                                struct ggml_tensor* context  = nullptr,
-                                std::vector<int> skip_layers = std::vector<int>()) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x,
+                         ggml_tensor* t,
+                         ggml_tensor* y               = nullptr,
+                         ggml_tensor* context         = nullptr,
+                         std::vector<int> skip_layers = std::vector<int>()) {
        // Forward pass of DiT.
        // x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
        // t: (N,) tensor of diffusion timesteps
@@ -822,11 +789,11 @@ public:
        auto x_embedder = std::dynamic_pointer_cast<PatchEmbed>(blocks["x_embedder"]);
        auto t_embedder = std::dynamic_pointer_cast<TimestepEmbedder>(blocks["t_embedder"]);

-        int64_t w = x->ne[0];
-        int64_t h = x->ne[1];
+        int64_t W = x->ne[0];
+        int64_t H = x->ne[1];

        auto patch_embed = x_embedder->forward(ctx, x);                      // [N, H*W, hidden_size]
-        auto pos_embed   = cropped_pos_embed(ctx->ggml_ctx, h, w);           // [1, H*W, hidden_size]
+        auto pos_embed   = cropped_pos_embed(ctx->ggml_ctx, H, W);           // [1, H*W, hidden_size]
        x                = ggml_add(ctx->ggml_ctx, patch_embed, pos_embed);  // [N, H*W, hidden_size]

        auto c = t_embedder->forward(ctx, t);  // [N, hidden_size]
@@ -845,7 +812,7 @@ public:

        x = forward_core_with_concat(ctx, x, c, context, skip_layers);  // (N, H*W, patch_size ** 2 * out_channels)

-        x = unpatchify(ctx->ggml_ctx, x, h, w);  // [N, C, H, W]
+        x = DiT::unpatchify_and_crop(ctx->ggml_ctx, x, H, W, patch_size, patch_size, /*patch_last*/ false);  // [N, C, H, W]

        return x;
    }
@@ -865,89 +832,93 @@ struct MMDiTRunner : public GGMLRunner {
        return "mmdit";
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
        mmdit.get_param_tensors(tensors, prefix);
    }

-    struct ggml_cgraph* build_graph(struct ggml_tensor* x,
-                                    struct ggml_tensor* timesteps,
-                                    struct ggml_tensor* context,
-                                    struct ggml_tensor* y,
-                                    std::vector<int> skip_layers = std::vector<int>()) {
-        struct ggml_cgraph* gf = new_graph_custom(MMDIT_GRAPH_SIZE);
+    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();
-        struct ggml_tensor* out = mmdit.forward(&runner_ctx,
-                                                x,
-                                                timesteps,
-                                                y,
-                                                context,
-                                                skip_layers);
+        auto runner_ctx  = get_context();
+        ggml_tensor* out = mmdit.forward(&runner_ctx,
+                                         x,
+                                         timesteps,
+                                         y,
+                                         context,
+                                         skip_layers);

        ggml_build_forward_expand(gf, out);

        return gf;
    }

-    bool compute(int n_threads,
-                 struct ggml_tensor* x,
-                 struct ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
-                 struct ggml_tensor* y,
-                 struct ggml_tensor** output     = nullptr,
-                 struct 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]
        // y: [N, adm_in_channels] or [1, adm_in_channels]
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
            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() {
-        struct ggml_init_params params;
+        ggml_init_params params;
        params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
        params.mem_buffer = nullptr;
        params.no_alloc   = false;

-        struct 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);

-            struct 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);
        }
    }
@@ -16,10 +16,6 @@
 #include "model.h"
 #include "stable-diffusion.h"
 #include "util.h"
-#include "vocab.hpp"
-#include "vocab_mistral.hpp"
-#include "vocab_qwen.hpp"
-#include "vocab_umt5.hpp"

 #include "ggml-alloc.h"
 #include "ggml-backend.h"
@@ -166,43 +162,7 @@ uint16_t f8_e4m3_to_f16(uint8_t f8) {
 }

 uint16_t f8_e5m2_to_f16(uint8_t fp8) {
-    uint8_t sign     = (fp8 >> 7) & 0x1;
-    uint8_t exponent = (fp8 >> 2) & 0x1F;
-    uint8_t mantissa = fp8 & 0x3;
-
-    uint16_t fp16_sign = sign << 15;
-    uint16_t fp16_exponent;
-    uint16_t fp16_mantissa;
-
-    if (exponent == 0 && mantissa == 0) {  // zero
-        return fp16_sign;
-    }
-
-    if (exponent == 0x1F) {  // NAN and INF
-        fp16_exponent = 0x1F;
-        fp16_mantissa = mantissa ? (mantissa << 8) : 0;
-        return fp16_sign | (fp16_exponent << 10) | fp16_mantissa;
-    }
-
-    if (exponent == 0) {  // subnormal numbers
-        fp16_mantissa = (mantissa << 8);
-        return fp16_sign | fp16_mantissa;
-    }
-
-    // normal numbers
-    int16_t true_exponent = (int16_t)exponent - 15 + 15;
-    if (true_exponent <= 0) {
-        fp16_exponent = 0;
-        fp16_mantissa = (mantissa << 8);
-    } else if (true_exponent >= 0x1F) {
-        fp16_exponent = 0x1F;
-        fp16_mantissa = 0;
-    } else {
-        fp16_exponent = (uint16_t)true_exponent;
-        fp16_mantissa = mantissa << 8;
-    }
-
-    return fp16_sign | (fp16_exponent << 10) | fp16_mantissa;
+    return static_cast<uint16_t>(fp8) << 8;
 }

 void f8_e4m3_to_f16_vec(uint8_t* src, uint16_t* dst, int64_t n) {
@@ -291,7 +251,7 @@ void ModelLoader::add_tensor_storage(const TensorStorage& tensor_storage) {
 }

 bool is_zip_file(const std::string& file_path) {
-    struct zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
+    zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
    if (zip == nullptr) {
        return false;
    }
@@ -376,7 +336,11 @@ bool ModelLoader::init_from_file(const std::string& file_path, const std::string
        LOG_INFO("load %s using checkpoint format", file_path.c_str());
        return init_from_ckpt_file(file_path, prefix);
    } else {
-        LOG_WARN("unknown format %s", file_path.c_str());
+        if (file_exists(file_path)) {
+            LOG_WARN("unknown format %s", file_path.c_str());
+        } else {
+            LOG_WARN("file %s not found", file_path.c_str());
+        }
        return false;
    }
 }
@@ -453,9 +417,9 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
    size_t total_size  = 0;
    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);
-        struct ggml_tensor* dummy = ggml_get_tensor(ctx_meta_, name.c_str());
-        size_t offset             = data_offset + gguf_get_tensor_offset(ctx_gguf_, 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);

        // LOG_DEBUG("%s", name.c_str());

@@ -812,7 +776,7 @@ struct PickleTensorReader {
        }
    }

-    void read_string(const std::string& str, struct zip_t* zip, std::string dir) {
+    void read_string(const std::string& str, zip_t* zip, std::string dir) {
        if (str == "storage") {
            read_global_type = true;
        } else if (str != "state_dict") {
@@ -995,7 +959,7 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
    file_paths_.push_back(file_path);
    size_t file_index = file_paths_.size() - 1;

-    struct zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
+    zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
    if (zip == nullptr) {
        LOG_ERROR("failed to open '%s'", file_path.c_str());
        return false;
@@ -1034,10 +998,14 @@ SDVersion ModelLoader::get_sd_version() {

    bool is_xl                       = false;
    bool is_flux                     = false;
+    bool is_flux2                    = false;
+    bool has_single_block_47         = false;
    bool is_wan                      = false;
    int64_t patch_embedding_channels = 0;
    bool has_img_emb                 = false;
    bool has_middle_block_1          = false;
+    bool has_output_block_311        = false;
+    bool has_output_block_71         = false;

    for (auto& [name, tensor_storage] : tensor_storage_map) {
        if (!(is_xl)) {
@@ -1053,8 +1021,14 @@ SDVersion ModelLoader::get_sd_version() {
            if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
                return VERSION_QWEN_IMAGE;
            }
+            if (tensor_storage.name.find("llm_adapter.blocks.0.cross_attn.q_proj.weight") != std::string::npos) {
+                return VERSION_ANIMA;
+            }
            if (tensor_storage.name.find("model.diffusion_model.double_stream_modulation_img.lin.weight") != std::string::npos) {
-                return VERSION_FLUX2;
+                is_flux2 = true;
+            }
+            if (tensor_storage.name.find("single_blocks.47.linear1.weight") != std::string::npos) {
+                has_single_block_47 = true;
            }
            if (tensor_storage.name.find("model.diffusion_model.double_blocks.0.img_mlp.gate_proj.weight") != std::string::npos) {
                return VERSION_OVIS_IMAGE;
@@ -1094,6 +1068,14 @@ SDVersion ModelLoader::get_sd_version() {
            tensor_storage.name.find("unet.mid_block.resnets.1.") != std::string::npos) {
            has_middle_block_1 = true;
        }
+        if (tensor_storage.name.find("model.diffusion_model.output_blocks.3.1.transformer_blocks.1") != std::string::npos ||
+            tensor_storage.name.find("unet.up_blocks.1.attentions.0.transformer_blocks.1") != std::string::npos) {
+            has_output_block_311 = true;
+        }
+        if (tensor_storage.name.find("model.diffusion_model.output_blocks.7.1") != std::string::npos ||
+            tensor_storage.name.find("unet.up_blocks.2.attentions.1") != std::string::npos) {
+            has_output_block_71 = true;
+        }
        if (tensor_storage.name == "cond_stage_model.transformer.text_model.embeddings.token_embedding.weight" ||
            tensor_storage.name == "cond_stage_model.model.token_embedding.weight" ||
            tensor_storage.name == "text_model.embeddings.token_embedding.weight" ||
@@ -1129,12 +1111,15 @@ SDVersion ModelLoader::get_sd_version() {
            return VERSION_SDXL_PIX2PIX;
        }
        if (!has_middle_block_1) {
+            if (!has_output_block_311) {
+                return VERSION_SDXL_VEGA;
+            }
            return VERSION_SDXL_SSD1B;
        }
        return VERSION_SDXL;
    }

-    if (is_flux) {
+    if (is_flux && !is_flux2) {
        if (input_block_weight.ne[0] == 384) {
            return VERSION_FLUX_FILL;
        }
@@ -1147,6 +1132,13 @@ SDVersion ModelLoader::get_sd_version() {
        return VERSION_FLUX;
    }

+    if (is_flux2) {
+        if (has_single_block_47) {
+            return VERSION_FLUX2;
+        }
+        return VERSION_FLUX2_KLEIN;
+    }
+
    if (token_embedding_weight.ne[0] == 768) {
        if (is_inpaint) {
            return VERSION_SD1_INPAINT;
@@ -1155,6 +1147,9 @@ SDVersion ModelLoader::get_sd_version() {
            return VERSION_SD1_PIX2PIX;
        }
        if (!has_middle_block_1) {
+            if (!has_output_block_71) {
+                return VERSION_SDXS;
+            }
            return VERSION_SD1_TINY_UNET;
        }
        return VERSION_SD1;
@@ -1310,42 +1305,13 @@ void ModelLoader::set_wtype_override(ggml_type wtype, std::string tensor_type_ru
    }
 }

-std::string ModelLoader::load_merges() {
-    std::string merges_utf8_str(reinterpret_cast<const char*>(merges_utf8_c_str), sizeof(merges_utf8_c_str));
-    return merges_utf8_str;
-}
-
-std::string ModelLoader::load_qwen2_merges() {
-    std::string merges_utf8_str(reinterpret_cast<const char*>(qwen2_merges_utf8_c_str), sizeof(qwen2_merges_utf8_c_str));
-    return merges_utf8_str;
-}
-
-std::string ModelLoader::load_mistral_merges() {
-    std::string merges_utf8_str(reinterpret_cast<const char*>(mistral_merges_utf8_c_str), sizeof(mistral_merges_utf8_c_str));
-    return merges_utf8_str;
-}
-
-std::string ModelLoader::load_mistral_vocab_json() {
-    std::string json_str(reinterpret_cast<const char*>(mistral_vocab_json_utf8_c_str), sizeof(mistral_vocab_json_utf8_c_str));
-    return json_str;
-}
-
-std::string ModelLoader::load_t5_tokenizer_json() {
-    std::string json_str(reinterpret_cast<const char*>(t5_tokenizer_json_str), sizeof(t5_tokenizer_json_str));
-    return json_str;
-}
-
-std::string ModelLoader::load_umt5_tokenizer_json() {
-    std::string json_str(reinterpret_cast<const char*>(umt5_tokenizer_json_str), sizeof(umt5_tokenizer_json_str));
-    return json_str;
-}
-
 bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads_p, bool enable_mmap) {
    int64_t process_time_ms = 0;
    std::atomic<int64_t> read_time_ms(0);
    std::atomic<int64_t> memcpy_time_ms(0);
    std::atomic<int64_t> copy_to_backend_time_ms(0);
    std::atomic<int64_t> convert_time_ms(0);
+    std::atomic<uint64_t> bytes_processed(0);

    int num_threads_to_use = n_threads_p > 0 ? n_threads_p : sd_get_num_physical_cores();
    LOG_DEBUG("using %d threads for model loading", num_threads_to_use);
@@ -1412,7 +1378,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
        for (int i = 0; i < n_threads; ++i) {
            workers.emplace_back([&, file_path, is_zip]() {
                std::ifstream file;
-                struct zip_t* zip = nullptr;
+                zip_t* zip = nullptr;
                if (is_zip) {
                    zip = zip_open(file_path.c_str(), 0, 'r');
                    if (zip == nullptr) {
@@ -1557,6 +1523,8 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                        t1 = ggml_time_ms();
                        copy_to_backend_time_ms.fetch_add(t1 - t0);
                    }
+
+                    bytes_processed.fetch_add((uint64_t)nbytes_to_read);
                }
                if (zip != nullptr) {
                    zip_close(zip);
@@ -1569,8 +1537,12 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
            if (current_idx >= file_tensors.size() || failed) {
                break;
            }
-            size_t curr_num = total_tensors_processed + current_idx;
-            pretty_progress(static_cast<int>(curr_num), static_cast<int>(total_tensors_to_process), (ggml_time_ms() - t_start) / 1000.0f / (curr_num + 1e-6f));
+            size_t curr_num       = total_tensors_processed + current_idx;
+            float elapsed_seconds = (ggml_time_ms() - t_start) / 1000.0f;
+            pretty_bytes_progress(static_cast<int>(curr_num),
+                                  static_cast<int>(total_tensors_to_process),
+                                  bytes_processed.load(),
+                                  elapsed_seconds);
            std::this_thread::sleep_for(std::chrono::milliseconds(200));
        }

@@ -1583,7 +1555,10 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
            break;
        }
        total_tensors_processed += file_tensors.size();
-        pretty_progress(static_cast<int>(total_tensors_processed), static_cast<int>(total_tensors_to_process), (ggml_time_ms() - t_start) / 1000.0f / (total_tensors_processed + 1e-6f));
+        pretty_bytes_progress(static_cast<int>(total_tensors_processed),
+                              static_cast<int>(total_tensors_to_process),
+                              bytes_processed.load(),
+                              (ggml_time_ms() - t_start) / 1000.0f);
        if (total_tensors_processed < total_tensors_to_process) {
            printf("\n");
        }
@@ -1600,7 +1575,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
    return success;
 }

-bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
+bool ModelLoader::load_tensors(std::map<std::string, ggml_tensor*>& tensors,
                               std::set<std::string> ignore_tensors,
                               int n_threads,
                               bool enable_mmap) {
@@ -1614,7 +1589,7 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
            tensor_names_in_file.insert(name);
        }

-        struct ggml_tensor* real;
+        ggml_tensor* real;
        if (tensors.find(name) != tensors.end()) {
            real = tensors[name];
        } else {
@@ -28,9 +28,11 @@ enum SDVersion {
    VERSION_SD2,
    VERSION_SD2_INPAINT,
    VERSION_SD2_TINY_UNET,
+    VERSION_SDXS,
    VERSION_SDXL,
    VERSION_SDXL_INPAINT,
    VERSION_SDXL_PIX2PIX,
+    VERSION_SDXL_VEGA,
    VERSION_SDXL_SSD1B,
    VERSION_SVD,
    VERSION_SD3,
@@ -43,14 +45,16 @@ enum SDVersion {
    VERSION_WAN2_2_I2V,
    VERSION_WAN2_2_TI2V,
    VERSION_QWEN_IMAGE,
+    VERSION_ANIMA,
    VERSION_FLUX2,
+    VERSION_FLUX2_KLEIN,
    VERSION_Z_IMAGE,
    VERSION_OVIS_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) {
+    if (version == VERSION_SD1 || version == VERSION_SD1_INPAINT || version == VERSION_SD1_PIX2PIX || version == VERSION_SD1_TINY_UNET || version == VERSION_SDXS) {
        return true;
    }
    return false;
@@ -64,7 +68,7 @@ static inline bool sd_version_is_sd2(SDVersion version) {
 }

 static inline bool sd_version_is_sdxl(SDVersion version) {
-    if (version == VERSION_SDXL || version == VERSION_SDXL_INPAINT || version == VERSION_SDXL_PIX2PIX || version == VERSION_SDXL_SSD1B) {
+    if (version == VERSION_SDXL || version == VERSION_SDXL_INPAINT || version == VERSION_SDXL_PIX2PIX || version == VERSION_SDXL_SSD1B || version == VERSION_SDXL_VEGA) {
        return true;
    }
    return false;
@@ -99,7 +103,7 @@ static inline bool sd_version_is_flux(SDVersion version) {
 }

 static inline bool sd_version_is_flux2(SDVersion version) {
-    if (version == VERSION_FLUX2) {
+    if (version == VERSION_FLUX2 || version == VERSION_FLUX2_KLEIN) {
        return true;
    }
    return false;
@@ -119,6 +123,13 @@ static inline bool sd_version_is_qwen_image(SDVersion version) {
    return false;
 }

