feat: add Qwen Image Edit support (#877)

* add ref latent support for qwen image

* optimize clip_preprocess and fix get_first_stage_encoding

* add qwen2vl vit support

* add qwen image edit support

* fix qwen image edit pipeline

* add mmproj file support

* support dynamic number of Qwen image transformer blocks

* set prompt_template_encode_start_idx every time

* to_add_out precision fix

* to_out.0 precision fix

* update docs

.clang-format

@@ -3,7 +3,6 @@ UseTab: Never
 IndentWidth: 4
 TabWidth: 4
 AllowShortIfStatementsOnASingleLine: false
-IndentCaseLabels: false
 ColumnLimit: 0
 AccessModifierOffset: -4
 NamespaceIndentation: All

.github/workflows/build.yml

@@ -4,17 +4,36 @@ on:
   workflow_dispatch: # allows manual triggering
     inputs:
       create_release:
-        description: 'Create new release'
+        description: "Create new release"
         required: true
         type: boolean
   push:
     branches:
       - master
       - ci
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+    paths:
+      [
+        ".github/workflows/**",
+        "**/CMakeLists.txt",
+        "**/Makefile",
+        "**/*.h",
+        "**/*.hpp",
+        "**/*.c",
+        "**/*.cpp",
+        "**/*.cu",
+      ]
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu']
+    paths:
+      [
+        "**/CMakeLists.txt",
+        "**/Makefile",
+        "**/*.h",
+        "**/*.hpp",
+        "**/*.c",
+        "**/*.cpp",
+        "**/*.cu",
+      ]
 
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
@@ -30,7 +49,6 @@ jobs:
         with:
           submodules: recursive
 
-
       - name: Dependencies
         id: depends
         run: |
@@ -42,14 +60,37 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake ..
+          cmake .. -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON
           cmake --build . --config Release
 
-      #- name: Test
-        #id: cmake_test
-        #run: |
-          #cd build
-          #ctest --verbose --timeout 900
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/main' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Fetch system info
+        id: system-info
+        run: |
+          echo "CPU_ARCH=`uname -m`" >> "$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: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp ggml/LICENSE ./build/bin/ggml.txt
+          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
+          zip -j sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.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 }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.zip
+          path: |
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.zip
 
   macOS-latest-cmake:
     runs-on: macos-latest
@@ -63,9 +104,8 @@ jobs:
 
       - name: Dependencies
         id: depends
-        continue-on-error: true
         run: |
-          brew update
+          brew install zip
 
       - name: Build
         id: cmake_build
@@ -73,30 +113,61 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake ..
+          cmake .. -DGGML_AVX2=ON -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" -DSD_BUILD_SHARED_LIBS=ON
           cmake --build . --config Release
 
-      #- name: Test
-        #id: cmake_test
-        #run: |
-          #cd build
-          #ctest --verbose --timeout 900
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/main' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Fetch system info
+        id: system-info
+        run: |
+          echo "CPU_ARCH=`uname -m`" >> "$GITHUB_OUTPUT"
+          echo "OS_NAME=`sw_vers -productName`" >> "$GITHUB_OUTPUT"
+          echo "OS_VERSION=`sw_vers -productVersion`" >> "$GITHUB_OUTPUT"
+          echo "OS_TYPE=`uname -s`" >> "$GITHUB_OUTPUT"
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp ggml/LICENSE ./build/bin/ggml.txt
+          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
+          zip -j sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.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 }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.zip
+          path: |
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.zip
 
   windows-latest-cmake:
-    runs-on: windows-latest
+    runs-on: windows-2025
+
+    env:
+      VULKAN_VERSION: 1.4.328.1
 
     strategy:
       matrix:
         include:
-          - build: 'noavx'
-            defines: '-DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF'
-          - build: 'avx2'
-            defines: '-DGGML_AVX2=ON'
-          - build: 'avx'
-            defines: '-DGGML_AVX2=OFF'
-          - build: 'avx512'
-            defines: '-DGGML_AVX512=ON'
-
+          - build: "noavx"
+            defines: "-DGGML_NATIVE=OFF -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DSD_BUILD_SHARED_LIBS=ON"
+          - build: "avx2"
+            defines: "-DGGML_NATIVE=OFF -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
+          - build: "avx"
+            defines: "-DGGML_NATIVE=OFF -DGGML_AVX=ON -DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS=ON"
+          - build: "avx512"
+            defines: "-DGGML_NATIVE=OFF -DGGML_AVX512=ON -DGGML_AVX=ON -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
+          - build: "cuda12"
+            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES=90;89;86;80;75"
+          # - build: "rocm5.5"
+          #   defines: '-G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS="gfx1100;gfx1102;gfx1030" -DSD_BUILD_SHARED_LIBS=ON'
+          - build: 'vulkan'
+            defines: "-DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON"
     steps:
       - name: Clone
         id: checkout
@@ -104,6 +175,37 @@ jobs:
         with:
           submodules: recursive
 
+      - name: Install cuda-toolkit
+        id: cuda-toolkit
+        if: ${{ matrix.build == 'cuda12' }}
+        uses: Jimver/cuda-toolkit@v0.2.19
+        with:
+          cuda: "12.6.2"
+          method: "network"
+          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
+
+      - name: Install rocm-toolkit
+        id: rocm-toolkit
+        if: ${{ matrix.build == 'rocm5.5' }}
+        uses: Cyberhan123/rocm-toolkit@v0.1.0
+        with:
+          rocm: "5.5.0"
+
+      - name: Install Ninja
+        id: install-ninja
+        if: ${{ matrix.build == 'rocm5.5' }}
+        uses: urkle/action-get-ninja@v1
+        with:
+          version: 1.11.1
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        if: ${{ matrix.build == 'vulkan' }} https://sdk.lunarg.com/sdk/download/1.4.328.1/windows/vulkansdk-windows-X64-1.4.328.1.exe
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
+          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
+
       - name: Build
         id: cmake_build
         run: |
@@ -125,12 +227,6 @@ jobs:
           & $cl /O2 /GS- /kernel avx512f.c /link /nodefaultlib /entry:main
           .\avx512f.exe && echo "AVX512F: YES" && ( echo HAS_AVX512F=1 >> $env:GITHUB_ENV ) || echo "AVX512F: NO"
 
-      #- name: Test
-        #id: cmake_test
-        #run: |
-          #cd build
-          #ctest -C Release --verbose --timeout 900
-
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -140,14 +236,44 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
-          Copy-Item ggml/LICENSE .\build\bin\Release\ggml.txt
-          Copy-Item LICENSE .\build\bin\Release\stable-diffusion.cpp.txt
-          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
+          $filePath = ".\build\bin\Release\*"
+          if (Test-Path $filePath) {
+            echo "Exists at path $filePath"
+            Copy-Item ggml/LICENSE .\build\bin\Release\ggml.txt
+            Copy-Item LICENSE .\build\bin\Release\stable-diffusion.cpp.txt
+          } elseif (Test-Path ".\build\bin\stable-diffusion.dll") {
+          $filePath = ".\build\bin\*"
+            echo "Exists at path $filePath"
+            Copy-Item ggml/LICENSE .\build\bin\ggml.txt
+            Copy-Item LICENSE .\build\bin\stable-diffusion.cpp.txt
+          } else {
+            ls .\build\bin
+            throw "Can't find stable-diffusion.dll"
+          }
+          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip $filePath
+
+      - name: Copy and pack Cuda runtime
+        id: pack_cuda_runtime
+        if: ${{ matrix.build == 'cuda12' && (github.event_name == 'push' && github.ref == 'refs/heads/master' || github.event.inputs.create_release == 'true') }}
+        run: |
+          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
+          $dst='.\build\bin\cudart\'
+          robocopy "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          7z a cudart-sd-bin-win-cu12-x64.zip $dst\*
+
+      - name: Upload Cuda runtime
+        if: ${{ matrix.build == 'cuda12' && (github.event_name == 'push' && github.ref == 'refs/heads/master' || github.event.inputs.create_release == 'true') }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: sd-cudart-sd-bin-win-cu12-x64.zip
+          path: |
+            cudart-sd-bin-win-cu12-x64.zip
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip
           path: |
             sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip
 
@@ -162,9 +288,23 @@ jobs:
       - windows-latest-cmake
 
     steps:
+      - name: Clone
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+
       - name: Download artifacts
         id: download-artifact
-        uses: actions/download-artifact@v3
+        uses: actions/download-artifact@v4
+        with:
+          path: ./artifact
+          pattern: sd-*
+          merge-multiple: true
+
+      - name: Get commit count
+        id: commit_count
+        run: |
+          echo "count=$(git rev-list --count HEAD)" >> $GITHUB_OUTPUT
 
       - name: Get commit hash
         id: commit
@@ -172,14 +312,16 @@ jobs:
 
       - name: Create release
         id: create_release
+        if: ${{ github.event_name == 'workflow_dispatch' || github.ref_name == 'master' }}
         uses: anzz1/action-create-release@v1
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
-          tag_name: ${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}
+          tag_name: ${{ format('{0}-{1}-{2}', env.BRANCH_NAME, steps.commit_count.outputs.count, steps.commit.outputs.short) }}
 
       - name: Upload release
         id: upload_release
+        if: ${{ github.event_name == 'workflow_dispatch' || github.ref_name == 'master' }}
         uses: actions/github-script@v3
         with:
           github-token: ${{secrets.GITHUB_TOKEN}}

.gitignore

@@ -1,14 +1,14 @@
 build*/
+cmake-build-*/
 test/
 .vscode/
+.idea/
 .cache/
 *.swp
-.vscode/
 *.bat
 *.bin
 *.exe
 *.gguf
 output*.png
 models*
-!taesd-model.gguf
-*.log
+*.log

.gitmodules

@@ -1,3 +1,3 @@
 [submodule "ggml"]
     path = ggml
-	url = https://github.com/leejet/ggml.git
+	url = https://github.com/ggml-org/ggml.git

CMakeLists.txt

@@ -24,49 +24,132 @@ endif()
 # general
 #option(SD_BUILD_TESTS                "sd: build tests"    ${SD_STANDALONE})
 option(SD_BUILD_EXAMPLES             "sd: build examples" ${SD_STANDALONE})
-option(SD_CUBLAS                     "sd: cuda backend" OFF)
+option(SD_CUDA                       "sd: cuda backend" OFF)
+option(SD_HIPBLAS                    "sd: rocm backend" OFF)
 option(SD_METAL                      "sd: metal backend" OFF)
-option(SD_FLASH_ATTN                 "sd: use flash attention for x4 less memory usage" OFF)
+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(BUILD_SHARED_LIBS             "sd: build shared libs" OFF)
+option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
+option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
 
-if(SD_CUBLAS)
-	message("Use CUBLAS as backend stable-diffusion")
-    set(GGML_CUBLAS ON)
-    add_definitions(-DSD_USE_CUBLAS)
+if(SD_CUDA)
+    message("-- Use CUDA as backend stable-diffusion")
+    set(GGML_CUDA ON)
+    add_definitions(-DSD_USE_CUDA)
+endif()
+
+if(SD_METAL)
+    message("-- Use Metal as backend stable-diffusion")
+    set(GGML_METAL ON)
+    add_definitions(-DSD_USE_METAL)
+endif()
+
+if (SD_VULKAN)
+    message("-- Use Vulkan as backend stable-diffusion")
+    set(GGML_VULKAN ON)
+    add_definitions(-DSD_USE_VULKAN)
+endif ()
+
+if (SD_OPENCL)
+    message("-- Use OpenCL as backend stable-diffusion")
+    set(GGML_OPENCL ON)
+    add_definitions(-DSD_USE_OPENCL)
+endif ()
+
+if (SD_HIPBLAS)
+    message("-- Use HIPBLAS as backend stable-diffusion")
+    set(GGML_HIP ON)
+    add_definitions(-DSD_USE_CUDA)
+    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()
 endif()
 
-if(SD_METAL)
-	message("Use Metal as backend stable-diffusion")
-    set(GGML_METAL ON)
-    add_definitions(-DSD_USE_METAL)
+set(SD_LIB stable-diffusion)
+
+file(GLOB SD_LIB_SOURCES
+    "*.h"
+    "*.cpp"
+    "*.hpp"
+)
+
+# we can get only one share lib
+if(SD_BUILD_SHARED_LIBS)
+    message("-- Build shared library")
+    message(${SD_LIB_SOURCES})
+    set(BUILD_SHARED_LIBS OFF)
+    add_library(${SD_LIB} SHARED ${SD_LIB_SOURCES})
+    add_definitions(-DSD_BUILD_SHARED_LIB)
+    target_compile_definitions(${SD_LIB} PRIVATE -DSD_BUILD_DLL)
+    set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+else()
+    message("-- Build static library")
+    set(BUILD_SHARED_LIBS OFF)
+    add_library(${SD_LIB} STATIC ${SD_LIB_SOURCES})
 endif()
 
-if(SD_FLASH_ATTN)
-    message("Use Flash Attention for memory optimization")
-    add_definitions(-DSD_USE_FLASH_ATTENTION)
+if(SD_SYCL)
+    message("-- Use SYCL as backend stable-diffusion")
+    set(GGML_SYCL ON)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
+    add_definitions(-DSD_USE_SYCL)
+    # disable fast-math on host, see:
+    # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html
+    if (WIN32)
+        set(SYCL_COMPILE_OPTIONS /fp:precise)
+    else()
+        set(SYCL_COMPILE_OPTIONS -fp-model=precise)
+    endif()
+    message("-- Turn off fast-math for host in SYCL backend")
+    target_compile_options(${SD_LIB} PRIVATE ${SYCL_COMPILE_OPTIONS})
 endif()
 
-
 set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
+
+if (NOT SD_USE_SYSTEM_GGML)
+    # see https://github.com/ggerganov/ggml/pull/682
+    add_definitions(-DGGML_MAX_NAME=128)
+endif()
+
 # deps
-add_subdirectory(ggml)
+# Only add ggml if it hasn't been added yet
+if (NOT TARGET ggml)
+    if (SD_USE_SYSTEM_GGML)
+        find_package(ggml REQUIRED)
+        if (NOT ggml_FOUND)
+            message(FATAL_ERROR "System-installed GGML library not found.")
+        endif()
+        add_library(ggml ALIAS ggml::ggml)
+    else()
+        add_subdirectory(ggml)
+    endif()
+endif()
 
 add_subdirectory(thirdparty)
 
-set(SD_LIB stable-diffusion)
-
-add_library(${SD_LIB} stable-diffusion.h stable-diffusion.cpp model.h model.cpp util.h util.cpp)
 target_link_libraries(${SD_LIB} PUBLIC ggml zip)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
-target_compile_features(${SD_LIB} PUBLIC cxx_std_11)
+target_compile_features(${SD_LIB} PUBLIC c_std_11 cxx_std_17)
 
 
 if (SD_BUILD_EXAMPLES)
     add_subdirectory(examples)
 endif()
 
+set(SD_PUBLIC_HEADERS stable-diffusion.h)
+set_target_properties(${SD_LIB} PROPERTIES PUBLIC_HEADER "${SD_PUBLIC_HEADERS}")
+
+install(TARGETS ${SD_LIB} LIBRARY PUBLIC_HEADER)

Dockerfile

@@ -1,16 +1,21 @@
 ARG UBUNTU_VERSION=22.04
 
-FROM ubuntu:$UBUNTU_VERSION as build
+FROM ubuntu:$UBUNTU_VERSION AS build
 
-RUN apt-get update && apt-get install -y build-essential git cmake
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake
 
 WORKDIR /sd.cpp
 
 COPY . .
 
-RUN mkdir build && cd build && cmake .. && cmake --build . --config Release
+RUN cmake . -B ./build
+RUN cmake --build ./build --config Release --parallel
 
-FROM ubuntu:$UBUNTU_VERSION as runtime
+FROM ubuntu:$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 /sd
 

Dockerfile.musa

@@ -0,0 +1,23 @@
+ARG MUSA_VERSION=rc4.2.0
+ARG UBUNTU_VERSION=22.04
+
+FROM mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64 as build
+
+RUN apt-get update && apt-get install -y ccache cmake git
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && \
+    cmake .. -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
+        -DCMAKE_C_FLAGS="${CMAKE_C_FLAGS} -fopenmp -I/usr/lib/llvm-14/lib/clang/14.0.0/include -L/usr/lib/llvm-14/lib" \
+        -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -fopenmp -I/usr/lib/llvm-14/lib/clang/14.0.0/include -L/usr/lib/llvm-14/lib" \
+        -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build . --config Release
+
+FROM mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64 as runtime
+
+COPY --from=build /sd.cpp/build/bin/sd /sd
+
+ENTRYPOINT [ "/sd" ]

Dockerfile.sycl

@@ -0,0 +1,19 @@
+ARG SYCL_VERSION=2025.1.0-0
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS build
+
+RUN apt-get update && apt-get install -y cmake
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && \
+    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DSD_SYCL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build . --config Release -j$(nproc)
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS runtime
+
+COPY --from=build /sd.cpp/build/bin/sd /sd
+
+ENTRYPOINT [ "/sd" ]

README.md

@@ -1,33 +1,49 @@
 <p align="center">
-  <img src="./assets/a%20lovely%20cat.png" width="256x">
+  <img src="./assets/cat_with_sd_cpp_42.png" width="360x">
 </p>
 
 # stable-diffusion.cpp
 
-Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in pure C/C++
+Diffusion model(SD,Flux,Wan,...) inference in pure C/C++
+
+***Note that this project is under active development. \
+API and command-line option may change frequently.***
 
 ## Features
 
 - Plain C/C++ implementation based on [ggml](https://github.com/ggerganov/ggml), working in the same way as [llama.cpp](https://github.com/ggerganov/llama.cpp)
 - Super lightweight and without external dependencies
-- SD1.x, SD2.x and SDXL support
-- [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo) and [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo) support
+- Supported models
+  - Image Models
+    - SD1.x, SD2.x, [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo)
+    - SDXL, [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo)
+      - !!!The VAE in SDXL encounters NaN issues under FP16, but unfortunately, the ggml_conv_2d only operates under FP16. Hence, a parameter is needed to specify the VAE that has fixed the FP16 NaN issue. You can find it here: [SDXL VAE FP16 Fix](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors).
+    - [SD3/SD3.5](./docs/sd3.md)
+    - [Flux-dev/Flux-schnell](./docs/flux.md)
+    - [Chroma](./docs/chroma.md)
+    - [Qwen Image](./docs/qwen_image.md)
+  - Image Edit Models
+    - [FLUX.1-Kontext-dev](./docs/kontext.md)
+    - [Qwen Image Edit/Qwen Image Edit 2509](./docs/qwen_image_edit.md)
+  - Video Models
+    - [Wan2.1/Wan2.2](./docs/wan.md)
+  - [PhotoMaker](https://github.com/TencentARC/PhotoMaker) support.
+  - Control Net support with SD 1.5
+  - LoRA support, same as [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#lora)
+  - Latent Consistency Models support (LCM/LCM-LoRA)
+  - Faster and memory efficient latent decoding with [TAESD](https://github.com/madebyollin/taesd)
+  - Upscale images generated with [ESRGAN](https://github.com/xinntao/Real-ESRGAN)
 - 16-bit, 32-bit float support
-- 4-bit, 5-bit and 8-bit integer quantization support
+- 2-bit, 3-bit, 4-bit, 5-bit and 8-bit integer quantization support
 - Accelerated memory-efficient CPU inference
     - Only requires ~2.3GB when using txt2img with fp16 precision to generate a 512x512 image, enabling Flash Attention just requires ~1.8GB.
 - AVX, AVX2 and AVX512 support for x86 architectures
-- Full CUDA and Metal backend for GPU acceleration.
+- Full CUDA, Metal, Vulkan, OpenCL and SYCL backend for GPU acceleration.
 - Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs models
     - No need to convert to `.ggml` or `.gguf` anymore!
-- Flash Attention for memory usage optimization (only cpu for now)
-- Original `txt2img` and `img2img` mode
+- Flash Attention for memory usage optimization
 - Negative prompt
 - [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) style tokenizer (not all the features, only token weighting for now)
-- LoRA support, same as [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#lora)
-- Latent Consistency Models support (LCM/LCM-LoRA)
-- Faster and memory efficient latent decoding with [TAESD](https://github.com/madebyollin/taesd)
-- Upscale images generated with [ESRGAN](https://github.com/xinntao/Real-ESRGAN)
 - VAE tiling processing for reduce memory usage
 - Sampling method
     - `Euler A`
@@ -44,21 +60,13 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
     - Linux
     - Mac OS
     - Windows
-    - Android (via Termux)
-
-### TODO
-
-- [ ] More sampling methods
-- [ ] Make inference faster
-    - The current implementation of ggml_conv_2d is slow and has high memory usage
-    - Implement Winograd Convolution 2D for 3x3 kernel filtering
-- [ ] Continuing to reduce memory usage (quantizing the weights of ggml_conv_2d)
-- [ ] Implement Textual Inversion (embeddings)
-- [ ] Implement Inpainting support
-- [ ] k-quants support
+    - Android (via Termux, [Local Diffusion](https://github.com/rmatif/Local-Diffusion))
 
 ## Usage
 
+For most users, you can download the built executable program from the latest [release](https://github.com/leejet/stable-diffusion.cpp/releases/latest).
+If the built product does not meet your requirements, you can choose to build it manually.
+
 ### Get the Code
 
 ```
@@ -81,11 +89,13 @@ git submodule update
     - Stable Diffusion v1.4 from https://huggingface.co/CompVis/stable-diffusion-v-1-4-original
     - Stable Diffusion v1.5 from https://huggingface.co/runwayml/stable-diffusion-v1-5
     - Stable Diffuison v2.1 from https://huggingface.co/stabilityai/stable-diffusion-2-1
+    - Stable Diffusion 3 2B from https://huggingface.co/stabilityai/stable-diffusion-3-medium
 
     ```shell
     curl -L -O https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt
     # curl -L -O https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
     # curl -L -O https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-nonema-pruned.safetensors
+    # curl -L -O https://huggingface.co/stabilityai/stable-diffusion-3-medium/resolve/main/sd3_medium_incl_clips_t5xxlfp16.safetensors
     ```
 
 ### Build
@@ -106,12 +116,34 @@ cmake .. -DGGML_OPENBLAS=ON
 cmake --build . --config Release
 ```
 
-##### Using CUBLAS
+##### Using CUDA
 
 This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). Recommended to have at least 4 GB of VRAM.
 
 ```
-cmake .. -DSD_CUBLAS=ON
+cmake .. -DSD_CUDA=ON
+cmake --build . --config Release
+```
+
+##### Using HipBLAS
+This provides BLAS acceleration using the ROCm cores of your AMD GPU. Make sure to have the ROCm toolkit installed.
+To build for another GPU architecture than installed in your system, set `$GFX_NAME` manually to the desired architecture (replace first command). This is also necessary if your GPU is not officially supported by ROCm, for example you have to set `$GFX_NAME` manually to `gfx1030` for consumer RDNA2 cards.
+
+Windows User Refer to [docs/hipBLAS_on_Windows.md](docs%2FhipBLAS_on_Windows.md) for a comprehensive guide.
+
+```
+if command -v rocminfo; then export GFX_NAME=$(rocminfo | awk '/ *Name: +gfx[1-9]/ {print $2; exit}'); else echo "rocminfo missing!"; fi
+if [ -z "${GFX_NAME}" ]; then echo "Error: Couldn't detect GPU!"; else echo "Building for GPU: ${GFX_NAME}"; fi
+cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS=$GFX_NAME -DAMDGPU_TARGETS=$GFX_NAME -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+cmake --build . --config Release
+```
+
+##### Using MUSA
+
+This provides BLAS acceleration using the MUSA cores of your Moore Threads GPU. Make sure to have the MUSA toolkit installed.
+
+```bash
+cmake .. -DCMAKE_C_COMPILER=/usr/local/musa/bin/clang -DCMAKE_CXX_COMPILER=/usr/local/musa/bin/clang++ -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release
 cmake --build . --config Release
 ```
 
@@ -124,15 +156,130 @@ cmake .. -DSD_METAL=ON
 cmake --build . --config Release
 ```
 
-### Using Flash Attention
+##### Using Vulkan
 
-Enabling flash attention reduces memory usage by at least 400 MB. At the moment, it is not supported when CUBLAS is enabled because the kernel implementation is missing.
+Install Vulkan SDK from https://www.lunarg.com/vulkan-sdk/.
 