+static inline bool sd_version_is_anima(SDVersion version) {
+    if (version == VERSION_ANIMA) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_z_image(SDVersion version) {
    if (version == VERSION_Z_IMAGE) {
        return true;
@@ -143,6 +154,7 @@ static inline bool sd_version_is_dit(SDVersion version) {
        sd_version_is_sd3(version) ||
        sd_version_is_wan(version) ||
        sd_version_is_qwen_image(version) ||
+        sd_version_is_anima(version) ||
        sd_version_is_z_image(version)) {
        return true;
    }
@@ -311,7 +323,7 @@ public:
    String2TensorStorage& get_tensor_storage_map() { return tensor_storage_map; }
    void set_wtype_override(ggml_type wtype, std::string tensor_type_rules = "");
    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads = 0, bool use_mmap = false);
-    bool load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
+    bool load_tensors(std::map<std::string, ggml_tensor*>& tensors,
                      std::set<std::string> ignore_tensors = {},
                      int n_threads                        = 0,
                      bool use_mmap                        = false);
@@ -328,13 +340,6 @@ public:
    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;
-
-    static std::string load_merges();
-    static std::string load_qwen2_merges();
-    static std::string load_mistral_merges();
-    static std::string load_mistral_vocab_json();
-    static std::string load_t5_tokenizer_json();
-    static std::string load_umt5_tokenizer_json();
 };

 #endif  // __MODEL_H__
@@ -653,6 +653,14 @@ std::string convert_diffusers_dit_to_original_lumina2(std::string name) {
    return name;
 }

+std::string convert_other_dit_to_original_anima(std::string name) {
+    static const std::string anima_net_prefix = "net.";
+    if (!starts_with(name, anima_net_prefix)) {
+        name = anima_net_prefix + name;
+    }
+    return name;
+}
+
 std::string convert_diffusion_model_name(std::string name, std::string prefix, SDVersion version) {
    if (sd_version_is_sd1(version) || sd_version_is_sd2(version)) {
        name = convert_diffusers_unet_to_original_sd1(name);
@@ -664,6 +672,8 @@ std::string convert_diffusion_model_name(std::string name, std::string prefix, S
        name = convert_diffusers_dit_to_original_flux(name);
    } else if (sd_version_is_z_image(version)) {
        name = convert_diffusers_dit_to_original_lumina2(name);
+    } else if (sd_version_is_anima(version)) {
+        name = convert_other_dit_to_original_anima(name);
    }
    return name;
 }
@@ -842,6 +852,7 @@ std::string convert_sep_to_dot(std::string name) {
        "conv_in",
        "conv_out",
        "lora_down",
+        "lora_mid",
        "lora_up",
        "diff_b",
        "hada_w1_a",
@@ -997,10 +1008,13 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
    if (is_lora) {
        std::map<std::string, std::string> lora_suffix_map = {
            {".lora_down.weight", ".weight.lora_down"},
+            {".lora_mid.weight", ".weight.lora_mid"},
            {".lora_up.weight", ".weight.lora_up"},
            {".lora.down.weight", ".weight.lora_down"},
+            {".lora.mid.weight", ".weight.lora_mid"},
            {".lora.up.weight", ".weight.lora_up"},
            {"_lora.down.weight", ".weight.lora_down"},
+            {"_lora.mid.weight", ".weight.lora_mid"},
            {"_lora.up.weight", ".weight.lora_up"},
            {".lora_A.weight", ".weight.lora_down"},
            {".lora_B.weight", ".weight.lora_up"},
@@ -1106,7 +1120,11 @@ 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);
-                name = prefix + name;
+                if (version == VERSION_SDXS) {
+                    name = "tae." + name;
+                } else {
+                    name = prefix + name;
+                }
                break;
            }
        }
@@ -21,19 +21,19 @@ public:
        blocks["layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(in_dim));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
        // x: [N, channels, h, w]

        auto fc1        = std::dynamic_pointer_cast<Linear>(blocks["fc1"]);
        auto fc2        = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);
        auto layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["layernorm"]);

-        struct ggml_tensor* r = x;
+        ggml_tensor* r = x;
        // x = ggml_ext_layer_norm(ctx, x, ln_w, ln_b);
        x = layer_norm->forward(ctx, x);
        // x = ggml_add(ctx, ggml_mul_mat(ctx, fc1_w, x),  fc1_b);
        x = fc1->forward(ctx, x);
-        x = ggml_gelu_inplace(ctx->ggml_ctx, x);
+        x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
        x = fc2->forward(ctx, x);
        // x = ggml_add(ctx, ggml_mul_mat(ctx, fc2_w, x),  fc2_b);
        if (use_residue)
@@ -54,8 +54,8 @@ public:
        blocks["1"]   = std::shared_ptr<GGMLBlock>(new Mlp(dim, inner_dim, dim, false));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x) {
        auto norm = std::dynamic_pointer_cast<LayerNorm>(blocks["0"]);
        auto ff   = std::dynamic_pointer_cast<Mlp>(blocks["1"]);

@@ -81,9 +81,9 @@ public:
        blocks["to_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim, false));
    }

-    struct ggml_tensor* reshape_tensor(struct ggml_context* ctx,
-                                       struct ggml_tensor* x,
-                                       int heads) {
+    ggml_tensor* reshape_tensor(ggml_context* ctx,
+                                ggml_tensor* x,
+                                int heads) {
        int64_t ne[4];
        for (int i = 0; i < 4; ++i)
            ne[i] = x->ne[i];
@@ -92,17 +92,17 @@ public:
        return x;
    }

-    std::vector<struct ggml_tensor*> chunk_half(struct ggml_context* ctx,
-                                                struct ggml_tensor* x) {
+    std::vector<ggml_tensor*> chunk_half(ggml_context* ctx,
+                                         ggml_tensor* x) {
        auto tlo = ggml_view_4d(ctx, x, x->ne[0] / 2, x->ne[1], x->ne[2], x->ne[3], x->nb[1], x->nb[2], x->nb[3], 0);
        auto tli = ggml_view_4d(ctx, x, x->ne[0] / 2, x->ne[1], x->ne[2], x->ne[3], x->nb[1], x->nb[2], x->nb[3], x->nb[0] * x->ne[0] / 2);
        return {ggml_cont(ctx, tlo),
                ggml_cont(ctx, tli)};
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
-                                struct ggml_tensor* latents) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x,
+                         ggml_tensor* latents) {
        // x (torch.Tensor): image features
        //     shape (b, n1, D)
        // latent (torch.Tensor): latent features
@@ -129,8 +129,8 @@ public:
        k             = reshape_tensor(ctx->ggml_ctx, k, heads);
        v             = reshape_tensor(ctx->ggml_ctx, v, heads);
        scale         = 1.f / sqrt(sqrt((float)dim_head));
-        k             = ggml_scale_inplace(ctx->ggml_ctx, k, scale);
-        q             = ggml_scale_inplace(ctx->ggml_ctx, q, scale);
+        k             = ggml_ext_scale(ctx->ggml_ctx, k, scale, true);
+        q             = ggml_ext_scale(ctx->ggml_ctx, q, scale, true);
        // auto weight = ggml_mul_mat(ctx, q, k);
        auto weight = ggml_mul_mat(ctx->ggml_ctx, k, q);  // NOTE order of mul is opposite to pytorch

@@ -176,9 +176,9 @@ public:
        }
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* latents,
-                                struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* latents,
+                         ggml_tensor* x) {
        // x: [N, channels, h, w]
        auto proj_in  = std::dynamic_pointer_cast<Linear>(blocks["proj_in"]);
        auto proj_out = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
@@ -225,19 +225,19 @@ public:
            4));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
-                                struct ggml_tensor* last_hidden_state) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x,
+                         ggml_tensor* last_hidden_state) {
        // x: [N, channels, h, w]
        auto token_proj          = std::dynamic_pointer_cast<Mlp>(blocks["token_proj"]);
        auto token_norm          = std::dynamic_pointer_cast<LayerNorm>(blocks["token_norm"]);
        auto perceiver_resampler = std::dynamic_pointer_cast<FacePerceiverResampler>(blocks["perceiver_resampler"]);

-        x                       = token_proj->forward(ctx, x);
-        int64_t nel             = ggml_nelements(x);
-        x                       = ggml_reshape_3d(ctx->ggml_ctx, x, cross_attention_dim, num_tokens, nel / (cross_attention_dim * num_tokens));
-        x                       = token_norm->forward(ctx, x);
-        struct ggml_tensor* out = perceiver_resampler->forward(ctx, x, last_hidden_state);
+        x                = token_proj->forward(ctx, x);
+        int64_t nel      = ggml_nelements(x);
+        x                = ggml_reshape_3d(ctx->ggml_ctx, x, cross_attention_dim, num_tokens, nel / (cross_attention_dim * num_tokens));
+        x                = token_norm->forward(ctx, x);
+        ggml_tensor* out = perceiver_resampler->forward(ctx, x, last_hidden_state);
        if (use_residul)
            out = ggml_add(ctx->ggml_ctx, x, out);
        return out;
@@ -256,9 +256,9 @@ public:
        blocks["layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(embed_dim));
    }

-    struct ggml_tensor* fuse_fn(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* prompt_embeds,
-                                struct ggml_tensor* id_embeds) {
+    ggml_tensor* fuse_fn(GGMLRunnerContext* ctx,
+                         ggml_tensor* prompt_embeds,
+                         ggml_tensor* id_embeds) {
        auto mlp1       = std::dynamic_pointer_cast<FuseBlock>(blocks["mlp1"]);
        auto mlp2       = std::dynamic_pointer_cast<FuseBlock>(blocks["mlp2"]);
        auto layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm"]);
@@ -273,24 +273,24 @@ public:
        return stacked_id_embeds;
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* prompt_embeds,
-                                struct ggml_tensor* id_embeds,
-                                struct ggml_tensor* class_tokens_mask,
-                                struct ggml_tensor* class_tokens_mask_pos,
-                                struct ggml_tensor* left,
-                                struct ggml_tensor* right) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* prompt_embeds,
+                         ggml_tensor* id_embeds,
+                         ggml_tensor* class_tokens_mask,
+                         ggml_tensor* class_tokens_mask_pos,
+                         ggml_tensor* left,
+                         ggml_tensor* right) {
        // x: [N, channels, h, w]

-        struct ggml_tensor* valid_id_embeds = id_embeds;
+        ggml_tensor* valid_id_embeds = id_embeds;
        // # slice out the image token embeddings
        ggml_set_name(class_tokens_mask_pos, "class_tokens_mask_pos");
        ggml_set_name(prompt_embeds, "prompt_embeds");
-        struct ggml_tensor* image_token_embeds = ggml_get_rows(ctx->ggml_ctx, prompt_embeds, class_tokens_mask_pos);
+        ggml_tensor* image_token_embeds = ggml_get_rows(ctx->ggml_ctx, prompt_embeds, class_tokens_mask_pos);
        ggml_set_name(image_token_embeds, "image_token_embeds");
-        valid_id_embeds                       = ggml_reshape_2d(ctx->ggml_ctx, valid_id_embeds, valid_id_embeds->ne[0],
-                                                                ggml_nelements(valid_id_embeds) / valid_id_embeds->ne[0]);
-        struct ggml_tensor* stacked_id_embeds = fuse_fn(ctx, image_token_embeds, valid_id_embeds);
+        valid_id_embeds                = ggml_reshape_2d(ctx->ggml_ctx, valid_id_embeds, valid_id_embeds->ne[0],
+                                                         ggml_nelements(valid_id_embeds) / valid_id_embeds->ne[0]);
+        ggml_tensor* stacked_id_embeds = fuse_fn(ctx, image_token_embeds, valid_id_embeds);

        if (left && right) {
            stacked_id_embeds = ggml_concat(ctx->ggml_ctx, left, stacked_id_embeds, 1);
@@ -301,10 +301,10 @@ public:
            stacked_id_embeds = ggml_concat(ctx->ggml_ctx, stacked_id_embeds, right, 1);
        }

-        class_tokens_mask                         = ggml_cont(ctx->ggml_ctx, ggml_transpose(ctx->ggml_ctx, class_tokens_mask));
-        class_tokens_mask                         = ggml_repeat(ctx->ggml_ctx, class_tokens_mask, prompt_embeds);
-        prompt_embeds                             = ggml_mul(ctx->ggml_ctx, prompt_embeds, class_tokens_mask);
-        struct ggml_tensor* updated_prompt_embeds = ggml_add(ctx->ggml_ctx, prompt_embeds, stacked_id_embeds);
+        class_tokens_mask                  = ggml_cont(ctx->ggml_ctx, ggml_transpose(ctx->ggml_ctx, class_tokens_mask));
+        class_tokens_mask                  = ggml_repeat(ctx->ggml_ctx, class_tokens_mask, prompt_embeds);
+        prompt_embeds                      = ggml_mul(ctx->ggml_ctx, prompt_embeds, class_tokens_mask);
+        ggml_tensor* updated_prompt_embeds = ggml_add(ctx->ggml_ctx, prompt_embeds, stacked_id_embeds);
        ggml_set_name(updated_prompt_embeds, "updated_prompt_embeds");
        return updated_prompt_embeds;
    }
@@ -317,22 +317,22 @@ struct PhotoMakerIDEncoderBlock : public CLIPVisionModelProjection {
        blocks["fuse_module"]         = std::shared_ptr<GGMLBlock>(new FuseModule(2048));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* id_pixel_values,
-                                struct ggml_tensor* prompt_embeds,
-                                struct ggml_tensor* class_tokens_mask,
-                                struct ggml_tensor* class_tokens_mask_pos,
-                                struct ggml_tensor* left,
-                                struct ggml_tensor* right) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* id_pixel_values,
+                         ggml_tensor* prompt_embeds,
+                         ggml_tensor* class_tokens_mask,
+                         ggml_tensor* class_tokens_mask_pos,
+                         ggml_tensor* left,
+                         ggml_tensor* right) {
        // x: [N, channels, h, w]
        auto vision_model        = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
        auto visual_projection   = std::dynamic_pointer_cast<CLIPProjection>(blocks["visual_projection"]);
        auto visual_projection_2 = std::dynamic_pointer_cast<Linear>(blocks["visual_projection_2"]);
        auto fuse_module         = std::dynamic_pointer_cast<FuseModule>(blocks["fuse_module"]);

-        struct ggml_tensor* shared_id_embeds = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
-        struct ggml_tensor* id_embeds        = visual_projection->forward(ctx, shared_id_embeds);    // [N, proj_dim(768)]
-        struct ggml_tensor* id_embeds_2      = visual_projection_2->forward(ctx, shared_id_embeds);  // [N, 1280]
+        ggml_tensor* shared_id_embeds = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
+        ggml_tensor* id_embeds        = visual_projection->forward(ctx, shared_id_embeds);    // [N, proj_dim(768)]
+        ggml_tensor* id_embeds_2      = visual_projection_2->forward(ctx, shared_id_embeds);  // [N, 1280]

        id_embeds   = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, id_embeds, 2, 0, 1, 3));
        id_embeds_2 = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, id_embeds_2, 2, 0, 1, 3));
@@ -340,12 +340,12 @@ struct PhotoMakerIDEncoderBlock : public CLIPVisionModelProjection {
        id_embeds = ggml_concat(ctx->ggml_ctx, id_embeds, id_embeds_2, 2);  // [batch_size, seq_length, 1, 2048] check whether concat at dim 2 is right
        id_embeds = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, id_embeds, 1, 2, 0, 3));

-        struct ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
-                                                                         prompt_embeds,
-                                                                         id_embeds,
-                                                                         class_tokens_mask,
-                                                                         class_tokens_mask_pos,
-                                                                         left, right);
+        ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
+                                                                  prompt_embeds,
+                                                                  id_embeds,
+                                                                  class_tokens_mask,
+                                                                  class_tokens_mask_pos,
+                                                                  left, right);
        return updated_prompt_embeds;
    }
 };
@@ -365,29 +365,29 @@ struct PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock : public CLIPVisionMo
                                                                                        num_tokens));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* id_pixel_values,
-                                struct ggml_tensor* prompt_embeds,
-                                struct ggml_tensor* class_tokens_mask,
-                                struct ggml_tensor* class_tokens_mask_pos,
-                                struct ggml_tensor* id_embeds,
-                                struct ggml_tensor* left,
-                                struct ggml_tensor* right) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* id_pixel_values,
+                         ggml_tensor* prompt_embeds,
+                         ggml_tensor* class_tokens_mask,
+                         ggml_tensor* class_tokens_mask_pos,
+                         ggml_tensor* id_embeds,
+                         ggml_tensor* left,
+                         ggml_tensor* right) {
        // x: [N, channels, h, w]
        auto vision_model      = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
        auto fuse_module       = std::dynamic_pointer_cast<FuseModule>(blocks["fuse_module"]);
        auto qformer_perceiver = std::dynamic_pointer_cast<QFormerPerceiver>(blocks["qformer_perceiver"]);

-        // struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
-        struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values, false);  // [N, hidden_size]
-        id_embeds                             = qformer_perceiver->forward(ctx, id_embeds, last_hidden_state);
+        // ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
+        ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values, false);  // [N, hidden_size]
+        id_embeds                      = qformer_perceiver->forward(ctx, id_embeds, last_hidden_state);

-        struct ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
-                                                                         prompt_embeds,
-                                                                         id_embeds,
-                                                                         class_tokens_mask,
-                                                                         class_tokens_mask_pos,
-                                                                         left, right);
+        ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
+                                                                  prompt_embeds,
+                                                                  id_embeds,
+                                                                  class_tokens_mask,
+                                                                  class_tokens_mask_pos,
+                                                                  left, right);
        return updated_prompt_embeds;
    }
 };
@@ -436,18 +436,17 @@ public:
        return pm_version;
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
        if (pm_version == PM_VERSION_1)
            id_encoder.get_param_tensors(tensors, prefix);
        else if (pm_version == PM_VERSION_2)
            id_encoder2.get_param_tensors(tensors, prefix);
    }

-    struct ggml_cgraph* build_graph(  // struct ggml_allocr* allocr,
-        struct ggml_tensor* id_pixel_values,
-        struct ggml_tensor* prompt_embeds,
-        std::vector<bool>& class_tokens_mask,
-        struct 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();
@@ -458,20 +457,20 @@ public:

        auto runner_ctx = get_context();

-        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+        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;

-        struct ggml_tensor* class_tokens_mask_d = ggml_new_tensor_1d(runner_ctx.ggml_ctx, type, class_tokens_mask.size());
+        ggml_tensor* class_tokens_mask_d = ggml_new_tensor_1d(runner_ctx.ggml_ctx, type, class_tokens_mask.size());

-        struct ggml_tensor* id_pixel_values_d = to_backend(id_pixel_values);
-        struct ggml_tensor* prompt_embeds_d   = to_backend(prompt_embeds);
-        struct ggml_tensor* id_embeds_d       = to_backend(id_embeds);
-
-        struct ggml_tensor* left  = nullptr;
-        struct ggml_tensor* right = nullptr;
+        ggml_tensor* left  = nullptr;
+        ggml_tensor* right = nullptr;
        for (int i = 0; i < class_tokens_mask.size(); i++) {
            if (class_tokens_mask[i]) {
                // printf(" 1,");
@@ -495,7 +494,7 @@ public:
            right = ggml_new_tensor_3d(runner_ctx.ggml_ctx, type,
                                       hidden_size, seq_length - ctmpos[ctmpos.size() - 1] - 1, 1);
        }
-        struct ggml_tensor* class_tokens_mask_pos = ggml_new_tensor_1d(runner_ctx.ggml_ctx, GGML_TYPE_I32, ctmpos.size());
+        ggml_tensor* class_tokens_mask_pos = ggml_new_tensor_1d(runner_ctx.ggml_ctx, GGML_TYPE_I32, ctmpos.size());

        {
            if (type == GGML_TYPE_F16)
@@ -526,21 +525,21 @@ public:
                }
            }
        }
-        struct ggml_tensor* updated_prompt_embeds = nullptr;
+        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);
@@ -548,25 +547,21 @@ public:
        return gf;
    }

-    bool compute(const int n_threads,
-                 struct ggml_tensor* id_pixel_values,
-                 struct ggml_tensor* prompt_embeds,
-                 struct ggml_tensor* id_embeds,
-                 std::vector<bool>& class_tokens_mask,
-                 struct ggml_tensor** updated_prompt_embeds,
-                 ggml_context* output_ctx) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
-            // return build_graph(compute_allocr, id_pixel_values, prompt_embeds, class_tokens_mask);
+    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(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));
    }
 };

 struct PhotoMakerIDEmbed : public GGMLRunner {
-    std::map<std::string, struct ggml_tensor*> tensors;
+    std::map<std::string, ggml_tensor*> tensors;
    std::string file_path;
    ModelLoader* model_loader;
    bool load_failed = false;
@@ -606,11 +601,11 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
            }
            if (dry_run) {
                std::lock_guard<std::mutex> lock(tensor_mutex);
-                struct ggml_tensor* real = ggml_new_tensor(params_ctx,
-                                                           tensor_storage.type,
-                                                           tensor_storage.n_dims,
-                                                           tensor_storage.ne);
-                tensors[name]            = real;
+                ggml_tensor* real = ggml_new_tensor(params_ctx,
+                                                    tensor_storage.type,
+                                                    tensor_storage.n_dims,
+                                                    tensor_storage.ne);
+                tensors[name]     = real;
            } else {
                auto real   = tensors[name];
                *dst_tensor = real;
@@ -629,8 +624,8 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
        return true;
    }

-    struct ggml_tensor* get() {
-        std::map<std::string, struct ggml_tensor*>::iterator pos;
+    ggml_tensor* get() {
+        std::map<std::string, ggml_tensor*>::iterator pos;
        pos = tensors.find("pmid.id_embeds");
        if (pos != tensors.end())
            return pos->second;
@@ -0,0 +1,278 @@
+#ifndef __PREPROCESSING_HPP__
+#define __PREPROCESSING_HPP__
+
+#include <cmath>
+#include <limits>
+
+#include "ggml_extend.hpp"
+
+#define M_PI_ 3.14159265358979323846f
+
+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;
+}
+
+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 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);
+    GGML_ASSERT(image_data != nullptr);
+
+    int width   = static_cast<int>(tensor.shape()[0]);
+    int height  = static_cast<int>(tensor.shape()[1]);
+    int channel = static_cast<int>(tensor.shape()[2]);
+    for (int y = 0; y < height; ++y) {
+        for (int x = 0; x < width; ++x) {
+            for (int c = 0; c < channel; ++c) {
+                float value                               = preprocessing_get_4d(tensor, x, y, c, 0);
+                value                                     = std::min(1.0f, std::max(0.0f, value));
+                image_data[(y * width + x) * channel + c] = static_cast<uint8_t>(std::round(value * 255.0f));
+            }
+        }
+    }
+}
+
+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_ * 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_size; ++x) {
+            float gy = static_cast<float>(-ks_mid + x);
+            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;
+}
+
+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;
+            preprocessing_set_4d(grayscale, gray, ix, iy, 0, 0);
+        }
+    }
+    return grayscale;
+}
+
+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;
+}
+
+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;
+}
+
+static inline void normalize_tensor(sd::Tensor<float>* g) {
+    GGML_ASSERT(g != nullptr);
+    if (g->empty()) {
+        return;
+    }
+    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);
+}
+
+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;
+
+            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 = 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;
+}
+
+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 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 (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 (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);
+            }
+        }
+    }
+
+    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) {
+                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) {
+    float kX[9] = {
+        -1, 0, 1,
+        -2, 0, 2,
+        -1, 0, 1};
+
+    float kY[9] = {
+        1, 2, 1,
+        0, 0, 0,
+        -1, -2, -1};
+
+    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;
+            for (uint32_t c = 0; c < img.channel; ++c) {
+                preprocessing_set_4d(image, gray, ix, iy, c, 0);
+            }
+        }
+    }
+
+    preprocessing_tensor_to_sd_image(image, img.data);
+    return true;
+}
+
+#endif  // __PREPROCESSING_HPP__
@@ -3,9 +3,8 @@

 #include <memory>

-#include "common.hpp"
+#include "common_block.hpp"
 #include "flux.hpp"
-#include "ggml_extend.hpp"

 namespace Qwen {
    constexpr int QWEN_IMAGE_GRAPH_SIZE = 20480;
@@ -27,9 +26,9 @@ namespace Qwen {
            blocks["linear_2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, out_dim, sample_proj_bias));
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* sample,
-                                    struct ggml_tensor* condition = nullptr) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* sample,
+                             ggml_tensor* condition = nullptr) {
            if (condition != nullptr) {
                auto cond_proj = std::dynamic_pointer_cast<Linear>(blocks["cond_proj"]);
                sample         = ggml_add(ctx->ggml_ctx, sample, cond_proj->forward(ctx, condition));
@@ -50,8 +49,8 @@ namespace Qwen {
            blocks["timestep_embedder"] = std::shared_ptr<GGMLBlock>(new TimestepEmbedding(256, embedding_dim));
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* timesteps) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* timesteps) {
            // timesteps: [N,]
            // return: [N, embedding_dim]
            auto timestep_embedder = std::dynamic_pointer_cast<TimestepEmbedding>(blocks["timestep_embedder"]);
@@ -108,10 +107,10 @@ namespace Qwen {
        }

        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                      struct ggml_tensor* img,
-                                                      struct ggml_tensor* txt,
-                                                      struct ggml_tensor* pe,
-                                                      struct ggml_tensor* mask = nullptr) {
+                                                      ggml_tensor* img,
+                                                      ggml_tensor* txt,
+                                                      ggml_tensor* pe,
+                                                      ggml_tensor* mask = nullptr) {
            // img: [N, n_img_token, hidden_size]
            // txt: [N, n_txt_token, hidden_size]
            // pe: [n_img_token + n_txt_token, d_head/2, 2, 2]
@@ -162,26 +161,25 @@ namespace Qwen {
            auto k = ggml_concat(ctx->ggml_ctx, txt_k, img_k, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
            auto v = ggml_concat(ctx->ggml_ctx, txt_v, img_v, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]

-            auto attn         = Rope::attention(ctx, q, k, v, pe, mask, (1.0f / 128.f));                  // [N, n_txt_token + n_img_token, n_head*d_head]
-            attn              = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, attn, 0, 2, 1, 3));  // [n_txt_token + n_img_token, N, hidden_size]
+            auto attn         = Rope::attention(ctx, q, k, v, pe, mask, (1.0f / 128.f));  // [N, n_txt_token + n_img_token, n_head*d_head]
            auto txt_attn_out = ggml_view_3d(ctx->ggml_ctx,
                                             attn,
                                             attn->ne[0],
-                                             attn->ne[1],
                                             txt->ne[1],
+                                             attn->ne[2],
                                             attn->nb[1],
                                             attn->nb[2],
-                                             0);                                                                  // [n_txt_token, N, hidden_size]
-            txt_attn_out      = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, txt_attn_out, 0, 2, 1, 3));  // [N, n_txt_token, hidden_size]
+                                             0);  // [N, n_txt_token, n_head*d_head]
            auto img_attn_out = ggml_view_3d(ctx->ggml_ctx,
                                             attn,
                                             attn->ne[0],
-                                             attn->ne[1],
                                             img->ne[1],
+                                             attn->ne[2],
                                             attn->nb[1],
                                             attn->nb[2],
-                                             attn->nb[2] * txt->ne[1]);                                           // [n_img_token, N, hidden_size]
-            img_attn_out      = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, img_attn_out, 0, 2, 1, 3));  // [N, n_img_token, hidden_size]
+                                             txt->ne[1] * attn->nb[1]);  // [N, n_img_token, n_head*d_head]
+            img_attn_out      = ggml_cont(ctx->ggml_ctx, img_attn_out);
+            txt_attn_out      = ggml_cont(ctx->ggml_ctx, txt_attn_out);

            img_attn_out = to_out_0->forward(ctx, img_attn_out);
            txt_attn_out = to_add_out->forward(ctx, txt_attn_out);
@@ -213,7 +211,7 @@ namespace Qwen {

            blocks["txt_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
            blocks["txt_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
-            blocks["txt_mlp"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim, 4, FeedForward::Activation::GELU));
+            blocks["txt_mlp"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim, 4, FeedForward::Activation::GELU, true));

            blocks["attn"] = std::shared_ptr<GGMLBlock>(new QwenImageAttention(dim,
                                                                               attention_head_dim,
@@ -251,11 +249,11 @@ namespace Qwen {
        }

        virtual std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                              struct ggml_tensor* img,
-                                                              struct ggml_tensor* txt,
-                                                              struct ggml_tensor* t_emb,
-                                                              struct ggml_tensor* pe,
-                                                              struct ggml_tensor* modulate_index = nullptr) {
+                                                              ggml_tensor* img,
+                                                              ggml_tensor* txt,
+                                                              ggml_tensor* t_emb,
+                                                              ggml_tensor* pe,
+                                                              ggml_tensor* modulate_index = nullptr) {
            // img: [N, n_img_token, hidden_size]
            // txt: [N, n_txt_token, hidden_size]
            // pe: [n_img_token + n_txt_token, d_head/2, 2, 2]
@@ -327,9 +325,9 @@ namespace Qwen {
            blocks["linear"] = std::shared_ptr<GGMLBlock>(new Linear(conditioning_embedding_dim, embedding_dim * 2, bias));
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
-                                    struct ggml_tensor* c) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* c) {
            // x: [N, n_token, hidden_size]
            // c: [N, hidden_size]
            // return: [N, n_token, patch_size * patch_size * out_channels]
@@ -391,75 +389,12 @@ namespace Qwen {
            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, params.patch_size * params.patch_size * params.out_channels));
        }

-        struct ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
-                                              struct ggml_tensor* x) {
-            int64_t W = x->ne[0];
-            int64_t H = x->ne[1];
-
-            int pad_h = (params.patch_size - H % params.patch_size) % params.patch_size;
-            int pad_w = (params.patch_size - W % params.patch_size) % params.patch_size;
-            x         = ggml_ext_pad(ctx->ggml_ctx, x, pad_w, pad_h, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
-            return x;
-        }
-
-        struct ggml_tensor* patchify(struct ggml_context* ctx,
-                                     struct ggml_tensor* x) {
-            // x: [N, C, H, W]
-            // return: [N, h*w, C * patch_size * patch_size]
-            int64_t N = x->ne[3];
-            int64_t C = x->ne[2];
-            int64_t H = x->ne[1];
-            int64_t W = x->ne[0];
-            int64_t p = params.patch_size;
-            int64_t h = H / params.patch_size;
-            int64_t w = W / params.patch_size;
-
-            GGML_ASSERT(h * p == H && w * p == W);
-
-            x = ggml_reshape_4d(ctx, x, p, w, p, h * C * N);       // [N*C*h, p, w, p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, w, p, p]
-            x = ggml_reshape_4d(ctx, x, p * p, w * h, C, N);       // [N, C, h*w, p*p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, h*w, C, p*p]
-            x = ggml_reshape_3d(ctx, x, p * p * C, w * h, N);      // [N, h*w, C*p*p]
-            return x;
-        }
-
-        struct ggml_tensor* process_img(GGMLRunnerContext* ctx,
-                                        struct ggml_tensor* x) {
-            x = pad_to_patch_size(ctx, x);
-            x = patchify(ctx->ggml_ctx, x);
-            return x;
-        }
-
-        struct ggml_tensor* unpatchify(struct ggml_context* ctx,
-                                       struct ggml_tensor* x,
-                                       int64_t h,
-                                       int64_t w) {
-            // x: [N, h*w, C*patch_size*patch_size]
-            // return: [N, C, H, W]
-            int64_t N = x->ne[2];
-            int64_t C = x->ne[0] / params.patch_size / params.patch_size;
-            int64_t H = h * params.patch_size;
-            int64_t W = w * params.patch_size;
-            int64_t p = params.patch_size;
-
-            GGML_ASSERT(C * p * p == x->ne[0]);
-
-            x = ggml_reshape_4d(ctx, x, p * p, C, w * h, N);       // [N, h*w, C, p*p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, C, h*w, p*p]
-            x = ggml_reshape_4d(ctx, x, p, p, w, h * C * N);       // [N*C*h, w, p, p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, p, w, p]
-            x = ggml_reshape_4d(ctx, x, W, H, C, N);               // [N, C, h*p, w*p]
-
-            return x;
-        }
-
-        struct ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* x,
-                                         struct ggml_tensor* timestep,
-                                         struct ggml_tensor* context,
-                                         struct ggml_tensor* pe,
-                                         struct ggml_tensor* modulate_index = nullptr) {
+        ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
+                                  ggml_tensor* x,
+                                  ggml_tensor* timestep,
+                                  ggml_tensor* context,
+                                  ggml_tensor* pe,
+                                  ggml_tensor* modulate_index = nullptr) {
            auto time_text_embed = std::dynamic_pointer_cast<QwenTimestepProjEmbeddings>(blocks["time_text_embed"]);
            auto txt_norm        = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm"]);
            auto img_in          = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
@@ -469,7 +404,7 @@ namespace Qwen {

            auto t_emb = time_text_embed->forward(ctx, timestep);
            if (params.zero_cond_t) {
-                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros(ctx->ggml_ctx, timestep->ne[0], timestep->ne[1], timestep->ne[2], timestep->ne[3]));
+                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros_like(ctx->ggml_ctx, timestep));
                t_emb        = ggml_concat(ctx->ggml_ctx, t_emb, t_emb_0, 1);
            }
            auto img = img_in->forward(ctx, x);
@@ -494,13 +429,13 @@ namespace Qwen {
            return img;
        }

-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
-                                    struct ggml_tensor* timestep,
-                                    struct ggml_tensor* context,
-                                    struct ggml_tensor* pe,
-                                    std::vector<ggml_tensor*> ref_latents = {},
-                                    struct ggml_tensor* modulate_index    = nullptr) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timestep,
+                             ggml_tensor* context,
+                             ggml_tensor* pe,
+                             std::vector<ggml_tensor*> ref_latents = {},
+                             ggml_tensor* modulate_index           = nullptr) {
            // Forward pass of DiT.
            // x: [N, C, H, W]
            // timestep: [N,]
@@ -513,19 +448,16 @@ namespace Qwen {
            int64_t C = x->ne[2];
            int64_t N = x->ne[3];

-            auto img           = process_img(ctx, x);
+            auto img           = DiT::pad_and_patchify(ctx, x, params.patch_size, params.patch_size);
            int64_t img_tokens = img->ne[1];

            if (ref_latents.size() > 0) {
                for (ggml_tensor* ref : ref_latents) {
-                    ref = process_img(ctx, ref);
+                    ref = DiT::pad_and_patchify(ctx, ref, params.patch_size, params.patch_size);
                    img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
                }
            }

-            int64_t h_len = ((H + (params.patch_size / 2)) / params.patch_size);
-            int64_t w_len = ((W + (params.patch_size / 2)) / params.patch_size);
-
            auto out = forward_orig(ctx, img, timestep, context, pe, modulate_index);  // [N, h_len*w_len, ph*pw*C]

            if (out->ne[1] > img_tokens) {
@@ -534,11 +466,7 @@ namespace Qwen {
                out = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, out, 0, 2, 1, 3));  // [N, h*w, C * patch_size * patch_size]
            }

-            out = unpatchify(ctx->ggml_ctx, out, h_len, w_len);  // [N, C, H + pad_h, W + pad_w]
-
-            // slice
-            out = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, H);  // [N, C, H, W + pad_w]
-            out = ggml_ext_slice(ctx->ggml_ctx, out, 0, 0, W);  // [N, C, H, W]
+            out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, params.patch_size, params.patch_size);  // [N, C, H, W]

            return out;
        }
@@ -593,24 +521,25 @@ namespace Qwen {
            return "qwen_image";
        }

-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
            qwen_image.get_param_tensors(tensors, prefix);
        }

-        struct ggml_cgraph* build_graph(struct ggml_tensor* x,
-                                        struct ggml_tensor* timesteps,
-                                        struct 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);
-            struct 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]),
@@ -659,67 +588,72 @@ namespace Qwen {

            auto runner_ctx = get_context();

-            struct ggml_tensor* out = qwen_image.forward(&runner_ctx,
-                                                         x,
-                                                         timesteps,
-                                                         context,
-                                                         pe,
-                                                         ref_latents,
-                                                         modulate_index);
+            ggml_tensor* out = qwen_image.forward(&runner_ctx,
+                                                  x,
+                                                  timesteps,
+                                                  context,
+                                                  pe,
+                                                  ref_latents,
+                                                  modulate_index);

            ggml_build_forward_expand(gf, out);

            return gf;
        }

-        bool compute(int n_threads,
-                     struct ggml_tensor* x,
-                     struct ggml_tensor* timesteps,
-                     struct ggml_tensor* context,
-                     std::vector<ggml_tensor*> ref_latents = {},
-                     bool increase_ref_index               = false,
-                     struct ggml_tensor** output           = nullptr,
-                     struct 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]
-            auto get_graph = [&]() -> struct ggml_cgraph* {
+            auto get_graph = [&]() -> ggml_cgraph* {
                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() {
-            struct ggml_init_params params;
+            ggml_init_params params;
            params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
            params.mem_buffer = nullptr;
            params.no_alloc   = false;

-            struct 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);

-                struct 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);
            }
        }
@@ -43,7 +43,7 @@ namespace Rope {

    __STATIC_INLINE__ std::vector<std::vector<float>> rope(const std::vector<float>& pos,
                                                           int dim,
-                                                           int theta,
+                                                           float theta,
                                                           const std::vector<int>& axis_wrap_dims = {}) {
        assert(dim % 2 == 0);
        int half_dim = dim / 2;
@@ -167,7 +167,7 @@ namespace Rope {