 ```
-cmake .. -DSD_FLASH_ATTN=ON
+cmake .. -DSD_VULKAN=ON
 cmake --build . --config Release
 ```
 
+##### Using OpenCL (for Adreno GPU)
+
+Currently, it supports only Adreno GPUs and is primarily optimized for Q4_0 type
+
+To build for Windows ARM please refers to [Windows 11 Arm64
+](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/OPENCL.md#windows-11-arm64)
+
+Building for Android:
+
+  Android NDK:
+       Download and install the Android NDK from the [official Android developer site](https://developer.android.com/ndk/downloads).
+
+Setup OpenCL Dependencies for NDK:
+
+You need to provide OpenCL headers and the ICD loader library to your NDK sysroot.
+
+*   OpenCL Headers:
+    ```bash
+    # In a temporary working directory
+    git clone https://github.com/KhronosGroup/OpenCL-Headers
+    cd OpenCL-Headers
+    # Replace <YOUR_NDK_PATH> with your actual NDK installation path
+    # e.g., cp -r CL /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+    sudo cp -r CL <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+    cd ..
+    ```
+
+*   OpenCL ICD Loader:
+    ```bash
+    # In the same temporary working directory
+    git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
+    cd OpenCL-ICD-Loader
+    mkdir build_ndk && cd build_ndk
+
+    # Replace <YOUR_NDK_PATH> in the CMAKE_TOOLCHAIN_FILE and OPENCL_ICD_LOADER_HEADERS_DIR
+    cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \
+      -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
+      -DOPENCL_ICD_LOADER_HEADERS_DIR=<YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \
+      -DANDROID_ABI=arm64-v8a \
+      -DANDROID_PLATFORM=24 \
+      -DANDROID_STL=c++_shared
+
+    ninja
+    # Replace <YOUR_NDK_PATH>
+    # e.g., cp libOpenCL.so /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+    sudo cp libOpenCL.so <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+    cd ../..
+    ```
+
+Build `stable-diffusion.cpp` for Android with OpenCL:
+
+```bash
+mkdir build-android && cd build-android
+
+# Replace <YOUR_NDK_PATH> with your actual NDK installation path
+# e.g., -DCMAKE_TOOLCHAIN_FILE=/path/to/android-ndk-r26c/build/cmake/android.toolchain.cmake
+cmake .. -G Ninja \
+  -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
+  -DANDROID_ABI=arm64-v8a \
+  -DANDROID_PLATFORM=android-28 \
+  -DGGML_OPENMP=OFF \
+  -DSD_OPENCL=ON
+
+ninja
+```
+*(Note: Don't forget to include `LD_LIBRARY_PATH=/vendor/lib64` in your command line before running the binary)*
+
+##### Using SYCL
+
+Using SYCL makes the computation run on the Intel GPU. Please make sure you have installed the related driver and [Intel® oneAPI Base toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) before start. More details and steps can refer to [llama.cpp SYCL backend](https://github.com/ggerganov/llama.cpp/blob/master/docs/backend/SYCL.md#linux).
+
+```
+# Export relevant ENV variables
+source /opt/intel/oneapi/setvars.sh
+
+# Option 1: Use FP32 (recommended for better performance in most cases)
+cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+# Option 2: Use FP16
+cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+
+cmake --build . --config Release
+```
+
+Example of text2img by using SYCL backend:
+
+- download `stable-diffusion` model weight, refer to [download-weight](#download-weights).
+
+- run `./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors --cfg-scale 5 --steps 30 --sampling-method euler  -H 1024 -W 1024 --seed 42 -p "fantasy medieval village world inside a glass sphere , high detail, fantasy, realistic, light effect, hyper detail, volumetric lighting, cinematic, macro, depth of field, blur, red light and clouds from the back, highly detailed epic cinematic concept art cg render made in maya, blender and photoshop, octane render, excellent composition, dynamic dramatic cinematic lighting, aesthetic, very inspirational, world inside a glass sphere by james gurney by artgerm with james jean, joe fenton and tristan eaton by ross tran, fine details, 4k resolution"`
+
+<p align="center">
+  <img src="./assets/sycl_sd3_output.png" width="360x">
+</p>
+
+
+
+##### Using Flash Attention
+
+Enabling flash attention for the diffusion model reduces memory usage by varying amounts of MB.
+eg.:
+ - flux 768x768 ~600mb
+ - SD2 768x768 ~1400mb
+
+For most backends, it slows things down, but for cuda it generally speeds it up too.
+At the moment, it is only supported for some models and some backends (like cpu, cuda/rocm, metal).
+
+Run by adding `--diffusion-fa` to the arguments and watch for:
+```
+[INFO ] stable-diffusion.cpp:312  - Using flash attention in the diffusion model
+```
+and the compute buffer shrink in the debug log:
+```
+[DEBUG] ggml_extend.hpp:1004 - flux compute buffer size: 650.00 MB(VRAM)
+```
+
 ### Run
 
 ```
@@ -140,52 +287,120 @@ usage: ./bin/sd [arguments]
 
 arguments:
   -h, --help                         show this help message and exit
-  -M, --mode [txt2img or img2img]    generation mode (default: txt2img)
-  -t, --threads N                    number of threads to use during computation (default: -1).
+  -M, --mode [MODE]                  run mode, one of: [img_gen, vid_gen, upscale, convert], default: img_gen
+  -t, --threads N                    number of threads to use during computation (default: -1)
                                      If threads <= 0, then threads will be set to the number of CPU physical cores
-  -m, --model [MODEL]                path to model
+  --offload-to-cpu                   place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
+  -m, --model [MODEL]                path to full model
+  --diffusion-model                  path to the standalone diffusion model
+  --high-noise-diffusion-model       path to the standalone high noise diffusion model
+  --clip_l                           path to the clip-l text encoder
+  --clip_g                           path to the clip-g text encoder
+  --clip_vision                      path to the clip-vision encoder
+  --t5xxl                            path to the t5xxl text encoder
+  --qwen2vl                          path to the qwen2vl text encoder
+  --qwen2vl_vision                   path to the qwen2vl vit
   --vae [VAE]                        path to vae
   --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
-  --upscale-model [ESRGAN_PATH]      path to esrgan model. Upscale images after generate, just RealESRGAN_x4plus_anime_6B supported by now.
-  --type [TYPE]                      weight type (f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0)
-                                     If not specified, the default is the type of the weight file.
+  --control-net [CONTROL_PATH]       path to control net model
+  --embd-dir [EMBEDDING_PATH]        path to embeddings
+  --upscale-model [ESRGAN_PATH]      path to esrgan model. For img_gen mode, upscale images after generate, just RealESRGAN_x4plus_anime_6B supported by now
+  --upscale-repeats                  Run the ESRGAN upscaler this many times (default 1)
+  --type [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
+  --tensor-type-rules [EXPRESSION]   weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
   --lora-model-dir [DIR]             lora model directory
-  -i, --init-img [IMAGE]             path to the input image, required by img2img
+  -i, --init-img [IMAGE]             path to the init image, required by img2img
+  --mask [MASK]                      path to the mask image, required by img2img with mask
+  -i, --end-img [IMAGE]              path to the end image, required by flf2v
+  --control-image [IMAGE]            path to image condition, control net
+  -r, --ref-image [PATH]             reference image for Flux Kontext models (can be used multiple times)
+  --control-video [PATH]             path to control video frames, It must be a directory path.
+                                     The video frames inside should be stored as images in lexicographical (character) order
+                                     For example, if the control video path is `frames`, the directory contain images such as 00.png, 01.png, 鈥?etc.
+  --increase-ref-index               automatically increase the indices of references images based on the order they are listed (starting with 1).
   -o, --output OUTPUT                path to write result image to (default: ./output.png)
   -p, --prompt [PROMPT]              the prompt to render
   -n, --negative-prompt PROMPT       the negative prompt (default: "")
   --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)
+  --img-cfg-scale SCALE              image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --guidance SCALE                   distilled guidance scale for models with guidance input (default: 3.5)
+  --slg-scale SCALE                  skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+                                     0 means disabled, a value of 2.5 is nice for sd3.5 medium
+  --eta SCALE                        eta in DDIM, only for DDIM and TCD: (default: 0)
+  --skip-layers LAYERS               Layers to skip for SLG steps: (default: [7,8,9])
+  --skip-layer-start START           SLG enabling point: (default: 0.01)
+  --skip-layer-end END               SLG disabling point: (default: 0.2)
+  --scheduler {discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple} Denoiser sigma scheduler (default: discrete)
+  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
+                                     sampling method (default: "euler" for Flux/SD3/Wan, "euler_a" otherwise)
+  --timestep-shift N                 shift timestep for NitroFusion models, default: 0, recommended N for NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
+  --steps  STEPS                     number of sample steps (default: 20)
+  --high-noise-cfg-scale SCALE       (high noise) unconditional guidance scale: (default: 7.0)
+  --high-noise-img-cfg-scale SCALE   (high noise) image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --high-noise-guidance SCALE        (high noise) distilled guidance scale for models with guidance input (default: 3.5)
+  --high-noise-slg-scale SCALE       (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+                                     0 means disabled, a value of 2.5 is nice for sd3.5 medium
+  --high-noise-eta SCALE             (high noise) eta in DDIM, only for DDIM and TCD: (default: 0)
+  --high-noise-skip-layers LAYERS    (high noise) Layers to skip for SLG steps: (default: [7,8,9])
+  --high-noise-skip-layer-start      (high noise) SLG enabling point: (default: 0.01)
+  --high-noise-skip-layer-end END    (high noise) SLG disabling point: (default: 0.2)
+  --high-noise-scheduler {discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple} Denoiser sigma scheduler (default: discrete)
+  --high-noise-sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
+                                     (high noise) sampling method (default: "euler_a")
+  --high-noise-steps  STEPS          (high noise) number of sample steps (default: -1 = auto)
+                                     SLG will be enabled at step int([STEPS]*[START]) and disabled at int([STEPS]*[END])
   --strength STRENGTH                strength for noising/unnoising (default: 0.75)
+  --control-strength STRENGTH        strength to apply Control Net (default: 0.9)
                                      1.0 corresponds to full destruction of information in init image
   -H, --height H                     image height, in pixel space (default: 512)
   -W, --width W                      image width, in pixel space (default: 512)
-  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, lcm}
-                                     sampling method (default: "euler_a")
-  --steps  STEPS                     number of sample steps (default: 20)
   --rng {std_default, cuda}          RNG (default: cuda)
   -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)
-  -b, --batch-count COUNT            number of images to generate.
-  --schedule {discrete, karras}      Denoiser sigma schedule (default: discrete)
-  --clip-skip N                      number of layers to skip of clip model (default: 0)
+  -b, --batch-count COUNT            number of images to generate
+  --clip-skip N                      ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)
+                                     <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
   --vae-tiling                       process vae in tiles to reduce memory usage
+  --vae-tile-size [X]x[Y]            tile size for vae tiling (default: 32x32)
+  --vae-relative-tile-size [X]x[Y]   relative tile size for vae tiling, in fraction of image size if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
+  --vae-tile-overlap OVERLAP         tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --vae-on-cpu                       keep vae in cpu (for low vram)
+  --clip-on-cpu                      keep clip in cpu (for low vram)
+  --diffusion-fa                     use flash attention in the diffusion model (for low vram)
+                                     Might lower quality, since it implies converting k and v to f16.
+                                     This might crash if it is not supported by the backend.
+  --diffusion-conv-direct            use Conv2d direct in the diffusion model
+                                     This might crash if it is not supported by the backend.
+  --vae-conv-direct                  use Conv2d direct in the vae model (should improve the performance)
+                                     This might crash if it is not supported by the backend.
+  --control-net-cpu                  keep controlnet in cpu (for low vram)
+  --canny                            apply canny preprocessor (edge detection)
+  --color                            colors the logging tags according to level
+  --chroma-disable-dit-mask          disable dit mask for chroma
+  --chroma-enable-t5-mask            enable t5 mask for chroma
+  --chroma-t5-mask-pad  PAD_SIZE     t5 mask pad size of chroma
+  --video-frames                     video frames (default: 1)
+  --fps                              fps (default: 24)
+  --moe-boundary BOUNDARY            timestep boundary for Wan2.2 MoE model. (default: 0.875)
+                                     only enabled if `--high-noise-steps` is set to -1
+  --flow-shift SHIFT                 shift value for Flow models like SD3.x or WAN (default: auto)
+  --vace-strength                    wan vace strength
+  --photo-maker                      path to PHOTOMAKER model
+  --pm-id-images-dir [DIR]           path to PHOTOMAKER input id images dir
+  --pm-id-embed-path [PATH]          path to PHOTOMAKER v2 id embed
+  --pm-style-strength                strength for keeping PHOTOMAKER input identity (default: 20)
   -v, --verbose                      print extra info
 ```
 
-#### Quantization
-
-You can specify the model weight type using the `--type` parameter. The weights are automatically converted when loading the model.
-
-- `f16` for 16-bit floating-point
-- `f32` for 32-bit floating-point
-- `q8_0` for 8-bit integer quantization
-- `q5_0` or `q5_1` for 5-bit integer quantization
-- `q4_0` or `q4_1` for 4-bit integer quantization
-
 #### txt2img example
 