    __STATIC_INLINE__ std::vector<float> embed_nd(const std::vector<std::vector<float>>& ids,
                                                  int bs,
-                                                  int theta,
+                                                  const std::vector<float>& axis_thetas,
                                                  const std::vector<int>& axes_dim,
                                                  const std::vector<std::vector<int>>& wrap_dims = {}) {
        std::vector<std::vector<float>> trans_ids = transpose(ids);
@@ -188,8 +188,12 @@ namespace Rope {
            if (!wrap_dims.empty() && i < (int)wrap_dims.size()) {
                axis_wrap_dims = wrap_dims[i];
            }
+            float axis_theta = 10000.0f;
+            if (!axis_thetas.empty()) {
+                axis_theta = axis_thetas[std::min(i, axis_thetas.size() - 1)];
+            }
            std::vector<std::vector<float>> rope_emb =
-                rope(trans_ids[i], axes_dim[i], theta, axis_wrap_dims);  // [bs*pos_len, axes_dim[i]/2 * 2 * 2]
+                rope(trans_ids[i], axes_dim[i], axis_theta, axis_wrap_dims);  // [bs*pos_len, axes_dim[i]/2 * 2 * 2]
            for (int b = 0; b < bs; ++b) {
                for (int j = 0; j < pos_len; ++j) {
                    for (int k = 0; k < rope_emb[0].size(); ++k) {
@@ -203,6 +207,15 @@ namespace Rope {
        return flatten(emb);
    }

+    __STATIC_INLINE__ std::vector<float> embed_nd(const std::vector<std::vector<float>>& ids,
+                                                  int bs,
+                                                  float theta,
+                                                  const std::vector<int>& axes_dim,
+                                                  const std::vector<std::vector<int>>& wrap_dims = {}) {
+        std::vector<float> axis_thetas(axes_dim.size(), theta);
+        return embed_nd(ids, bs, axis_thetas, axes_dim, wrap_dims);
+    }
+
    __STATIC_INLINE__ std::vector<std::vector<float>> gen_refs_ids(int patch_size,
                                                                   int bs,
                                                                   int axes_dim_num,
@@ -332,7 +345,7 @@ namespace Rope {
                }
            }
        }
-        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
    }

    __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen_image_ids(int h,
@@ -421,7 +434,7 @@ namespace Rope {
                }
            }
        }
-        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
    }

    __STATIC_INLINE__ std::vector<std::vector<float>> gen_vid_ids(int t,
@@ -475,7 +488,7 @@ namespace Rope {
                                                    int theta,
                                                    const std::vector<int>& axes_dim) {
        std::vector<std::vector<float>> ids = gen_vid_ids(t, h, w, pt, ph, pw, bs);
-        return embed_nd(ids, bs, theta, axes_dim);
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim);
    }

    __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen2vl_ids(int grid_h,
@@ -511,7 +524,7 @@ namespace Rope {
                                                        int theta,
                                                        const std::vector<int>& axes_dim) {
        std::vector<std::vector<float>> ids = gen_qwen2vl_ids(grid_h, grid_w, merge_size, window_index);
-        return embed_nd(ids, 1, theta, axes_dim);
+        return embed_nd(ids, 1, static_cast<float>(theta), axes_dim);
    }

    __STATIC_INLINE__ int bound_mod(int a, int m) {
@@ -584,13 +597,13 @@ namespace Rope {
            }
        }

-        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
    }

-    __STATIC_INLINE__ struct ggml_tensor* apply_rope(struct ggml_context* ctx,
-                                                     struct ggml_tensor* x,
-                                                     struct ggml_tensor* pe,
-                                                     bool rope_interleaved = true) {
+    __STATIC_INLINE__ ggml_tensor* apply_rope(ggml_context* ctx,
+                                              ggml_tensor* x,
+                                              ggml_tensor* pe,
+                                              bool rope_interleaved = true) {
        // x: [N, L, n_head, d_head]
        // pe: [L, d_head/2, 2, 2], [[cos, -sin], [sin, cos]]
        int64_t d_head = x->ne[0];
@@ -628,21 +641,21 @@ namespace Rope {
        return x_out;
    }

-    __STATIC_INLINE__ struct ggml_tensor* attention(GGMLRunnerContext* ctx,
-                                                    struct ggml_tensor* q,
-                                                    struct ggml_tensor* k,
-                                                    struct ggml_tensor* v,
-                                                    struct ggml_tensor* pe,
-                                                    struct ggml_tensor* mask,
-                                                    float kv_scale        = 1.0f,
-                                                    bool rope_interleaved = true) {
+    __STATIC_INLINE__ ggml_tensor* attention(GGMLRunnerContext* ctx,
+                                             ggml_tensor* q,
+                                             ggml_tensor* k,
+                                             ggml_tensor* v,
+                                             ggml_tensor* pe,
+                                             ggml_tensor* mask,
+                                             float kv_scale        = 1.0f,
+                                             bool rope_interleaved = true) {
        // q,k,v: [N, L, n_head, d_head]
        // pe: [L, d_head/2, 2, 2]
        // return: [N, L, n_head*d_head]
        q = apply_rope(ctx->ggml_ctx, q, pe, rope_interleaved);  // [N*n_head, L, d_head]
        k = apply_rope(ctx->ggml_ctx, k, pe, rope_interleaved);  // [N*n_head, L, d_head]

-        auto x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, v->ne[1], mask, false, true, ctx->flash_attn_enabled, kv_scale);  // [N, L, n_head*d_head]
+        auto x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, v->ne[1], mask, true, ctx->flash_attn_enabled, kv_scale);  // [N, L, n_head*d_head]
        return x;
    }
 };  // namespace Rope
@@ -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
@@ -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__
@@ -0,0 +1,187 @@
+#ifndef __SPECTRUM_HPP__
+#define __SPECTRUM_HPP__
+
+#include <cmath>
+#include <cstring>
+#include <vector>
+
+#include "ggml_extend.hpp"
+#include "tensor.hpp"
+
+struct SpectrumConfig {
+    float w            = 0.40f;
+    int m              = 3;
+    float lam          = 1.0f;
+    int window_size    = 2;
+    float flex_window  = 0.50f;
+    int warmup_steps   = 4;
+    float stop_percent = 0.9f;
+};
+
+struct SpectrumState {
+    SpectrumConfig config;
+    int cnt                 = 0;
+    int num_cached          = 0;
+    float curr_ws           = 2.0f;
+    int K                   = 6;
+    int stop_step           = 0;
+    int total_steps_skipped = 0;
+
+    std::vector<std::vector<float>> H_buf;
+    std::vector<float> T_buf;
+
+    void init(const SpectrumConfig& cfg, size_t total_steps) {
+        config              = cfg;
+        cnt                 = 0;
+        num_cached          = 0;
+        curr_ws             = (float)cfg.window_size;
+        K                   = std::max(cfg.m + 1, 6);
+        stop_step           = (int)(cfg.stop_percent * (float)total_steps);
+        total_steps_skipped = 0;
+        H_buf.clear();
+        T_buf.clear();
+    }
+
+    float taus(int step_cnt) const {
+        return (step_cnt / 50.0f) * 2.0f - 1.0f;
+    }
+
+    bool should_predict() {
+        if (cnt < config.warmup_steps)
+            return false;
+        if (stop_step > 0 && cnt >= stop_step)
+            return false;
+        if ((int)H_buf.size() < 2)
+            return false;
+
+        int ws = std::max(1, (int)std::floor(curr_ws));
+        return (num_cached + 1) % ws != 0;
+    }
+
+    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) {
+            H_buf.erase(H_buf.begin());
+            T_buf.erase(T_buf.begin());
+        }
+
+        if (cnt >= config.warmup_steps)
+            curr_ws += config.flex_window;
+
+        num_cached = 0;
+        cnt++;
+    }
+
+    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);
+
+        std::vector<float> X(K_curr * M1);
+        for (int i = 0; i < K_curr; i++) {
+            X[i * M1] = 1.0f;
+            if (M1 > 1)
+                X[i * M1 + 1] = T_buf[i];
+            for (int j = 2; j < M1; j++)
+                X[i * M1 + j] = 2.0f * T_buf[i] * X[i * M1 + j - 1] - X[i * M1 + j - 2];
+        }
+
+        std::vector<float> x_star(M1);
+        x_star[0] = 1.0f;
+        if (M1 > 1)
+            x_star[1] = tau_at;
+        for (int j = 2; j < M1; j++)
+            x_star[j] = 2.0f * tau_at * x_star[j - 1] - x_star[j - 2];
+
+        std::vector<float> XtX(M1 * M1, 0.0f);
+        for (int i = 0; i < M1; i++) {
+            for (int j = 0; j < M1; j++) {
+                float sum = 0.0f;
+                for (int k = 0; k < K_curr; k++)
+                    sum += X[k * M1 + i] * X[k * M1 + j];
+                XtX[i * M1 + j] = sum + (i == j ? config.lam : 0.0f);
+            }
+        }
+
+        std::vector<float> L(M1 * M1, 0.0f);
+        if (!cholesky_decompose(XtX.data(), L.data(), M1)) {
+            float trace = 0.0f;
+            for (int i = 0; i < M1; i++)
+                trace += XtX[i * M1 + i];
+            for (int i = 0; i < M1; i++)
+                XtX[i * M1 + i] += 1e-4f * trace / M1;
+            cholesky_decompose(XtX.data(), L.data(), M1);
+        }
+
+        std::vector<float> v(M1);
+        cholesky_solve(L.data(), x_star.data(), v.data(), M1);
+
+        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];
+
+        float* out          = denoised->data();
+        float w_cheb        = config.w;
+        float w_taylor      = 1.0f - w_cheb;
+        const float* h_last = H_buf.back().data();
+        const float* h_prev = H_buf[H_buf.size() - 2].data();
+
+        for (int64_t f = 0; f < F; f++) {
+            float pred_cheb = 0.0f;
+            for (int k = 0; k < K_curr; k++)
+                pred_cheb += weights[k] * H_buf[k][f];
+
+            float pred_taylor = h_last[f] + 0.5f * (h_last[f] - h_prev[f]);
+
+            out[f] = w_taylor * pred_taylor + w_cheb * pred_cheb;
+        }
+
+        num_cached++;
+        total_steps_skipped++;
+        cnt++;
+    }
+
+private:
+    static bool cholesky_decompose(const float* A, float* L, int n) {
+        std::memset(L, 0, n * n * sizeof(float));
+        for (int i = 0; i < n; i++) {
+            for (int j = 0; j <= i; j++) {
+                float sum = 0.0f;
+                for (int k = 0; k < j; k++)
+                    sum += L[i * n + k] * L[j * n + k];
+                if (i == j) {
+                    float diag = A[i * n + i] - sum;
+                    if (diag <= 0.0f)
+                        return false;
+                    L[i * n + j] = std::sqrt(diag);
+                } else {
+                    L[i * n + j] = (A[i * n + j] - sum) / L[j * n + j];
+                }
+            }
+        }
+        return true;
+    }
+
+    static void cholesky_solve(const float* L, const float* b, float* x, int n) {
+        std::vector<float> y(n);
+        for (int i = 0; i < n; i++) {
+            float sum = 0.0f;
+            for (int j = 0; j < i; j++)
+                sum += L[i * n + j] * y[j];
+            y[i] = (b[i] - sum) / L[i * n + i];
+        }
+        for (int i = n - 1; i >= 0; i--) {
+            float sum = 0.0f;
+            for (int j = i + 1; j < n; j++)
+                sum += L[j * n + i] * x[j];
+            x[i] = (y[i] - sum) / L[i * n + i];
+        }
+    }
+};
+
+#endif  // __SPECTRUM_HPP__
@@ -17,22 +17,43 @@ class TAEBlock : public UnaryBlock {
 protected:
    int n_in;
    int n_out;
+    bool use_midblock_gn;

 public:
-    TAEBlock(int n_in, int n_out)
-        : n_in(n_in), n_out(n_out) {
+    TAEBlock(int n_in, int n_out, bool use_midblock_gn = false)
+        : n_in(n_in), n_out(n_out), use_midblock_gn(use_midblock_gn) {
        blocks["conv.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_in, n_out, {3, 3}, {1, 1}, {1, 1}));
        blocks["conv.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_out, n_out, {3, 3}, {1, 1}, {1, 1}));
        blocks["conv.4"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_out, n_out, {3, 3}, {1, 1}, {1, 1}));
        if (n_in != n_out) {
            blocks["skip"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_in, n_out, {1, 1}, {1, 1}, {1, 1}, {1, 1}, false));
        }
+        if (use_midblock_gn) {
+            int n_gn         = n_in * 4;
+            blocks["pool.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_in, n_gn, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+            blocks["pool.1"] = std::shared_ptr<GGMLBlock>(new GroupNorm(4, n_gn));
+            // pool.2 is ReLU, handled in forward
+            blocks["pool.3"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_gn, n_in, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+        }
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
        // x: [n, n_in, h, w]
        // return: [n, n_out, h, w]

+        if (use_midblock_gn) {
+            auto pool_0 = std::dynamic_pointer_cast<Conv2d>(blocks["pool.0"]);
+            auto pool_1 = std::dynamic_pointer_cast<GroupNorm>(blocks["pool.1"]);
+            auto pool_3 = std::dynamic_pointer_cast<Conv2d>(blocks["pool.3"]);
+
+            auto p = pool_0->forward(ctx, x);
+            p      = pool_1->forward(ctx, p);
+            p      = ggml_relu_inplace(ctx->ggml_ctx, p);
+            p      = pool_3->forward(ctx, p);
+
+            x = ggml_add(ctx->ggml_ctx, x, p);
+        }
+
        auto conv_0 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.0"]);
        auto conv_2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.2"]);
        auto conv_4 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.4"]);
@@ -62,7 +83,7 @@ class TinyEncoder : public UnaryBlock {
    int num_blocks  = 3;

 public:
-    TinyEncoder(int z_channels = 4)
+    TinyEncoder(int z_channels = 4, bool use_midblock_gn = false)
        : z_channels(z_channels) {
        int index                       = 0;
        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, channels, {3, 3}, {1, 1}, {1, 1}));
@@ -80,13 +101,13 @@ public:

        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
        for (int i = 0; i < num_blocks; i++) {
-            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels, use_midblock_gn));
        }

        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, z_channels, {3, 3}, {1, 1}, {1, 1}));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
        // x: [n, in_channels, h, w]
        // return: [n, z_channels, h/8, w/8]

@@ -107,7 +128,7 @@ class TinyDecoder : public UnaryBlock {
    int num_blocks   = 3;

 public:
-    TinyDecoder(int z_channels = 4)
+    TinyDecoder(int z_channels = 4, bool use_midblock_gn = false)
        : z_channels(z_channels) {
        int index = 0;

@@ -115,7 +136,7 @@ public:
        index++;  // nn.ReLU()

        for (int i = 0; i < num_blocks; i++) {
-            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels, use_midblock_gn));
        }
        index++;  // nn.Upsample()
        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
@@ -136,13 +157,13 @@ public:
        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* z) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* z) override {
        // z: [n, z_channels, h, w]
        // return: [n, out_channels, h*8, w*8]

-        auto h = ggml_scale(ctx->ggml_ctx, z, 1.0f / 3.0f);
+        auto h = ggml_ext_scale(ctx->ggml_ctx, z, 1.0f / 3.0f);
        h      = ggml_tanh_inplace(ctx->ggml_ctx, h);
-        h      = ggml_scale(ctx->ggml_ctx, h, 3.0f);
+        h      = ggml_ext_scale(ctx->ggml_ctx, h, 3.0f);

        for (int i = 0; i < num_blocks * 3 + 10; i++) {
            if (blocks.find(std::to_string(i)) == blocks.end()) {
@@ -171,7 +192,7 @@ public:
        blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels * stride, channels, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
        auto conv = std::dynamic_pointer_cast<UnaryBlock>(blocks["conv"]);
        auto h    = x;
        if (stride != 1) {
@@ -191,7 +212,7 @@ public:
        blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels * stride, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
        auto conv = std::dynamic_pointer_cast<UnaryBlock>(blocks["conv"]);
        auto h    = conv->forward(ctx, x);
        if (stride != 1) {
@@ -215,7 +236,7 @@ public:
        }
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x, struct ggml_tensor* past) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* past) {
        // x: [n, channels, h, w]
        auto conv0 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.0"]);
        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.2"]);
@@ -239,10 +260,10 @@ public:
    }
 };

-struct ggml_tensor* patchify(struct ggml_context* ctx,
-                             struct ggml_tensor* x,
-                             int64_t patch_size,
-                             int64_t b = 1) {
+ggml_tensor* patchify(ggml_context* ctx,
+                      ggml_tensor* x,
+                      int64_t patch_size,
+                      int64_t b = 1) {
    // x: [f, b*c, h*q, w*r]
    // return: [f, b*c*r*q, h, w]
    if (patch_size == 1) {
@@ -268,10 +289,10 @@ struct ggml_tensor* patchify(struct ggml_context* ctx,
    return x;
 }

-struct ggml_tensor* unpatchify(struct ggml_context* ctx,
-                               struct ggml_tensor* x,
-                               int64_t patch_size,
-                               int64_t b = 1) {
+ggml_tensor* unpatchify(ggml_context* ctx,
+                        ggml_tensor* x,
+                        int64_t patch_size,
+                        int64_t b = 1) {
    // x: [f, b*c*r*q, h, w]
    // return: [f, b*c, h*q, w*r]
    if (patch_size == 1) {
@@ -318,7 +339,7 @@ public:
        blocks[std::to_string(index)] = std::shared_ptr<GGMLBlock>(new Conv2d(hidden, z_channels, {3, 3}, {1, 1}, {1, 1}));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* z) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* z) override {
        auto first_conv = std::dynamic_pointer_cast<Conv2d>(blocks["0"]);

        if (patch_size > 1) {
@@ -375,14 +396,15 @@ public:
        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels[num_layers], out_channels * patch_size * patch_size, {3, 3}, {1, 1}, {1, 1}));
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* z) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* z) override {
        auto first_conv = std::dynamic_pointer_cast<Conv2d>(blocks["1"]);

        // Clamp()
-        auto h = ggml_scale_inplace(ctx->ggml_ctx,
-                                    ggml_tanh_inplace(ctx->ggml_ctx,
-                                                      ggml_scale(ctx->ggml_ctx, z, 1.0f / 3.0f)),
-                                    3.0f);
+        auto h = ggml_ext_scale(ctx->ggml_ctx,
+                                ggml_tanh_inplace(ctx->ggml_ctx,
+                                                  ggml_ext_scale(ctx->ggml_ctx, z, 1.0f / 3.0f)),
+                                3.0f,
+                                true);

        h         = first_conv->forward(ctx, h);
        h         = ggml_relu_inplace(ctx->ggml_ctx, h);
@@ -420,11 +442,13 @@ protected:
    bool decode_only;
    SDVersion version;

+public:
+    int z_channels = 16;
+
 public:
    TAEHV(bool decode_only = true, SDVersion version = VERSION_WAN2)
        : decode_only(decode_only), version(version) {
-        int z_channels = 16;
-        int patch      = 1;
+        int patch = 1;
        if (version == VERSION_WAN2_2_TI2V) {
            z_channels = 48;
            patch      = 2;
@@ -435,7 +459,7 @@ public:
        }
    }

-    struct ggml_tensor* decode(GGMLRunnerContext* ctx, struct ggml_tensor* z) {
+    ggml_tensor* decode(GGMLRunnerContext* ctx, ggml_tensor* z) {
        auto decoder = std::dynamic_pointer_cast<TinyVideoDecoder>(blocks["decoder"]);
        if (sd_version_is_wan(version)) {
            // (W, H, C, T) -> (W, H, T, C)
@@ -449,7 +473,7 @@ public:
        return result;
    }

-    struct ggml_tensor* encode(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* encode(GGMLRunnerContext* ctx, ggml_tensor* x) {
        auto encoder = std::dynamic_pointer_cast<TinyVideoEncoder>(blocks["encoder"]);
        // (W, H, T, C) -> (W, H, C, T)
        x                  = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
@@ -470,45 +494,50 @@ public:
 class TAESD : public GGMLBlock {
 protected:
    bool decode_only;
+    bool taef2 = false;
+
+public:
+    int z_channels = 4;

 public:
    TAESD(bool decode_only = true, SDVersion version = VERSION_SD1)
        : decode_only(decode_only) {
-        int z_channels = 4;
+        bool use_midblock_gn = false;
+        taef2                = sd_version_is_flux2(version);
+
        if (sd_version_is_dit(version)) {
            z_channels = 16;
        }
-        blocks["decoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyDecoder(z_channels));
+        if (taef2) {
+            z_channels      = 32;
+            use_midblock_gn = true;
+        }
+        blocks["decoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyDecoder(z_channels, use_midblock_gn));

        if (!decode_only) {
-            blocks["encoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyEncoder(z_channels));
+            blocks["encoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyEncoder(z_channels, use_midblock_gn));
        }
    }

-    struct ggml_tensor* decode(GGMLRunnerContext* ctx, struct ggml_tensor* z) {
+    ggml_tensor* decode(GGMLRunnerContext* ctx, ggml_tensor* z) {
        auto decoder = std::dynamic_pointer_cast<TinyDecoder>(blocks["decoder.layers"]);
+        if (taef2) {
+            z = unpatchify(ctx->ggml_ctx, z, 2);
+        }
        return decoder->forward(ctx, z);
    }

-    struct ggml_tensor* encode(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* encode(GGMLRunnerContext* ctx, ggml_tensor* x) {
        auto encoder = std::dynamic_pointer_cast<TinyEncoder>(blocks["encoder.layers"]);
-        return encoder->forward(ctx, x);
+        auto z       = encoder->forward(ctx, x);
+        if (taef2) {
+            z = patchify(ctx->ggml_ctx, z, 2);
+        }
+        return z;
    }
 };

-struct TinyAutoEncoder : public GGMLRunner {
-    TinyAutoEncoder(ggml_backend_t backend, bool offload_params_to_cpu)
-        : GGMLRunner(backend, offload_params_to_cpu) {}
-    virtual bool compute(const int n_threads,
-                         struct ggml_tensor* z,
-                         bool decode_graph,
-                         struct ggml_tensor** output,
-                         struct ggml_context* output_ctx = nullptr) = 0;
-
-    virtual bool load_from_file(const std::string& file_path, int n_threads) = 0;
-};
-
-struct TinyImageAutoEncoder : public TinyAutoEncoder {
+struct TinyImageAutoEncoder : public VAE {
    TAESD taesd;
    bool decode_only = false;

@@ -520,7 +549,8 @@ struct TinyImageAutoEncoder : public TinyAutoEncoder {
                         SDVersion version = VERSION_SD1)
        : decode_only(decoder_only),
          taesd(decoder_only, version),
-          TinyAutoEncoder(backend, offload_params_to_cpu) {
+          VAE(version, backend, offload_params_to_cpu) {
+        scale_input = false;
        taesd.init(params_ctx, tensor_storage_map, prefix);
    }

@@ -528,56 +558,48 @@ struct TinyImageAutoEncoder : public TinyAutoEncoder {
        return "taesd";
    }

-    bool load_from_file(const std::string& file_path, int n_threads) {
-        LOG_INFO("loading taesd from '%s', decode_only = %s", file_path.c_str(), decode_only ? "true" : "false");
-        alloc_params_buffer();
-        std::map<std::string, ggml_tensor*> taesd_tensors;
-        taesd.get_param_tensors(taesd_tensors);
-        std::set<std::string> ignore_tensors;
-        if (decode_only) {
-            ignore_tensors.insert("encoder.");
-        }
-
-        ModelLoader model_loader;
-        if (!model_loader.init_from_file_and_convert_name(file_path)) {
-            LOG_ERROR("init taesd model loader from file failed: '%s'", file_path.c_str());
-            return false;
-        }
-
-        bool success = model_loader.load_tensors(taesd_tensors, ignore_tensors, n_threads);
-
-        if (!success) {
-            LOG_ERROR("load tae tensors from model loader failed");
-            return false;
-        }
-
-        LOG_INFO("taesd model loaded");
-        return success;
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        taesd.get_param_tensors(tensors, prefix);
    }

-    struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
-        struct ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
-        z                       = to_backend(z);
-        auto runner_ctx         = get_context();
-        struct ggml_tensor* out = decode_graph ? taesd.decode(&runner_ctx, z) : taesd.encode(&runner_ctx, z);
+    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;
+    }
+
+    sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
+        return 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(const sd::Tensor<float>& z_tensor, bool decode_graph) {
+        ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        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,
-                 struct ggml_tensor* z,
-                 bool decode_graph,
-                 struct ggml_tensor** output,
-                 struct ggml_context* output_ctx = nullptr) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(z, decode_graph);
+    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_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());
    }
 };

-struct TinyVideoAutoEncoder : public TinyAutoEncoder {
+struct TinyVideoAutoEncoder : public VAE {
    TAEHV taehv;
    bool decode_only = false;

@@ -589,7 +611,8 @@ struct TinyVideoAutoEncoder : public TinyAutoEncoder {
                         SDVersion version = VERSION_WAN2)
        : decode_only(decoder_only),
          taehv(decoder_only, version),
-          TinyAutoEncoder(backend, offload_params_to_cpu) {
+          VAE(version, backend, offload_params_to_cpu) {
+        scale_input = false;
        taehv.init(params_ctx, tensor_storage_map, prefix);
    }

@@ -597,53 +620,45 @@ struct TinyVideoAutoEncoder : public TinyAutoEncoder {
        return "taehv";
    }

-    bool load_from_file(const std::string& file_path, int n_threads) {
-        LOG_INFO("loading taehv from '%s', decode_only = %s", file_path.c_str(), decode_only ? "true" : "false");
-        alloc_params_buffer();
-        std::map<std::string, ggml_tensor*> taehv_tensors;
-        taehv.get_param_tensors(taehv_tensors);
-        std::set<std::string> ignore_tensors;
-        if (decode_only) {
-            ignore_tensors.insert("encoder.");
-        }
-
-        ModelLoader model_loader;
-        if (!model_loader.init_from_file(file_path)) {
-            LOG_ERROR("init taehv model loader from file failed: '%s'", file_path.c_str());
-            return false;
-        }
-
-        bool success = model_loader.load_tensors(taehv_tensors, ignore_tensors, n_threads);
-
-        if (!success) {
-            LOG_ERROR("load tae tensors from model loader failed");
-            return false;
-        }
-
-        LOG_INFO("taehv model loaded");
-        return success;
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        taehv.get_param_tensors(tensors, prefix);
    }

-    struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
-        struct ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
-        z                       = to_backend(z);
-        auto runner_ctx         = get_context();
-        struct ggml_tensor* out = decode_graph ? taehv.decode(&runner_ctx, z) : taehv.encode(&runner_ctx, z);
+    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;
+    }
+
+    sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
+        return 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(const sd::Tensor<float>& z_tensor, bool decode_graph) {
+        ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        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,
-                 struct ggml_tensor* z,
-                 bool decode_graph,
-                 struct ggml_tensor** output,
-                 struct ggml_context* output_ctx = nullptr) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(z, decode_graph);
+    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_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__
@@ -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
@@ -6,8 +6,10 @@
 #include <unordered_map>
 #include <vector>

+#include "condition_cache_utils.hpp"
 #include "denoiser.hpp"
 #include "ggml_extend.hpp"
+#include "tensor.hpp"

 struct UCacheConfig {
    bool enabled                = false;
@@ -19,6 +21,7 @@ struct UCacheConfig {
    bool adaptive_threshold     = true;
    float early_step_multiplier = 0.5f;
    float late_step_multiplier  = 1.5f;
+    float relative_norm_gain    = 1.6f;
    bool reset_error_on_compute = true;
 };

@@ -28,15 +31,15 @@ struct UCacheCacheEntry {

 struct UCacheState {
    UCacheConfig 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*, UCacheCacheEntry> 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*, UCacheCacheEntry> cache_diffs;
    std::vector<float> prev_input;
    std::vector<float> prev_output;
    float output_prev_norm                = 0.0f;
@@ -45,14 +48,16 @@ struct UCacheState {
    bool has_output_prev_norm             = false;
    bool has_relative_transformation_rate = false;
    float relative_transformation_rate    = 0.0f;
-    float cumulative_change_rate          = 0.0f;
    float last_input_change               = 0.0f;
    bool has_last_input_change            = false;
+    float output_change_ema               = 0.0f;
+    bool has_output_change_ema            = false;
    int total_steps_skipped               = 0;
    int current_step_index                = -1;
    int steps_computed_since_active       = 0;
+    int expected_total_steps              = 0;
+    int consecutive_skipped_steps         = 0;
    float accumulated_error               = 0.0f;
-    float reference_output_norm           = 0.0f;

    struct BlockMetrics {
        float sum_transformation_rate = 0.0f;
@@ -106,14 +111,16 @@ struct UCacheState {
        has_output_prev_norm             = false;
        has_relative_transformation_rate = false;
        relative_transformation_rate     = 0.0f;
-        cumulative_change_rate           = 0.0f;
        last_input_change                = 0.0f;
        has_last_input_change            = false;
+        output_change_ema                = 0.0f;
+        has_output_change_ema            = false;
        total_steps_skipped              = 0;
        current_step_index               = -1;
        steps_computed_since_active      = 0;
+        expected_total_steps             = 0;
+        consecutive_skipped_steps        = 0;
        accumulated_error                = 0.0f;
-        reference_output_norm            = 0.0f;
        block_metrics.reset();
        total_active_steps = 0;
    }
@@ -133,7 +140,8 @@ struct UCacheState {
        if (!initialized || sigmas.size() < 2) {
            return;
        }
-        size_t n_steps = sigmas.size() - 1;
+        size_t n_steps       = sigmas.size() - 1;
+        expected_total_steps = static_cast<int>(n_steps);

        size_t start_step = static_cast<size_t>(config.start_percent * n_steps);
        size_t end_step   = static_cast<size_t>(config.end_percent * n_steps);
@@ -207,11 +215,15 @@ struct UCacheState {
        }

        int effective_total = estimated_total_steps;
+        if (effective_total <= 0) {
+            effective_total = expected_total_steps;
+        }
        if (effective_total <= 0) {
            effective_total = std::max(20, steps_computed_since_active * 2);
        }

        float progress = (effective_total > 0) ? (static_cast<float>(steps_computed_since_active) / effective_total) : 0.0f;
+        progress       = std::max(0.0f, std::min(1.0f, progress));

        float multiplier = 1.0f;
        if (progress < 0.2f) {
@@ -223,43 +235,30 @@ struct UCacheState {
        return base_threshold * multiplier;
    }

-    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) {
        UCacheCacheEntry& 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) {
@@ -292,13 +291,13 @@ struct UCacheState {
            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]);
        }
@@ -309,17 +308,31 @@ struct UCacheState {

        if (has_output_prev_norm && has_relative_transformation_rate &&
            last_input_change > 0.0f && output_prev_norm > 0.0f) {
-            float approx_output_change_rate = (relative_transformation_rate * last_input_change) / output_prev_norm;
-            accumulated_error               = accumulated_error * config.error_decay_rate + approx_output_change_rate;
+            float approx_output_change = relative_transformation_rate * last_input_change;
+            float approx_output_change_rate;
+            if (config.use_relative_threshold) {
+                float base_scale          = std::max(output_prev_norm, 1e-6f);
+                float dyn_scale           = has_output_change_ema
+                                                ? std::max(output_change_ema * std::max(1.0f, config.relative_norm_gain), 1e-6f)
+                                                : base_scale;
+                float scale               = std::sqrt(base_scale * dyn_scale);
+                approx_output_change_rate = approx_output_change / scale;
+            } else {
+                approx_output_change_rate = approx_output_change;
+            }
+            // Increase estimated error with skip horizon to avoid long extrapolation streaks
+            approx_output_change_rate *= (1.0f + 0.50f * consecutive_skipped_steps);
+            accumulated_error = accumulated_error * config.error_decay_rate + approx_output_change_rate;

            float effective_threshold = get_adaptive_threshold();
-            if (config.use_relative_threshold && reference_output_norm > 0.0f) {
-                effective_threshold = effective_threshold * reference_output_norm;
+            if (!config.use_relative_threshold && output_prev_norm > 0.0f) {
+                effective_threshold = effective_threshold * output_prev_norm;
            }

            if (accumulated_error < effective_threshold) {
                skip_current_step = true;
                total_steps_skipped++;
+                consecutive_skipped_steps++;
                apply_cache(cond, input, output);
                return true;
            } else if (config.reset_error_on_compute) {
@@ -330,7 +343,7 @@ struct UCacheState {
        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;
        }
@@ -340,17 +353,19 @@ struct UCacheState {
        if (cond != anchor_condition) {
            return;
        }
+        steps_computed_since_active++;
+        consecutive_skipped_steps = 0;

-        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]);
@@ -359,6 +374,14 @@ struct UCacheState {
                output_change /= static_cast<float>(ne);
            }
        }
+        if (std::isfinite(output_change) && output_change > 0.0f) {
+            if (!has_output_change_ema) {
+                output_change_ema     = output_change;
+                has_output_change_ema = true;
+            } else {
+                output_change_ema = 0.8f * output_change_ema + 0.2f * output_change;
+            }
+        }

        prev_output.resize(ne);
        for (size_t i = 0; i < ne; ++i) {
@@ -373,10 +396,6 @@ struct UCacheState {
        output_prev_norm     = (ne > 0) ? (mean_abs / static_cast<float>(ne)) : 0.0f;
        has_output_prev_norm = output_prev_norm > 0.0f;

-        if (reference_output_norm == 0.0f) {
-            reference_output_norm = output_prev_norm;
-        }
-
        if (has_last_input_change && last_input_change > 0.0f && output_change > 0.0f) {
            float rate = output_change / last_input_change;
            if (std::isfinite(rate)) {
@@ -1,8 +1,7 @@
 #ifndef __UNET_HPP__
 #define __UNET_HPP__

-#include "common.hpp"
-#include "ggml_extend.hpp"
+#include "common_block.hpp"
 #include "model.h"

 /*==================================================== UnetModel =====================================================*/
@@ -61,10 +60,10 @@ public:
        blocks["time_mixer"] = std::shared_ptr<GGMLBlock>(new AlphaBlender());
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
-                                struct ggml_tensor* context,
-                                int timesteps) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x,
+                         ggml_tensor* context,
+                         int timesteps) {
        // x: [N, in_channels, h, w] aka [b*t, in_channels, h, w], t == timesteps
        // context: [N, max_position(aka n_context), hidden_size(aka context_dim)] aka [b*t, n_context, context_dim], t == timesteps
        // t_emb: [N, in_channels] aka [b*t, in_channels]
@@ -201,6 +200,9 @@ public:
            num_head_channels     = 64;
            num_heads             = -1;
            use_linear_projection = true;
+            if (version == VERSION_SDXL_VEGA) {
+                transformer_depth = {1, 1, 2};
+            }
        } else if (version == VERSION_SVD) {
            in_channels           = 8;
            out_channels          = 4;
@@ -215,10 +217,13 @@ public:
        } else if (sd_version_is_unet_edit(version)) {
            in_channels = 8;
        }
-        if (version == VERSION_SD1_TINY_UNET || version == VERSION_SD2_TINY_UNET) {
+        if (version == VERSION_SD1_TINY_UNET || version == VERSION_SD2_TINY_UNET || version == VERSION_SDXS) {
            num_res_blocks = 1;
            channel_mult   = {1, 2, 4};
            tiny_unet      = true;
+            if (version == VERSION_SDXS) {
+                attention_resolutions = {4, 2};  // here just like SDXL
+            }
        }

        // dims is always 2
@@ -316,7 +321,7 @@ public:
        }
        if (!tiny_unet) {
            blocks["middle_block.0"] = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, ch));
-            if (version != VERSION_SDXL_SSD1B) {
+            if (version != VERSION_SDXL_SSD1B && version != VERSION_SDXL_VEGA) {
                blocks["middle_block.1"] = std::shared_ptr<GGMLBlock>(get_attention_layer(ch,
                                                                                          n_head,
                                                                                          d_head,
@@ -383,11 +388,11 @@ public:
        blocks["out.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(model_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
    }

-    struct ggml_tensor* resblock_forward(std::string name,
-                                         GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* x,
-                                         struct ggml_tensor* emb,
-                                         int num_video_frames) {
+    ggml_tensor* resblock_forward(std::string name,
+                                  GGMLRunnerContext* ctx,
+                                  ggml_tensor* x,
+                                  ggml_tensor* emb,
+                                  int num_video_frames) {
        if (version == VERSION_SVD) {
            auto block = std::dynamic_pointer_cast<VideoResBlock>(blocks[name]);