 ```sh
 ./bin/sd -m ../models/sd-v1-4.ckpt -p "a lovely cat"
 # ./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
+# ./bin/sd -m ../models/sd_xl_base_1.0.safetensors --vae ../models/sdxl_vae-fp16-fix.safetensors -H 1024 -W 1024 -p "a lovely cat" -v
+# ./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable Diffusion CPP\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd --diffusion-model  ../models/flux1-dev-q3_k.gguf --vae ../models/ae.sft --clip_l ../models/clip_l.safetensors --t5xxl ../models/t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
 ```
 
 Using formats of different precisions will yield results of varying quality.
@@ -200,93 +415,45 @@ Using formats of different precisions will yield results of varying quality.
 
 
 ```
-./bin/sd --mode img2img -m ../models/sd-v1-4.ckpt -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
+./bin/sd -m ../models/sd-v1-4.ckpt -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
 ```
 
 <p align="center">
   <img src="./assets/img2img_output.png" width="256x">
 </p>
 
-#### with LoRA
+## More Guides
 
-- You can specify the directory where the lora weights are stored via `--lora-model-dir`. If not specified, the default is the current working directory.
+- [LoRA](./docs/lora.md)
+- [LCM/LCM-LoRA](./docs/lcm.md)
+- [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)
+- [Using ESRGAN to upscale results](./docs/esrgan.md)
+- [Using TAESD to faster decoding](./docs/taesd.md)
+- [Docker](./docs/docker.md)
+- [Quantization and GGUF](./docs/quantization_and_gguf.md)
 
-- LoRA is specified via prompt, just like [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#lora).
+## Bindings
 
-Here's a simple example:
+These projects wrap `stable-diffusion.cpp` for easier use in other languages/frameworks.
 
-```
-./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:marblesh:1>" --lora-model-dir ../models
-```
+* Golang (non-cgo): [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
+* Golang (cgo): [Binozo/GoStableDiffusion](https://github.com/Binozo/GoStableDiffusion)
+* C#: [DarthAffe/StableDiffusion.NET](https://github.com/DarthAffe/StableDiffusion.NET)
+* Python: [william-murray1204/stable-diffusion-cpp-python](https://github.com/william-murray1204/stable-diffusion-cpp-python)
+* Rust: [newfla/diffusion-rs](https://github.com/newfla/diffusion-rs)
+* Flutter/Dart: [rmatif/Local-Diffusion](https://github.com/rmatif/Local-Diffusion)
 
-`../models/marblesh.safetensors` or `../models/marblesh.ckpt` will be applied to the model
+## UIs
 
-#### LCM/LCM-LoRA
+These projects use `stable-diffusion.cpp` as a backend for their image generation.
 
-- Download LCM-LoRA form https://huggingface.co/latent-consistency/lcm-lora-sdv1-5
-- Specify LCM-LoRA by adding `<lora:lcm-lora-sdv1-5:1>` to prompt
-- It's advisable to set `--cfg-scale` to `1.0` instead of the default `7.0`. For `--steps`, a range of `2-8` steps is recommended. For `--sampling-method`, `lcm`/`euler_a` is recommended.
-
-Here's a simple example:
-
-```
-./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:lcm-lora-sdv1-5:1>" --steps 4 --lora-model-dir ../models -v --cfg-scale 1
-```
-
-| without LCM-LoRA (--cfg-scale 7)  | with LCM-LoRA (--cfg-scale 1)  |
-| ----  |----    |
-| ![](./assets/without_lcm.png) |![](./assets/with_lcm.png)  |
-
-## Using TAESD to faster decoding
-
-You can use TAESD to accelerate the decoding of latent images by following these steps:
-
-- Download the model [weights](https://huggingface.co/madebyollin/taesd/blob/main/diffusion_pytorch_model.safetensors).
-
-Or curl
-
-```bash
-curl -L -O https://huggingface.co/madebyollin/taesd/blob/main/diffusion_pytorch_model.safetensors
-```
-
-- Specify the model path using the `--taesd PATH` parameter. example:
-
-```bash
-sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --taesd ../models/diffusion_pytorch_model.safetensors
-```
-
-## 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.
-
-- Specify the model path using the `--upscale-model PATH` parameter. example:
-
-```bash
-sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --upscale-model ../models/RealESRGAN_x4plus_anime_6B.pth
-```
-
-### Docker
-
-#### Building using Docker
-
-```shell
-docker build -t sd .
-```
-
-#### Run
-
-```shell
-docker run -v /path/to/models:/models -v /path/to/output/:/output sd [args...]
-# For example
-# docker run -v ./models:/models -v ./build:/output sd -m /models/sd-v1-4.ckpt -p "a lovely cat" -v -o /output/output.png
-```
-
-## Memory Requirements
-
-| precision | f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
-| ----         | ----  |----  |----  |----  |----  |----  |----  |
-|  **Memory** (txt2img - 512 x 512) | ~2.8G | ~2.3G | ~2.1G | ~2.0G | ~2.0G | ~2.0G | ~2.0G |
-|  **Memory** (txt2img - 512 x 512) *with Flash Attention* | ~2.4G | ~1.9G | ~1.6G | ~1.5G | ~1.5G | ~1.5G | ~1.5G |
+- [Jellybox](https://jellybox.com)
+- [Stable Diffusion GUI](https://github.com/fszontagh/sd.cpp.gui.wx)
+- [Stable Diffusion CLI-GUI](https://github.com/piallai/stable-diffusion.cpp)
+- [Local Diffusion](https://github.com/rmatif/Local-Diffusion)
+- [sd.cpp-webui](https://github.com/daniandtheweb/sd.cpp-webui)
+- [LocalAI](https://github.com/mudler/LocalAI)
+- [Neural-Pixel](https://github.com/Luiz-Alcantara/Neural-Pixel)
 
 ## Contributors
 
@@ -294,12 +461,22 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp
 
 [![Contributors](https://contrib.rocks/image?repo=leejet/stable-diffusion.cpp)](https://github.com/leejet/stable-diffusion.cpp/graphs/contributors)
 
+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=leejet/stable-diffusion.cpp&type=Date)](https://star-history.com/#leejet/stable-diffusion.cpp&Date)
+
 ## References
 
 - [ggml](https://github.com/ggerganov/ggml)
+- [diffusers](https://github.com/huggingface/diffusers)
 - [stable-diffusion](https://github.com/CompVis/stable-diffusion)
+- [sd3-ref](https://github.com/Stability-AI/sd3-ref)
 - [stable-diffusion-stability-ai](https://github.com/Stability-AI/stablediffusion)
 - [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui)
+- [ComfyUI](https://github.com/comfyanonymous/ComfyUI)
 - [k-diffusion](https://github.com/crowsonkb/k-diffusion)
 - [latent-consistency-model](https://github.com/luosiallen/latent-consistency-model)
 - [generative-models](https://github.com/Stability-AI/generative-models/)
+- [PhotoMaker](https://github.com/TencentARC/PhotoMaker)
+- [Wan2.1](https://github.com/Wan-Video/Wan2.1)
+- [Wan2.2](https://github.com/Wan-Video/Wan2.2)

clip.hpp

@@ -0,0 +1,969 @@
+#ifndef __CLIP_HPP__
+#define __CLIP_HPP__
+
+#include "ggml_extend.hpp"
+#include "model.h"
+
+/*================================================== CLIPTokenizer ===================================================*/
+
+__STATIC_INLINE__ std::pair<std::unordered_map<std::string, float>, std::string> extract_and_remove_lora(std::string text) {
+    std::regex re("<lora:([^:]+):([^>]+)>");
+    std::smatch matches;
+    std::unordered_map<std::string, float> filename2multiplier;
+
+    while (std::regex_search(text, matches, re)) {
+        std::string filename = matches[1].str();
+        float multiplier     = std::stof(matches[2].str());
+
+        text = std::regex_replace(text, re, "", std::regex_constants::format_first_only);
+
+        if (multiplier == 0.f) {
+            continue;
+        }
+
+        if (filename2multiplier.find(filename) == filename2multiplier.end()) {
+            filename2multiplier[filename] = multiplier;
+        } else {
+            filename2multiplier[filename] += multiplier;
+        }
+    }
+
+    return std::make_pair(filename2multiplier, text);
+}
+
+__STATIC_INLINE__ std::vector<std::pair<int, std::u32string>> bytes_to_unicode() {
+    std::vector<std::pair<int, std::u32string>> byte_unicode_pairs;
+    std::set<int> byte_set;
+    for (int b = static_cast<int>('!'); b <= static_cast<int>('~'); ++b) {
+        byte_set.insert(b);
+        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
+    }
+    for (int b = 161; b <= 172; ++b) {
+        byte_set.insert(b);
+        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
+    }
+    for (int b = 174; b <= 255; ++b) {
+        byte_set.insert(b);
+        byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(b)));
+    }
+    int n = 0;
+    for (int b = 0; b < 256; ++b) {
+        if (byte_set.find(b) == byte_set.end()) {
+            byte_unicode_pairs.push_back(std::pair<int, std::u32string>(b, unicode_value_to_utf32(n + 256)));
+            ++n;
+        }
+    }
+    // LOG_DEBUG("byte_unicode_pairs %d", byte_unicode_pairs.size());
+    return byte_unicode_pairs;
+}
+
+// Ref: https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py
+
+typedef std::function<bool(std::string&, std::vector<int32_t>&)> on_new_token_cb_t;
+
+class CLIPTokenizer {
+private:
+    std::map<int, std::u32string> byte_encoder;
+    std::map<std::u32string, int> byte_decoder;
+    std::map<std::u32string, int> encoder;
+    std::map<int, std::u32string> decoder;
+    std::map<std::pair<std::u32string, std::u32string>, int> bpe_ranks;
+    std::regex pat;
+    int encoder_len;
+    int bpe_len;
+
+public:
+    const std::string UNK_TOKEN = "<|endoftext|>";
+    const std::string BOS_TOKEN = "<|startoftext|>";
+    const std::string EOS_TOKEN = "<|endoftext|>";
+    const std::string PAD_TOKEN = "<|endoftext|>";
+
+    const int UNK_TOKEN_ID = 49407;
+    const int BOS_TOKEN_ID = 49406;
+    const int EOS_TOKEN_ID = 49407;
+    const int PAD_TOKEN_ID = 49407;
+
+private:
+    static std::string strip(const std::string& str) {
+        std::string::size_type start = str.find_first_not_of(" \t\n\r\v\f");
+        std::string::size_type end   = str.find_last_not_of(" \t\n\r\v\f");
+
+        if (start == std::string::npos) {
+            // String contains only whitespace characters
+            return "";
+        }
+
+        return str.substr(start, end - start + 1);
+    }
+
+    static std::string whitespace_clean(std::string text) {
+        text = std::regex_replace(text, std::regex(R"(\s+)"), " ");
+        text = strip(text);
+        return text;
+    }
+
+    static std::set<std::pair<std::u32string, std::u32string>> get_pairs(const std::vector<std::u32string>& subwords) {
+        std::set<std::pair<std::u32string, std::u32string>> pairs;
+        if (subwords.size() == 0) {
+            return pairs;
+        }
+        std::u32string prev_subword = subwords[0];
+        for (int i = 1; i < subwords.size(); i++) {
+            std::u32string subword = subwords[i];
+            std::pair<std::u32string, std::u32string> pair(prev_subword, subword);
+            pairs.insert(pair);
+            prev_subword = subword;
+        }
+        return pairs;
+    }
+
+public:
+    CLIPTokenizer(int pad_token_id = 49407, const std::string& merges_utf8_str = "")
+        : PAD_TOKEN_ID(pad_token_id) {
+        if (merges_utf8_str.size() > 0) {
+            load_from_merges(merges_utf8_str);
+        } else {
+            load_from_merges(ModelLoader::load_merges());
+        }
+    }
+
+    void load_from_merges(const std::string& merges_utf8_str) {
+        auto byte_unicode_pairs = bytes_to_unicode();
+        // printf("byte_unicode_pairs have %lu pairs \n", byte_unicode_pairs.size());
+        byte_encoder = std::map<int, std::u32string>(byte_unicode_pairs.begin(), byte_unicode_pairs.end());
+        for (auto& pair : byte_unicode_pairs) {
+            byte_decoder[pair.second] = pair.first;
+        }
+        // for (auto & pair: byte_unicode_pairs) {
+        //     std::cout << pair.first << ": " << pair.second << std::endl;
+        // }
+        std::vector<std::u32string> merges;
+        size_t start = 0;
+        size_t pos;
+        std::u32string merges_utf32_str = utf8_to_utf32(merges_utf8_str);
+        while ((pos = merges_utf32_str.find('\n', start)) != std::string::npos) {
+            merges.push_back(merges_utf32_str.substr(start, pos - start));
+            start = pos + 1;
+        }
+        // LOG_DEBUG("merges size %llu", merges.size());
+        GGML_ASSERT(merges.size() == 48895);
+        merges = std::vector<std::u32string>(merges.begin() + 1, merges.end());
+        std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
+        for (const auto& merge : merges) {
+            size_t space_pos = merge.find(' ');
+            merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
+            // LOG_DEBUG("%s", utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
+            // printf("%s :: %s | %s \n", utf32_to_utf8(merge).c_str(), utf32_to_utf8(merge.substr(0, space_pos)).c_str(),
+            //                     utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
+        }
+        std::vector<std::u32string> vocab;
+        for (const auto& pair : byte_unicode_pairs) {
+            vocab.push_back(pair.second);
+        }
+        for (const auto& pair : byte_unicode_pairs) {
+            vocab.push_back(pair.second + utf8_to_utf32("</w>"));
+        }
+        for (const auto& merge : merge_pairs) {
+            vocab.push_back(merge.first + merge.second);
+        }
+        vocab.push_back(utf8_to_utf32("<|startoftext|>"));
+        vocab.push_back(utf8_to_utf32("<|endoftext|>"));
+        LOG_DEBUG("vocab size: %llu", vocab.size());
+        int i = 0;
+        for (const auto& token : vocab) {
+            encoder[token] = i;
+            decoder[i]     = token;
+            i++;
+        }
+        encoder_len = i;
+
+        auto it = encoder.find(utf8_to_utf32("img</w>"));
+        if (it != encoder.end()) {
+            LOG_DEBUG("trigger word img already in vocab");
+        } else {
+            LOG_DEBUG("trigger word img not in vocab yet");
+        }
+
+        int rank = 0;
+        for (const auto& merge : merge_pairs) {
+            bpe_ranks[merge] = rank++;
+        }
+        bpe_len = rank;
+    };
+
+    void add_token(const std::string& text) {
+        std::u32string token = utf8_to_utf32(text);
+        auto it              = encoder.find(token);
+        if (it != encoder.end()) {
+            encoder[token]       = encoder_len;
+            decoder[encoder_len] = token;
+            encoder_len++;
+        }
+    }
+
+    std::u32string bpe(const std::u32string& token) {
+        std::vector<std::u32string> word;
+
+        for (int i = 0; i < token.size() - 1; i++) {
+            word.emplace_back(1, token[i]);
+        }
+        word.push_back(token.substr(token.size() - 1) + utf8_to_utf32("</w>"));
+
+        std::set<std::pair<std::u32string, std::u32string>> pairs = get_pairs(word);
+
+        if (pairs.empty()) {
+            return token + utf8_to_utf32("</w>");
+        }
+
+        while (true) {
+            auto min_pair_iter = std::min_element(pairs.begin(),
+                                                  pairs.end(),
+                                                  [&](const std::pair<std::u32string, std::u32string>& a,
+                                                      const std::pair<std::u32string, std::u32string>& b) {
+                                                      if (bpe_ranks.find(a) == bpe_ranks.end()) {
+                                                          return false;
+                                                      } else if (bpe_ranks.find(b) == bpe_ranks.end()) {
+                                                          return true;
+                                                      }
+                                                      return bpe_ranks.at(a) < bpe_ranks.at(b);
+                                                  });
+
+            const std::pair<std::u32string, std::u32string>& bigram = *min_pair_iter;
+
+            if (bpe_ranks.find(bigram) == bpe_ranks.end()) {
+                break;
+            }
+
+            std::u32string first  = bigram.first;
+            std::u32string second = bigram.second;
+            std::vector<std::u32string> new_word;
+            int32_t i = 0;
+
+            while (i < word.size()) {
+                auto it = std::find(word.begin() + i, word.end(), first);
+                if (it == word.end()) {
+                    new_word.insert(new_word.end(), word.begin() + i, word.end());
+                    break;
+                }
+                new_word.insert(new_word.end(), word.begin() + i, it);
+                i = static_cast<int32_t>(std::distance(word.begin(), it));
+
+                if (word[i] == first && i < static_cast<int32_t>(word.size()) - 1 && word[i + 1] == second) {
+                    new_word.push_back(first + second);
+                    i += 2;
+                } else {
+                    new_word.push_back(word[i]);
+                    i += 1;
+                }
+            }
+
+            word = new_word;
+
+            if (word.size() == 1) {
+                break;
+            }
+            pairs = get_pairs(word);
+        }
+
+        std::u32string result;
+        for (int i = 0; i < word.size(); i++) {
+            result += word[i];
+            if (i != word.size() - 1) {
+                result += utf8_to_utf32(" ");
+            }
+        }
+
+        return result;
+    }
+
+    std::vector<int> tokenize(std::string text,
+                              on_new_token_cb_t on_new_token_cb,
+                              size_t max_length = 0,
+                              bool padding      = false) {
+        std::vector<int32_t> tokens = encode(text, on_new_token_cb);
+
+        tokens.insert(tokens.begin(), BOS_TOKEN_ID);
+        if (max_length > 0) {
+            if (tokens.size() > max_length - 1) {
+                tokens.resize(max_length - 1);
+                tokens.push_back(EOS_TOKEN_ID);
+            } else {
+                tokens.push_back(EOS_TOKEN_ID);
+                if (padding) {
+                    tokens.insert(tokens.end(), max_length - tokens.size(), PAD_TOKEN_ID);
+                }
+            }
+        }
+
+        return tokens;
+    }
+
+    void pad_tokens(std::vector<int>& tokens,
+                    std::vector<float>& weights,
+                    size_t max_length = 0,
+                    bool padding      = false) {
+        if (max_length > 0 && padding) {
+            size_t n = std::ceil(tokens.size() * 1.0 / (max_length - 2));
+            if (n == 0) {
+                n = 1;
+            }
+            size_t length = max_length * n;
+            LOG_DEBUG("token length: %llu", length);
+            std::vector<int> new_tokens;
+            std::vector<float> new_weights;
+            new_tokens.push_back(BOS_TOKEN_ID);
+            new_weights.push_back(1.0);
+            int token_idx = 0;
+            for (int i = 1; i < length; i++) {
+                if (token_idx >= tokens.size()) {
+                    break;
+                }
+                if (i % max_length == 0) {
+                    new_tokens.push_back(BOS_TOKEN_ID);
+                    new_weights.push_back(1.0);
+                } else if (i % max_length == max_length - 1) {
+                    new_tokens.push_back(EOS_TOKEN_ID);
+                    new_weights.push_back(1.0);
+                } else {
+                    new_tokens.push_back(tokens[token_idx]);
+                    new_weights.push_back(weights[token_idx]);
+                    token_idx++;
+                }
+            }
+
+            new_tokens.push_back(EOS_TOKEN_ID);
+            new_weights.push_back(1.0);
+            tokens  = new_tokens;
+            weights = new_weights;
+
+            if (padding) {
+                tokens.insert(tokens.end(), length - tokens.size(), PAD_TOKEN_ID);
+                weights.insert(weights.end(), length - weights.size(), 1.0);
+            }
+        }
+    }
+
+    std::string clean_up_tokenization(std::string& text) {
+        std::regex pattern(R"( ,)");
+        // Replace " ," with ","
+        std::string result = std::regex_replace(text, pattern, ",");
+        return result;
+    }
+
+    std::string decode(const std::vector<int>& tokens) {
+        std::string text = "";
+        for (int t : tokens) {
+            if (t == 49406 || t == 49407)
+                continue;
+            std::u32string ts = decoder[t];
+            // printf("%d, %s \n", t,  utf32_to_utf8(ts).c_str());
+            std::string s = utf32_to_utf8(ts);
+            if (s.length() >= 4) {
+                if (ends_with(s, "</w>")) {
+                    text += s.replace(s.length() - 4, s.length() - 1, "") + " ";
+                } else {
+                    text += s;
+                }
+            } else {
+                text += " " + s;
+            }
+        }
+        // std::vector<unsigned char> bytes;
+        // for (auto c : text){
+        //     bytes.push_back(byte_decoder[c]);
+        // }
+
+        // std::string s((char *)bytes.data());
+        // std::string s = "";
+        text = clean_up_tokenization(text);
+        return trim(text);
+    }
+
+    std::vector<int> encode(std::string text, on_new_token_cb_t on_new_token_cb) {
+        std::string original_text = text;
+        std::vector<int32_t> bpe_tokens;
+        text = whitespace_clean(text);
+        std::transform(text.begin(), text.end(), text.begin(), [](unsigned char c) { return std::tolower(c); });
+
+        std::regex pat(R"(<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[[:alpha:]]+|[[:digit:]]|[^[:space:][:alpha:][:digit:]]+)",
+                       std::regex::icase);
+
+        std::smatch matches;
+        std::string str = text;
+        std::vector<std::string> token_strs;
+        while (std::regex_search(str, matches, pat)) {
+            bool skip = on_new_token_cb(str, bpe_tokens);
+            if (skip) {
+                continue;
+            }
+            for (auto& token : matches) {
+                std::string token_str = token.str();
+                std::u32string utf32_token;
+                for (int i = 0; i < token_str.length(); i++) {
+                    unsigned char b = token_str[i];
+                    utf32_token += byte_encoder[b];
+                }
+                auto bpe_strs = bpe(utf32_token);
+                size_t start  = 0;
+                size_t pos;
+                while ((pos = bpe_strs.find(' ', start)) != std::u32string::npos) {
+                    auto bpe_str = bpe_strs.substr(start, pos - start);
+                    bpe_tokens.push_back(encoder[bpe_str]);
+                    token_strs.push_back(utf32_to_utf8(bpe_str));
+
+                    start = pos + 1;
+                }
+                auto bpe_str = bpe_strs.substr(start, bpe_strs.size() - start);
+                bpe_tokens.push_back(encoder[bpe_str]);
+                token_strs.push_back(utf32_to_utf8(bpe_str));
+            }
+            str = matches.suffix();
+        }
+        std::stringstream ss;
+        ss << "[";
+        for (auto token : token_strs) {
+            ss << "\"" << token << "\", ";
+        }
+        ss << "]";
+        // LOG_DEBUG("split prompt \"%s\" to tokens %s", original_text.c_str(), ss.str().c_str());
+        // printf("split prompt \"%s\" to tokens %s \n", original_text.c_str(), ss.str().c_str());
+        return bpe_tokens;
+    }
+};
+
+/*================================================ FrozenCLIPEmbedder ================================================*/
+
+// Ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/clip/modeling_clip.py
+
+struct CLIPMLP : public GGMLBlock {
+protected:
+    bool use_gelu;
+
+public:
+    CLIPMLP(int64_t d_model, int64_t intermediate_size) {
+        blocks["fc1"] = std::shared_ptr<GGMLBlock>(new Linear(d_model, intermediate_size));
+        blocks["fc2"] = std::shared_ptr<GGMLBlock>(new Linear(intermediate_size, d_model));
+
+        if (d_model == 1024 || d_model == 1280) {  // SD 2.x
+            use_gelu = true;
+        } else {  // SD 1.x
+            use_gelu = false;
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct 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, x);
+        } else {
+            x = ggml_gelu_quick_inplace(ctx, x);
+        }
+        x = fc2->forward(ctx, x);
+        return x;
+    }
+};
+
+struct CLIPLayer : public GGMLBlock {
+protected:
+    int64_t d_model;  // hidden_size/embed_dim
+    int64_t n_head;
+    int64_t intermediate_size;
+
+public:
+    CLIPLayer(int64_t d_model,
+              int64_t n_head,
+              int64_t intermediate_size)
+        : d_model(d_model),
+          n_head(n_head),
+          intermediate_size(intermediate_size) {
+        blocks["self_attn"] = std::shared_ptr<GGMLBlock>(new MultiheadAttention(d_model, n_head, true, true));
+
+        blocks["layer_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(d_model));
+        blocks["layer_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(d_model));
+
+        blocks["mlp"] = std::shared_ptr<GGMLBlock>(new CLIPMLP(d_model, intermediate_size));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, ggml_backend_t backend, struct ggml_tensor* x, bool mask = true) {
+        // 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"]);
+        auto layer_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm2"]);
+        auto mlp         = std::dynamic_pointer_cast<CLIPMLP>(blocks["mlp"]);
+
+        x = ggml_add(ctx, x, self_attn->forward(ctx, backend, layer_norm1->forward(ctx, x), mask));
+        x = ggml_add(ctx, x, mlp->forward(ctx, layer_norm2->forward(ctx, x)));
+        return x;
+    }
+};
+
+struct CLIPEncoder : public GGMLBlock {
+protected:
+    int64_t n_layer;
+
+public:
+    CLIPEncoder(int64_t n_layer,
+                int64_t d_model,
+                int64_t n_head,
+                int64_t intermediate_size)
+        : n_layer(n_layer) {
+        for (int i = 0; i < n_layer; i++) {
+            std::string name = "layers." + std::to_string(i);
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new CLIPLayer(d_model, n_head, intermediate_size));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* x,
+                                int clip_skip = -1,
+                                bool mask     = true) {
+        // x: [N, n_token, d_model]
+        int layer_idx = n_layer - 1;
+        // LOG_DEBUG("clip_skip %d", clip_skip);
+        if (clip_skip > 0) {
+            layer_idx = n_layer - clip_skip;
+        }
+
+        for (int i = 0; i < n_layer; i++) {
+            // LOG_DEBUG("layer %d", i);
+            if (i == layer_idx + 1) {
+                break;
+            }
+            std::string name = "layers." + std::to_string(i);
+            auto layer       = std::dynamic_pointer_cast<CLIPLayer>(blocks[name]);
+            x                = layer->forward(ctx, backend, x, mask);  // [N, n_token, d_model]
+            // LOG_DEBUG("layer %d", i);
+        }
+        return x;
+    }
+};
+
+class CLIPEmbeddings : public GGMLBlock {
+protected:
+    int64_t embed_dim;
+    int64_t vocab_size;
+    int64_t num_positions;
+    bool force_clip_f32;
+
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        enum ggml_type token_wtype = GGML_TYPE_F32;
+        if (!force_clip_f32) {
+            token_wtype = get_type(prefix + "token_embedding.weight", tensor_types, GGML_TYPE_F32);
+            if (!support_get_rows(token_wtype)) {
+                token_wtype = GGML_TYPE_F32;
+            }
+        }
+        enum ggml_type position_wtype       = GGML_TYPE_F32;
+        params["token_embedding.weight"]    = ggml_new_tensor_2d(ctx, token_wtype, embed_dim, vocab_size);
+        params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
+    }
+
+public:
+    CLIPEmbeddings(int64_t embed_dim,
+                   int64_t vocab_size    = 49408,
+                   int64_t num_positions = 77,
+                   bool force_clip_f32   = false)
+        : embed_dim(embed_dim),
+          vocab_size(vocab_size),
+          num_positions(num_positions),
+          force_clip_f32(force_clip_f32) {
+    }
+
+    struct ggml_tensor* get_token_embed_weight() {
+        return params["token_embedding.weight"];
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* input_ids,
+                                struct 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"];
+
+        GGML_ASSERT(input_ids->ne[0] == position_embed_weight->ne[1]);
+        input_ids            = ggml_reshape_3d(ctx, input_ids, input_ids->ne[0], 1, input_ids->ne[1]);
+        auto token_embedding = ggml_get_rows(ctx, custom_embed_weight != NULL ? custom_embed_weight : token_embed_weight, input_ids);
+        token_embedding      = ggml_reshape_3d(ctx, token_embedding, token_embedding->ne[0], token_embedding->ne[1], token_embedding->ne[3]);
+
+        // token_embedding + position_embedding
+        auto x = ggml_add(ctx,
+                          token_embedding,
+                          position_embed_weight);  // [N, n_token, embed_dim]
+        return x;
+    }
+};
+
+class CLIPVisionEmbeddings : public GGMLBlock {
+protected:
+    int64_t embed_dim;
+    int64_t num_channels;
+    int64_t patch_size;
+    int64_t image_size;
+    int64_t num_patches;
+    int64_t num_positions;
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        enum ggml_type patch_wtype    = GGML_TYPE_F16;
+        enum ggml_type class_wtype    = GGML_TYPE_F32;
+        enum ggml_type position_wtype = GGML_TYPE_F32;
+
+        params["patch_embedding.weight"]    = ggml_new_tensor_4d(ctx, patch_wtype, patch_size, patch_size, num_channels, embed_dim);
+        params["class_embedding"]           = ggml_new_tensor_1d(ctx, class_wtype, embed_dim);
+        params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
+    }
+
+public:
+    CLIPVisionEmbeddings(int64_t embed_dim,
+                         int64_t num_channels = 3,
+                         int64_t patch_size   = 14,
+                         int64_t image_size   = 224)
+        : embed_dim(embed_dim),
+          num_channels(num_channels),
+          patch_size(patch_size),
+          image_size(image_size) {
+        num_patches   = (image_size / patch_size) * (image_size / patch_size);
+        num_positions = num_patches + 1;
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct 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);
+
+        auto patch_embed_weight    = params["patch_embedding.weight"];
+        auto class_embed_weight    = params["class_embedding"];
+        auto position_embed_weight = params["position_embedding.weight"];
+
+        // concat(patch_embedding, class_embedding) + position_embedding
+        struct ggml_tensor* patch_embedding;
+        int64_t N       = pixel_values->ne[3];
+        patch_embedding = ggml_nn_conv_2d(ctx, pixel_values, patch_embed_weight, NULL, patch_size, patch_size);  // [N, embed_dim, image_size // pacht_size, image_size // pacht_size]
+        patch_embedding = ggml_reshape_3d(ctx, patch_embedding, num_patches, embed_dim, N);                      // [N, embed_dim, num_patches]
+        patch_embedding = ggml_cont(ctx, ggml_permute(ctx, patch_embedding, 1, 0, 2, 3));                        // [N, num_patches, embed_dim]
+        patch_embedding = ggml_reshape_4d(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_TYPE_F32, embed_dim, N);
+        class_embedding                     = ggml_repeat(ctx, class_embed_weight, class_embedding);      // [N, embed_dim]
+        class_embedding                     = ggml_reshape_4d(ctx, class_embedding, 1, embed_dim, 1, N);  // [N, 1, embed_dim, 1]
+
+        struct ggml_tensor* x = ggml_concat(ctx, class_embedding, patch_embedding, 2);  // [N, num_positions, embed_dim, 1]
+        x                     = ggml_reshape_3d(ctx, x, embed_dim, num_positions, N);   // [N, num_positions, embed_dim]
+        x                     = ggml_add(ctx, x, position_embed_weight);
+        return x;  // [N, num_positions, embed_dim]
+    }
+};
+
+// OPENAI_CLIP_VIT_L_14: https://huggingface.co/openai/clip-vit-large-patch14/blob/main/config.json
+// OPEN_CLIP_VIT_H_14: https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/blob/main/config.json
+// OPEN_CLIP_VIT_BIGG_14: https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k/blob/main/config.json (CLIPTextModelWithProjection)
+
+enum CLIPVersion {
+    OPENAI_CLIP_VIT_L_14,   // SD 1.x and SDXL
+    OPEN_CLIP_VIT_H_14,     // SD 2.x
+    OPEN_CLIP_VIT_BIGG_14,  // SDXL
+};
+
+class CLIPTextModel : public GGMLBlock {
+protected:
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        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);
+        }
+    }
+
+public:
+    CLIPVersion version = OPENAI_CLIP_VIT_L_14;
+    // network hparams
+    int32_t vocab_size        = 49408;
+    int32_t n_token           = 77;  // max_position_embeddings
+    int32_t hidden_size       = 768;
+    int32_t intermediate_size = 3072;
+    int32_t n_head            = 12;
+    int32_t n_layer           = 12;    // num_hidden_layers
+    int32_t projection_dim    = 1280;  // only for OPEN_CLIP_VIT_BIGG_14
+    bool with_final_ln        = true;
+
+    CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
+                  bool with_final_ln  = true,
+                  bool force_clip_f32 = false)
+        : version(version), with_final_ln(with_final_ln) {
+        if (version == OPEN_CLIP_VIT_H_14) {
+            hidden_size       = 1024;
+            intermediate_size = 4096;
+            n_head            = 16;
+            n_layer           = 24;
+        } else if (version == OPEN_CLIP_VIT_BIGG_14) {  // CLIPTextModelWithProjection
+            hidden_size       = 1280;
+            intermediate_size = 5120;
+            n_head            = 20;
+            n_layer           = 32;
+        }
+
+        blocks["embeddings"]       = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token, force_clip_f32));
+        blocks["encoder"]          = std::shared_ptr<GGMLBlock>(new CLIPEncoder(n_layer, hidden_size, n_head, intermediate_size));
+        blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
+    }
+
+    struct 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(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* input_ids,
+                                struct ggml_tensor* tkn_embeddings,
+                                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, backend, x, return_pooled ? -1 : clip_skip, true);
+        if (return_pooled || with_final_ln) {
+            x = final_layer_norm->forward(ctx, x);
+        }
+
+        if (return_pooled) {
+            auto text_projection = params["text_projection"];
+            ggml_tensor* pooled  = ggml_view_1d(ctx, x, hidden_size, x->nb[1] * max_token_idx);
+            if (text_projection != NULL) {
+                pooled = ggml_nn_linear(ctx, pooled, text_projection, NULL);
+            } else {
+                LOG_DEBUG("identity projection");
+            }
+            return pooled;  // [hidden_size, 1, 1]
+        }
+
+        return x;  // [N, n_token, hidden_size]
+    }
+};
+
+class CLIPVisionModel : public GGMLBlock {
+public:
+    // network hparams
+    int32_t num_channels      = 3;
+    int32_t patch_size        = 14;
+    int32_t image_size        = 224;
+    int32_t num_positions     = 257;  // (image_size / patch_size)^2 + 1
+    int32_t hidden_size       = 1024;
+    int32_t intermediate_size = 4096;
+    int32_t n_head            = 16;
+    int32_t n_layer           = 24;
+
+public:
+    CLIPVisionModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14) {
+        if (version == OPEN_CLIP_VIT_H_14) {
+            hidden_size       = 1280;
+            intermediate_size = 5120;
+            n_head            = 16;
+            n_layer           = 32;
+        } else if (version == OPEN_CLIP_VIT_BIGG_14) {
+            hidden_size       = 1664;
+            intermediate_size = 8192;
+            n_head            = 16;
+            n_layer           = 48;
+        }
+
+        blocks["embeddings"]     = std::shared_ptr<GGMLBlock>(new CLIPVisionEmbeddings(hidden_size, num_channels, patch_size, image_size));
+        blocks["pre_layernorm"]  = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
+        blocks["encoder"]        = std::shared_ptr<GGMLBlock>(new CLIPEncoder(n_layer, hidden_size, n_head, intermediate_size));
+        blocks["post_layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct 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"]);
+        auto encoder        = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
+        auto post_layernorm = std::dynamic_pointer_cast<LayerNorm>(blocks["post_layernorm"]);
+
+        auto x = embeddings->forward(ctx, pixel_values);  // [N, num_positions, embed_dim]
+        x      = pre_layernorm->forward(ctx, x);
+        x      = encoder->forward(ctx, backend, x, clip_skip, false);
+        // print_ggml_tensor(x, true, "ClipVisionModel x: ");
+        auto last_hidden_state = x;
+        x                      = post_layernorm->forward(ctx, x);  // [N, n_token, hidden_size]
+
+        GGML_ASSERT(x->ne[3] == 1);
+        if (return_pooled) {
+            ggml_tensor* pooled = ggml_cont(ctx, ggml_view_2d(ctx, x, x->ne[0], x->ne[2], x->nb[2], 0));
+            return pooled;  // [N, hidden_size]
+        } else {
+            // return x;  // [N, n_token, hidden_size]
+            return last_hidden_state;  // [N, n_token, hidden_size]
+        }
+    }
+};
+
+class CLIPProjection : public UnaryBlock {
+protected:
+    int64_t in_features;
+    int64_t out_features;
+    bool transpose_weight;
+
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
+        if (transpose_weight) {
+            params["weight"] = ggml_new_tensor_2d(ctx, wtype, out_features, in_features);
+        } else {
+            params["weight"] = ggml_new_tensor_2d(ctx, wtype, in_features, out_features);
+        }
+    }
+
+public:
+    CLIPProjection(int64_t in_features,
+                   int64_t out_features,
+                   bool transpose_weight = false)
+        : in_features(in_features),
+          out_features(out_features),
+          transpose_weight(transpose_weight) {}
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        struct ggml_tensor* w = params["weight"];
+        if (transpose_weight) {
+            w = ggml_cont(ctx, ggml_transpose(ctx, w));
+        }
+        return ggml_nn_linear(ctx, x, w, NULL);
+    }
+};
+
+class CLIPVisionModelProjection : public GGMLBlock {
+public:
+    int32_t hidden_size    = 1024;
+    int32_t projection_dim = 768;
+    int32_t image_size     = 224;
+
+public:
+    CLIPVisionModelProjection(CLIPVersion version   = OPENAI_CLIP_VIT_L_14,
+                              bool transpose_proj_w = false) {
+        if (version == OPEN_CLIP_VIT_H_14) {
+            hidden_size    = 1280;
+            projection_dim = 1024;
+        } else if (version == OPEN_CLIP_VIT_BIGG_14) {
+            hidden_size = 1664;
+        }
+
+        blocks["vision_model"]      = std::shared_ptr<GGMLBlock>(new CLIPVisionModel(version));
+        blocks["visual_projection"] = std::shared_ptr<GGMLBlock>(new CLIPProjection(hidden_size, projection_dim, transpose_proj_w));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct 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"]);
+        auto visual_projection = std::dynamic_pointer_cast<CLIPProjection>(blocks["visual_projection"]);
+
+        auto x = vision_model->forward(ctx, backend, pixel_values, return_pooled, clip_skip);  // [N, hidden_size] or [N, n_token, hidden_size]
+
+        if (return_pooled) {
+            x = visual_projection->forward(ctx, x);  // [N, projection_dim]
+        }
+
+        return x;
+    }
+};
+
+struct CLIPTextModelRunner : public GGMLRunner {
+    CLIPTextModel model;
+
+    CLIPTextModelRunner(ggml_backend_t backend,
+                        bool offload_params_to_cpu,
+                        const String2GGMLType& tensor_types,
+                        const std::string prefix,
+                        CLIPVersion version = OPENAI_CLIP_VIT_L_14,
+                        bool with_final_ln  = true,
+                        bool force_clip_f32 = false)
+        : GGMLRunner(backend, offload_params_to_cpu), model(version, with_final_ln, force_clip_f32) {
+        model.init(params_ctx, tensor_types, prefix);
+    }
+
+    std::string get_desc() {
+        return "clip";
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        model.get_param_tensors(tensors, prefix);
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* input_ids,
+                                struct ggml_tensor* embeddings,
+                                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) {
+            GGML_ASSERT(input_ids->ne[0] % model.n_token == 0);
+            input_ids = ggml_reshape_2d(ctx, input_ids, model.n_token, input_ids->ne[0] / model.n_token);
+        }
+
+        return model.forward(ctx, backend, input_ids, embeddings, 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 = NULL,
+                                    size_t max_token_idx         = 0,
+                                    bool return_pooled           = false,
+                                    int clip_skip                = -1) {
+        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+
+        input_ids = to_backend(input_ids);
+
+        struct ggml_tensor* embeddings = NULL;
+
+        if (num_custom_embeddings > 0 && custom_embeddings_data != NULL) {
+            auto token_embed_weight = model.get_token_embed_weight();
+            auto custom_embeddings  = ggml_new_tensor_2d(compute_ctx,
+                                                         token_embed_weight->type,
+                                                         model.hidden_size,
+                                                         num_custom_embeddings);
+            set_backend_tensor_data(custom_embeddings, custom_embeddings_data);
+
+            // concatenate custom embeddings
+            embeddings = ggml_concat(compute_ctx, token_embed_weight, custom_embeddings, 1);
+        }
+
+        struct ggml_tensor* hidden_states = forward(compute_ctx, runtime_backend, input_ids, embeddings, max_token_idx, return_pooled, clip_skip);
+
+        ggml_build_forward_expand(gf, hidden_states);
+
+        return gf;
+    }
+
+    void 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 = NULL) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled, clip_skip);
+        };
+        GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
+    }
+};
+
+#endif  // __CLIP_HPP__

common.hpp

@@ -0,0 +1,566 @@
+#ifndef __COMMON_HPP__
+#define __COMMON_HPP__
+
+#include "ggml_extend.hpp"
+
+class DownSampleBlock : public GGMLBlock {
+protected:
+    int channels;
+    int out_channels;
+    bool vae_downsample;
+
+public:
+    DownSampleBlock(int channels,
+                    int out_channels,
+                    bool vae_downsample = false)
+        : channels(channels),
+          out_channels(out_channels),
+          vae_downsample(vae_downsample) {
+        if (vae_downsample) {
+            blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {2, 2}, {0, 0}));
+        } else {
+            blocks["op"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {2, 2}, {1, 1}));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, channels, h, w]
+        if (vae_downsample) {
+            auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
+
+            x = ggml_pad(ctx, x, 1, 1, 0, 0);
+            x = conv->forward(ctx, x);
+        } else {
+            auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["op"]);
+
+            x = conv->forward(ctx, x);
+        }
+        return x;  // [N, out_channels, h/2, w/2]
+    }
+};
+
+class UpSampleBlock : public GGMLBlock {
+protected:
+    int channels;
+    int out_channels;
+
+public:
+    UpSampleBlock(int channels,
+                  int out_channels)
+        : channels(channels),
+          out_channels(out_channels) {
+        blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, channels, h, w]
+        auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
+
+        x = ggml_upscale(ctx, x, 2, GGML_SCALE_MODE_NEAREST);  // [N, channels, h*2, w*2]
+        x = conv->forward(ctx, x);                             // [N, out_channels, h*2, w*2]
+        return x;
+    }
+};
+
+class ResBlock : public GGMLBlock {
+protected:
+    // network hparams
+    int64_t channels;      // model_channels * (1, 1, 1, 2, 2, 4, 4, 4)
+    int64_t emb_channels;  // time_embed_dim
+    int64_t out_channels;  // mult * model_channels
+    std::pair<int, int> kernel_size;
+    int dims;
+    bool skip_t_emb;
+    bool exchange_temb_dims;
+
+    std::shared_ptr<GGMLBlock> conv_nd(int dims,
+                                       int64_t in_channels,
+                                       int64_t out_channels,
+                                       std::pair<int, int> kernel_size,
+                                       std::pair<int, int> padding) {
+        GGML_ASSERT(dims == 2 || dims == 3);
+        if (dims == 3) {
+            return std::shared_ptr<GGMLBlock>(new Conv3dnx1x1(in_channels, out_channels, kernel_size.first, 1, padding.first));
+        } else {
+            return std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, kernel_size, {1, 1}, padding));
+        }
+    }
+
+public:
+    ResBlock(int64_t channels,
+             int64_t emb_channels,
+             int64_t out_channels,
+             std::pair<int, int> kernel_size = {3, 3},
+             int dims                        = 2,
+             bool exchange_temb_dims         = false,
+             bool skip_t_emb                 = false)
+        : channels(channels),
+          emb_channels(emb_channels),
+          out_channels(out_channels),
+          kernel_size(kernel_size),
+          dims(dims),
+          skip_t_emb(skip_t_emb),
+          exchange_temb_dims(exchange_temb_dims) {
+        std::pair<int, int> padding = {kernel_size.first / 2, kernel_size.second / 2};
+        blocks["in_layers.0"]       = std::shared_ptr<GGMLBlock>(new GroupNorm32(channels));
+        // in_layer_1 is nn.SILU()
+        blocks["in_layers.2"] = conv_nd(dims, channels, out_channels, kernel_size, padding);
+
+        if (!skip_t_emb) {
+            // emb_layer_0 is nn.SILU()
+            blocks["emb_layers.1"] = std::shared_ptr<GGMLBlock>(new Linear(emb_channels, out_channels));
+        }
+
+        blocks["out_layers.0"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(out_channels));
+        // out_layer_1 is nn.SILU()
+        // out_layer_2 is nn.Dropout(), skip for inference
+        blocks["out_layers.3"] = conv_nd(dims, out_channels, out_channels, kernel_size, padding);
+
+        if (out_channels != channels) {
+            blocks["skip_connection"] = conv_nd(dims, channels, out_channels, {1, 1}, {0, 0});
+        }
+    }
+
+    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* emb = NULL) {
+        // 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]
+        // emb: [N, emb_channels] if dims == 2 else [N, t, emb_channels]
+        auto in_layers_0  = std::dynamic_pointer_cast<GroupNorm32>(blocks["in_layers.0"]);
+        auto in_layers_2  = std::dynamic_pointer_cast<UnaryBlock>(blocks["in_layers.2"]);
+        auto out_layers_0 = std::dynamic_pointer_cast<GroupNorm32>(blocks["out_layers.0"]);
+        auto out_layers_3 = std::dynamic_pointer_cast<UnaryBlock>(blocks["out_layers.3"]);
+
+        if (emb == NULL) {
+            GGML_ASSERT(skip_t_emb);
+        }
+
+        // in_layers
+        auto h = in_layers_0->forward(ctx, x);
+        h      = ggml_silu_inplace(ctx, h);
+        h      = in_layers_2->forward(ctx, h);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
+
+        // emb_layers
+        if (!skip_t_emb) {
+            auto emb_layer_1 = std::dynamic_pointer_cast<Linear>(blocks["emb_layers.1"]);
+
+            auto emb_out = ggml_silu(ctx, emb);
+            emb_out      = emb_layer_1->forward(ctx, emb_out);  // [N, out_channels] if dims == 2 else [N, t, out_channels]
+
+            if (dims == 2) {
+                emb_out = ggml_reshape_4d(ctx, emb_out, 1, 1, emb_out->ne[0], emb_out->ne[1]);  // [N, out_channels, 1, 1]
+            } else {
+                emb_out = ggml_reshape_4d(ctx, emb_out, 1, emb_out->ne[0], emb_out->ne[1], emb_out->ne[2]);  // [N, t, out_channels, 1]
+                if (exchange_temb_dims) {
+                    // emb_out = rearrange(emb_out, "b t c ... -> b c t ...")
+                    emb_out = ggml_cont(ctx, ggml_permute(ctx, emb_out, 0, 2, 1, 3));  // [N, out_channels, t, 1]
+                }
+            }
+
+            h = ggml_add(ctx, h, emb_out);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
+        }
+
+        // out_layers
+        h = out_layers_0->forward(ctx, h);
+        h = ggml_silu_inplace(ctx, h);
+        // dropout, skip for inference
+        h = out_layers_3->forward(ctx, h);
+
+        // skip connection
+        if (out_channels != channels) {
+            auto skip_connection = std::dynamic_pointer_cast<UnaryBlock>(blocks["skip_connection"]);
+            x                    = skip_connection->forward(ctx, x);  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
+        }
+
+        h = ggml_add(ctx, h, x);
+        return h;  // [N, out_channels, h, w] if dims == 2 else [N, out_channels, t, h, w]
+    }
+};
+
+class GEGLU : public UnaryBlock {
+protected:
+    int64_t dim_in;
+    int64_t dim_out;
+
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
+        enum ggml_type wtype      = get_type(prefix + "proj.weight", tensor_types, GGML_TYPE_F32);
+        enum ggml_type bias_wtype = GGML_TYPE_F32;
+        params["proj.weight"]     = ggml_new_tensor_2d(ctx, wtype, dim_in, dim_out * 2);
+        params["proj.bias"]       = ggml_new_tensor_1d(ctx, bias_wtype, dim_out * 2);
+    }
+
+public:
+    GEGLU(int64_t dim_in, int64_t dim_out)
+        : dim_in(dim_in), dim_out(dim_out) {}
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [ne3, ne2, ne1, dim_in]
+        // return: [ne3, ne2, ne1, dim_out]
+        struct ggml_tensor* w = params["proj.weight"];
+        struct ggml_tensor* b = params["proj.bias"];
+
+        auto x_w    = ggml_view_2d(ctx, w, w->ne[0], w->ne[1] / 2, w->nb[1], 0);                        // [dim_out, dim_in]
+        auto x_b    = ggml_view_1d(ctx, b, b->ne[0] / 2, 0);                                            // [dim_out, dim_in]
+        auto gate_w = ggml_view_2d(ctx, w, w->ne[0], w->ne[1] / 2, w->nb[1], w->nb[1] * w->ne[1] / 2);  // [dim_out, ]
+        auto gate_b = ggml_view_1d(ctx, b, b->ne[0] / 2, b->nb[0] * b->ne[0] / 2);                      // [dim_out, ]
+
+        auto x_in = x;
+        x         = ggml_nn_linear(ctx, x_in, x_w, x_b);        // [ne3, ne2, ne1, dim_out]
+        auto gate = ggml_nn_linear(ctx, x_in, gate_w, gate_b);  // [ne3, ne2, ne1, dim_out]
+
+        gate = ggml_gelu_inplace(ctx, gate);
+
+        x = ggml_mul(ctx, x, gate);  // [ne3, ne2, ne1, dim_out]
+
+        return x;
+    }
+};
+
+class GELU : public UnaryBlock {
+public:
+    GELU(int64_t dim_in, int64_t dim_out, bool bias = true) {
+        blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim_in, dim_out, bias));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // 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, x);
+        return x;
+    }
+};
+
+class FeedForward : public GGMLBlock {
+public:
+    enum class Activation {
+        GEGLU,
+        GELU
+    };
+    FeedForward(int64_t dim,
+                int64_t dim_out,
+                int64_t mult          = 4,
+                Activation activation = Activation::GEGLU,
+                bool precision_fix    = false) {
+        int64_t inner_dim = dim * mult;
+        if (activation == Activation::GELU) {
+            blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GELU(dim, inner_dim));
+        } else {
+            blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GEGLU(dim, inner_dim));
+        }
+
+        // net_1 is nn.Dropout(), skip for inference
+        float scale = 1.f;
+        if (precision_fix) {
+            scale = 1.f / 128.f;
+        }
+        // The purpose of the scale here is to prevent NaN issues in certain situations.
+        // For example, when using Vulkan without enabling force_prec_f32,
+        // or when using CUDA but the weights are k-quants.
+        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out, true, false, false, scale));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [ne3, ne2, ne1, dim]
+        // return: [ne3, ne2, ne1, dim_out]
+
+        auto net_0 = std::dynamic_pointer_cast<UnaryBlock>(blocks["net.0"]);
+        auto net_2 = std::dynamic_pointer_cast<Linear>(blocks["net.2"]);
+
+        x = net_0->forward(ctx, x);  // [ne3, ne2, ne1, inner_dim]
+        x = net_2->forward(ctx, x);  // [ne3, ne2, ne1, dim_out]
+        return x;
+    }
+};
+
+class CrossAttention : public GGMLBlock {
+protected:
+    int64_t query_dim;
+    int64_t context_dim;
+    int64_t n_head;
+    int64_t d_head;
+    bool flash_attn;
+
+public:
+    CrossAttention(int64_t query_dim,
+                   int64_t context_dim,
+                   int64_t n_head,
+                   int64_t d_head,
+                   bool flash_attn = false)
+        : n_head(n_head),
+          d_head(d_head),
+          query_dim(query_dim),
+          context_dim(context_dim),
+          flash_attn(flash_attn) {
+        int64_t inner_dim = d_head * n_head;
+
+        blocks["to_q"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, false));
+        blocks["to_k"] = std::shared_ptr<GGMLBlock>(new Linear(context_dim, inner_dim, false));
+        blocks["to_v"] = std::shared_ptr<GGMLBlock>(new Linear(context_dim, inner_dim, false));
+
+        blocks["to_out.0"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, query_dim));
+        // to_out_1 is nn.Dropout(), skip for inference
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* context) {
+        // x: [N, n_token, query_dim]
+        // context: [N, n_context, context_dim]
+        // return: [N, n_token, query_dim]
+        auto to_q     = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
+        auto to_k     = std::dynamic_pointer_cast<Linear>(blocks["to_k"]);
+        auto to_v     = std::dynamic_pointer_cast<Linear>(blocks["to_v"]);
+        auto to_out_0 = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+
+        int64_t n         = x->ne[2];
+        int64_t n_token   = x->ne[1];
+        int64_t n_context = context->ne[1];
+        int64_t inner_dim = d_head * n_head;
+
+        auto q = to_q->forward(ctx, x);        // [N, n_token, inner_dim]
+        auto 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_nn_attention_ext(ctx, backend, q, k, v, n_head, NULL, false, false, flash_attn);  // [N, n_token, inner_dim]
+
+        x = to_out_0->forward(ctx, x);  // [N, n_token, query_dim]
+        return x;
+    }
+};
+
+class BasicTransformerBlock : public GGMLBlock {
+protected:
+    int64_t n_head;
+    int64_t d_head;
+    bool ff_in;
+
+public:
+    BasicTransformerBlock(int64_t dim,
+                          int64_t n_head,
+                          int64_t d_head,
+                          int64_t context_dim,
+                          bool ff_in      = false,
+                          bool flash_attn = false)
+        : n_head(n_head), d_head(d_head), ff_in(ff_in) {
+        // disable_self_attn is always False
+        // disable_temporal_crossattention is always False
+        // switch_temporal_ca_to_sa is always False
+        // inner_dim is always None or equal to dim
+        // gated_ff is always True
+        blocks["attn1"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, dim, n_head, d_head, flash_attn));
+        blocks["attn2"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, context_dim, n_head, d_head, flash_attn));
+        blocks["ff"]    = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim));
+        blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["norm3"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+
+        if (ff_in) {
+            blocks["norm_in"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+            blocks["ff_in"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* context) {
+        // x: [N, n_token, query_dim]
+        // context: [N, n_context, context_dim]
+        // return: [N, n_token, query_dim]
+
+        auto attn1 = std::dynamic_pointer_cast<CrossAttention>(blocks["attn1"]);
+        auto attn2 = std::dynamic_pointer_cast<CrossAttention>(blocks["attn2"]);
+        auto ff    = std::dynamic_pointer_cast<FeedForward>(blocks["ff"]);
+        auto norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
+        auto norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm2"]);
+        auto norm3 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm3"]);
+
+        if (ff_in) {
+            auto norm_in = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_in"]);
+            auto ff_in   = std::dynamic_pointer_cast<FeedForward>(blocks["ff_in"]);
+
+            auto x_skip = x;
+            x           = norm_in->forward(ctx, x);
+            x           = ff_in->forward(ctx, x);
+            // self.is_res is always True
+            x = ggml_add(ctx, x, x_skip);
+        }
+
+        auto r = x;
+        x      = norm1->forward(ctx, x);
+        x      = attn1->forward(ctx, backend, x, x);  // self-attention
+        x      = ggml_add(ctx, x, r);
+        r      = x;
+        x      = norm2->forward(ctx, x);
+        x      = attn2->forward(ctx, backend, x, context);  // cross-attention
+        x      = ggml_add(ctx, x, r);
+        r      = x;
+        x      = norm3->forward(ctx, x);
+        x      = ff->forward(ctx, x);
+        x      = ggml_add(ctx, x, r);
+
+        return x;
+    }
+};
+
+class SpatialTransformer : public GGMLBlock {
+protected:
+    int64_t in_channels;  // mult * model_channels
+    int64_t n_head;
+    int64_t d_head;
+    int64_t depth       = 1;    // 1
+    int64_t context_dim = 768;  // hidden_size, 1024 for VERSION_SD2
+
+public:
+    SpatialTransformer(int64_t in_channels,
+                       int64_t n_head,
+                       int64_t d_head,
+                       int64_t depth,
+                       int64_t context_dim,
+                       bool flash_attn = false)
+        : in_channels(in_channels),
+          n_head(n_head),
+          d_head(d_head),
+          depth(depth),
+          context_dim(context_dim) {
+        // We will convert unet transformer linear to conv2d 1x1 when loading the weights, so use_linear is always False
+        // disable_self_attn is always False
+        int64_t inner_dim = n_head * d_head;  // in_channels
+        blocks["norm"]    = std::shared_ptr<GGMLBlock>(new GroupNorm32(in_channels));
+        blocks["proj_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, inner_dim, {1, 1}));
+
+        for (int i = 0; i < depth; i++) {
+            std::string name = "transformer_blocks." + std::to_string(i);
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new BasicTransformerBlock(inner_dim, n_head, d_head, context_dim, false, flash_attn));
+        }
+
+        blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(inner_dim, in_channels, {1, 1}));
+    }
+
+    virtual struct ggml_tensor* forward(struct ggml_context* ctx,
+                                        ggml_backend_t backend,
+                                        struct ggml_tensor* x,
+                                        struct 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"]);
+        auto proj_in  = std::dynamic_pointer_cast<Conv2d>(blocks["proj_in"]);
+        auto proj_out = std::dynamic_pointer_cast<Conv2d>(blocks["proj_out"]);
+
+        auto x_in         = x;
+        int64_t n         = x->ne[3];
+        int64_t h         = x->ne[1];
+        int64_t w         = x->ne[0];
+        int64_t inner_dim = n_head * d_head;
+
+        x = norm->forward(ctx, x);
+        x = proj_in->forward(ctx, x);  // [N, inner_dim, h, w]
+
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 1, 2, 0, 3));  // [N, h, w, inner_dim]
+        x = ggml_reshape_3d(ctx, x, inner_dim, w * h, n);      // [N, h * w, inner_dim]
+
+        for (int i = 0; i < depth; i++) {
+            std::string name       = "transformer_blocks." + std::to_string(i);
+            auto transformer_block = std::dynamic_pointer_cast<BasicTransformerBlock>(blocks[name]);
+
+            x = transformer_block->forward(ctx, backend, x, context);
+        }
+
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 1, 0, 2, 3));  // [N, inner_dim, h * w]
+        x = ggml_reshape_4d(ctx, x, w, h, inner_dim, n);       // [N, inner_dim, h, w]
+
+        // proj_out
+        x = proj_out->forward(ctx, x);  // [N, in_channels, h, w]
+
+        x = ggml_add(ctx, x, x_in);
+        return x;
+    }
+};
+
+class AlphaBlender : public GGMLBlock {
+protected:
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
+        // Get the type of the "mix_factor" tensor from the input tensors map with the specified prefix
+        enum ggml_type wtype = GGML_TYPE_F32;
+        params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
+    }
+
+    float get_alpha() {
+        // image_only_indicator is always tensor([0.]) and since mix_factor.shape is [1,]
+        // so learned_with_images is same as learned
+        float alpha = ggml_backend_tensor_get_f32(params["mix_factor"]);
+        return sigmoid(alpha);
+    }
+
+public:
+    AlphaBlender() {
+        // merge_strategy is always learned_with_images
+        // for inference, we don't need to set alpha
+        // since mix_factor.shape is [1,], we don't need rearrange using  rearrange_pattern
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x_spatial,
+                                struct ggml_tensor* x_temporal) {
+        // image_only_indicator is always tensor([0.])
+        float alpha = get_alpha();
+        auto x      = ggml_add(ctx,
+                               ggml_scale(ctx, x_spatial, alpha),
+                               ggml_scale(ctx, x_temporal, 1.0f - alpha));
+        return x;
+    }
+};
+
+class VideoResBlock : public ResBlock {
+public:
+    VideoResBlock(int channels,
+                  int emb_channels,
+                  int out_channels,
+                  std::pair<int, int> kernel_size = {3, 3},
+                  int64_t video_kernel_size       = 3,
+                  int dims                        = 2)  // always 2
+        : ResBlock(channels, emb_channels, out_channels, kernel_size, dims) {
+        blocks["time_stack"] = std::shared_ptr<GGMLBlock>(new ResBlock(out_channels, emb_channels, out_channels, kernel_size, 3, true));
+        blocks["time_mixer"] = std::shared_ptr<GGMLBlock>(new AlphaBlender());
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct 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.])
+        auto time_stack = std::dynamic_pointer_cast<ResBlock>(blocks["time_stack"]);
+        auto time_mixer = std::dynamic_pointer_cast<AlphaBlender>(blocks["time_mixer"]);
+
+        x = ResBlock::forward(ctx, x, emb);
+
+        int64_t T = num_video_frames;
+        int64_t B = x->ne[3] / T;
+        int64_t C = x->ne[2];
+        int64_t H = x->ne[1];
+        int64_t W = x->ne[0];
+
+        x          = ggml_reshape_4d(ctx, x, W * H, C, T, B);           // (b t) c h w -> b t c (h w)
+        x          = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // b t c (h w) -> b c t (h w)
+        auto x_mix = x;
+
+        emb = ggml_reshape_4d(ctx, emb, emb->ne[0], T, B, emb->ne[3]);  // (b t) ... -> b t ...
+
+        x = time_stack->forward(ctx, x, emb);  // b t c (h w)
+
+        x = time_mixer->forward(ctx, x_mix, x);  // b t c (h w)
+
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
+        x = ggml_reshape_4d(ctx, x, W, H, C, T * B);           // b t c (h w) -> (b t) c h w
+
+        return x;
+    }
+};
+
+#endif  // __COMMON_HPP__

control.hpp

@@ -0,0 +1,473 @@
+#ifndef __CONTROL_HPP__
+#define __CONTROL_HPP__
+
+#include "common.hpp"
+#include "ggml_extend.hpp"
+#include "model.h"
+
+#define CONTROL_NET_GRAPH_SIZE 1536
+
+/*
+    =================================== ControlNet ===================================
+    Reference: https://github.com/comfyanonymous/ComfyUI/blob/master/comfy/cldm/cldm.py
+
+*/
+class ControlNetBlock : public GGMLBlock {
+protected:
+    SDVersion version = VERSION_SD1;
+    // network hparams
+    int in_channels                        = 4;
+    int out_channels                       = 4;
+    int hint_channels                      = 3;
+    int num_res_blocks                     = 2;
+    std::vector<int> attention_resolutions = {4, 2, 1};
+    std::vector<int> channel_mult          = {1, 2, 4, 4};
+    std::vector<int> transformer_depth     = {1, 1, 1, 1};
+    int time_embed_dim                     = 1280;  // model_channels*4
+    int num_heads                          = 8;
+    int num_head_channels                  = -1;   // channels // num_heads
+    int context_dim                        = 768;  // 1024 for VERSION_SD2, 2048 for VERSION_SDXL
+
+public:
+    int model_channels  = 320;
+    int adm_in_channels = 2816;  // only for VERSION_SDXL
+
+    ControlNetBlock(SDVersion version = VERSION_SD1)
+        : version(version) {
+        if (sd_version_is_sd2(version)) {
+            context_dim       = 1024;
+            num_head_channels = 64;
+            num_heads         = -1;
+        } else if (sd_version_is_sdxl(version)) {
+            context_dim           = 2048;
+            attention_resolutions = {4, 2};
+            channel_mult          = {1, 2, 4};
+            transformer_depth     = {1, 2, 10};
+            num_head_channels     = 64;