@@ -399,11 +404,11 @@ public:
        }
    }

-    struct ggml_tensor* attention_layer_forward(std::string name,
-                                                GGMLRunnerContext* ctx,
-                                                struct ggml_tensor* x,
-                                                struct ggml_tensor* context,
-                                                int timesteps) {
+    ggml_tensor* attention_layer_forward(std::string name,
+                                         GGMLRunnerContext* ctx,
+                                         ggml_tensor* x,
+                                         ggml_tensor* context,
+                                         int timesteps) {
        if (version == VERSION_SVD) {
            auto block = std::dynamic_pointer_cast<SpatialVideoTransformer>(blocks[name]);

@@ -415,15 +420,15 @@ public:
        }
    }

-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
-                                struct ggml_tensor* timesteps,
-                                struct ggml_tensor* context,
-                                struct ggml_tensor* c_concat              = nullptr,
-                                struct ggml_tensor* y                     = nullptr,
-                                int num_video_frames                      = -1,
-                                std::vector<struct ggml_tensor*> controls = {},
-                                float control_strength                    = 0.f) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x,
+                         ggml_tensor* timesteps,
+                         ggml_tensor* context,
+                         ggml_tensor* c_concat              = nullptr,
+                         ggml_tensor* y                     = nullptr,
+                         int num_video_frames               = -1,
+                         std::vector<ggml_tensor*> controls = {},
+                         float control_strength             = 0.f) {
        // x: [N, in_channels, h, w] or [N, in_channels/2, h, w]
        // timesteps: [N,]
        // context: [N, max_position, hidden_size] or [1, max_position, hidden_size]. for example, [N, 77, 768]
@@ -475,7 +480,7 @@ public:
        }

        // input_blocks
-        std::vector<struct ggml_tensor*> hs;
+        std::vector<ggml_tensor*> hs;

        // input block 0
        auto h = input_blocks_0_0->forward(ctx, x);
@@ -517,13 +522,13 @@ public:
        // middle_block
        if (!tiny_unet) {
            h = resblock_forward("middle_block.0", ctx, h, emb, num_video_frames);  // [N, 4*model_channels, h/8, w/8]
-            if (version != VERSION_SDXL_SSD1B) {
+            if (version != VERSION_SDXL_SSD1B && version != VERSION_SDXL_VEGA) {
                h = attention_layer_forward("middle_block.1", ctx, h, context, num_video_frames);  // [N, 4*model_channels, h/8, w/8]
                h = resblock_forward("middle_block.2", ctx, h, emb, num_video_frames);             // [N, 4*model_channels, h/8, w/8]
            }
        }
        if (controls.size() > 0) {
-            auto cs = ggml_scale_inplace(ctx->ggml_ctx, controls[controls.size() - 1], control_strength);
+            auto cs = ggml_ext_scale(ctx->ggml_ctx, controls[controls.size() - 1], control_strength, true);
            h       = ggml_add(ctx->ggml_ctx, h, cs);  // middle control
        }
        int control_offset = static_cast<int>(controls.size() - 2);
@@ -536,7 +541,7 @@ public:
                hs.pop_back();

                if (controls.size() > 0) {
-                    auto cs = ggml_scale_inplace(ctx->ggml_ctx, controls[control_offset], control_strength);
+                    auto cs = ggml_ext_scale(ctx->ggml_ctx, controls[control_offset], control_strength, true);
                    h_skip  = ggml_add(ctx->ggml_ctx, h_skip, cs);  // control net condition
                    control_offset--;
                }
@@ -600,82 +605,81 @@ struct UNetModelRunner : public GGMLRunner {
        return "unet";
    }

-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
        unet.get_param_tensors(tensors, prefix);
    }

-    struct ggml_cgraph* build_graph(struct ggml_tensor* x,
-                                    struct ggml_tensor* timesteps,
-                                    struct ggml_tensor* context,
-                                    struct ggml_tensor* c_concat              = nullptr,
-                                    struct ggml_tensor* y                     = nullptr,
-                                    int num_video_frames                      = -1,
-                                    std::vector<struct ggml_tensor*> controls = {},
-                                    float control_strength                    = 0.f) {
-        struct ggml_cgraph* gf = new_graph_custom(UNET_GRAPH_SIZE);
+    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                     = {},
+                             int num_video_frames                                  = -1,
+                             const std::vector<sd::Tensor<float>>& controls_tensor = {},
+                             float control_strength                                = 0.f) {
+        ggml_cgraph* gf = new_graph_custom(UNET_GRAPH_SIZE);
+
+        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);
+        std::vector<ggml_tensor*> controls;
+        controls.reserve(controls_tensor.size());
+        for (const auto& control_tensor : controls_tensor) {
+            controls.push_back(make_input(control_tensor));
+        }

        if (num_video_frames == -1) {
            num_video_frames = static_cast<int>(x->ne[3]);
        }

-        x         = to_backend(x);
-        context   = to_backend(context);
-        y         = to_backend(y);
-        timesteps = to_backend(timesteps);
-        c_concat  = to_backend(c_concat);
-
-        for (int i = 0; i < controls.size(); i++) {
-            controls[i] = to_backend(controls[i]);
-        }
-
        auto runner_ctx = get_context();

-        struct ggml_tensor* out = unet.forward(&runner_ctx,
-                                               x,
-                                               timesteps,
-                                               context,
-                                               c_concat,
-                                               y,
-                                               num_video_frames,
-                                               controls,
-                                               control_strength);
+        ggml_tensor* out = unet.forward(&runner_ctx,
+                                        x,
+                                        timesteps,
+                                        context,
+                                        c_concat,
+                                        y,
+                                        num_video_frames,
+                                        controls,
+                                        control_strength);

        ggml_build_forward_expand(gf, out);

        return gf;
    }

-    bool compute(int n_threads,
-                 struct ggml_tensor* x,
-                 struct ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
-                 struct ggml_tensor* c_concat,
-                 struct ggml_tensor* y,
-                 int num_video_frames                      = -1,
-                 std::vector<struct ggml_tensor*> controls = {},
-                 float control_strength                    = 0.f,
-                 struct ggml_tensor** output               = nullptr,
-                 struct 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<float>& c_concat              = {},
+                              const sd::Tensor<float>& y                     = {},
+                              int num_video_frames                           = -1,
+                              const std::vector<sd::Tensor<float>>& controls = {},
+                              float control_strength                         = 0.f) {
        // x: [N, in_channels, h, w]
        // timesteps: [N, ]
        // context: [N, max_position, hidden_size]([N, 77, 768]) or [1, max_position, hidden_size]
        // c_concat: [N, in_channels, h, w] or [1, in_channels, h, w]
        // y: [N, adm_in_channels] or [1, adm_in_channels]
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
            return build_graph(x, timesteps, context, c_concat, y, num_video_frames, controls, control_strength);
        };

-        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() {
-        struct ggml_init_params params;
+        ggml_init_params params;
        params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
        params.mem_buffer = nullptr;
        params.no_alloc   = false;

-        struct ggml_context* work_ctx = ggml_init(params);
-        GGML_ASSERT(work_ctx != nullptr);
+        ggml_context* ctx = ggml_init(params);
+        GGML_ASSERT(ctx != nullptr);

        {
            // CPU, num_video_frames = 1, x{num_video_frames, 8, 8, 8}: Pass
@@ -684,27 +688,37 @@ struct UNetModelRunner : public GGMLRunner {
            // CUDA, num_video_frames = 3, x{num_video_frames, 8, 8, 8}: nan
            int num_video_frames = 3;

-            auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 8, 8, 8, num_video_frames);
+            sd::Tensor<float> x({8, 8, 8, num_video_frames});
            std::vector<float> timesteps_vec(num_video_frames, 999.f);
-            auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
-            ggml_set_f32(x, 0.5f);
+            auto timesteps = sd::Tensor<float>::from_vector(timesteps_vec);
+            x.fill_(0.5f);
            // print_ggml_tensor(x);

-            auto context = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 1024, 1, num_video_frames);
-            ggml_set_f32(context, 0.5f);
+            sd::Tensor<float> context({1024, 1, num_video_frames});
+            context.fill_(0.5f);
            // print_ggml_tensor(context);

-            auto y = ggml_new_tensor_2d(work_ctx, GGML_TYPE_F32, 768, num_video_frames);
-            ggml_set_f32(y, 0.5f);
+            sd::Tensor<float> y({768, num_video_frames});
+            y.fill_(0.5f);
            // print_ggml_tensor(y);

-            struct ggml_tensor* out = nullptr;
+            sd::Tensor<float> out;

-            int64_t t0 = ggml_time_ms();
-            compute(8, x, timesteps, context, nullptr, y, num_video_frames, {}, 0.f, &out, work_ctx);
-            int64_t t1 = ggml_time_ms();
+            int64_t t0   = ggml_time_ms();
+            auto out_opt = compute(8,
+                                   x,
+                                   timesteps,
+                                   context,
+                                   {},
+                                   y,
+                                   num_video_frames,
+                                   {},
+                                   0.f);
+            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("unet test done in %lldms", t1 - t0);
        }
    }
@@ -2,6 +2,7 @@
 #include "ggml_extend.hpp"
 #include "model.h"
 #include "stable-diffusion.h"
+#include "util.h"

 struct UpscalerGGML {
    ggml_backend_t backend    = nullptr;  // general backend
@@ -64,6 +65,39 @@ struct UpscalerGGML {
        return true;
    }

+    sd::Tensor<float> upscale_tensor(const sd::Tensor<float>& input_tensor) {
+        sd::Tensor<float> upscaled;
+        if (tile_size <= 0 || (input_tensor.shape()[0] <= tile_size && input_tensor.shape()[1] <= tile_size)) {
+            upscaled = esrgan_upscaler->compute(n_threads, input_tensor);
+        } else {
+            auto on_processing = [&](const sd::Tensor<float>& input_tile) -> sd::Tensor<float> {
+                auto output_tile = esrgan_upscaler->compute(n_threads, input_tile);
+                if (output_tile.empty()) {
+                    LOG_ERROR("esrgan compute failed while processing a tile");
+                    return {};
+                }
+                return output_tile;
+            };
+
+            upscaled = process_tiles_2d(input_tensor,
+                                        static_cast<int>(input_tensor.shape()[0] * esrgan_upscaler->scale),
+                                        static_cast<int>(input_tensor.shape()[1] * esrgan_upscaler->scale),
+                                        esrgan_upscaler->scale,
+                                        tile_size,
+                                        tile_size,
+                                        0.25f,
+                                        false,
+                                        false,
+                                        on_processing);
+        }
+        esrgan_upscaler->free_compute_buffer();
+        if (upscaled.empty()) {
+            LOG_ERROR("esrgan compute failed");
+            return {};
+        }
+        return upscaled;
+    }
+
    sd_image_t upscale(sd_image_t input_image, uint32_t upscale_factor) {
        // upscale_factor, unused for RealESRGAN_x4plus_anime_6B.pth
        sd_image_t upscaled_image = {0, 0, 0, nullptr};
@@ -72,39 +106,17 @@ struct UpscalerGGML {
        LOG_INFO("upscaling from (%i x %i) to (%i x %i)",
                 input_image.width, input_image.height, output_width, output_height);

-        struct ggml_init_params params;
-        params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
-        params.mem_buffer = nullptr;
-        params.no_alloc   = false;
-
-        // draft context
-        struct ggml_context* upscale_ctx = ggml_init(params);
-        if (!upscale_ctx) {
-            LOG_ERROR("ggml_init() failed");
+        sd::Tensor<float> input_tensor = sd_image_to_tensor(input_image);
+        sd::Tensor<float> upscaled;
+        int64_t t0 = ggml_time_ms();
+        upscaled   = upscale_tensor(input_tensor);
+        if (upscaled.empty()) {
            return upscaled_image;
        }
-        // LOG_DEBUG("upscale work buffer size: %.2f MB", params.mem_size / 1024.f / 1024.f);
-        ggml_tensor* input_image_tensor = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, input_image.width, input_image.height, 3, 1);
-        sd_image_to_ggml_tensor(input_image, input_image_tensor);
-
-        ggml_tensor* upscaled = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, output_width, output_height, 3, 1);
-        auto on_tiling        = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
-            esrgan_upscaler->compute(n_threads, in, &out);
-        };
-        int64_t t0 = ggml_time_ms();
-        sd_tiling(input_image_tensor, upscaled, esrgan_upscaler->scale, esrgan_upscaler->tile_size, 0.25f, on_tiling);
-        esrgan_upscaler->free_compute_buffer();
-        ggml_ext_tensor_clamp_inplace(upscaled, 0.f, 1.f);
-        uint8_t* upscaled_data = ggml_tensor_to_sd_image(upscaled);
-        ggml_free(upscale_ctx);
-        int64_t t3 = ggml_time_ms();
+        sd_image_t upscaled_data = tensor_to_sd_image(upscaled);
+        int64_t t3               = ggml_time_ms();
        LOG_INFO("input_image_tensor upscaled, taking %.2fs", (t3 - t0) / 1000.0f);
-        upscaled_image = {
-            (uint32_t)output_width,
-            (uint32_t)output_height,
-            3,
-            upscaled_data,
-        };
+        upscaled_image = upscaled_data;
        return upscaled_image;
    }
 };
@@ -337,17 +337,13 @@ std::vector<std::string> split_string(const std::string& str, char delimiter) {
    return result;
 }

-void pretty_progress(int step, int steps, float time) {
-    if (sd_progress_cb) {
-        sd_progress_cb(step, steps, time, sd_progress_cb_data);
-        return;
-    }
-    if (step == 0) {
-        return;
-    }
+static std::string build_progress_bar(int step, int steps) {
    std::string progress = "  |";
    int max_progress     = 50;
-    int32_t current      = (int32_t)(step * 1.f * max_progress / steps);
+    int32_t current      = 0;
+    if (steps > 0) {
+        current = (int32_t)(step * 1.f * max_progress / steps);
+    }
    for (int i = 0; i < 50; i++) {
        if (i > current) {
            progress += " ";
@@ -358,16 +354,57 @@ void pretty_progress(int step, int steps, float time) {
        }
    }
    progress += "|";
+    return progress;
+}

-    const char* lf   = (step == steps ? "\n" : "");
+static void print_progress_line(int step, int steps, const std::string& speed_text) {
+    if (step == 0) {
+        return;
+    }
+    std::string progress = build_progress_bar(step, steps);
+    const char* lf       = (step == steps ? "\n" : "");
+    printf("\r%s %i/%i - %s\033[K%s", progress.c_str(), step, steps, speed_text.c_str(), lf);
+    fflush(stdout);  // for linux
+}
+
+void pretty_progress(int step, int steps, float time) {
+    if (sd_progress_cb) {
+        sd_progress_cb(step, steps, time, sd_progress_cb_data);
+        return;
+    }
+    if (step == 0) {
+        return;
+    }
    const char* unit = "s/it";
    float speed      = time;
    if (speed < 1.0f && speed > 0.f) {
        speed = 1.0f / speed;
        unit  = "it/s";
    }
-    printf("\r%s %i/%i - %.2f%s\033[K%s", progress.c_str(), step, steps, speed, unit, lf);
-    fflush(stdout);  // for linux
+    print_progress_line(step, steps, sd_format("%.2f%s", speed, unit));
+}
+
+void pretty_bytes_progress(int step, int steps, uint64_t bytes_processed, float elapsed_seconds) {
+    if (sd_progress_cb) {
+        float time = elapsed_seconds / (step + 1e-6f);
+        sd_progress_cb(step, steps, time, sd_progress_cb_data);
+        return;
+    }
+    if (step == 0) {
+        return;
+    }
+
+    double bytes_per_second = 0.0;
+    if (elapsed_seconds > 0.0f) {
+        bytes_per_second = bytes_processed / (double)elapsed_seconds;
+    }
+
+    double speed_mb = bytes_per_second / (1024.0 * 1024.0);
+    if (speed_mb >= 1024.0) {
+        print_progress_line(step, steps, sd_format("%.2fGB/s", speed_mb / 1024.0));
+    } else {
+        print_progress_line(step, steps, sd_format("%.2fMB/s", speed_mb));
+    }
 }

 std::string ltrim(const std::string& s) {
@@ -479,158 +516,96 @@ const char* sd_get_system_info() {
    return buffer;
 }

-sd_image_f32_t sd_image_t_to_sd_image_f32_t(sd_image_t image) {
-    sd_image_f32_t converted_image;
-    converted_image.width   = image.width;
-    converted_image.height  = image.height;
-    converted_image.channel = image.channel;
+sd_image_t tensor_to_sd_image(const sd::Tensor<float>& tensor, int frame_index) {
+    const auto& shape = tensor.shape();
+    GGML_ASSERT(shape.size() == 4 || shape.size() == 5);
+    int width     = static_cast<int>(shape[0]);
+    int height    = static_cast<int>(shape[1]);
+    int channel   = static_cast<int>(shape[shape.size() == 5 ? 3 : 2]);
+    uint8_t* data = (uint8_t*)malloc(static_cast<size_t>(width * height * channel));
+    GGML_ASSERT(data != nullptr);

-    // Allocate memory for float data
-    converted_image.data = (float*)malloc(image.width * image.height * image.channel * sizeof(float));
-
-    for (uint32_t i = 0; i < image.width * image.height * image.channel; i++) {
-        // Convert uint8_t to float
-        converted_image.data[i] = (float)image.data[i];
-    }
-
-    return converted_image;
-}
-
-// Function to perform double linear interpolation
-float interpolate(float v1, float v2, float v3, float v4, float x_ratio, float y_ratio) {
-    return v1 * (1 - x_ratio) * (1 - y_ratio) + v2 * x_ratio * (1 - y_ratio) + v3 * (1 - x_ratio) * y_ratio + v4 * x_ratio * y_ratio;
-}
-
-sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int target_height) {
-    sd_image_f32_t resized_image;
-    resized_image.width   = target_width;
-    resized_image.height  = target_height;
-    resized_image.channel = image.channel;
-
-    // Allocate memory for resized float data
-    resized_image.data = (float*)malloc(target_width * target_height * image.channel * sizeof(float));
-
-    for (int y = 0; y < target_height; y++) {
-        for (int x = 0; x < target_width; x++) {
-            float original_x = (float)x * image.width / target_width;
-            float original_y = (float)y * image.height / target_height;
-
-            uint32_t x1 = (uint32_t)original_x;
-            uint32_t y1 = (uint32_t)original_y;
-            uint32_t x2 = std::min(x1 + 1, image.width - 1);
-            uint32_t y2 = std::min(y1 + 1, image.height - 1);
-
-            for (uint32_t k = 0; k < image.channel; k++) {
-                float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
-                float v2 = *(image.data + y1 * image.width * image.channel + x2 * image.channel + k);
-                float v3 = *(image.data + y2 * image.width * image.channel + x1 * image.channel + k);
-                float v4 = *(image.data + y2 * image.width * image.channel + x2 * image.channel + k);
-
-                float x_ratio = original_x - x1;
-                float y_ratio = original_y - y1;
-
-                float value = interpolate(v1, v2, v3, v4, x_ratio, y_ratio);
-
-                *(resized_image.data + y * target_width * image.channel + x * image.channel + k) = value;
+    for (int iw = 0; iw < width; ++iw) {
+        for (int ih = 0; ih < height; ++ih) {
+            for (int ic = 0; ic < channel; ++ic) {
+                float value                            = shape.size() == 5 ? tensor.index(iw, ih, frame_index, ic, 0)
+                                                                           : tensor.index(iw, ih, ic, frame_index);
+                value                                  = std::clamp(value, 0.0f, 1.0f);
+                data[(ih * width + iw) * channel + ic] = static_cast<uint8_t>(std::round(value * 255.0f));
            }
        }
    }
-
-    return resized_image;
+    return {
+        static_cast<uint32_t>(width),
+        static_cast<uint32_t>(height),
+        static_cast<uint32_t>(channel),
+        data,
+    };
 }