+            num_heads             = -1;
+        } else if (version == VERSION_SVD) {
+            in_channels       = 8;
+            out_channels      = 4;
+            context_dim       = 1024;
+            adm_in_channels   = 768;
+            num_head_channels = 64;
+            num_heads         = -1;
+        }
+
+        blocks["time_embed.0"] = std::shared_ptr<GGMLBlock>(new Linear(model_channels, time_embed_dim));
+        // time_embed_1 is nn.SiLU()
+        blocks["time_embed.2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, time_embed_dim));
+
+        if (sd_version_is_sdxl(version) || version == VERSION_SVD) {
+            blocks["label_emb.0.0"] = std::shared_ptr<GGMLBlock>(new Linear(adm_in_channels, time_embed_dim));
+            // label_emb_1 is nn.SiLU()
+            blocks["label_emb.0.2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, time_embed_dim));
+        }
+
+        // input_blocks
+        blocks["input_blocks.0.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, model_channels, {3, 3}, {1, 1}, {1, 1}));
+
+        std::vector<int> input_block_chans;
+        input_block_chans.push_back(model_channels);
+        int ch              = model_channels;
+        int input_block_idx = 0;
+        int ds              = 1;
+
+        auto get_resblock = [&](int64_t channels, int64_t emb_channels, int64_t out_channels) -> ResBlock* {
+            return new ResBlock(channels, emb_channels, out_channels);
+        };
+
+        auto get_attention_layer = [&](int64_t in_channels,
+                                       int64_t n_head,
+                                       int64_t d_head,
+                                       int64_t depth,
+                                       int64_t context_dim) -> SpatialTransformer* {
+            return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim);
+        };
+
+        auto make_zero_conv = [&](int64_t channels) {
+            return new Conv2d(channels, channels, {1, 1});
+        };
+
+        blocks["zero_convs.0.0"] = std::shared_ptr<GGMLBlock>(make_zero_conv(model_channels));
+
+        blocks["input_hint_block.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(hint_channels, 16, {3, 3}, {1, 1}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(16, 16, {3, 3}, {1, 1}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.4"] = std::shared_ptr<GGMLBlock>(new Conv2d(16, 32, {3, 3}, {2, 2}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.6"] = std::shared_ptr<GGMLBlock>(new Conv2d(32, 32, {3, 3}, {1, 1}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.8"] = std::shared_ptr<GGMLBlock>(new Conv2d(32, 96, {3, 3}, {2, 2}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.10"] = std::shared_ptr<GGMLBlock>(new Conv2d(96, 96, {3, 3}, {1, 1}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.12"] = std::shared_ptr<GGMLBlock>(new Conv2d(96, 256, {3, 3}, {2, 2}, {1, 1}));
+        // nn.SiLU()
+        blocks["input_hint_block.14"] = std::shared_ptr<GGMLBlock>(new Conv2d(256, model_channels, {3, 3}, {1, 1}, {1, 1}));
+
+        size_t len_mults = channel_mult.size();
+        for (int i = 0; i < len_mults; i++) {
+            int mult = channel_mult[i];
+            for (int j = 0; j < num_res_blocks; j++) {
+                input_block_idx += 1;
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, mult * model_channels));
+
+                ch = mult * model_channels;
+                if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
+                    int n_head = num_heads;
+                    int d_head = ch / num_heads;
+                    if (num_head_channels != -1) {
+                        d_head = num_head_channels;
+                        n_head = ch / d_head;
+                    }
+                    std::string name = "input_blocks." + std::to_string(input_block_idx) + ".1";
+                    blocks[name]     = std::shared_ptr<GGMLBlock>(get_attention_layer(ch,
+                                                                                      n_head,
+                                                                                      d_head,
+                                                                                      transformer_depth[i],
+                                                                                      context_dim));
+                }
+                blocks["zero_convs." + std::to_string(input_block_idx) + ".0"] = std::shared_ptr<GGMLBlock>(make_zero_conv(ch));
+                input_block_chans.push_back(ch);
+            }
+            if (i != len_mults - 1) {
+                input_block_idx += 1;
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new DownSampleBlock(ch, ch));
+
+                blocks["zero_convs." + std::to_string(input_block_idx) + ".0"] = std::shared_ptr<GGMLBlock>(make_zero_conv(ch));
+
+                input_block_chans.push_back(ch);
+                ds *= 2;
+            }
+        }
+
+        // middle blocks
+        int n_head = num_heads;
+        int d_head = ch / num_heads;
+        if (num_head_channels != -1) {
+            d_head = num_head_channels;
+            n_head = ch / d_head;
+        }
+        blocks["middle_block.0"] = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, ch));
+        blocks["middle_block.1"] = std::shared_ptr<GGMLBlock>(get_attention_layer(ch,
+                                                                                  n_head,
+                                                                                  d_head,
+                                                                                  transformer_depth[transformer_depth.size() - 1],
+                                                                                  context_dim));
+        blocks["middle_block.2"] = std::shared_ptr<GGMLBlock>(get_resblock(ch, time_embed_dim, ch));
+
+        // middle_block_out
+        blocks["middle_block_out.0"] = std::shared_ptr<GGMLBlock>(make_zero_conv(ch));
+    }
+
+    struct ggml_tensor* resblock_forward(std::string name,
+                                         struct ggml_context* ctx,
+                                         struct ggml_tensor* x,
+                                         struct 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,
+                                                struct ggml_context* ctx,
+                                                ggml_backend_t backend,
+                                                struct ggml_tensor* x,
+                                                struct ggml_tensor* context) {
+        auto block = std::dynamic_pointer_cast<SpatialTransformer>(blocks[name]);
+        return block->forward(ctx, backend, x, context);
+    }
+
+    struct ggml_tensor* input_hint_block_forward(struct ggml_context* ctx,
+                                                 struct ggml_tensor* hint,
+                                                 struct ggml_tensor* emb,
+                                                 struct ggml_tensor* context) {
+        int num_input_blocks = 15;
+        auto h               = hint;
+        for (int i = 0; i < num_input_blocks; i++) {
+            if (i % 2 == 0) {
+                auto block = std::dynamic_pointer_cast<Conv2d>(blocks["input_hint_block." + std::to_string(i)]);
+
+                h = block->forward(ctx, h);
+            } else {
+                h = ggml_silu_inplace(ctx, h);
+            }
+        }
+        return h;
+    }
+
+    std::vector<struct ggml_tensor*> forward(struct ggml_context* ctx,
+                                             ggml_backend_t backend,
+                                             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 = NULL) {
+        // 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]
+        // y: [N, adm_in_channels] or [1, adm_in_channels]
+        if (context != NULL) {
+            if (context->ne[2] != x->ne[3]) {
+                context = ggml_repeat(ctx, context, ggml_new_tensor_3d(ctx, GGML_TYPE_F32, context->ne[0], context->ne[1], x->ne[3]));
+            }
+        }
+
+        if (y != NULL) {
+            if (y->ne[1] != x->ne[3]) {
+                y = ggml_repeat(ctx, y, ggml_new_tensor_2d(ctx, GGML_TYPE_F32, y->ne[0], x->ne[3]));
+            }
+        }
+
+        auto time_embed_0     = std::dynamic_pointer_cast<Linear>(blocks["time_embed.0"]);
+        auto time_embed_2     = std::dynamic_pointer_cast<Linear>(blocks["time_embed.2"]);
+        auto input_blocks_0_0 = std::dynamic_pointer_cast<Conv2d>(blocks["input_blocks.0.0"]);
+        auto zero_convs_0     = std::dynamic_pointer_cast<Conv2d>(blocks["zero_convs.0.0"]);
+
+        auto middle_block_out = std::dynamic_pointer_cast<Conv2d>(blocks["middle_block_out.0"]);
+
+        auto t_emb = ggml_nn_timestep_embedding(ctx, timesteps, model_channels);  // [N, model_channels]
+
+        auto emb = time_embed_0->forward(ctx, t_emb);
+        emb      = ggml_silu_inplace(ctx, emb);
+        emb      = time_embed_2->forward(ctx, emb);  // [N, time_embed_dim]
+
+        // SDXL/SVD
+        if (y != NULL) {
+            auto label_embed_0 = std::dynamic_pointer_cast<Linear>(blocks["label_emb.0.0"]);
+            auto label_embed_2 = std::dynamic_pointer_cast<Linear>(blocks["label_emb.0.2"]);
+
+            auto label_emb = label_embed_0->forward(ctx, y);
+            label_emb      = ggml_silu_inplace(ctx, label_emb);
+            label_emb      = label_embed_2->forward(ctx, label_emb);  // [N, time_embed_dim]
+
+            emb = ggml_add(ctx, emb, label_emb);  // [N, time_embed_dim]
+        }
+
+        std::vector<struct ggml_tensor*> outs;
+
+        if (guided_hint == NULL) {
+            guided_hint = input_hint_block_forward(ctx, hint, emb, context);
+        }
+        outs.push_back(guided_hint);
+
+        // input_blocks
+
+        // input block 0
+        auto h = input_blocks_0_0->forward(ctx, x);
+        h      = ggml_add(ctx, h, guided_hint);
+        outs.push_back(zero_convs_0->forward(ctx, h));
+
+        // input block 1-11
+        size_t len_mults    = channel_mult.size();
+        int input_block_idx = 0;
+        int ds              = 1;
+        for (int i = 0; i < len_mults; i++) {
+            int mult = channel_mult[i];
+            for (int j = 0; j < num_res_blocks; j++) {
+                input_block_idx += 1;
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                h                = resblock_forward(name, ctx, h, emb);  // [N, mult*model_channels, h, w]
+                if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
+                    std::string name = "input_blocks." + std::to_string(input_block_idx) + ".1";
+                    h                = attention_layer_forward(name, ctx, backend, h, context);  // [N, mult*model_channels, h, w]
+                }
+
+                auto zero_conv = std::dynamic_pointer_cast<Conv2d>(blocks["zero_convs." + std::to_string(input_block_idx) + ".0"]);
+
+                outs.push_back(zero_conv->forward(ctx, h));
+            }
+            if (i != len_mults - 1) {
+                ds *= 2;
+                input_block_idx += 1;
+
+                std::string name = "input_blocks." + std::to_string(input_block_idx) + ".0";
+                auto block       = std::dynamic_pointer_cast<DownSampleBlock>(blocks[name]);
+
+                h = block->forward(ctx, h);  // [N, mult*model_channels, h/(2^(i+1)), w/(2^(i+1))]
+
+                auto zero_conv = std::dynamic_pointer_cast<Conv2d>(blocks["zero_convs." + std::to_string(input_block_idx) + ".0"]);
+
+                outs.push_back(zero_conv->forward(ctx, h));
+            }
+        }
+        // [N, 4*model_channels, h/8, w/8]
+
+        // middle_block
+        h = resblock_forward("middle_block.0", ctx, h, emb);                      // [N, 4*model_channels, h/8, w/8]
+        h = attention_layer_forward("middle_block.1", ctx, backend, h, context);  // [N, 4*model_channels, h/8, w/8]
+        h = resblock_forward("middle_block.2", ctx, h, emb);                      // [N, 4*model_channels, h/8, w/8]
+
+        // out
+        outs.push_back(middle_block_out->forward(ctx, h));
+        return outs;
+    }
+};
+
+struct ControlNet : public GGMLRunner {
+    SDVersion version = VERSION_SD1;
+    ControlNetBlock control_net;
+
+    ggml_backend_buffer_t control_buffer = NULL;  // keep control output tensors in backend memory
+    ggml_context* control_ctx            = NULL;
+    std::vector<struct ggml_tensor*> controls;  // (12 input block outputs, 1 middle block output) SD 1.5
+    struct ggml_tensor* guided_hint = NULL;     // guided_hint cache, for faster inference
+    bool guided_hint_cached         = false;
+
+    ControlNet(ggml_backend_t backend,
+               bool offload_params_to_cpu,
+               const String2GGMLType& tensor_types = {},
+               SDVersion version                   = VERSION_SD1)
+        : GGMLRunner(backend, offload_params_to_cpu), control_net(version) {
+        control_net.init(params_ctx, tensor_types, "");
+    }
+
+    void enable_conv2d_direct() {
+        std::vector<GGMLBlock*> blocks;
+        control_net.get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                auto conv_block = (Conv2d*)block;
+                conv_block->enable_direct();
+            }
+        }
+    }
+
+    ~ControlNet() {
+        free_control_ctx();
+    }
+
+    void alloc_control_ctx(std::vector<struct ggml_tensor*> outs) {
+        struct ggml_init_params params;
+        params.mem_size   = static_cast<size_t>(outs.size() * ggml_tensor_overhead()) + 1024 * 1024;
+        params.mem_buffer = NULL;
+        params.no_alloc   = true;
+        control_ctx       = ggml_init(params);
+
+        controls.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);
+
+        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_buffer = ggml_backend_alloc_ctx_tensors(control_ctx, runtime_backend);
+
+        LOG_DEBUG("control buffer size %.2fMB", control_buffer_size * 1.f / 1024.f / 1024.f);
+    }
+
+    void free_control_ctx() {
+        if (control_buffer != NULL) {
+            ggml_backend_buffer_free(control_buffer);
+            control_buffer = NULL;
+        }
+        if (control_ctx != NULL) {
+            ggml_free(control_ctx);
+            control_ctx = NULL;
+        }
+        guided_hint        = NULL;
+        guided_hint_cached = false;
+        controls.clear();
+    }
+
+    std::string get_desc() {
+        return "control_net";
+    }
+
+    void get_param_tensors(std::map<std::string, struct 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 = NULL) {
+        struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, CONTROL_NET_GRAPH_SIZE, false);
+
+        x = to_backend(x);
+        if (guided_hint_cached) {
+            hint = NULL;
+        } else {
+            hint = to_backend(hint);
+        }
+        context   = to_backend(context);
+        y         = to_backend(y);
+        timesteps = to_backend(timesteps);
+
+        auto outs = control_net.forward(compute_ctx,
+                                        runtime_backend,
+                                        x,
+                                        hint,
+                                        guided_hint_cached ? guided_hint : NULL,
+                                        timesteps,
+                                        context,
+                                        y);
+
+        if (control_ctx == NULL) {
+            alloc_control_ctx(outs);
+        }
+
+        ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[0], guided_hint));
+        for (int i = 0; i < outs.size() - 1; i++) {
+            ggml_build_forward_expand(gf, ggml_cpy(compute_ctx, outs[i + 1], controls[i]));
+        }
+
+        return gf;
+    }
+
+    void 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     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        // 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* {
+            return build_graph(x, hint, timesteps, context, y);
+        };
+
+        GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+        guided_hint_cached = true;
+    }
+
+    bool load_from_file(const std::string& file_path, int n_threads) {
+        LOG_INFO("loading control net from '%s'", file_path.c_str());
+        alloc_params_buffer();
+        std::map<std::string, ggml_tensor*> tensors;
+        control_net.get_param_tensors(tensors);
+        std::set<std::string> ignore_tensors;
+
+        ModelLoader model_loader;
+        if (!model_loader.init_from_file(file_path)) {
+            LOG_ERROR("init control net model loader from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        bool success = model_loader.load_tensors(tensors, ignore_tensors, n_threads);
+
+        if (!success) {
+            LOG_ERROR("load control net tensors from model loader failed");
+            return false;
+        }
+
+        LOG_INFO("control net model loaded");
+        return success;
+    }
+};
+
+#endif  // __CONTROL_HPP__

diffusion_model.hpp

@@ -0,0 +1,323 @@
+#ifndef __DIFFUSION_MODEL_H__
+#define __DIFFUSION_MODEL_H__
+
+#include "flux.hpp"
+#include "mmdit.hpp"
+#include "qwen_image.hpp"
+#include "unet.hpp"
+#include "wan.hpp"
+
+struct DiffusionParams {
+    struct ggml_tensor* x                     = NULL;
+    struct ggml_tensor* timesteps             = NULL;
+    struct ggml_tensor* context               = NULL;
+    struct ggml_tensor* c_concat              = NULL;
+    struct ggml_tensor* y                     = NULL;
+    struct ggml_tensor* guidance              = NULL;
+    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          = NULL;
+    float vace_strength                       = 1.f;
+    std::vector<int> skip_layers              = {};
+};
+
+struct DiffusionModel {
+    virtual std::string get_desc()                                                      = 0;
+    virtual void compute(int n_threads,
+                         DiffusionParams diffusion_params,
+                         struct ggml_tensor** output     = NULL,
+                         struct ggml_context* output_ctx = NULL)                        = 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 int64_t get_adm_in_channels()                                               = 0;
+};
+
+struct UNetModel : public DiffusionModel {
+    UNetModelRunner unet;
+
+    UNetModel(ggml_backend_t backend,
+              bool offload_params_to_cpu,
+              const String2GGMLType& tensor_types = {},
+              SDVersion version                   = VERSION_SD1,
+              bool flash_attn                     = false)
+        : unet(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model", version, flash_attn) {
+    }
+
+    std::string get_desc() {
+        return unet.get_desc();
+    }
+
+    void alloc_params_buffer() {
+        unet.alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        unet.free_params_buffer();
+    }
+
+    void free_compute_buffer() {
+        unet.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        unet.get_param_tensors(tensors, "model.diffusion_model");
+    }
+
+    size_t get_params_buffer_size() {
+        return unet.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() {
+        return unet.unet.adm_in_channels;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return unet.compute(n_threads,
+                            diffusion_params.x,
+                            diffusion_params.timesteps,
+                            diffusion_params.context,
+                            diffusion_params.c_concat,
+                            diffusion_params.y,
+                            diffusion_params.num_video_frames,
+                            diffusion_params.controls,
+                            diffusion_params.control_strength, output, output_ctx);
+    }
+};
+
+struct MMDiTModel : public DiffusionModel {
+    MMDiTRunner mmdit;
+
+    MMDiTModel(ggml_backend_t backend,
+               bool offload_params_to_cpu,
+               bool flash_attn                     = false,
+               const String2GGMLType& tensor_types = {})
+        : mmdit(backend, offload_params_to_cpu, flash_attn, tensor_types, "model.diffusion_model") {
+    }
+
+    std::string get_desc() {
+        return mmdit.get_desc();
+    }
+
+    void alloc_params_buffer() {
+        mmdit.alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        mmdit.free_params_buffer();
+    }
+
+    void free_compute_buffer() {
+        mmdit.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        mmdit.get_param_tensors(tensors, "model.diffusion_model");
+    }
+
+    size_t get_params_buffer_size() {
+        return mmdit.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() {
+        return 768 + 1280;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return mmdit.compute(n_threads,
+                             diffusion_params.x,
+                             diffusion_params.timesteps,
+                             diffusion_params.context,
+                             diffusion_params.y,
+                             output,
+                             output_ctx,
+                             diffusion_params.skip_layers);
+    }
+};
+
+struct FluxModel : public DiffusionModel {
+    Flux::FluxRunner flux;
+
+    FluxModel(ggml_backend_t backend,
+              bool offload_params_to_cpu,
+              const String2GGMLType& tensor_types = {},
+              SDVersion version                   = VERSION_FLUX,
+              bool flash_attn                     = false,
+              bool use_mask                       = false)
+        : flux(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model", version, flash_attn, use_mask) {
+    }
+
+    std::string get_desc() {
+        return flux.get_desc();
+    }
+
+    void alloc_params_buffer() {
+        flux.alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        flux.free_params_buffer();
+    }
+
+    void free_compute_buffer() {
+        flux.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        flux.get_param_tensors(tensors, "model.diffusion_model");
+    }
+
+    size_t get_params_buffer_size() {
+        return flux.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() {
+        return 768;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        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.increase_ref_index,
+                            output,
+                            output_ctx,
+                            diffusion_params.skip_layers);
+    }
+};
+
+struct WanModel : public DiffusionModel {
+    std::string prefix;
+    WAN::WanRunner wan;
+
+    WanModel(ggml_backend_t backend,
+             bool offload_params_to_cpu,
+             const String2GGMLType& tensor_types = {},
+             const std::string prefix            = "model.diffusion_model",
+             SDVersion version                   = VERSION_WAN2,
+             bool flash_attn                     = false)
+        : prefix(prefix), wan(backend, offload_params_to_cpu, tensor_types, prefix, version, flash_attn) {
+    }
+
+    std::string get_desc() {
+        return wan.get_desc();
+    }
+
+    void alloc_params_buffer() {
+        wan.alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        wan.free_params_buffer();
+    }
+
+    void free_compute_buffer() {
+        wan.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        wan.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() {
+        return wan.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() {
+        return 768;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return wan.compute(n_threads,
+                           diffusion_params.x,
+                           diffusion_params.timesteps,
+                           diffusion_params.context,
+                           diffusion_params.y,
+                           diffusion_params.c_concat,
+                           NULL,
+                           diffusion_params.vace_context,
+                           diffusion_params.vace_strength,
+                           output,
+                           output_ctx);
+    }
+};
+
+struct QwenImageModel : public DiffusionModel {
+    std::string prefix;
+    Qwen::QwenImageRunner qwen_image;
+
+    QwenImageModel(ggml_backend_t backend,
+                   bool offload_params_to_cpu,
+                   const String2GGMLType& tensor_types = {},
+                   const std::string prefix            = "model.diffusion_model",
+                   SDVersion version                   = VERSION_QWEN_IMAGE,
+                   bool flash_attn                     = false)
+        : prefix(prefix), qwen_image(backend, offload_params_to_cpu, tensor_types, prefix, version, flash_attn) {
+    }
+
+    std::string get_desc() {
+        return qwen_image.get_desc();
+    }
+
+    void alloc_params_buffer() {
+        qwen_image.alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        qwen_image.free_params_buffer();
+    }
+
+    void free_compute_buffer() {
+        qwen_image.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        qwen_image.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() {
+        return qwen_image.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() {
+        return 768;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        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);
+    }
+};
+
+#endif

docs/chroma.md

@@ -0,0 +1,33 @@
+# How to Use
+
+You can run Chroma using stable-diffusion.cpp with a GPU that has 6GB or even 4GB of VRAM, without needing to offload to RAM.
+
+## Download weights
+
+- Download Chroma
+    - If you don't want to do the conversion yourself, download the preconverted gguf model from [silveroxides/Chroma-GGUF](https://huggingface.co/silveroxides/Chroma-GGUF)
+    - Otherwise, download chroma's safetensors from [lodestones/Chroma](https://huggingface.co/lodestones/Chroma)
+- Download vae from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download t5xxl from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
+
+## Convert Chroma weights
+
+You can download the preconverted gguf weights from [silveroxides/Chroma-GGUF](https://huggingface.co/silveroxides/Chroma-GGUF), this way you don't have to do the conversion yourself.
+
+```
+.\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\chroma-unlocked-v40.safetensors -o ..\models\chroma-unlocked-v40-q8_0.gguf -v --type q8_0
+```
+
+## Run
+
+### Example
+For example:
+
+```
+ .\bin\Release\sd.exe --diffusion-model  ..\models\chroma-unlocked-v40-q8_0.gguf --vae ..\models\ae.sft --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma.cpp'" --cfg-scale 4.0 --sampling-method euler -v --chroma-disable-dit-mask --clip-on-cpu
+```
+
+![](../assets/flux/chroma_v40.png)
+
+
+

docs/docker.md

@@ -0,0 +1,15 @@
+## Docker
+
+### Building using Docker
+
+```shell
+docker build -t sd .
+```
+
+### Run
+
+```shell
+docker run -v /path/to/models:/models -v /path/to/output/:/output sd [args...]
+# For example
+# docker run -v ./models:/models -v ./build:/output sd -m /models/sd-v1-4.ckpt -p "a lovely cat" -v -o /output/output.png
+```

docs/esrgan.md

@@ -0,0 +1,9 @@
+## 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.
+
+- Specify the model path using the `--upscale-model PATH` parameter. example:
+
+```bash
+sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --upscale-model ../models/RealESRGAN_x4plus_anime_6B.pth
+```

docs/flux.md

@@ -0,0 +1,66 @@
+# How to Use
+
+You can run Flux using stable-diffusion.cpp with a GPU that has 6GB or even 4GB of VRAM, without needing to offload to RAM.
+
+## Download weights
+
+- Download flux
+    - If you don't want to do the conversion yourself, download the preconverted gguf model from [FLUX.1-dev-gguf](https://huggingface.co/leejet/FLUX.1-dev-gguf) or [FLUX.1-schnell](https://huggingface.co/leejet/FLUX.1-schnell-gguf)
+    - Otherwise, download flux-dev from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/flux1-dev.safetensors or flux-schnell from https://huggingface.co/black-forest-labs/FLUX.1-schnell/blob/main/flux1-schnell.safetensors
+- Download vae from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download clip_l from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/clip_l.safetensors
+- Download t5xxl from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
+
+## Convert flux weights
+
+You can download the preconverted gguf weights from [FLUX.1-dev-gguf](https://huggingface.co/leejet/FLUX.1-dev-gguf) or [FLUX.1-schnell](https://huggingface.co/leejet/FLUX.1-schnell-gguf), this way you don't have to do the conversion yourself.
+
+Using fp16 will lead to overflow, but ggml's support for bf16 is not yet fully developed. Therefore, we need to convert flux to gguf format here, which also saves VRAM. For example:
+```
+.\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-dev.sft -o ..\models\flux1-dev-q8_0.gguf -v --type q8_0
+```
+
+## Run
+
+- `--cfg-scale` is recommended to be set to 1. 
+
+### Flux-dev
+For example:
+
+```
+ .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+```
+
+Using formats of different precisions will yield results of varying quality.
+
+| Type | q8_0  | q4_0  | q4_k  | q3_k  | q2_k |
+|---- | ----  |----  |----  |----  |----  |
+| **Memory** | 12068.09 MB  | 6394.53 MB | 6395.17 MB | 4888.16 MB  | 3735.73 MB |
+| **Result** | ![](../assets/flux/flux1-dev-q8_0.png) |![](../assets/flux/flux1-dev-q4_0.png) |![](../assets/flux/flux1-dev-q4_k.png) |![](../assets/flux/flux1-dev-q3_k.png) |![](../assets/flux/flux1-dev-q2_k.png)|
+
+
+
+### Flux-schnell
+
+
+```
+ .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-schnell-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --steps 4 --clip-on-cpu
+```
+
+| q8_0  |
+| ----  |
+|![](../assets/flux/flux1-schnell-q8_0.png) |
+
+## Run with LoRA
+
+Since many flux LoRA training libraries have used various LoRA naming formats, it is possible that not all flux LoRA naming formats are supported. It is recommended to use LoRA with naming formats compatible with ComfyUI.
+
+### Flux-dev q8_0 with LoRA
+
+- LoRA model from https://huggingface.co/XLabs-AI/flux-lora-collection/tree/main (using comfy converted version!!!)
+
+```
+.\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ...\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'<lora:realism_lora_comfy_converted:1>" --cfg-scale 1.0 --sampling-method euler -v --lora-model-dir ../models --clip-on-cpu
+```
+
+![output](../assets/flux/flux1-dev-q8_0%20with%20lora.png)

docs/hipBLAS_on_Windows.md

@@ -0,0 +1,85 @@
+# Using hipBLAS on Windows
+
+To get hipBLAS in `stable-diffusion.cpp` working on Windows, go through this guide section by section.
+
+## Build Tools for Visual Studio 2022
+
+Skip this step if you already have Build Tools installed.
+
+To install Build Tools, go to [Visual Studio Downloads](https://visualstudio.microsoft.com/vs/), download `Visual Studio 2022 and other Products` and run the installer.
+
+## CMake
+
+Skip this step if you already have CMake installed: running `cmake --version` should output `cmake version x.y.z`.
+
+Download latest `Windows x64 Installer` from [Download | CMake](https://cmake.org/download/) and run it.
+
+## ROCm
+
+Skip this step if you already have Build Tools installed.
+
+The [validation tools](https://rocm.docs.amd.com/en/latest/reference/validation_tools.html) not support on Windows. So you should confirm the Version of `ROCM` by yourself.
+
+Fortunately, `AMD` provides complete help documentation, you can use the help documentation to install [ROCM](https://rocm.docs.amd.com/en/latest/deploy/windows/quick_start.html)
+
+>**If you encounter an error, if it is [AMD ROCm Windows Installation Error 215](https://github.com/RadeonOpenCompute/ROCm/issues/2363), don't worry about this error. ROCM has been installed correctly, but the vs studio plugin installation failed, we can ignore it.**
+
+Then we must set `ROCM` as environment variables before running cmake.
+
+Usually if you install according to the official tutorial and do not modify the ROCM path, then there is a high probability that it is here `C:\Program Files\AMD\ROCm\5.5\bin`
+
+This is what I use to set the clang:
+```Commandline
+set CC=C:\Program Files\AMD\ROCm\5.5\bin\clang.exe
+set CXX=C:\Program Files\AMD\ROCm\5.5\bin\clang++.exe
+```
+
+## Ninja
+
+Skip this step if you already have Ninja installed: running `ninja --version` should output `1.11.1`.
+
+Download latest `ninja-win.zip` from [GitHub Releases Page](https://github.com/ninja-build/ninja/releases/tag/v1.11.1) and unzip. Then set as environment variables. I unzipped it in `C:\Program Files\ninja`, so I set it like this:
+
+```Commandline
+set ninja=C:\Program Files\ninja\ninja.exe
+```
+## Building stable-diffusion.cpp
+