-void normalize_sd_image_f32_t(sd_image_f32_t image, float means[3], float stds[3]) {
-    for (uint32_t y = 0; y < image.height; y++) {
-        for (uint32_t x = 0; x < image.width; x++) {
-            for (uint32_t k = 0; k < image.channel; k++) {
-                int index         = (y * image.width + x) * image.channel + k;
-                image.data[index] = (image.data[index] - means[k]) / stds[k];
+sd::Tensor<float> sd_image_to_tensor(sd_image_t image,
+                                     int target_width,
+                                     int target_height,
+                                     bool scale) {
+    sd::Tensor<float> tensor = sd::zeros<float>({static_cast<int64_t>(image.width),
+                                                 static_cast<int64_t>(image.height),
+                                                 static_cast<int64_t>(image.channel),
+                                                 1});
+    for (uint32_t iw = 0; iw < image.width; ++iw) {
+        for (uint32_t ih = 0; ih < image.height; ++ih) {
+            for (uint32_t ic = 0; ic < image.channel; ++ic) {
+                tensor.index(iw, ih, ic, 0) = sd_image_get_f32(image, iw, ih, ic, scale);
            }
        }
    }
+    if (target_width >= 0 && target_height >= 0 &&
+        (tensor.shape()[0] != target_width || tensor.shape()[1] != target_height)) {
+        tensor = sd::ops::interpolate(tensor,
+                                      {target_width,
+                                       target_height,
+                                       tensor.shape()[2],
+                                       tensor.shape()[3]});
+    }
+    return tensor;
 }

 // Constants for means and std
 float means[3] = {0.48145466f, 0.4578275f, 0.40821073f};
 float stds[3]  = {0.26862954f, 0.26130258f, 0.27577711f};

-// Function to clip and preprocess sd_image_f32_t
-sd_image_f32_t clip_preprocess(sd_image_f32_t image, int target_width, int target_height) {
-    float width_scale  = (float)target_width / image.width;
-    float height_scale = (float)target_height / image.height;
+sd::Tensor<float> clip_preprocess(const sd::Tensor<float>& image, int target_width, int target_height) {
+    GGML_ASSERT(image.dim() == 4);
+    GGML_ASSERT(image.shape()[2] == 3);
+    GGML_ASSERT(image.shape()[3] == 1);
+    GGML_ASSERT(target_width > 0 && target_height > 0);

-    float scale = std::fmax(width_scale, height_scale);
+    float width_scale  = static_cast<float>(target_width) / static_cast<float>(image.shape()[0]);
+    float height_scale = static_cast<float>(target_height) / static_cast<float>(image.shape()[1]);
+    float scale        = std::fmax(width_scale, height_scale);

-    // Interpolation
-    int resized_width   = (int)(scale * image.width);
-    int resized_height  = (int)(scale * image.height);
-    float* resized_data = (float*)malloc(resized_width * resized_height * image.channel * sizeof(float));
+    int64_t resized_width  = static_cast<int64_t>(scale * static_cast<float>(image.shape()[0]));
+    int64_t resized_height = static_cast<int64_t>(scale * static_cast<float>(image.shape()[1]));

-    for (int y = 0; y < resized_height; y++) {
-        for (int x = 0; x < resized_width; x++) {
-            float original_x = (float)x * image.width / resized_width;
-            float original_y = (float)y * image.height / resized_height;
+    sd::Tensor<float> resized = sd::ops::interpolate(
+        image,
+        {resized_width, resized_height, image.shape()[2], image.shape()[3]});

-            uint32_t x1 = (uint32_t)original_x;
-            uint32_t y1 = (uint32_t)original_y;
-            uint32_t x2 = std::min(x1 + 1, image.width - 1);
-            uint32_t y2 = std::min(y1 + 1, image.height - 1);
+    int64_t h_offset = std::max<int64_t>((resized_height - target_height) / 2, 0);
+    int64_t w_offset = std::max<int64_t>((resized_width - target_width) / 2, 0);

-            for (uint32_t k = 0; k < image.channel; k++) {
-                float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
-                float v2 = *(image.data + y1 * image.width * image.channel + x2 * image.channel + k);
-                float v3 = *(image.data + y2 * image.width * image.channel + x1 * image.channel + k);
-                float v4 = *(image.data + y2 * image.width * image.channel + x2 * image.channel + k);
-
-                float x_ratio = original_x - x1;
-                float y_ratio = original_y - y1;
-
-                float value = interpolate(v1, v2, v3, v4, x_ratio, y_ratio);
-
-                *(resized_data + y * resized_width * image.channel + x * image.channel + k) = value;
+    sd::Tensor<float> cropped({target_width, target_height, image.shape()[2], image.shape()[3]});
+    for (int64_t y = 0; y < target_height; ++y) {
+        for (int64_t x = 0; x < target_width; ++x) {
+            for (int64_t c = 0; c < image.shape()[2]; ++c) {
+                cropped.index(x, y, c, 0) = resized.index(x + w_offset, y + h_offset, c, 0);
            }
        }
    }

-    // Clip and preprocess
-    int h_offset = std::max((int)(resized_height - target_height) / 2, 0);
-    int w_offset = std::max((int)(resized_width - target_width) / 2, 0);
-
-    sd_image_f32_t result;
-    result.width   = target_width;
-    result.height  = target_height;
-    result.channel = image.channel;
-    result.data    = (float*)malloc(target_height * target_width * image.channel * sizeof(float));
-
-    for (uint32_t k = 0; k < image.channel; k++) {
-        for (uint32_t i = 0; i < result.height; i++) {
-            for (uint32_t j = 0; j < result.width; j++) {
-                int src_y = std::min(static_cast<int>(i + h_offset), resized_height - 1);
-                int src_x = std::min(static_cast<int>(j + w_offset), resized_width - 1);
-                *(result.data + i * result.width * image.channel + j * image.channel + k) =
-                    fmin(fmax(*(resized_data + src_y * resized_width * image.channel + src_x * image.channel + k), 0.0f), 255.0f) / 255.0f;
-            }
-        }
-    }
-
-    // Free allocated memory
-    free(resized_data);
-
-    // Normalize
-    for (uint32_t k = 0; k < image.channel; k++) {
-        for (uint32_t i = 0; i < result.height; i++) {
-            for (uint32_t j = 0; j < result.width; j++) {
-                // *(result.data + i * size * image.channel + j * image.channel + k) = 0.5f;
-                int offset  = i * result.width * image.channel + j * image.channel + k;
-                float value = *(result.data + offset);
-                value       = (value - means[k]) / stds[k];
-                // value = 0.5f;
-                *(result.data + offset) = value;
-            }
-        }
-    }
-
-    return result;
+    sd::Tensor<float> normalized = sd::ops::clamp(cropped, 0.0f, 1.0f);
+    sd::Tensor<float> mean({1, 1, 3, 1}, {means[0], means[1], means[2]});
+    sd::Tensor<float> std({1, 1, 3, 1}, {stds[0], stds[1], stds[2]});
+    return (normalized - mean) / std;
 }

 // Ref: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/cad87bf4e3e0b0a759afa94e933527c3123d59bc/modules/prompt_parser.py#L345
@@ -7,6 +7,7 @@
 #include <vector>

 #include "stable-diffusion.h"
+#include "tensor.hpp"

 #define SAFE_STR(s) ((s) ? (s) : "")
 #define BOOL_STR(b) ((b) ? "true" : "false")
@@ -29,20 +30,14 @@ std::string utf32_to_utf8(const std::u32string& utf32_str);
 std::u32string unicode_value_to_utf32(int unicode_value);
 // std::string sd_basename(const std::string& path);

-typedef struct {
-    uint32_t width;
-    uint32_t height;
-    uint32_t channel;
-    float* data;
-} sd_image_f32_t;
+sd_image_t tensor_to_sd_image(const sd::Tensor<float>& tensor, int frame_index = 0);

-void normalize_sd_image_f32_t(sd_image_f32_t image, float means[3], float stds[3]);
+sd::Tensor<float> sd_image_to_tensor(sd_image_t image,
+                                     int target_width  = -1,
+                                     int target_height = -1,
+                                     bool scale        = true);

-sd_image_f32_t sd_image_t_to_sd_image_f32_t(sd_image_t image);
-
-sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int target_height);
-
-sd_image_f32_t clip_preprocess(sd_image_f32_t image, int target_width, int target_height);
+sd::Tensor<float> clip_preprocess(const sd::Tensor<float>& image, int target_width, int target_height);

 class MmapWrapper {
 public:
@@ -69,6 +64,7 @@ protected:
 std::string path_join(const std::string& p1, const std::string& p2);
 std::vector<std::string> split_string(const std::string& str, char delimiter);
 void pretty_progress(int step, int steps, float time);
+void pretty_bytes_progress(int step, int steps, uint64_t bytes_processed, float elapsed_seconds);

 void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...);

@@ -0,0 +1,253 @@
+#ifndef __VAE_HPP__
+#define __VAE_HPP__
+
+#include "common_block.hpp"
+#include "tensor_ggml.hpp"
+
+struct VAE : public GGMLRunner {
+protected:
+    SDVersion version;
+    bool scale_input                                      = true;
+    virtual sd::Tensor<float> _compute(const int n_threads,
+                                       const sd::Tensor<float>& z,
+                                       bool decode_graph) = 0;
+
+    static inline void scale_tensor_to_minus1_1(sd::Tensor<float>* tensor) {
+        GGML_ASSERT(tensor != nullptr);
+        for (int64_t i = 0; i < tensor->numel(); ++i) {
+            (*tensor)[i] = (*tensor)[i] * 2.0f - 1.0f;
+        }
+    }
+
+    static inline void scale_tensor_to_0_1(sd::Tensor<float>* tensor) {
+        GGML_ASSERT(tensor != nullptr);
+        for (int64_t i = 0; i < tensor->numel(); ++i) {
+            float value  = ((*tensor)[i] + 1.0f) * 0.5f;
+            (*tensor)[i] = std::max(0.0f, std::min(1.0f, value));
+        }
+    }
+
+    sd::Tensor<float> tiled_compute(const sd::Tensor<float>& input,
+                                    int n_threads,
+                                    int output_width,
+                                    int output_height,
+                                    int scale,
+                                    int p_tile_size_x,
+                                    int p_tile_size_y,
+                                    float tile_overlap_factor,
+                                    bool circular_x,
+                                    bool circular_y,
+                                    bool decode_graph,
+                                    const char* error_message,
+                                    bool silent = false) {
+        auto on_processing = [&](const sd::Tensor<float>& input_tile) {
+            auto output_tile = _compute(n_threads, input_tile, decode_graph);
+            if (output_tile.empty()) {
+                LOG_ERROR("%s", error_message);
+                return sd::Tensor<float>();
+            }
+            return output_tile;
+        };
+        return ::process_tiles_2d(input,
+                                  output_width,
+                                  output_height,
+                                  scale,
+                                  p_tile_size_x,
+                                  p_tile_size_y,
+                                  tile_overlap_factor,
+                                  circular_x,
+                                  circular_y,
+                                  on_processing,
+                                  silent);
+    }
+
+public:
+    VAE(SDVersion version, ggml_backend_t backend, bool offload_params_to_cpu)
+        : version(version), GGMLRunner(backend, offload_params_to_cpu) {}
+
+    int get_scale_factor() {
+        int scale_factor = 8;
+        if (version == VERSION_WAN2_2_TI2V) {
+            scale_factor = 16;
+        } else if (sd_version_is_flux2(version)) {
+            scale_factor = 16;
+        } else if (version == VERSION_CHROMA_RADIANCE) {
+            scale_factor = 1;
+        }
+        return scale_factor;
+    }
+
+    virtual int get_encoder_output_channels(int input_channels) = 0;
+
+    void get_tile_sizes(int& tile_size_x,
+                        int& tile_size_y,
+                        float& tile_overlap,
+                        const sd_tiling_params_t& params,
+                        int64_t latent_x,
+                        int64_t latent_y,
+                        float encoding_factor = 1.0f) {
+        tile_overlap       = std::max(std::min(params.target_overlap, 0.5f), 0.0f);
+        auto get_tile_size = [&](int requested_size, float factor, int64_t latent_size) {
+            const int default_tile_size  = 32;
+            const int min_tile_dimension = 4;
+            int tile_size                = default_tile_size;
+            // factor <= 1 means simple fraction of the latent dimension
+            // factor > 1 means number of tiles across that dimension
+            if (factor > 0.f) {
+                if (factor > 1.0)
+                    factor = 1 / (factor - factor * tile_overlap + tile_overlap);
+                tile_size = static_cast<int>(std::round(latent_size * factor));
+            } else if (requested_size >= min_tile_dimension) {
+                tile_size = requested_size;
+            }
+            tile_size = static_cast<int>(tile_size * encoding_factor);
+            return std::max(std::min(tile_size, static_cast<int>(latent_size)), min_tile_dimension);
+        };
+
+        tile_size_x = get_tile_size(params.tile_size_x, params.rel_size_x, latent_x);
+        tile_size_y = get_tile_size(params.tile_size_y, params.rel_size_y, latent_y);
+    }
+
+    sd::Tensor<float> encode(int n_threads,
+                             const sd::Tensor<float>& x,
+                             sd_tiling_params_t tiling_params,
+                             bool circular_x = false,
+                             bool circular_y = false) {
+        int64_t t0              = ggml_time_ms();
+        sd::Tensor<float> input = x;
+        sd::Tensor<float> output;
+        if (scale_input) {
+            scale_tensor_to_minus1_1(&input);
+        }
+
+        if (tiling_params.enabled) {
+            const int scale_factor = get_scale_factor();
+            int64_t W              = input.shape()[0] / scale_factor;
+            int64_t H              = input.shape()[1] / scale_factor;
+            float tile_overlap;
+            int tile_size_x, tile_size_y;
+            get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, tiling_params, W, H, 1.30539f);
+            LOG_DEBUG("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
+            output = tiled_compute(input,
+                                   n_threads,
+                                   static_cast<int>(W),
+                                   static_cast<int>(H),
+                                   scale_factor,
+                                   tile_size_x,
+                                   tile_size_y,
+                                   tile_overlap,
+                                   circular_x,
+                                   circular_y,
+                                   false,
+                                   "vae encode compute failed while processing a tile");
+        } else {
+            output = _compute(n_threads, input, false);
+            free_compute_buffer();
+        }
+
+        if (output.empty()) {
+            LOG_ERROR("vae encode compute failed");
+            return {};
+        }
+        int64_t t1 = ggml_time_ms();
+        LOG_DEBUG("computing vae encode graph completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
+        return std::move(output);
+    }
+
+    sd::Tensor<float> decode(int n_threads,
+                             const sd::Tensor<float>& x,
+                             sd_tiling_params_t tiling_params,
+                             bool decode_video = false,
+                             bool circular_x   = false,
+                             bool circular_y   = false,
+                             bool silent       = false) {
+        int64_t t0              = ggml_time_ms();
+        sd::Tensor<float> input = x;
+        sd::Tensor<float> output;
+
+        if (tiling_params.enabled) {
+            const int scale_factor = get_scale_factor();
+            int64_t W              = input.shape()[0] * scale_factor;
+            int64_t H              = input.shape()[1] * scale_factor;
+            float tile_overlap;
+            int tile_size_x, tile_size_y;
+            get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, tiling_params, input.shape()[0], input.shape()[1]);
+            if (!silent) {
+                LOG_DEBUG("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
+            }
+            output = tiled_compute(
+                input,
+                n_threads,
+                static_cast<int>(W),
+                static_cast<int>(H),
+                scale_factor,
+                tile_size_x,
+                tile_size_y,
+                tile_overlap,
+                circular_x,
+                circular_y,
+                true,
+                "vae decode compute failed while processing a tile",
+                silent);
+        } else {
+            output = _compute(n_threads, input, true);
+        }
+
+        free_compute_buffer();
+
+        if (output.empty()) {
+            LOG_ERROR("vae decode compute failed");
+            return {};
+        }
+        if (scale_input) {
+            scale_tensor_to_0_1(&output);
+        }
+        int64_t t1 = ggml_time_ms();
+        LOG_DEBUG("computing vae decode graph completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
+        return std::move(output);
+    }
+
+    virtual sd::Tensor<float> vae_output_to_latents(const sd::Tensor<float>& vae_output, std::shared_ptr<RNG> rng) = 0;
+    virtual sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents)                           = 0;
+    virtual sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents)                           = 0;
+    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix)         = 0;
+    virtual void set_conv2d_scale(float scale) { SD_UNUSED(scale); };
+};
+
+struct FakeVAE : public VAE {
+    FakeVAE(SDVersion version, ggml_backend_t backend, bool offload_params_to_cpu)
+        : VAE(version, backend, offload_params_to_cpu) {}
+
+    int get_encoder_output_channels(int input_channels) {
+        return input_channels;
+    }
+
+    sd::Tensor<float> _compute(const int n_threads,
+                               const sd::Tensor<float>& z,
+                               bool decode_graph) override {
+        SD_UNUSED(n_threads);
+        SD_UNUSED(decode_graph);
+        return z;
+    }
+
+    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;
+    }
+
+    sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents) override {
+        return latents;
+    }
+
+    sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents) override {
+        return latents;
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) override {}
+
+    std::string get_desc() override {
+        return "fake_vae";
+    }
+};
+
+#endif  // __VAE_HPP__
@@ -1,4 +1,4 @@
-static unsigned char merges_utf8_c_str[] = {
+static const unsigned char clip_merges_utf8_c_str[] = {
    0x23,
    0x76,
    0x65,
@@ -524620,7 +524620,7 @@ static unsigned char merges_utf8_c_str[] = {
    0x0a,
 };

-static unsigned char t5_tokenizer_json_str[] = {
+static const unsigned char t5_tokenizer_json_str[] = {
    0x7b,
    0x0a,
    0x20,
@@ -1,4 +1,4 @@
-unsigned char mistral_merges_utf8_c_str[] = {
+static const unsigned char mistral_merges_utf8_c_str[] = {
  0xc4, 0xa0, 0x20, 0xc4, 0xa0, 0x0a, 0xc4, 0xa0, 0x20, 0x74, 0x0a, 0x65,
  0x20, 0x72, 0x0a, 0x69, 0x20, 0x6e, 0x0a, 0xc4, 0xa0, 0x20, 0xc4, 0xa0,
  0xc4, 0xa0, 0xc4, 0xa0, 0x0a, 0xc4, 0xa0, 0xc4, 0xa0, 0x20, 0xc4, 0xa0,
@@ -260614,7 +260614,7 @@ unsigned char mistral_merges_utf8_c_str[] = {
  0xc3, 0xa5, 0xc4, 0xb2, 0xc4, 0xb0, 0x20, 0xc3, 0xa6, 0xc2, 0xb1, 0xc4,
  0xab, 0xc3, 0xa4, 0xc2, 0xb9, 0xc2, 0xa6, 0x0a,
 };
-unsigned char mistral_vocab_json_utf8_c_str[] = {
+static const unsigned char mistral_vocab_json_utf8_c_str[] = {
  0x7b, 0x22, 0x3c, 0x75, 0x6e, 0x6b, 0x3e, 0x22, 0x3a, 0x20, 0x30, 0x2c,
  0x20, 0x22, 0x3c, 0x73, 0x3e, 0x22, 0x3a, 0x20, 0x31, 0x2c, 0x20, 0x22,
  0x3c, 0x2f, 0x73, 0x3e, 0x22, 0x3a, 0x20, 0x32, 0x2c, 0x20, 0x22, 0x5b,
@@ -1,4 +1,4 @@
-unsigned char qwen2_merges_utf8_c_str[] = {
+static const unsigned char qwen2_merges_utf8_c_str[] = {
  0xc4, 0xa0, 0x20, 0xc4, 0xa0, 0x0a, 0xc4, 0xa0, 0xc4, 0xa0, 0x20, 0xc4,
  0xa0, 0xc4, 0xa0, 0x0a, 0x69, 0x20, 0x6e, 0x0a, 0xc4, 0xa0, 0x20, 0x74,
  0x0a, 0xc4, 0xa0, 0xc4, 0xa0, 0xc4, 0xa0, 0xc4, 0xa0, 0x20, 0xc4, 0xa0,
@@ -1,4 +1,4 @@
-unsigned char umt5_tokenizer_json_str[] = {
+static const unsigned char umt5_tokenizer_json_str[] = {
  0x7b, 0x22, 0x76, 0x65, 0x72, 0x73, 0x69, 0x6f, 0x6e, 0x22, 0x3a, 0x20,
  0x22, 0x31, 0x2e, 0x30, 0x22, 0x2c, 0x20, 0x22, 0x74, 0x72, 0x75, 0x6e,
  0x63, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3a, 0x20, 0x6e, 0x75, 0x6c,
@@ -0,0 +1,35 @@
+#include "vocab.h"
+#include "clip_t5.hpp"
+#include "mistral.hpp"
+#include "qwen.hpp"
+#include "umt5.hpp"
+
+std::string load_clip_merges() {
+    std::string merges_utf8_str(reinterpret_cast<const char*>(clip_merges_utf8_c_str), sizeof(clip_merges_utf8_c_str));
+    return merges_utf8_str;
+}
+
+std::string load_qwen2_merges() {
+    std::string merges_utf8_str(reinterpret_cast<const char*>(qwen2_merges_utf8_c_str), sizeof(qwen2_merges_utf8_c_str));
+    return merges_utf8_str;
+}
+
+std::string load_mistral_merges() {
+    std::string merges_utf8_str(reinterpret_cast<const char*>(mistral_merges_utf8_c_str), sizeof(mistral_merges_utf8_c_str));
+    return merges_utf8_str;
+}
+
+std::string load_mistral_vocab_json() {
+    std::string json_str(reinterpret_cast<const char*>(mistral_vocab_json_utf8_c_str), sizeof(mistral_vocab_json_utf8_c_str));
+    return json_str;
+}
+
+std::string load_t5_tokenizer_json() {
+    std::string json_str(reinterpret_cast<const char*>(t5_tokenizer_json_str), sizeof(t5_tokenizer_json_str));
+    return json_str;
+}
+
+std::string load_umt5_tokenizer_json() {
+    std::string json_str(reinterpret_cast<const char*>(umt5_tokenizer_json_str), sizeof(umt5_tokenizer_json_str));
+    return json_str;
+}
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Wagner Bruna	be9f51b25c	refactor: simplify DiscreteFlowDenoiser (#1405 )	2026-04-11 17:18:23 +08:00
leejet	e8323cabb0	feat: add flux2 small decoder support (#1402 )	2026-04-08 23:13:25 +08:00
Wagner Bruna	dd753729cc	fix: correct double increment on flow denoisers sigma calculations (#1372 )	2026-04-08 23:13:05 +08:00
leejet	8afbeb6ba9	chore: normalize text files to utf-8 without bom (#1394 )	2026-04-06 21:25:34 +08:00
leejet	5bf438d568	refactor: split examples common into header and source (#1393 )	2026-04-06 21:11:57 +08:00
leejet	359eb8b8de	refactor: apply RAII ownership to examples (#1392 )	2026-04-06 20:33:46 +08:00
leejet	7397ddaa86	feat: add webm support (#1391 )	2026-04-06 01:49:28 +08:00
stduhpfandleejet	9369ab759f	feat: inpaint improvements (#1357 ) * inpaint: get max pixel max instead of single sample * inpaint: masked diffusion for inpainting models with inflated mask * refactor tensor interpolate nearest-like reduction paths and generalize max_pool_2d --------- Co-authored-by: leejet <leejet714@gmail.com>	2026-04-06 00:44:26 +08:00
Wagner Brunaandleejet	687a81f251	chore: make libwebp optional and support system libwebp (#1387 ) Co-authored-by: leejet <leejet714@gmail.com>	2026-04-05 23:52:05 +08:00
leejet	87ecb95cbc	feat: add webp support (#1384 )	2026-04-02 01:36:11 +08:00
Wagner Brunaandleejet	99c1de379b	feat: ancestral sampler implementations for flow models (#1374 ) * feat: add support for the eta parameter to ancestral samplers * feat: Euler Ancestral sampler implementation for flow models * refine flow ancestral sampling and normalize eta defaults --------- Co-authored-by: leejet <leejet714@gmail.com>	2026-04-02 01:35:29 +08:00
leejet	09b12d5f6d	feat(cli): add metadata inspection mode (#1381 )	2026-04-01 00:52:03 +08:00
leejet	6dfe945958	fix: use resolved image size in embedded metadata (#1382 )	2026-03-31 23:55:49 +08:00
leejet	bf0216765a	feat: show tensor loading progress in MB/s or GB/s (#1380 )	2026-03-31 23:06:44 +08:00
Wagner Bruna	4fe7a35939	feat(server): add generation metadata to png images (#1217 )	2026-03-31 23:06:27 +08:00
Jan Ekström	4d5232083f	chore(server): link winsock2 for non-MSVC windows (#1378 )	2026-03-31 22:10:34 +08:00
leejet	1d6cb0f8c3	refactor: split and simplify sample_k_diffusion samplers (#1377 )	2026-03-31 00:32:14 +08:00
leejet	83e8f6f0af	refactor(server): split server endpoint registration (#1376 )	2026-03-31 00:02:03 +08:00
leejet	8d878872d9	fix: disable extra T5 mask padding for Wan (#1375 )	2026-03-30 23:10:07 +08:00
Wagner Bruna	02dd5e5dd2	refactor: move VAE tiling parameters to SDGenerationParams (#1261 )	2026-03-30 23:09:38 +08:00
leejet	8f2967c006	sync: update ggml	2026-03-30 23:00:53 +08:00
leejet	f16a110f87	refactor: migrate generation pipeline to sd::Tensor (#1373 )	2026-03-30 00:19:25 +08:00
stduhpf	ed88e215a2	refactor: simplify f8_e5m2_to_f16 function a little bit (#1358 )	2026-03-30 00:14:33 +08:00
Wagner Bruna	6293ab5aaf	docs: update Spectrum info about DiT models (#1360 )	2026-03-30 00:12:57 +08:00
leejet	545fac4f3f	refactor: simplify sample cache flow (#1350 )	2026-03-17 00:28:03 +08:00
Tay	5265a5efa1	perf(z-image): switch to fused SwiGLU kernel (#1302 )	2026-03-17 00:27:46 +08:00
leejet	84cbd88df1	style: remove redundant struct qualifiers for consistent C/C++ type usage (#1349 )	2026-03-16 22:17:22 +08:00
Daniele	997bb11fb6	fix: correct encoder channels for flux2 (#1346 )	2026-03-16 22:16:43 +08:00
leejet	862a6586cb	feat: add embedded WebUI (#1207 )	2026-03-16 00:26:57 +08:00
leejet	61d8331ef3	ci: avoid cuda docker build timeout by using -j16	2026-03-15 18:39:29 +08:00
leejet	acc3bf1fdc	refactor: optimize the VAE architecture (#1345 )	2026-03-15 16:57:42 +08:00
Kevin Nause	83eabd7c01	ci: add CUDA Dockerfile (#1314 )	2026-03-15 16:46:01 +08:00
Wagner Bruna	630ee03f23	refactor: move all cache parameter defaults to the library (#1327 )	2026-03-15 16:43:46 +08:00
Wagner Bruna	f6968bc589	chore: remove SD_FAST_SOFTMAX build flag (#1338 )	2026-03-15 16:42:47 +08:00
rmatif	adfef62900	feat: add generic DiT support to spectrum cache (#1336 )	2026-03-15 16:41:05 +08:00
JusteLeo	6fa7ca9317	docs: add Anima2 gguf download link to anima.md (#1335 )	2026-03-15 16:40:14 +08:00
leejet	d6dd6d7b55	refactor: remove ununsed encode_video (#1332 )	2026-03-10 00:36:09 +08:00
rmatif	dea4980f4e	feat: add spectrum caching method (#1322 )	2026-03-10 00:35:32 +08:00
leejet	c8fb3d2458	fix: resolve SD1 Pix2Pix issue (#1329 )	2026-03-08 00:28:05 +08:00
stduhpf	3d33caaef8	fix: make tiling work better when using circular (#1299 )	2026-03-08 00:25:07 +08:00
WinkelCode	9b424db0f4	ci: change workflow owner of "actions-commit-hash" from "pr-mpt" to "prompt" (#1323 )	2026-03-08 00:23:23 +08:00
rmatif	d95062737e	fix: ucache: normalize reuse error (#1313 )	2026-03-04 23:50:45 +08:00
Korsar13	7c880f80c7	fix: avoid sd-server memory leak (#1316 )	2026-03-04 23:47:38 +08:00
leejet	aaa8a51bd8	docs: update sd-cli/sd-server docs	2026-03-04 00:41:17 +08:00
leejet	ba35dd734e	refactor: introduce ggml_ext_zeros_like/ggml_ext_ones_like (#1312 )	2026-03-04 00:36:52 +08:00
bssrdf	d41f5fff69	perf: improved flux attention qkv unpacking (#1306 )	2026-03-04 00:36:32 +08:00
Korsar13	810ef0cf76	fix: reset weight adapter for models if no loras in request (#1307 )	2026-03-04 00:34:07 +08:00
leejet	5792c66879	feat: support some non-standard Anima weight names (#1305 )	2026-03-01 22:01:29 +08:00
Wagner Bruna	39d54702a6	feat: accept legacy image parameter on v1/images/edits (#1270 )	2026-03-01 22:00:50 +08:00
Wagner Bruna	60889bc9a1	fix: correct sdapi LoRA file handling (#1276 )	2026-03-01 21:57:06 +08:00
leejet	e64baa3611	refactor: reuse DiT's patchify/unpatchify functions (#1304 )	2026-03-01 21:44:51 +08:00
leejet	cec4aedcfd	docs: add anima docs	2026-03-01 15:32:25 +08:00
rmatif	4cdfff5ff2	feat: add Anima support (#1296 )	2026-03-01 15:23:18 +08:00
leejet	0752cc9d3a	fix: resolve image quality degradation issue (#1297 )	2026-02-26 00:26:21 +08:00
Wagner Brunaandleejet	b314d80ad0	feat: turn flow_shift into a generation parameter (#1289 ) * feat: turn flow_shift into a generation parameter * format code * simplify set_shift/set_parameters * fix sd_sample_params_to_str * remove unused variable * update docs --------- Co-authored-by: leejet <leejet714@gmail.com>	2026-02-26 00:26:04 +08:00
leejet	c9cd49701a	fix: safely handle whitespace and consecutive newlines (#1288 )	2026-02-19 20:54:42 +08:00
akleine	c5eb1e4137	fix: avoid black images if using an invalid VAE (for SDXL) (#1273 )	2026-02-19 20:54:18 +08:00
leejet	636d3cb6ff	refactor: reorganize the vocab file structure (#1271 )	2026-02-11 00:44:17 +08:00
Wagner Bruna	adea272225	feat(server): use image and command-line dimensions by default on server (#1262 )	2026-02-11 00:42:50 +08:00
Mario Limonciello	45ce78a3ae	ci: correct rocm artifact of linux (#1269 )	2026-02-10 23:19:28 +08:00
leejet	28ef93c0e1	refactor: reorganize the file structure (#1266 )	2026-02-10 23:13:35 +08:00
leejet	3296545090	feat: add extra_c_crossattns support for llm embedder (#1265 )	2026-02-10 00:00:17 +08:00
akleine	d60fb27560	fix: avoid unwanted file extension changes (#1257 )	2026-02-09 23:59:43 +08:00
Wagner Bruna	c7ccafbd6f	fix: correct sdapi handling of cfg_scale and steps (#1260 )	2026-02-09 23:34:19 +08:00
stduhpf	aa0b899397	fix: improve handling of VAE decode failures (#1222 )	2026-02-09 23:29:41 +08:00
Mario Limonciello	5e264372ce	ci: add a github action to generate a Linux ROCm artifact (#1258 )	2026-02-09 23:23:06 +08:00
leejet	f0f641a142	feat(server): add lora support to sdapi (#1256 )	2026-02-08 00:11:16 +08:00
stduhpf	9f56833e14	feat: optimize LoKr at runtime (#1233 )	2026-02-08 00:08:09 +08:00
Roj234	65891d74cc	fix: avoid the issue of NaN for qwen-image on certain devices (#1249 )	2026-02-04 23:49:05 +08:00
leejet	f957fa3d2a	feat: add --fa option (#1242 )	2026-02-01 21:44:54 +08:00
leejet	c252e03c6b	sync: update ggml	2026-02-01 20:54:23 +08:00
rmatif	e63daba33d	feat: add res_multistep, res_2s sampler and bong tangent scheduler (#1234 )	2026-02-01 20:05:27 +08:00
stduhpf	3959109281	fix: improve LoCon support with other naming conventions (#1239 )	2026-02-01 20:00:16 +08:00
leejet	e411520407	docs: add z-image-base example	2026-01-28 21:47:36 +08:00
leejet	43e829f219	refactor: unify the processing of attention mask (#1230 )	2026-01-26 00:33:34 +08:00
leejet	7837232631	perf: make dit faster (#1228 )	2026-01-25 22:50:10 +08:00
Equious	4ccce027b2	fix: correct mask and control image loading in cli (#1229 )	2026-01-25 22:47:52 +08:00
leejet	fa61ea744d	fix: set default lora_model_dir to . (#1224 )	2026-01-23 22:13:59 +08:00
leejet	5e4579c11d	feat: use image width and height when not explicitly set (#1206 )	2026-01-22 23:54:41 +08:00
Wagner Bruna	329571131d	chore: clarify warning about missing model files (#1219 )	2026-01-21 22:34:11 +08:00
leejet	a48b4a3ade	docs: add FLUX.2-klein support to news	2026-01-19 23:56:50 +08:00
stduhpf	b87fe13afd	feat: support new chroma radiance "x0_x32_proto" (#1209 )	2026-01-19 23:51:26 +08:00
Oleg Skutte	e50e1f253d	feat: add taef2 support (#1211 )	2026-01-19 23:39:36 +08:00
leejet	c6206fb351	fix: set VAE conv scale for all SDXL variants	2026-01-19 23:21:48 +08:00
akleine	639091fbe9	feat: add support for Segmind's Vega model (#1195 )	2026-01-19 23:15:47 +08:00
leejet	9293016c9d	docs: update esrgan.md	2026-01-19 23:00:50 +08:00
leejet	2efd19978d	fix: use Unix timestamp for field instead of ISO string (#1205 )	2026-01-19 00:21:29 +08:00
Wagner Brunaandleejet	61659ef299	feat: add basic sdapi support to sd-server (#1197 ) * feat: add basic sdapi support to sd-server Compatible with AUTOMATIC1111 / Forge. * fix img2img with no mask * add more parameter validation * eliminate MSVC warnings --------- Co-authored-by: leejet <leejet714@gmail.com>	2026-01-19 00:21:11 +08:00
leejet	9565c7f6bd	add support for flux2 klein (#1193 ) * add support for flux2 klein 4b * add support for flux2 klein 8b * use attention_mask in Flux.2 klein LLMEmbedder * update docs	2026-01-18 01:17:33 +08:00
Wagner Bruna	fbce16e02d	fix: avoid undefined behavior on image mask allocation failure (#1198 )	2026-01-18 01:14:56 +08:00
akleineandleejet	7010bb4dff	feat: support for SDXS-512 model (#1180 ) * feat: add U-Net specials of SDXS * docs: update distilled_sd.md for SDXS-512 * feat: for SDXS use AutoencoderTiny as the primary VAE * docs: update distilled_sd.md for SDXS-512 * fix: SDXS code cleaning after review by stduhpf * format code * fix sdxs with --taesd-preview-only --------- Co-authored-by: leejet <leejet714@gmail.com>	2026-01-14 01:14:57 +08:00
Wagner Bruna	48d3161a8d	feat: add sd-server API support for steps, sampler and scheduler (#1173 )	2026-01-14 00:34:27 +08:00
Weiqi Gao	271b594e74	sync: update ggml (#1187 )	2026-01-14 00:28:55 +08:00