+The thing different from the regular CPU build is `-DSD_HIPBLAS=ON` ,
+`-G "Ninja"`, `-DCMAKE_C_COMPILER=clang`, `-DCMAKE_CXX_COMPILER=clang++`, `-DAMDGPU_TARGETS=gfx1100`
+
+>**Notice**: check the `clang` and `clang++` information:
+```Commandline
+clang --version
+clang++ --version
+```
+
+If you see like this, we can continue:
+```
+clang version 17.0.0 (git@github.amd.com:Compute-Mirrors/llvm-project e3201662d21c48894f2156d302276eb1cf47c7be)
+Target: x86_64-pc-windows-msvc
+Thread model: posix
+InstalledDir: C:\Program Files\AMD\ROCm\5.5\bin
+```
+
+```
+clang version 17.0.0 (git@github.amd.com:Compute-Mirrors/llvm-project e3201662d21c48894f2156d302276eb1cf47c7be)
+Target: x86_64-pc-windows-msvc
+Thread model: posix
+InstalledDir: C:\Program Files\AMD\ROCm\5.5\bin
+```
+
+>**Notice** that the `gfx1100` is the GPU architecture of my GPU, you can change it to your GPU architecture. Click here to see your architecture [LLVM Target](https://rocm.docs.amd.com/en/latest/release/windows_support.html#windows-supported-gpus)
+
+My GPU is AMD Radeon™ RX 7900 XTX Graphics, so I set it to `gfx1100`.
+
+option:
+
+```commandline
+mkdir build
+cd build
+cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=gfx1100
+cmake --build . --config Release
+```
+
+If everything went OK, `build\bin\sd.exe` file should appear.

docs/kontext.md

@@ -0,0 +1,39 @@
+# How to Use
+
+You can run Kontext using stable-diffusion.cpp with a GPU that has 6GB or even 4GB of VRAM, without needing to offload to RAM.
+
+## Download weights
+
+- Download Kontext
+    - If you don't want to do the conversion yourself, download the preconverted gguf model from [FLUX.1-Kontext-dev-GGUF](https://huggingface.co/QuantStack/FLUX.1-Kontext-dev-GGUF)
+    - Otherwise, download FLUX.1-Kontext-dev from https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev/blob/main/flux1-kontext-dev.safetensors
+- Download vae from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download clip_l from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/clip_l.safetensors
+- Download t5xxl from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
+
+## Convert Kontext weights
+
+You can download the preconverted gguf weights from [FLUX.1-Kontext-dev-GGUF](https://huggingface.co/QuantStack/FLUX.1-Kontext-dev-GGUF), this way you don't have to do the conversion yourself.
+
+```
+.\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-kontext-dev.safetensors -o ..\models\flux1-kontext-dev-q8_0.gguf -v --type q8_0
+```
+
+## Run
+
+- `--cfg-scale` is recommended to be set to 1. 
+
+### Example
+For example:
+
+```
+ .\bin\Release\sd.exe -r .\flux1-dev-q8_0.png --diffusion-model  ..\models\flux1-kontext-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors -p "change 'flux.cpp' to 'kontext.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+```
+
+
+| ref_image | prompt  | output  |
+| ---- | ----  |----  |
+| ![](../assets/flux/flux1-dev-q8_0.png) | change 'flux.cpp' to 'kontext.cpp' |![](../assets/flux/kontext1_dev_output.png) |
+
+
+

docs/lcm.md

@@ -0,0 +1,15 @@
+## LCM/LCM-LoRA
+
+- Download LCM-LoRA form https://huggingface.co/latent-consistency/lcm-lora-sdv1-5
+- Specify LCM-LoRA by adding `<lora:lcm-lora-sdv1-5:1>` to prompt
+- It's advisable to set `--cfg-scale` to `1.0` instead of the default `7.0`. For `--steps`, a range of `2-8` steps is recommended. For `--sampling-method`, `lcm`/`euler_a` is recommended.
+
+Here's a simple example:
+
+```
+./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:lcm-lora-sdv1-5:1>" --steps 4 --lora-model-dir ../models -v --cfg-scale 1
+```
+
+| without LCM-LoRA (--cfg-scale 7)  | with LCM-LoRA (--cfg-scale 1)  |
+| ----  |----    |
+| ![](../assets/without_lcm.png) |![](../assets/with_lcm.png)  |

docs/lora.md

@@ -0,0 +1,49 @@
+## LoRA
+
+- You can specify the directory where the lora weights are stored via `--lora-model-dir`. If not specified, the default is the current working directory.
+
+- LoRA is specified via prompt, just like [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#lora).
+
+Here's a simple example:
+
+```
+./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat<lora:marblesh:1>" --lora-model-dir ../models
+```
+
+`../models/marblesh.safetensors` or `../models/marblesh.ckpt` will be applied to the model
+
+# Support matrix
+
+> ℹ️ CUDA `get_rows` support is defined here:  
+> [ggml-org/ggml/src/ggml-cuda/getrows.cu#L156](https://github.com/ggml-org/ggml/blob/7dee1d6a1e7611f238d09be96738388da97c88ed/src/ggml-cuda/getrows.cu#L156)  
+> Currently only the basic types + Q4/Q5/Q8 are implemented. K-quants are **not** supported.
+
+NOTE: The other backends may have different support.
+
+| Quant / Type | CUDA | Vulkan |
+|--------------|------|--------|
+| F32          | ✔️   | ✔️   |
+| F16          | ✔️   | ✔️   |
+| BF16         | ✔️   | ✔️   |
+| I32          | ✔️   | ❌   |
+| Q4_0         | ✔️   | ✔️   |
+| Q4_1         | ✔️   | ✔️   |
+| Q5_0         | ✔️   | ✔️   |
+| Q5_1         | ✔️   | ✔️   |
+| Q8_0         | ✔️   | ✔️   |
+| Q2_K         | ❌   | ❌   |
+| Q3_K         | ❌   | ❌   |
+| Q4_K         | ❌   | ❌   |
+| Q5_K         | ❌   | ❌   |
+| Q6_K         | ❌   | ❌   |
+| Q8_K         | ❌   | ❌   |
+| IQ1_S        | ❌   | ✔️   |
+| IQ1_M        | ❌   | ✔️   |
+| IQ2_XXS      | ❌   | ✔️   |
+| IQ2_XS       | ❌   | ✔️   |
+| IQ2_S        | ❌   | ✔️   |
+| IQ3_XXS      | ❌   | ✔️   |
+| IQ3_S        | ❌   | ✔️   |
+| IQ4_XS       | ❌   | ✔️   |
+| IQ4_NL       | ❌   | ✔️   |
+| MXFP4        | ❌   | ✔️   |

docs/photo_maker.md

@@ -0,0 +1,53 @@
+## Using PhotoMaker to personalize image generation
+
+You can use [PhotoMaker](https://github.com/TencentARC/PhotoMaker) to personalize generated images with your own ID.
+
+**NOTE**, currently PhotoMaker **ONLY** works with **SDXL** (any SDXL model files will work).
+
+Download PhotoMaker model file (in safetensor format) [here](https://huggingface.co/bssrdf/PhotoMaker). The official release of the model file (in .bin format) does not work with ```stablediffusion.cpp```.
+
+- Specify the PhotoMaker model path using the `--photo-maker PATH` parameter.
+- Specify the input images path using the `--pm-id-images-dir PATH` parameter.
+
+In prompt, make sure you have a class word followed by the trigger word ```"img"``` (hard-coded for now). The class word could be one of ```"man, woman, girl, boy"```. If input ID images contain asian faces, add ```Asian``` before the class
+word.
+
+Another PhotoMaker specific parameter:
+
+- ```--pm-style-strength  (0-100)%```: default is 20 and 10-20 typically gets good results. Lower ratio means more faithfully following input ID (not necessarily better quality).
+
+Other parameters recommended for running Photomaker:
+
+- ```--cfg-scale 5.0```
+- ```-H 1024```
+- ```-W 1024```
+
+If on low memory GPUs (<= 8GB), recommend running with ```--vae-on-cpu``` option to get artifact free images.
+
+Example:
+
+```bash
+bin/sd -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --photo-maker ../models/photomaker-v1.safetensors --pm-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --pm-style-strength 10 --vae-on-cpu --steps 50
+```
+
+## PhotoMaker Version 2
+
+[PhotoMaker Version 2 (PMV2)](https://github.com/TencentARC/PhotoMaker/blob/main/README_pmv2.md) has some key improvements. Unfortunately it has a very heavy dependency which makes running it a bit involved in ```SD.cpp```. 
+
+Running PMV2 is now a two-step process:
+
+- Run a python script ```face_detect.py``` to obtain **id_embeds** for the given input images
+```
+python face_detect.py input_image_dir
+```
+An ```id_embeds.safetensors``` file will be generated in ```input_images_dir```
+
+**Note: this step is only needed to run once; the same ```id_embeds``` can be reused**
+
+- Run the same command as in version 1 but replacing ```photomaker-v1.safetensors``` with ```photomaker-v2.safetensors```.
+
+  You can download ```photomaker-v2.safetensors``` from [here](https://huggingface.co/bssrdf/PhotoMakerV2)
+
+- All the command line parameters from Version 1 remain the same for Version 2
+
+

docs/quantization_and_gguf.md

@@ -0,0 +1,27 @@
+## Quantization
+
+You can specify the model weight type using the `--type` parameter. The weights are automatically converted when loading the model.
+
+- `f16` for 16-bit floating-point
+- `f32` for 32-bit floating-point
+- `q8_0` for 8-bit integer quantization
+- `q5_0` or `q5_1` for 5-bit integer quantization
+- `q4_0` or `q4_1` for 4-bit integer quantization
+
+
+### Memory Requirements of Stable Diffusion 1.x
+
+| precision | f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
+| ----         | ----  |----  |----  |----  |----  |----  |----  |
+|  **Memory** (txt2img - 512 x 512) | ~2.8G | ~2.3G | ~2.1G | ~2.0G | ~2.0G | ~2.0G | ~2.0G |
+|  **Memory** (txt2img - 512 x 512) *with Flash Attention* | ~2.4G | ~1.9G | ~1.6G | ~1.5G | ~1.5G | ~1.5G | ~1.5G |
+
+## Convert to GGUF
+
+You can also convert weights in the formats `ckpt/safetensors/diffusers` to gguf and perform quantization in advance, avoiding the need for quantization every time you load them.
+
+For example:
+
+```sh
+./bin/sd -M convert -m ../models/v1-5-pruned-emaonly.safetensors -o  ../models/v1-5-pruned-emaonly.q8_0.gguf -v --type q8_0
+```

docs/qwen_image.md

@@ -0,0 +1,23 @@
+# How to Use
+
+## Download weights
+
+- Download Qwen Image
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/QuantStack/Qwen-Image-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae
+- Download qwen_2.5_vl 7b
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-GGUF/tree/main
+
+## Examples
+
+```
+.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --qwen2vl ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf  -p '一个穿着"QWEN"标志的T恤的中国美女正拿着黑色的马克笔面相镜头微笑。她身后的玻璃板上手写体写着 “一、Qwen-Image的技术路线： 探索视觉生成基础模型的极限，开创理解与生成一体化的未来。二、Qwen-Image的模型特色：1、复杂文字渲染。支持中英渲染、自动布局； 2、精准图像编辑。支持文字编辑、物体增减、风格变换。三、Qwen-Image的未来愿景：赋能专业内容创作、助力生成式AI发展。”' --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu -H 1024 -W 1024 --diffusion-fa --flow-shift 3
+```
+
+<img alt="qwen example" src="../assets/qwen/example.png" />
+
+
+

docs/qwen_image_edit.md

@@ -0,0 +1,35 @@
+# How to Use
+
+## Download weights
+
+- Download Qwen Image
+    - Qwen Image Edit
+        - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
+        - gguf: https://huggingface.co/QuantStack/Qwen-Image-Edit-GGUF/tree/main
+    - Qwen Image Edit 2509
+        - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
+        - gguf: https://huggingface.co/QuantStack/Qwen-Image-Edit-2509-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae
+- Download qwen_2.5_vl 7b
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-GGUF/tree/main
+
+## Examples
+
+### Qwen Image Edit
+
+```
+.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen_Image_Edit-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --qwen2vl ..\..\ComfyUI\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'" --seed 1118877715456453
+```
+
+<img alt="qwen_image_edit" src="../assets/qwen/qwen_image_edit.png" />
+
+
+### Qwen Image Edit 2509
+
+```
+.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen-Image-Edit-2509-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --qwen2vl ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf --qwen2vl_vision ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'Qwen Image Edit 2509'"
+```
+
+<img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2509.png" />

docs/sd3.md

@@ -0,0 +1,20 @@
+# How to Use
+
+## Download weights
+
+- Download sd3.5_large from https://huggingface.co/stabilityai/stable-diffusion-3.5-large/blob/main/sd3.5_large.safetensors
+- Download clip_g from https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_g.safetensors
+- Download clip_l from https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_l.safetensors
+- Download t5xxl from https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/t5xxl_fp16.safetensors
+
+
+## Run
+
+### SD3.5 Large
+For example:
+
+```
+.\bin\Release\sd.exe -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+```
+
+![](../assets/sd3.5_large.png)

docs/taesd.md

@@ -0,0 +1,17 @@
+## Using TAESD to faster decoding
+
+You can use TAESD to accelerate the decoding of latent images by following these steps:
+
+- Download the model [weights](https://huggingface.co/madebyollin/taesd/blob/main/diffusion_pytorch_model.safetensors).
+
+Or curl
+
+```bash
+curl -L -O https://huggingface.co/madebyollin/taesd/blob/main/diffusion_pytorch_model.safetensors
+```
+
+- Specify the model path using the `--taesd PATH` parameter. example:
+
+```bash
+sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat" --taesd ../models/diffusion_pytorch_model.safetensors
+```

docs/wan.md

@@ -0,0 +1,204 @@
+# How to Use
+
+## Download weights
+
+- Download Wan
+    - Wan2.1
+        - Wan2.1 T2V 1.3B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+        - Wan2.1 T2V 14B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/city96/Wan2.1-T2V-14B-gguf/tree/main
+        - Wan2.1 I2V 14B 480P
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/city96/Wan2.1-I2V-14B-480P-gguf/tree/main
+        - Wan2.1 I2V 14B 720P
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/city96/Wan2.1-I2V-14B-720P-gguf/tree/main
+        - Wan2.1 FLF2V 14B 720P
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/city96/Wan2.1-FLF2V-14B-720P-gguf/tree/main
+        - Wan2.1 VACE 1.3B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/calcuis/wan-1.3b-gguf/tree/main
+        - Wan2.1 VACE 14B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/QuantStack/Wan2.1_14B_VACE-GGUF/tree/main
+    - Wan2.2
+        - Wan2.2 TI2V 5B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/QuantStack/Wan2.2-TI2V-5B-GGUF/tree/main
+        - Wan2.2 T2V A14B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/QuantStack/Wan2.2-T2V-A14B-GGUF/tree/main
+        - Wan2.2 I2V A14B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF/tree/main
+- Download vae
+    - wan_2.1_vae (for all the wan model except Wan2.2 TI2V 5B)
+        - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors
+    - wan_2.2_vae (for Wan2.2 TI2V 5B only)
+        - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/vae/wan2.2_vae.safetensors
+- Download umt5_xxl
+    - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/text_encoders/umt5_xxl_fp16.safetensors
+    - gguf: https://huggingface.co/city96/umt5-xxl-encoder-gguf/tree/main
+
+- Download clip_vison_h (for Wan2.1 I2V/FLF2V only)
+    - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/clip_vision/clip_vision_h.safetensors
+
+
+## Examples
+
+### Wan2.1 T2V 1.3B
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1_t2v_1.3B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --flow-shift 3.0
+```
+
+<video src=../assets/wan/Wan2.1_1.3B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.1 T2V 14B
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-t2v-14b-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa  --offload-to-cpu --video-frames 33 --flow-shift 3.0
+```
+
+<video src=../assets/wan/Wan2.1_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+
+### Wan2.1 I2V 14B
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-i2v-14b-480p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+```
+
+<video src=../assets/wan/Wan2.1_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.2 T2V A14B
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
+```
+
+<video src=../assets/wan/Wan2.2_14B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.2 I2V A14B
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --offload-to-cpu -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+```
+
+<video src=../assets/wan/Wan2.2_14B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.2 T2V A14B T2I
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --flow-shift 3.0
+```
+
+<img width="832" height="480" alt="Wan2 2_14B_t2i" src="../assets/wan/Wan2.2_14B_t2i.png" />
+
+### Wan2.2 T2V 14B with Lora
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat<lora:wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise:1><lora:|high_noise|wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise:1>" --cfg-scale 3.5 --sampling-method euler --steps 4 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 4 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --lora-model-dir ..\..\ComfyUI\models\loras --video-frames 33 --flow-shift 3.0
+```
+
+<video src=../assets/wan/Wan2.2_14B_t2v_lora.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+
+### Wan2.2 TI2V 5B
+
+#### T2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 --flow-shift 3.0
+```
+
+<video src=../assets/wan/Wan2.2_5B_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+#### I2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.2_ti2v_5B_fp16.safetensors --vae ..\..\ComfyUI\models\vae\wan2.2_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\cat_with_sd_cpp_42.png --flow-shift 3.0
+```
+
+<video src=../assets/wan/Wan2.2_5B_i2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.1 FLF2V 14B
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-flf2v-14b-720p-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --clip_vision ..\..\ComfyUI\models\clip_vision\clip_vision_h.safetensors -p "glass flower blossom" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
+```
+
+
+<video src=../assets/wan/Wan2.1_14B_flf2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.2 FLF2V 14B
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -p "glass flower blossom" -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa --video-frames 33 --offload-to-cpu --init-img ..\..\ComfyUI\input\start_image.png --end-img ..\..\ComfyUI\input\end_image.png --flow-shift 3.0
+```
+
+<video src=../assets/wan/Wan2.2_14B_flf2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.1 VACE 1.3B
+
+#### T2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 1 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_1.3B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+#### R2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_1.3B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+#### V2V
+
+```
+mkdir post+depth
+ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\frame_%04d.jpg
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_1.3B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.1 VACE 14B
+
+#### T2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_14B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+#### R2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_14B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+
+#### V2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_14B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>

esrgan.hpp

@@ -0,0 +1,378 @@
+#ifndef __ESRGAN_HPP__
+#define __ESRGAN_HPP__
+
+#include "ggml_extend.hpp"
+#include "model.h"
+
+/*
+    ===================================    ESRGAN  ===================================
+    References:
+    https://github.com/xinntao/Real-ESRGAN/blob/master/inference_realesrgan.py
+    https://github.com/XPixelGroup/BasicSR/blob/v1.4.2/basicsr/archs/rrdbnet_arch.py
+
+*/
+
+class ResidualDenseBlock : public GGMLBlock {
+protected:
+    int num_feat;
+    int num_grow_ch;
+
+public:
+    ResidualDenseBlock(int num_feat = 64, int num_grow_ch = 32)
+        : num_feat(num_feat), num_grow_ch(num_grow_ch) {
+        blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv3"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv4"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        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(struct ggml_context* ctx, struct ggml_tensor* x) {
+        return ggml_leaky_relu(ctx, x, 0.2f, true);
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, num_feat, h, w]
+        // return: [n, num_feat, h, w]
+
+        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv1"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv2"]);
+        auto conv3 = std::dynamic_pointer_cast<Conv2d>(blocks["conv3"]);
+        auto conv4 = std::dynamic_pointer_cast<Conv2d>(blocks["conv4"]);
+        auto conv5 = std::dynamic_pointer_cast<Conv2d>(blocks["conv5"]);
+
+        auto x1    = lrelu(ctx, conv1->forward(ctx, x));
+        auto x_cat = ggml_concat(ctx, x, x1, 2);
+        auto x2    = lrelu(ctx, conv2->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx, x_cat, x2, 2);
+        auto x3    = lrelu(ctx, conv3->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx, x_cat, x3, 2);
+        auto x4    = lrelu(ctx, conv4->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx, x_cat, x4, 2);
+        auto x5    = conv5->forward(ctx, x_cat);
+
+        x5 = ggml_add(ctx, ggml_scale(ctx, x5, 0.2f), x);
+        return x5;
+    }
+};
+
+class RRDB : public GGMLBlock {
+public:
+    RRDB(int num_feat, int num_grow_ch = 32) {
+        blocks["rdb1"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+        blocks["rdb2"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+        blocks["rdb3"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, num_feat, h, w]
+        // return: [n, num_feat, h, w]
+
+        auto rdb1 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb1"]);
+        auto rdb2 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb2"]);
+        auto rdb3 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb3"]);
+
+        auto out = rdb1->forward(ctx, x);
+        out      = rdb2->forward(ctx, out);
+        out      = rdb3->forward(ctx, out);
+
+        out = ggml_add(ctx, ggml_scale(ctx, out, 0.2f), x);
+        return out;
+    }
+};
+
+class RRDBNet : public GGMLBlock {
+protected:
+    int scale       = 4;
+    int num_block   = 23;
+    int num_in_ch   = 3;
+    int num_out_ch  = 3;
+    int num_feat    = 64;
+    int num_grow_ch = 32;
+
+public:
+    RRDBNet(int scale, int num_block, int num_in_ch, int num_out_ch, int num_feat, int num_grow_ch)
+        : scale(scale), num_block(num_block), num_in_ch(num_in_ch), num_out_ch(num_out_ch), num_feat(num_feat), num_grow_ch(num_grow_ch) {
+        blocks["conv_first"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_in_ch, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        for (int i = 0; i < num_block; i++) {
+            std::string name = "body." + std::to_string(i);
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new RRDB(num_feat, num_grow_ch));
+        }
+        blocks["conv_body"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        if (scale >= 2) {
+            blocks["conv_up1"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        }
+        if (scale == 4) {
+            blocks["conv_up2"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        }
+        blocks["conv_hr"]   = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv_last"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_out_ch, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    int get_scale() { return scale; }
+    int get_num_block() { return num_block; }
+
+    struct ggml_tensor* lrelu(struct ggml_context* ctx, struct ggml_tensor* x) {
+        return ggml_leaky_relu(ctx, x, 0.2f, true);
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct 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"]);
+        auto conv_body  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_body"]);
+        auto conv_hr    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_hr"]);
+        auto conv_last  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_last"]);
+
+        auto feat      = conv_first->forward(ctx, x);
+        auto body_feat = feat;
+        for (int i = 0; i < num_block; i++) {
+            std::string name = "body." + std::to_string(i);
+            auto block       = std::dynamic_pointer_cast<RRDB>(blocks[name]);
+
+            body_feat = block->forward(ctx, body_feat);
+        }
+        body_feat = conv_body->forward(ctx, body_feat);
+        feat      = ggml_add(ctx, feat, body_feat);
+        // upsample
+        if (scale >= 2) {
+            auto conv_up1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up1"]);
+            feat          = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+            if (scale == 4) {
+                auto conv_up2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up2"]);
+                feat          = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+            }
+        }
+        // for all scales
+        auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat)));
+        return out;
+    }
+};
+
+struct ESRGAN : public GGMLRunner {
+    std::unique_ptr<RRDBNet> rrdb_net;
+    int scale     = 4;
+    int tile_size = 128;  // avoid cuda OOM for 4gb VRAM
+
+    ESRGAN(ggml_backend_t backend,
+           bool offload_params_to_cpu,
+           const String2GGMLType& tensor_types = {})
+        : GGMLRunner(backend, offload_params_to_cpu) {
+        // rrdb_net will be created in load_from_file
+    }
+
+    void enable_conv2d_direct() {
+        if (!rrdb_net)
+            return;
+        std::vector<GGMLBlock*> blocks;
+        rrdb_net->get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                auto conv_block = (Conv2d*)block;
+                conv_block->enable_direct();
+            }
+        }
+    }
+
+    std::string get_desc() {
+        return "esrgan";
+    }
+
+    bool load_from_file(const std::string& file_path, int n_threads) {
+        LOG_INFO("loading esrgan from '%s'", file_path.c_str());
+
+        ModelLoader model_loader;
+        if (!model_loader.init_from_file(file_path)) {
+            LOG_ERROR("init esrgan model loader from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        // Get tensor names
+        auto tensor_names = model_loader.get_tensor_names();
+
+        // Detect if it's ESRGAN format
+        bool is_ESRGAN = std::find(tensor_names.begin(), tensor_names.end(), "model.0.weight") != tensor_names.end();
+
+        // Detect parameters from tensor names
+        int detected_num_block = 0;
+        if (is_ESRGAN) {
+            for (const auto& name : tensor_names) {
+                if (name.find("model.1.sub.") == 0) {
+                    size_t first_dot = name.find('.', 12);
+                    if (first_dot != std::string::npos) {
+                        size_t second_dot = name.find('.', first_dot + 1);
+                        if (second_dot != std::string::npos && name.substr(first_dot + 1, 3) == "RDB") {
+                            try {
+                                int idx            = std::stoi(name.substr(12, first_dot - 12));
+                                detected_num_block = std::max(detected_num_block, idx + 1);
+                            } catch (...) {
+                            }
+                        }
+                    }
+                }
+            }
+        } else {
+            // Original format
+            for (const auto& name : tensor_names) {
+                if (name.find("body.") == 0) {
+                    size_t pos = name.find('.', 5);
+                    if (pos != std::string::npos) {
+                        try {
+                            int idx            = std::stoi(name.substr(5, pos - 5));
+                            detected_num_block = std::max(detected_num_block, idx + 1);
+                        } catch (...) {
+                        }
+                    }
+                }
+            }
+        }
+
+        int detected_scale = 4;  // default
+        if (is_ESRGAN) {
+            // For ESRGAN format, detect scale by highest model number
+            int max_model_num = 0;
+            for (const auto& name : tensor_names) {
+                if (name.find("model.") == 0) {
+                    size_t dot_pos = name.find('.', 6);
+                    if (dot_pos != std::string::npos) {
+                        try {
+                            int num       = std::stoi(name.substr(6, dot_pos - 6));
+                            max_model_num = std::max(max_model_num, num);
+                        } catch (...) {
+                        }
+                    }
+                }
+            }
+            if (max_model_num <= 4) {
+                detected_scale = 1;
+            } else if (max_model_num <= 7) {
+                detected_scale = 2;
+            } else {
+                detected_scale = 4;
+            }
+        } else {
+            // Original format
+            bool has_conv_up2 = std::any_of(tensor_names.begin(), tensor_names.end(), [](const std::string& name) {
+                return name == "conv_up2.weight";
+            });
+            bool has_conv_up1 = std::any_of(tensor_names.begin(), tensor_names.end(), [](const std::string& name) {
+                return name == "conv_up1.weight";
+            });
+            if (has_conv_up2) {
+                detected_scale = 4;
+            } else if (has_conv_up1) {
+                detected_scale = 2;
+            } else {
+                detected_scale = 1;
+            }
+        }
+
+        int detected_num_in_ch   = 3;
+        int detected_num_out_ch  = 3;
+        int detected_num_feat    = 64;
+        int detected_num_grow_ch = 32;
+
+        // Create RRDBNet with detected parameters
+        rrdb_net = std::make_unique<RRDBNet>(detected_scale, detected_num_block, detected_num_in_ch, detected_num_out_ch, detected_num_feat, detected_num_grow_ch);
+        rrdb_net->init(params_ctx, {}, "");
+
+        alloc_params_buffer();
+        std::map<std::string, ggml_tensor*> esrgan_tensors;
+        rrdb_net->get_param_tensors(esrgan_tensors);
+
+        bool success;
+        if (is_ESRGAN) {
+            // Build name mapping for ESRGAN format
+            std::map<std::string, std::string> expected_to_model;
+            expected_to_model["conv_first.weight"] = "model.0.weight";
+            expected_to_model["conv_first.bias"]   = "model.0.bias";
+
+            for (int i = 0; i < detected_num_block; i++) {
+                for (int j = 1; j <= 3; j++) {
+                    for (int k = 1; k <= 5; k++) {
+                        std::string expected_weight        = "body." + std::to_string(i) + ".rdb" + std::to_string(j) + ".conv" + std::to_string(k) + ".weight";
+                        std::string model_weight           = "model.1.sub." + std::to_string(i) + ".RDB" + std::to_string(j) + ".conv" + std::to_string(k) + ".0.weight";
+                        expected_to_model[expected_weight] = model_weight;
+
+                        std::string expected_bias        = "body." + std::to_string(i) + ".rdb" + std::to_string(j) + ".conv" + std::to_string(k) + ".bias";
+                        std::string model_bias           = "model.1.sub." + std::to_string(i) + ".RDB" + std::to_string(j) + ".conv" + std::to_string(k) + ".0.bias";
+                        expected_to_model[expected_bias] = model_bias;
+                    }
+                }
+            }
+
+            if (detected_scale == 1) {
+                expected_to_model["conv_body.weight"] = "model.1.sub." + std::to_string(detected_num_block) + ".weight";
+                expected_to_model["conv_body.bias"]   = "model.1.sub." + std::to_string(detected_num_block) + ".bias";
+                expected_to_model["conv_hr.weight"]   = "model.2.weight";
+                expected_to_model["conv_hr.bias"]     = "model.2.bias";
+                expected_to_model["conv_last.weight"] = "model.4.weight";
+                expected_to_model["conv_last.bias"]   = "model.4.bias";
+            } else {
+                expected_to_model["conv_body.weight"] = "model.1.sub." + std::to_string(detected_num_block) + ".weight";
+                expected_to_model["conv_body.bias"]   = "model.1.sub." + std::to_string(detected_num_block) + ".bias";
+                if (detected_scale >= 2) {
+                    expected_to_model["conv_up1.weight"] = "model.3.weight";
+                    expected_to_model["conv_up1.bias"]   = "model.3.bias";
+                }
+                if (detected_scale == 4) {
+                    expected_to_model["conv_up2.weight"]  = "model.6.weight";
+                    expected_to_model["conv_up2.bias"]    = "model.6.bias";
+                    expected_to_model["conv_hr.weight"]   = "model.8.weight";
+                    expected_to_model["conv_hr.bias"]     = "model.8.bias";
+                    expected_to_model["conv_last.weight"] = "model.10.weight";
+                    expected_to_model["conv_last.bias"]   = "model.10.bias";
+                } else if (detected_scale == 2) {
+                    expected_to_model["conv_hr.weight"]   = "model.5.weight";
+                    expected_to_model["conv_hr.bias"]     = "model.5.bias";
+                    expected_to_model["conv_last.weight"] = "model.7.weight";
+                    expected_to_model["conv_last.bias"]   = "model.7.bias";
+                }
+            }
+
+            std::map<std::string, ggml_tensor*> model_tensors;
+            for (auto& p : esrgan_tensors) {
+                auto it = expected_to_model.find(p.first);
+                if (it != expected_to_model.end()) {
+                    model_tensors[it->second] = p.second;
+                }
+            }
+
+            success = model_loader.load_tensors(model_tensors, {}, n_threads);
+        } else {
+            success = model_loader.load_tensors(esrgan_tensors, {}, n_threads);
+        }
+
+        if (!success) {
+            LOG_ERROR("load esrgan tensors from model loader failed");
+            return false;
+        }
+
+        scale = rrdb_net->get_scale();
+        LOG_INFO("esrgan model loaded with scale=%d, num_block=%d", scale, detected_num_block);
+        return success;
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* x) {
+        if (!rrdb_net)
+            return nullptr;
+        constexpr int kGraphNodes = 1 << 16;  // 65k
+        struct ggml_cgraph* gf    = ggml_new_graph_custom(compute_ctx, kGraphNodes, /*grads*/ false);
+        x                         = to_backend(x);
+        struct ggml_tensor* out   = rrdb_net->forward(compute_ctx, x);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    void compute(const int n_threads,
+                 struct ggml_tensor* x,
+                 ggml_tensor** output,
+                 ggml_context* output_ctx = NULL) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(x);
+        };
+        GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+    }
+};
+
+#endif  // __ESRGAN_HPP__

examples/cli/CMakeLists.txt

@@ -3,4 +3,4 @@ set(TARGET sd)
 add_executable(${TARGET} main.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE stable-diffusion ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PUBLIC cxx_std_11)
+target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)

examples/cli/avi_writer.h

@@ -0,0 +1,217 @@
+#ifndef __AVI_WRITER_H__
+#define __AVI_WRITER_H__
+
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#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  = NULL;
+        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, jpeg_data.size);
+        index[i].offset = ftell(f) - 8;
+        index[i].size   = 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__

face_detect.py

@@ -0,0 +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)
+
+    

format-code.sh

@@ -1,2 +1,5 @@
-clang-format -style=file -i *.cpp *.h
-clang-format -style=file -i examples/cli/*.cpp
+for f in *.cpp *.h *.hpp examples/cli/*.cpp examples/cli/*.h; do
+  [[ "$f" == vocab* ]] && continue
+  echo "formatting '$f'"
+  clang-format -style=file -i "$f"
+done

gguf_reader.hpp

@@ -0,0 +1,231 @@
+#ifndef __GGUF_READER_HPP__
+#define __GGUF_READER_HPP__
+
+#include <cstdint>
+#include <fstream>
+#include <string>
+#include <vector>
+
+#include "ggml.h"
+#include "util.h"
+
+struct GGUFTensorInfo {
+    std::string name;
+    ggml_type type;
+    std::vector<int64_t> shape;
+    size_t offset;
+};
+
+enum class GGUFMetadataType : uint32_t {
+    UINT8   = 0,
+    INT8    = 1,
+    UINT16  = 2,
+    INT16   = 3,
+    UINT32  = 4,
+    INT32   = 5,
+    FLOAT32 = 6,
+    BOOL    = 7,
+    STRING  = 8,
+    ARRAY   = 9,
+    UINT64  = 10,
+    INT64   = 11,
+    FLOAT64 = 12,
+};
+
+class GGUFReader {
+private:
+    std::vector<GGUFTensorInfo> tensors_;
+    size_t data_offset_;
+    size_t alignment_ = 32;  // default alignment is 32
+
+    template <typename T>
+    bool safe_read(std::ifstream& fin, T& value) {
+        fin.read(reinterpret_cast<char*>(&value), sizeof(T));
+        return fin.good();
+    }
+
+    bool safe_read(std::ifstream& fin, char* buffer, size_t size) {
+        fin.read(buffer, size);
+        return fin.good();
+    }
+
+    bool safe_seek(std::ifstream& fin, std::streamoff offset, std::ios::seekdir dir) {
+        fin.seekg(offset, dir);
+        return fin.good();
+    }
+
+    bool read_metadata(std::ifstream& fin) {
+        uint64_t key_len = 0;
+        if (!safe_read(fin, key_len))
+            return false;
+
+        std::string key(key_len, '\0');
+        if (!safe_read(fin, (char*)key.data(), key_len))
+            return false;
+
+        uint32_t type = 0;
+        if (!safe_read(fin, type))
+            return false;
+
+        if (key == "general.alignment") {
+            uint32_t align_val = 0;
+            if (!safe_read(fin, align_val))
+                return false;
+
+            if (align_val != 0 && (align_val & (align_val - 1)) == 0) {
+                alignment_ = align_val;
+                LOG_DEBUG("Found alignment: %zu", alignment_);
+            } else {
+                LOG_ERROR("Invalid alignment value %u, fallback to default %zu", align_val, alignment_);
+            }
+            return true;
+        }
+
+        switch (static_cast<GGUFMetadataType>(type)) {
+            case GGUFMetadataType::UINT8:
+            case GGUFMetadataType::INT8:
+            case GGUFMetadataType::BOOL:
+                return safe_seek(fin, 1, std::ios::cur);
+
+            case GGUFMetadataType::UINT16:
+            case GGUFMetadataType::INT16:
+                return safe_seek(fin, 2, std::ios::cur);
+
+            case GGUFMetadataType::UINT32:
+            case GGUFMetadataType::INT32:
+            case GGUFMetadataType::FLOAT32:
+                return safe_seek(fin, 4, std::ios::cur);
+
+            case GGUFMetadataType::UINT64:
+            case GGUFMetadataType::INT64:
+            case GGUFMetadataType::FLOAT64:
+                return safe_seek(fin, 8, std::ios::cur);
+
+            case GGUFMetadataType::STRING: {
+                uint64_t len = 0;
+                if (!safe_read(fin, len))
+                    return false;
+                return safe_seek(fin, len, std::ios::cur);
+            }
+
+            case GGUFMetadataType::ARRAY: {
+                uint32_t elem_type = 0;
+                uint64_t len       = 0;
+                if (!safe_read(fin, elem_type))
+                    return false;
+                if (!safe_read(fin, len))
+                    return false;
+
+                for (uint64_t i = 0; i < len; i++) {
+                    if (!read_metadata(fin))
+                        return false;
+                }
+                return true;
+            }
+
+            default:
+                LOG_ERROR("Unknown metadata type=%u", type);
+                return false;
+        }
+    }
+
+    GGUFTensorInfo read_tensor_info(std::ifstream& fin) {
+        GGUFTensorInfo info;
+
+        uint64_t name_len;
+        if (!safe_read(fin, name_len))
+            throw std::runtime_error("read tensor name length failed");
+
+        info.name.resize(name_len);
+        if (!safe_read(fin, (char*)info.name.data(), name_len))
+            throw std::runtime_error("read tensor name failed");
+
+        uint32_t n_dims;
+        if (!safe_read(fin, n_dims))
+            throw std::runtime_error("read tensor dims failed");
+
+        info.shape.resize(n_dims);
+        for (uint32_t i = 0; i < n_dims; i++) {
+            if (!safe_read(fin, info.shape[i]))
+                throw std::runtime_error("read tensor shape failed");
+        }
+
+        if (n_dims > GGML_MAX_DIMS) {
+            for (int i = GGML_MAX_DIMS; i < n_dims; i++) {
+                info.shape[GGML_MAX_DIMS - 1] *= info.shape[i];  // stack to last dim;
+            }
+            info.shape.resize(GGML_MAX_DIMS);
+            n_dims = GGML_MAX_DIMS;
+        }
+
+        uint32_t type;
+        if (!safe_read(fin, type))
+            throw std::runtime_error("read tensor type failed");
+        info.type = static_cast<ggml_type>(type);
+
+        if (!safe_read(fin, info.offset))
+            throw std::runtime_error("read tensor offset failed");
+
+        return info;
+    }
+
+public:
+    bool load(const std::string& file_path) {
+        std::ifstream fin(file_path, std::ios::binary);
+        if (!fin) {
+            LOG_ERROR("failed to open '%s'", file_path.c_str());
+            return false;
+        }
+
+        // --- Header ---
+        char magic[4];
+        if (!safe_read(fin, magic, 4) || strncmp(magic, "GGUF", 4) != 0) {
+            LOG_ERROR("not a valid GGUF file");
+            return false;
+        }
+
+        uint32_t version;
+        if (!safe_read(fin, version))
+            return false;
+
+        uint64_t tensor_count, metadata_kv_count;
+        if (!safe_read(fin, tensor_count))
+            return false;
+        if (!safe_read(fin, metadata_kv_count))
+            return false;
+
+        LOG_DEBUG("GGUF v%u, tensor_count=%llu, metadata_kv_count=%llu",
+                  version, (unsigned long long)tensor_count, (unsigned long long)metadata_kv_count);
+
+        // --- Read Metadata ---
+        for (uint64_t i = 0; i < metadata_kv_count; i++) {
+            if (!read_metadata(fin)) {
+                LOG_ERROR("read meta data failed");
+                return false;
+            }
+        }
+
+        // --- Tensor Infos ---
+        tensors_.clear();
+        try {
+            for (uint64_t i = 0; i < tensor_count; i++) {
+                tensors_.push_back(read_tensor_info(fin));
+            }
+        } catch (const std::runtime_error& e) {
+            LOG_ERROR("%s", e.what());
+            return false;
+        }
+
+        data_offset_ = static_cast<size_t>(fin.tellg());
+        if ((data_offset_ % alignment_) != 0) {
+            data_offset_ = ((data_offset_ + alignment_ - 1) / alignment_) * alignment_;
+        }
+        fin.close();
+        return true;
+    }
+
+    const std::vector<GGUFTensorInfo>& tensors() const { return tensors_; }
+    size_t data_offset() const { return data_offset_; }
+};
+
+#endif  // __GGUF_READER_HPP__

gits_noise.inl

@@ -0,0 +1,349 @@
+#ifndef GITS_NOISE_INL
+#define GITS_NOISE_INL
+
+const std::vector<std::vector<float>> GITS_NOISE_0_80 = {
+    { 14.61464119f, 7.49001646f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 6.77309084f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 3.07277966f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 2.05039096f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 2.05039096f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.19567990f, 1.98035145f, 0.86115354f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.19567990f, 1.98035145f, 0.86115354f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.88507891f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.07277966f, 1.84880662f, 0.83188516f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_0_85 = {
+    { 14.61464119f, 7.49001646f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 1.84880662f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 6.77309084f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.11996698f, 3.07277966f, 1.24153244f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.09240818f, 2.84484982f, 0.95350921f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.09240818f, 2.84484982f, 0.95350921f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.58536053f, 3.19567990f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 8.75849152f, 7.49001646f, 5.58536053f, 3.19567990f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 8.75849152f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 8.75849152f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.60512662f, 2.63833880f, 1.56271636f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.88507891f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_0_90 = {
+    { 14.61464119f, 6.77309084f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 3.07277966f, 0.95350921f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.54230714f, 0.89115214f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 2.54230714f, 0.89115214f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.09240818f, 3.07277966f, 1.61558151f, 0.69515091f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.11996698f, 4.86714602f, 3.07277966f, 1.61558151f, 0.69515091f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 2.95596409f, 1.61558151f, 0.69515091f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.24153244f, 0.57119018f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.24153244f, 0.57119018f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.24153244f, 0.57119018f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.75677586f, 2.84484982f, 1.84880662f, 1.08895338f, 0.52423614f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.75677586f, 2.84484982f, 1.84880662f, 1.08895338f, 0.52423614f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.45427561f, 3.32507086f, 2.45070267f, 1.61558151f, 0.95350921f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.45427561f, 3.32507086f, 2.45070267f, 1.61558151f, 0.95350921f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.86714602f, 3.91689563f, 3.07277966f, 2.27973175f, 1.56271636f, 0.95350921f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.86714602f, 3.91689563f, 3.07277966f, 2.27973175f, 1.56271636f, 0.95350921f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.86714602f, 3.91689563f, 3.07277966f, 2.27973175f, 1.56271636f, 0.95350921f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.60512662f, 2.95596409f, 2.19988537f, 1.51179266f, 0.89115214f, 0.43325692f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_0_95 = {
+    { 14.61464119f, 6.77309084f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 2.84484982f, 0.89115214f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.36326075f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.95596409f, 1.56271636f, 0.64427125f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 2.95596409f, 1.56271636f, 0.64427125f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 3.07277966f, 1.91321158f, 1.08895338f, 0.50118381f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.07277966f, 1.91321158f, 1.08895338f, 0.50118381f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.07277966f, 1.91321158f, 1.08895338f, 0.50118381f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.41535246f, 0.803307f, 0.38853383f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.46139455f, 2.63833880f, 1.84880662f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.46139455f, 2.63833880f, 1.84880662f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 6.14220476f, 4.86714602f, 3.75677586f, 2.95596409f, 2.19988537f, 1.56271636f, 1.05362725f, 0.64427125f, 0.32104823f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 6.44769001f, 5.58536053f, 4.65472794f, 3.60512662f, 2.95596409f, 2.19988537f, 1.56271636f, 1.05362725f, 0.64427125f, 0.32104823f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.65472794f, 3.60512662f, 2.95596409f, 2.19988537f, 1.56271636f, 1.05362725f, 0.64427125f, 0.32104823f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.65472794f, 3.75677586f, 3.07277966f, 2.45070267f, 1.78698075f, 1.24153244f, 0.83188516f, 0.50118381f, 0.22545385f, 0.02916753f },
+    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.60512662f, 2.95596409f, 2.36326075f, 1.72759056f, 1.24153244f, 0.83188516f, 0.50118381f, 0.22545385f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.60512662f, 2.95596409f, 2.36326075f, 1.72759056f, 1.24153244f, 0.83188516f, 0.50118381f, 0.22545385f, 0.02916753f },
+    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.75677586f, 3.07277966f, 2.45070267f, 1.91321158f, 1.46270394f, 1.05362725f, 0.72133851f, 0.43325692f, 0.19894916f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_00 = {
+    { 14.61464119f, 1.56271636f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 0.95350921f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 2.36326075f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 7.11996698f, 3.07277966f, 1.56271636f, 0.59516323f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.41535246f, 0.57119018f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.86115354f, 0.38853383f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.86115354f, 0.38853383f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 3.07277966f, 1.98035145f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.07277966f, 1.98035145f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.27973175f, 1.51179266f, 0.95350921f, 0.54755926f, 0.25053367f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.36326075f, 1.61558151f, 1.08895338f, 0.72133851f, 0.41087446f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.36326075f, 1.61558151f, 1.08895338f, 0.72133851f, 0.41087446f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.12350607f, 1.56271636f, 1.08895338f, 0.72133851f, 0.41087446f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.19988537f, 1.61558151f, 1.162866f, 0.803307f, 0.50118381f, 0.27464288f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.75677586f, 3.07277966f, 2.45070267f, 1.84880662f, 1.36964464f, 1.01931262f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 6.14220476f, 5.09240818f, 4.26497746f, 3.46139455f, 2.84484982f, 2.19988537f, 1.67050016f, 1.24153244f, 0.92192322f, 0.64427125f, 0.43325692f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 6.14220476f, 5.09240818f, 4.26497746f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.12534678f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 5.09240818f, 4.26497746f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.12534678f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 12.23089790f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.26497746f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.12534678f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_05 = {
+    { 14.61464119f, 0.95350921f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 0.89115214f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 2.05039096f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 2.84484982f, 1.28281462f, 0.52423614f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.07277966f, 1.61558151f, 0.803307f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.56271636f, 0.803307f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.95350921f, 0.52423614f, 0.22545385f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 1.98035145f, 1.24153244f, 0.74807048f, 0.41087446f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.27973175f, 1.51179266f, 0.95350921f, 0.59516323f, 0.34370604f, 0.13792117f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.09240818f, 3.46139455f, 2.45070267f, 1.61558151f, 1.08895338f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.09240818f, 3.46139455f, 2.45070267f, 1.61558151f, 1.08895338f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.36326075f, 1.61558151f, 1.08895338f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.45070267f, 1.72759056f, 1.24153244f, 0.86115354f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.19988537f, 1.61558151f, 1.162866f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.19988537f, 1.67050016f, 1.28281462f, 0.95350921f, 0.72133851f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.36326075f, 1.84880662f, 1.41535246f, 1.08895338f, 0.83188516f, 0.61951244f, 0.45573691f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.20157266f, 0.95350921f, 0.74807048f, 0.57119018f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.30717278f, 7.11996698f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.20157266f, 0.95350921f, 0.74807048f, 0.57119018f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 8.30717278f, 7.11996698f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.45070267f, 1.98035145f, 1.61558151f, 1.32549286f, 1.08895338f, 0.86115354f, 0.69515091f, 0.54755926f, 0.41087446f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_10 = {
+    { 14.61464119f, 0.89115214f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 1.61558151f, 0.57119018f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 2.45070267f, 1.08895338f, 0.45573691f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 2.95596409f, 1.56271636f, 0.803307f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.07277966f, 1.61558151f, 0.89115214f, 0.4783645f, 0.19894916f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.07277966f, 1.84880662f, 1.08895338f, 0.64427125f, 0.34370604f, 0.13792117f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.95350921f, 0.54755926f, 0.27464288f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.95596409f, 1.91321158f, 1.24153244f, 0.803307f, 0.4783645f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.05039096f, 1.41535246f, 0.95350921f, 0.64427125f, 0.41087446f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.27973175f, 1.61558151f, 1.12534678f, 0.803307f, 0.54755926f, 0.36617002f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.32507086f, 2.45070267f, 1.72759056f, 1.24153244f, 0.89115214f, 0.64427125f, 0.45573691f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.09240818f, 3.60512662f, 2.84484982f, 2.05039096f, 1.51179266f, 1.08895338f, 0.803307f, 0.59516323f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.09240818f, 3.60512662f, 2.84484982f, 2.12350607f, 1.61558151f, 1.24153244f, 0.95350921f, 0.72133851f, 0.54755926f, 0.41087446f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.45070267f, 1.84880662f, 1.41535246f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.45070267f, 1.91321158f, 1.51179266f, 1.20157266f, 0.95350921f, 0.74807048f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 5.85520077f, 4.45427561f, 3.46139455f, 2.84484982f, 2.19988537f, 1.72759056f, 1.36964464f, 1.08895338f, 0.86115354f, 0.69515091f, 0.54755926f, 0.43325692f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.46139455f, 2.84484982f, 2.19988537f, 1.72759056f, 1.36964464f, 1.08895338f, 0.86115354f, 0.69515091f, 0.54755926f, 0.43325692f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.46139455f, 2.84484982f, 2.19988537f, 1.72759056f, 1.36964464f, 1.08895338f, 0.89115214f, 0.72133851f, 0.59516323f, 0.4783645f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_15 = {
+    { 14.61464119f, 0.83188516f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.59516323f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 1.56271636f, 0.52423614f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 1.91321158f, 0.83188516f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.24153244f, 0.59516323f, 0.25053367f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.51179266f, 0.803307f, 0.41087446f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.56271636f, 0.89115214f, 0.50118381f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.07277966f, 1.84880662f, 1.12534678f, 0.72133851f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.07277966f, 1.91321158f, 1.24153244f, 0.803307f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 2.95596409f, 1.91321158f, 1.24153244f, 0.803307f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.05039096f, 1.36964464f, 0.95350921f, 0.69515091f, 0.4783645f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.803307f, 0.59516323f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.803307f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.19988537f, 1.61558151f, 1.24153244f, 0.95350921f, 0.74807048f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.78698075f, 1.32549286f, 1.01931262f, 0.803307f, 0.64427125f, 0.50118381f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.78698075f, 1.32549286f, 1.01931262f, 0.803307f, 0.64427125f, 0.52423614f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.84880662f, 1.41535246f, 1.12534678f, 0.89115214f, 0.72133851f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.84880662f, 1.41535246f, 1.12534678f, 0.89115214f, 0.72133851f, 0.59516323f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_20 = {
+    { 14.61464119f, 0.803307f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.52423614f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 0.92192322f, 0.36617002f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.24153244f, 0.59516323f, 0.25053367f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.05039096f, 0.95350921f, 0.45573691f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.24153244f, 0.64427125f, 0.29807833f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.36964464f, 0.803307f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 0.95350921f, 0.59516323f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.67050016f, 1.08895338f, 0.74807048f, 0.50118381f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.84880662f, 1.24153244f, 0.83188516f, 0.59516323f, 0.41087446f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 3.07277966f, 1.98035145f, 1.36964464f, 0.95350921f, 0.69515091f, 0.50118381f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.46139455f, 2.36326075f, 1.56271636f, 1.08895338f, 0.803307f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 6.77309084f, 3.46139455f, 2.45070267f, 1.61558151f, 1.162866f, 0.86115354f, 0.64427125f, 0.50118381f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.41087446f, 0.34370604f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.19988537f, 1.61558151f, 1.20157266f, 0.92192322f, 0.72133851f, 0.57119018f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.19988537f, 1.61558151f, 1.24153244f, 0.95350921f, 0.74807048f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.19988537f, 1.61558151f, 1.24153244f, 0.95350921f, 0.74807048f, 0.59516323f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_25 = {
+    { 14.61464119f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.50118381f, 0.02916753f },
+    { 14.61464119f, 2.05039096f, 0.803307f, 0.32104823f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 0.95350921f, 0.43325692f, 0.17026083f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.24153244f, 0.59516323f, 0.27464288f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.51179266f, 0.803307f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.36326075f, 1.24153244f, 0.72133851f, 0.41087446f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.36964464f, 0.83188516f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 0.98595673f, 0.64427125f, 0.43325692f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.67050016f, 1.08895338f, 0.74807048f, 0.52423614f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.72759056f, 1.162866f, 0.803307f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.84880662f, 1.24153244f, 0.86115354f, 0.64427125f, 0.4783645f, 0.36617002f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.84880662f, 1.28281462f, 0.92192322f, 0.69515091f, 0.52423614f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.91321158f, 1.32549286f, 0.95350921f, 0.72133851f, 0.54755926f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.91321158f, 1.32549286f, 0.95350921f, 0.72133851f, 0.57119018f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.95596409f, 1.91321158f, 1.32549286f, 0.95350921f, 0.74807048f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 3.07277966f, 2.05039096f, 1.41535246f, 1.05362725f, 0.803307f, 0.61951244f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 3.07277966f, 2.05039096f, 1.41535246f, 1.05362725f, 0.803307f, 0.64427125f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 3.07277966f, 2.05039096f, 1.46270394f, 1.08895338f, 0.83188516f, 0.66947293f, 0.54755926f, 0.45573691f, 0.38853383f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_30 = {
+    { 14.61464119f, 0.72133851f, 0.02916753f },
+    { 14.61464119f, 1.24153244f, 0.43325692f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.59516323f, 0.22545385f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.803307f, 0.36617002f, 0.13792117f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 1.01931262f, 0.52423614f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.36964464f, 0.74807048f, 0.41087446f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.56271636f, 0.89115214f, 0.54755926f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.61558151f, 0.95350921f, 0.61951244f, 0.41087446f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.36964464f, 0.83188516f, 0.54755926f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.41535246f, 0.92192322f, 0.64427125f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.6383388f, 1.56271636f, 1.01931262f, 0.72133851f, 0.50118381f, 0.36617002f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.05362725f, 0.74807048f, 0.54755926f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.08895338f, 0.77538133f, 0.57119018f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.59516323f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.72759056f, 1.162866f, 0.83188516f, 0.64427125f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.72759056f, 1.162866f, 0.83188516f, 0.64427125f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.78698075f, 1.24153244f, 0.92192322f, 0.72133851f, 0.57119018f, 0.45573691f, 0.38853383f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.84484982f, 1.78698075f, 1.24153244f, 0.92192322f, 0.72133851f, 0.57119018f, 0.4783645f, 0.41087446f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_35 = {
+    { 14.61464119f, 0.69515091f, 0.02916753f },
+    { 14.61464119f, 0.95350921f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.57119018f, 0.19894916f, 0.02916753f },
+    { 14.61464119f, 1.61558151f, 0.69515091f, 0.29807833f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.83188516f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.162866f, 0.64427125f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.36964464f, 0.803307f, 0.50118381f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.41535246f, 0.83188516f, 0.54755926f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 0.95350921f, 0.64427125f, 0.45573691f, 0.32104823f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 0.95350921f, 0.64427125f, 0.45573691f, 0.34370604f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.61558151f, 1.01931262f, 0.72133851f, 0.52423614f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.61558151f, 1.01931262f, 0.72133851f, 0.52423614f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.61558151f, 1.05362725f, 0.74807048f, 0.54755926f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.72759056f, 1.12534678f, 0.803307f, 0.59516323f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 3.07277966f, 1.72759056f, 1.12534678f, 0.803307f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.45070267f, 1.51179266f, 1.01931262f, 0.74807048f, 0.57119018f, 0.45573691f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.6383388f, 1.61558151f, 1.08895338f, 0.803307f, 0.61951244f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.6383388f, 1.61558151f, 1.08895338f, 0.803307f, 0.64427125f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 5.85520077f, 2.6383388f, 1.61558151f, 1.08895338f, 0.803307f, 0.64427125f, 0.52423614f, 0.45573691f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_40 = {
+    { 14.61464119f, 0.59516323f, 0.02916753f },
+    { 14.61464119f, 0.95350921f, 0.34370604f, 0.02916753f },
+    { 14.61464119f, 1.08895338f, 0.43325692f, 0.13792117f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.64427125f, 0.27464288f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.61558151f, 0.803307f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.05039096f, 0.95350921f, 0.54755926f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.24153244f, 0.72133851f, 0.43325692f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.24153244f, 0.74807048f, 0.50118381f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.52423614f, 0.36617002f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.54755926f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.41535246f, 0.86115354f, 0.59516323f, 0.43325692f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.64427125f, 0.45573691f, 0.34370604f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.64427125f, 0.4783645f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 0.98595673f, 0.69515091f, 0.52423614f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 1.01931262f, 0.72133851f, 0.54755926f, 0.43325692f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.61558151f, 1.05362725f, 0.74807048f, 0.57119018f, 0.45573691f, 0.38853383f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.61951244f, 0.50118381f, 0.41087446f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.61951244f, 0.50118381f, 0.43325692f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.64427125f, 0.52423614f, 0.45573691f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_45 = {
+    { 14.61464119f, 0.59516323f, 0.02916753f },
+    { 14.61464119f, 0.803307f, 0.25053367f, 0.02916753f },
+    { 14.61464119f, 0.95350921f, 0.34370604f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.24153244f, 0.54755926f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.72133851f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.61558151f, 0.803307f, 0.45573691f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.91321158f, 0.95350921f, 0.57119018f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.19988537f, 1.08895338f, 0.64427125f, 0.41087446f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.24153244f, 0.74807048f, 0.50118381f, 0.34370604f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.24153244f, 0.74807048f, 0.50118381f, 0.36617002f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.54755926f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.57119018f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.83188516f, 0.59516323f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.28281462f, 0.83188516f, 0.59516323f, 0.45573691f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.69515091f, 0.52423614f, 0.41087446f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.69515091f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 0.98595673f, 0.72133851f, 0.54755926f, 0.45573691f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 1.01931262f, 0.74807048f, 0.57119018f, 0.4783645f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.84484982f, 1.56271636f, 1.01931262f, 0.74807048f, 0.59516323f, 0.50118381f, 0.43325692f, 0.38853383f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<std::vector<float>> GITS_NOISE_1_50 = {
+    { 14.61464119f, 0.54755926f, 0.02916753f },
+    { 14.61464119f, 0.803307f, 0.25053367f, 0.02916753f },
+    { 14.61464119f, 0.86115354f, 0.32104823f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.24153244f, 0.54755926f, 0.25053367f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.56271636f, 0.72133851f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.61558151f, 0.803307f, 0.45573691f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.61558151f, 0.83188516f, 0.52423614f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.95350921f, 0.59516323f, 0.38853383f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.95350921f, 0.59516323f, 0.41087446f, 0.29807833f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 1.84880662f, 0.95350921f, 0.61951244f, 0.43325692f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.19988537f, 1.12534678f, 0.72133851f, 0.50118381f, 0.36617002f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.19988537f, 1.12534678f, 0.72133851f, 0.50118381f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.57119018f, 0.43325692f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.57119018f, 0.43325692f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.59516323f, 0.45573691f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.59516323f, 0.45573691f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.32549286f, 0.86115354f, 0.64427125f, 0.50118381f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.36964464f, 0.92192322f, 0.69515091f, 0.54755926f, 0.45573691f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
+    { 14.61464119f, 2.45070267f, 1.41535246f, 0.95350921f, 0.72133851f, 0.57119018f, 0.4783645f, 0.43325692f, 0.38853383f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
+};
+
+const std::vector<const std::vector<std::vector<float>>*> GITS_NOISE = {
+    &GITS_NOISE_0_80,
+    &GITS_NOISE_0_85,
+    &GITS_NOISE_0_90,
+    &GITS_NOISE_0_95,
+    &GITS_NOISE_1_00,
+    &GITS_NOISE_1_05,
+    &GITS_NOISE_1_10,
+    &GITS_NOISE_1_15,
+    &GITS_NOISE_1_20,
+    &GITS_NOISE_1_25,
+    &GITS_NOISE_1_30,
+    &GITS_NOISE_1_35,
+    &GITS_NOISE_1_40,
+    &GITS_NOISE_1_45,
+    &GITS_NOISE_1_50
+};
+
+#endif // GITS_NOISE_INL

lora.hpp

@@ -0,0 +1,897 @@
+#ifndef __LORA_HPP__
+#define __LORA_HPP__
+
+#include <mutex>
+#include "ggml_extend.hpp"
+
+#define LORA_GRAPH_BASE_SIZE 10240
+
+struct LoraModel : public GGMLRunner {
+    enum lora_t {
+        REGULAR      = 0,
+        DIFFUSERS    = 1,
+        DIFFUSERS_2  = 2,
+        DIFFUSERS_3  = 3,
+        TRANSFORMERS = 4,
+        LORA_TYPE_COUNT
+    };
+
+    const std::string lora_ups[LORA_TYPE_COUNT] = {
+        ".lora_up",
+        "_lora.up",
+        ".lora_B",
+        ".lora.up",
+        ".lora_linear_layer.up",
+    };
+
+    const std::string lora_downs[LORA_TYPE_COUNT] = {
+        ".lora_down",
+        "_lora.down",
+        ".lora_A",
+        ".lora.down",
+        ".lora_linear_layer.down",
+    };
+
+    const std::string lora_pre[LORA_TYPE_COUNT] = {
+        "lora.",
+        "",
+        "",
+        "",
+        "",
+    };
+
+    const std::map<std::string, std::string> alt_names = {
+        // mmdit
+        {"final_layer.adaLN_modulation.1", "norm_out.linear"},
+        {"pos_embed", "pos_embed.proj"},
+        {"final_layer.linear", "proj_out"},
+        {"y_embedder.mlp.0", "time_text_embed.text_embedder.linear_1"},
+        {"y_embedder.mlp.2", "time_text_embed.text_embedder.linear_2"},
+        {"t_embedder.mlp.0", "time_text_embed.timestep_embedder.linear_1"},
+        {"t_embedder.mlp.2", "time_text_embed.timestep_embedder.linear_2"},
+        {"x_block.mlp.fc1", "ff.net.0.proj"},
+        {"x_block.mlp.fc2", "ff.net.2"},
+        {"context_block.mlp.fc1", "ff_context.net.0.proj"},
+        {"context_block.mlp.fc2", "ff_context.net.2"},
+        {"x_block.adaLN_modulation.1", "norm1.linear"},
+        {"context_block.adaLN_modulation.1", "norm1_context.linear"},
+        {"context_block.attn.proj", "attn.to_add_out"},
+        {"x_block.attn.proj", "attn.to_out.0"},
+        {"x_block.attn2.proj", "attn2.to_out.0"},
+        // flux
+        {"img_in", "x_embedder"},
+        // singlestream
+        {"linear2", "proj_out"},
+        {"modulation.lin", "norm.linear"},
+        // doublestream
+        {"txt_attn.proj", "attn.to_add_out"},
+        {"img_attn.proj", "attn.to_out.0"},
+        {"txt_mlp.0", "ff_context.net.0.proj"},
+        {"txt_mlp.2", "ff_context.net.2"},
+        {"img_mlp.0", "ff.net.0.proj"},
+        {"img_mlp.2", "ff.net.2"},
+        {"txt_mod.lin", "norm1_context.linear"},
+        {"img_mod.lin", "norm1.linear"},
+    };
+
+    const std::map<std::string, std::string> qkv_prefixes = {
+        // mmdit
+        {"context_block.attn.qkv", "attn.add_"},  // suffix "_proj"
+        {"x_block.attn.qkv", "attn.to_"},
+        {"x_block.attn2.qkv", "attn2.to_"},
+        // flux
+        // doublestream
+        {"txt_attn.qkv", "attn.add_"},  // suffix "_proj"
+        {"img_attn.qkv", "attn.to_"},
+    };
+    const std::map<std::string, std::string> qkvm_prefixes = {
+        // flux
+        // singlestream
+        {"linear1", ""},
+    };
+
+    const std::string* type_fingerprints = lora_ups;
+
+    float multiplier = 1.0f;
+    std::map<std::string, struct ggml_tensor*> lora_tensors;
+    std::map<ggml_tensor*, ggml_tensor*> original_tensor_to_final_tensor;
+    std::string file_path;
+    ModelLoader model_loader;
+    bool load_failed                = false;
+    bool applied                    = false;
+    std::vector<int> zero_index_vec = {0};
+    ggml_tensor* zero_index         = NULL;
+    enum lora_t type                = REGULAR;
+
+    LoraModel(ggml_backend_t backend,
+              const std::string& file_path = "",
+              const std::string prefix     = "")
+        : file_path(file_path), GGMLRunner(backend, false) {
+        if (!model_loader.init_from_file(file_path, prefix)) {
+            load_failed = true;
+        }
+    }
+
+    std::string get_desc() {
+        return "lora";
+    }
+
+    bool load_from_file(bool filter_tensor, int n_threads) {
+        LOG_INFO("loading LoRA from '%s'", file_path.c_str());
+
+        if (load_failed) {
+            LOG_ERROR("init lora model loader from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        std::unordered_map<std::string, TensorStorage> tensors_to_create;
+        std::mutex lora_mutex;
+        bool dry_run          = true;
+        auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
+            if (dry_run) {
+                const std::string& name = tensor_storage.name;
+
+                if (filter_tensor && !contains(name, "lora")) {
+                    return true;
+                }
+
+                {
+                    std::lock_guard<std::mutex> lock(lora_mutex);
+                    for (int i = 0; i < LORA_TYPE_COUNT; i++) {
+                        if (name.find(type_fingerprints[i]) != std::string::npos) {
+                            type = (lora_t)i;
+                            break;
+                        }
+                    }
+                    tensors_to_create[name] = tensor_storage;
+                }
+            } else {
+                const std::string& name = tensor_storage.name;
+                auto iter               = lora_tensors.find(name);
+                if (iter != lora_tensors.end()) {
+                    *dst_tensor = iter->second;
+                }
+            }
+            return true;
+        };
+
+        model_loader.load_tensors(on_new_tensor_cb, n_threads);
+
+        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;
+        }
+
+        alloc_params_buffer();
+
+        dry_run = false;
+        model_loader.load_tensors(on_new_tensor_cb, n_threads);
+
+        LOG_DEBUG("lora type: \"%s\"/\"%s\"", lora_downs[type].c_str(), lora_ups[type].c_str());
+
+        LOG_DEBUG("finished loaded lora");
+        return true;
+    }
+
+    ggml_tensor* to_f32(ggml_context* ctx, ggml_tensor* a) {
+        auto out = ggml_reshape_1d(ctx, a, ggml_nelements(a));
+        out      = ggml_get_rows(ctx, out, zero_index);
+        out      = ggml_reshape(ctx, out, a);
+        // auto out = ggml_cast(ctx, a, GGML_TYPE_F32);
+        return out;
+    }
+
+    std::vector<std::string> to_lora_keys(std::string blk_name, SDVersion version) {
+        std::vector<std::string> keys;
+        // if (!sd_version_is_sd3(version) || blk_name != "model.diffusion_model.pos_embed") {
+        size_t k_pos = blk_name.find(".weight");
+        if (k_pos == std::string::npos) {
+            return keys;
+        }
+        blk_name = blk_name.substr(0, k_pos);
+        // }
+        keys.push_back(blk_name);
+        keys.push_back("lora." + blk_name);
+        if (sd_version_is_dit(version)) {
+            if (blk_name.find("model.diffusion_model") != std::string::npos) {
+                blk_name.replace(blk_name.find("model.diffusion_model"), sizeof("model.diffusion_model") - 1, "transformer");
+            }
+
+            if (blk_name.find(".single_blocks") != std::string::npos) {
+                blk_name.replace(blk_name.find(".single_blocks"), sizeof(".single_blocks") - 1, ".single_transformer_blocks");
+            }
+            if (blk_name.find(".double_blocks") != std::string::npos) {
+                blk_name.replace(blk_name.find(".double_blocks"), sizeof(".double_blocks") - 1, ".transformer_blocks");
+            }
+
+            if (blk_name.find(".joint_blocks") != std::string::npos) {
+                blk_name.replace(blk_name.find(".joint_blocks"), sizeof(".joint_blocks") - 1, ".transformer_blocks");
+            }
+
+            if (blk_name.find("text_encoders.clip_l") != std::string::npos) {
+                blk_name.replace(blk_name.find("text_encoders.clip_l"), sizeof("text_encoders.clip_l") - 1, "cond_stage_model");
+            }
+
+            for (const auto& item : alt_names) {
+                size_t match = blk_name.find(item.first);
+                if (match != std::string::npos) {
+                    blk_name = blk_name.substr(0, match) + item.second;
+                }
+            }
+            for (const auto& prefix : qkv_prefixes) {
+                size_t match = blk_name.find(prefix.first);
+                if (match != std::string::npos) {
+                    std::string split_blk = "SPLIT|" + blk_name.substr(0, match) + prefix.second;
+                    keys.push_back(split_blk);
+                }
+            }
+            for (const auto& prefix : qkvm_prefixes) {
+                size_t match = blk_name.find(prefix.first);
+                if (match != std::string::npos) {
+                    std::string split_blk = "SPLIT_L|" + blk_name.substr(0, match) + prefix.second;
+                    keys.push_back(split_blk);
+                }
+            }
+            keys.push_back(blk_name);
+        }
+
+        std::vector<std::string> ret;
+        for (std::string& key : keys) {
+            ret.push_back(key);
+            replace_all_chars(key, '.', '_');
+            // fix for some sdxl lora, like lcm-lora-xl
+            if (key == "model_diffusion_model_output_blocks_2_2_conv") {
+                ret.push_back("model_diffusion_model_output_blocks_2_1_conv");
+            }
+            ret.push_back(key);
+        }
+        return ret;
+    }
+
+    struct ggml_cgraph* build_lora_graph(std::map<std::string, struct 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);
+
+        zero_index = ggml_new_tensor_1d(compute_ctx, GGML_TYPE_I32, 1);
+        set_backend_tensor_data(zero_index, zero_index_vec.data());
+        ggml_build_forward_expand(gf, zero_index);
+
+        original_tensor_to_final_tensor.clear();
+
+        std::set<std::string> applied_lora_tensors;
+        for (auto it : model_tensors) {
+            std::string model_tensor_name    = it.first;
+            struct ggml_tensor* model_tensor = model_tensors[it.first];
+
+            std::vector<std::string> keys = to_lora_keys(model_tensor_name, version);
+            bool is_bias                  = ends_with(model_tensor_name, ".bias");
+            if (keys.size() == 0) {
+                if (is_bias) {
+                    keys.push_back(model_tensor_name.substr(0, model_tensor_name.size() - 5));  // remove .bias
+                } else {
+                    continue;
+                }
+            }
+
+            for (auto& key : keys) {
+                bool is_qkv_split = starts_with(key, "SPLIT|");
+                if (is_qkv_split) {
+                    key = key.substr(sizeof("SPLIT|") - 1);
+                }
+                bool is_qkvm_split = starts_with(key, "SPLIT_L|");
+                if (is_qkvm_split) {
+                    key = key.substr(sizeof("SPLIT_L|") - 1);
+                }
+                struct ggml_tensor* updown = NULL;
+                float scale_value          = 1.0f;
+                std::string full_key       = lora_pre[type] + key;
+                if (is_bias) {
+                    if (lora_tensors.find(full_key + ".diff_b") != lora_tensors.end()) {
+                        std::string diff_name = full_key + ".diff_b";
+                        ggml_tensor* diff     = lora_tensors[diff_name];
+                        updown                = to_f32(compute_ctx, diff);
+                        applied_lora_tensors.insert(diff_name);
+                    } else {
+                        continue;
+                    }
+                } else if (lora_tensors.find(full_key + ".diff") != lora_tensors.end()) {
+                    std::string diff_name = full_key + ".diff";
+                    ggml_tensor* diff     = lora_tensors[diff_name];
+                    updown                = to_f32(compute_ctx, diff);
+                    applied_lora_tensors.insert(diff_name);
+                } else if (lora_tensors.find(full_key + ".hada_w1_a") != lora_tensors.end()) {
+                    // LoHa mode
+
+                    // TODO: split qkv convention for LoHas (is it ever used?)
+                    if (is_qkv_split || is_qkvm_split) {
+                        LOG_ERROR("Split qkv isn't supported for LoHa models.");
+                        break;
+                    }
+                    std::string alpha_name = "";
+
+                    ggml_tensor* hada_1_mid  = NULL;  // tau for tucker decomposition
+                    ggml_tensor* hada_1_up   = NULL;
+                    ggml_tensor* hada_1_down = NULL;
+
+                    ggml_tensor* hada_2_mid  = NULL;  // tau for tucker decomposition
+                    ggml_tensor* hada_2_up   = NULL;
+                    ggml_tensor* hada_2_down = NULL;
+
+                    std::string hada_1_mid_name  = "";
+                    std::string hada_1_down_name = "";
+                    std::string hada_1_up_name   = "";
+
+                    std::string hada_2_mid_name  = "";
+                    std::string hada_2_down_name = "";
+                    std::string hada_2_up_name   = "";
+
+                    hada_1_down_name = full_key + ".hada_w1_b";
+                    hada_1_up_name   = full_key + ".hada_w1_a";
+                    hada_1_mid_name  = full_key + ".hada_t1";
+                    if (lora_tensors.find(hada_1_down_name) != lora_tensors.end()) {
+                        hada_1_down = to_f32(compute_ctx, lora_tensors[hada_1_down_name]);
+                    }
+                    if (lora_tensors.find(hada_1_up_name) != lora_tensors.end()) {
+                        hada_1_up = to_f32(compute_ctx, lora_tensors[hada_1_up_name]);
+                    }
+                    if (lora_tensors.find(hada_1_mid_name) != lora_tensors.end()) {
+                        hada_1_mid = to_f32(compute_ctx, lora_tensors[hada_1_mid_name]);
+                        applied_lora_tensors.insert(hada_1_mid_name);
+                        hada_1_up = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, hada_1_up));
+                    }
+
+                    hada_2_down_name = full_key + ".hada_w2_b";
+                    hada_2_up_name   = full_key + ".hada_w2_a";
+                    hada_2_mid_name  = full_key + ".hada_t2";
+                    if (lora_tensors.find(hada_2_down_name) != lora_tensors.end()) {
+                        hada_2_down = to_f32(compute_ctx, lora_tensors[hada_2_down_name]);
+                    }
+                    if (lora_tensors.find(hada_2_up_name) != lora_tensors.end()) {
+                        hada_2_up = to_f32(compute_ctx, lora_tensors[hada_2_up_name]);
+                    }
+                    if (lora_tensors.find(hada_2_mid_name) != lora_tensors.end()) {
+                        hada_2_mid = to_f32(compute_ctx, lora_tensors[hada_2_mid_name]);
+                        applied_lora_tensors.insert(hada_2_mid_name);
+                        hada_2_up = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, hada_2_up));
+                    }
+
+                    alpha_name = full_key + ".alpha";
+
+                    applied_lora_tensors.insert(hada_1_down_name);
+                    applied_lora_tensors.insert(hada_1_up_name);
+                    applied_lora_tensors.insert(hada_2_down_name);
+                    applied_lora_tensors.insert(hada_2_up_name);
+
+                    applied_lora_tensors.insert(alpha_name);
+                    if (hada_1_up == NULL || hada_1_down == NULL || hada_2_up == NULL || hada_2_down == NULL) {
+                        continue;
+                    }
+
+                    struct ggml_tensor* updown_1 = ggml_merge_lora(compute_ctx, hada_1_down, hada_1_up, hada_1_mid);
+                    struct ggml_tensor* updown_2 = ggml_merge_lora(compute_ctx, hada_2_down, hada_2_up, hada_2_mid);
+                    updown                       = ggml_mul_inplace(compute_ctx, updown_1, updown_2);
+
+                    // calc_scale
+                    // TODO: .dora_scale?
+                    int64_t rank = hada_1_down->ne[ggml_n_dims(hada_1_down) - 1];
+                    if (lora_tensors.find(alpha_name) != lora_tensors.end()) {
+                        float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
+                        scale_value = alpha / rank;
+                    }
+                } else if (lora_tensors.find(full_key + ".lokr_w1") != lora_tensors.end() || lora_tensors.find(full_key + ".lokr_w1_a") != lora_tensors.end()) {
+                    // LoKr mode
+
+                    // TODO: split qkv convention for LoKrs (is it ever used?)
+                    if (is_qkv_split || is_qkvm_split) {
+                        LOG_ERROR("Split qkv isn't supported for LoKr models.");
+                        break;
+                    }
+
+                    std::string alpha_name = full_key + ".alpha";
+
+                    ggml_tensor* lokr_w1 = NULL;
+                    ggml_tensor* lokr_w2 = NULL;
+
+                    std::string lokr_w1_name = "";
+                    std::string lokr_w2_name = "";
+
+                    lokr_w1_name = full_key + ".lokr_w1";
+                    lokr_w2_name = full_key + ".lokr_w2";
+
+                    if (lora_tensors.find(lokr_w1_name) != lora_tensors.end()) {
+                        lokr_w1 = to_f32(compute_ctx, lora_tensors[lokr_w1_name]);
+                        applied_lora_tensors.insert(lokr_w1_name);
+                    } else {
+                        ggml_tensor* down     = NULL;
+                        ggml_tensor* up       = NULL;
+                        std::string down_name = lokr_w1_name + "_b";
+                        std::string up_name   = lokr_w1_name + "_a";
+                        if (lora_tensors.find(down_name) != lora_tensors.end()) {
+                            // w1 should not be low rank normally, sometimes w1 and w2 are swapped
+                            down = to_f32(compute_ctx, lora_tensors[down_name]);
+                            applied_lora_tensors.insert(down_name);
+
+                            int64_t rank = down->ne[ggml_n_dims(down) - 1];
+                            if (lora_tensors.find(alpha_name) != lora_tensors.end()) {
+                                float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
+                                scale_value = alpha / rank;
+                            }
+                        }
+                        if (lora_tensors.find(up_name) != lora_tensors.end()) {
+                            up = to_f32(compute_ctx, lora_tensors[up_name]);
+                            applied_lora_tensors.insert(up_name);
+                        }
+                        lokr_w1 = ggml_merge_lora(compute_ctx, down, up);
+                    }
+                    if (lora_tensors.find(lokr_w2_name) != lora_tensors.end()) {
+                        lokr_w2 = to_f32(compute_ctx, lora_tensors[lokr_w2_name]);
+                        applied_lora_tensors.insert(lokr_w2_name);
+                    } else {
+                        ggml_tensor* down     = NULL;
+                        ggml_tensor* up       = NULL;
+                        std::string down_name = lokr_w2_name + "_b";
+                        std::string up_name   = lokr_w2_name + "_a";
+                        if (lora_tensors.find(down_name) != lora_tensors.end()) {
+                            down = to_f32(compute_ctx, lora_tensors[down_name]);
+                            applied_lora_tensors.insert(down_name);
+
+                            int64_t rank = down->ne[ggml_n_dims(down) - 1];
+                            if (lora_tensors.find(alpha_name) != lora_tensors.end()) {
+                                float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
+                                scale_value = alpha / rank;
+                            }
+                        }
+                        if (lora_tensors.find(up_name) != lora_tensors.end()) {
+                            up = to_f32(compute_ctx, lora_tensors[up_name]);
+                            applied_lora_tensors.insert(up_name);
+                        }
+                        lokr_w2 = ggml_merge_lora(compute_ctx, down, up);
+                    }
+
+                    // Technically it might be unused, but I believe it's the expected behavior
+                    applied_lora_tensors.insert(alpha_name);
+
+                    updown = ggml_kronecker(compute_ctx, lokr_w1, lokr_w2);
+
+                } else {
+                    // LoRA mode
+                    ggml_tensor* lora_mid  = NULL;  // tau for tucker decomposition
+                    ggml_tensor* lora_up   = NULL;
+                    ggml_tensor* lora_down = NULL;
+
+                    std::string alpha_name         = "";
+                    std::string scale_name         = "";
+                    std::string split_q_scale_name = "";
+                    std::string lora_mid_name      = "";
+                    std::string lora_down_name     = "";
+                    std::string lora_up_name       = "";
+
+                    if (is_qkv_split) {
+                        std::string suffix  = "";
+                        auto split_q_d_name = full_key + "q" + suffix + lora_downs[type] + ".weight";
+
+                        if (lora_tensors.find(split_q_d_name) == lora_tensors.end()) {
+                            suffix         = "_proj";
+                            split_q_d_name = full_key + "q" + suffix + lora_downs[type] + ".weight";
+                        }
+                        if (lora_tensors.find(split_q_d_name) != lora_tensors.end()) {
+                            // print_ggml_tensor(it.second, true);  //[3072, 21504, 1, 1]
+                            // find qkv and mlp up parts in LoRA model
+                            auto split_k_d_name = full_key + "k" + suffix + lora_downs[type] + ".weight";
+                            auto split_v_d_name = full_key + "v" + suffix + lora_downs[type] + ".weight";
+
+                            auto split_q_u_name = full_key + "q" + suffix + lora_ups[type] + ".weight";
+                            auto split_k_u_name = full_key + "k" + suffix + lora_ups[type] + ".weight";
+                            auto split_v_u_name = full_key + "v" + suffix + lora_ups[type] + ".weight";
+
+                            auto split_q_scale_name = full_key + "q" + suffix + ".scale";
+                            auto split_k_scale_name = full_key + "k" + suffix + ".scale";
+                            auto split_v_scale_name = full_key + "v" + suffix + ".scale";
+
+                            auto split_q_alpha_name = full_key + "q" + suffix + ".alpha";
+                            auto split_k_alpha_name = full_key + "k" + suffix + ".alpha";
+                            auto split_v_alpha_name = full_key + "v" + suffix + ".alpha";
+
+                            ggml_tensor* lora_q_down = NULL;
+                            ggml_tensor* lora_q_up   = NULL;
+                            ggml_tensor* lora_k_down = NULL;
+                            ggml_tensor* lora_k_up   = NULL;
+                            ggml_tensor* lora_v_down = NULL;
+                            ggml_tensor* lora_v_up   = NULL;
+
+                            lora_q_down = to_f32(compute_ctx, lora_tensors[split_q_d_name]);
+
+                            if (lora_tensors.find(split_q_u_name) != lora_tensors.end()) {
+                                lora_q_up = to_f32(compute_ctx, lora_tensors[split_q_u_name]);
+                            }
+
+                            if (lora_tensors.find(split_k_d_name) != lora_tensors.end()) {
+                                lora_k_down = to_f32(compute_ctx, lora_tensors[split_k_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_k_u_name) != lora_tensors.end()) {
+                                lora_k_up = to_f32(compute_ctx, lora_tensors[split_k_u_name]);
+                            }
+
+                            if (lora_tensors.find(split_v_d_name) != lora_tensors.end()) {
+                                lora_v_down = to_f32(compute_ctx, lora_tensors[split_v_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_v_u_name) != lora_tensors.end()) {
+                                lora_v_up = to_f32(compute_ctx, lora_tensors[split_v_u_name]);
+                            }
+
+                            float q_rank = lora_q_up->ne[0];
+                            float k_rank = lora_k_up->ne[0];
+                            float v_rank = lora_v_up->ne[0];
+
+                            float lora_q_scale = 1;
+                            float lora_k_scale = 1;
+                            float lora_v_scale = 1;
+
+                            if (lora_tensors.find(split_q_scale_name) != lora_tensors.end()) {
+                                lora_q_scale = ggml_backend_tensor_get_f32(lora_tensors[split_q_scale_name]);
+                                applied_lora_tensors.insert(split_q_scale_name);
+                            }
+                            if (lora_tensors.find(split_k_scale_name) != lora_tensors.end()) {
+                                lora_k_scale = ggml_backend_tensor_get_f32(lora_tensors[split_k_scale_name]);
+                                applied_lora_tensors.insert(split_k_scale_name);
+                            }
+                            if (lora_tensors.find(split_v_scale_name) != lora_tensors.end()) {
+                                lora_v_scale = ggml_backend_tensor_get_f32(lora_tensors[split_v_scale_name]);
+                                applied_lora_tensors.insert(split_v_scale_name);
+                            }
+
+                            if (lora_tensors.find(split_q_alpha_name) != lora_tensors.end()) {
+                                float lora_q_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_q_alpha_name]);
+                                applied_lora_tensors.insert(split_q_alpha_name);
+                                lora_q_scale = lora_q_alpha / q_rank;
+                            }
+                            if (lora_tensors.find(split_k_alpha_name) != lora_tensors.end()) {
+                                float lora_k_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_k_alpha_name]);
+                                applied_lora_tensors.insert(split_k_alpha_name);
+                                lora_k_scale = lora_k_alpha / k_rank;
+                            }
+                            if (lora_tensors.find(split_v_alpha_name) != lora_tensors.end()) {
+                                float lora_v_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_v_alpha_name]);
+                                applied_lora_tensors.insert(split_v_alpha_name);
+                                lora_v_scale = lora_v_alpha / v_rank;
+                            }
+
+                            ggml_scale_inplace(compute_ctx, lora_q_down, lora_q_scale);
+                            ggml_scale_inplace(compute_ctx, lora_k_down, lora_k_scale);
+                            ggml_scale_inplace(compute_ctx, lora_v_down, lora_v_scale);
+
+                            // print_ggml_tensor(lora_q_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_k_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_v_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_q_up, true);    //[R, 3072, 1, 1]
+                            // print_ggml_tensor(lora_k_up, true);    //[R, 3072, 1, 1]
+                            // print_ggml_tensor(lora_v_up, true);    //[R, 3072, 1, 1]
+
+                            // these need to be stitched together this way:
+                            //                          |q_up,0   ,0   |
+                            //                          |0   ,k_up,0   |
+                            //                          |0   ,0   ,v_up|
+                            // (q_down,k_down,v_down) . (q   ,k   ,v)
+
+                            // up_concat will be [9216, R*3, 1, 1]
+                            // down_concat will be [R*3, 3072, 1, 1]
+                            ggml_tensor* lora_down_concat = ggml_concat(compute_ctx, ggml_concat(compute_ctx, lora_q_down, lora_k_down, 1), lora_v_down, 1);
+
+                            ggml_tensor* z = ggml_dup_tensor(compute_ctx, lora_q_up);
+                            ggml_scale(compute_ctx, z, 0);
+                            ggml_tensor* zz = ggml_concat(compute_ctx, z, z, 1);
+
+                            ggml_tensor* q_up = ggml_concat(compute_ctx, lora_q_up, zz, 1);
+                            ggml_tensor* k_up = ggml_concat(compute_ctx, ggml_concat(compute_ctx, z, lora_k_up, 1), z, 1);
+                            ggml_tensor* v_up = ggml_concat(compute_ctx, zz, lora_v_up, 1);
+                            // print_ggml_tensor(q_up, true);  //[R, 9216, 1, 1]
+                            // print_ggml_tensor(k_up, true);  //[R, 9216, 1, 1]
+                            // print_ggml_tensor(v_up, true);  //[R, 9216, 1, 1]
+                            ggml_tensor* lora_up_concat = ggml_concat(compute_ctx, ggml_concat(compute_ctx, q_up, k_up, 0), v_up, 0);
+                            // print_ggml_tensor(lora_up_concat, true);  //[R*3, 9216, 1, 1]
+
+                            lora_down = ggml_cont(compute_ctx, lora_down_concat);
+                            lora_up   = ggml_cont(compute_ctx, lora_up_concat);
+
+                            applied_lora_tensors.insert(split_q_u_name);
+                            applied_lora_tensors.insert(split_k_u_name);
+                            applied_lora_tensors.insert(split_v_u_name);
+
+                            applied_lora_tensors.insert(split_q_d_name);
+                            applied_lora_tensors.insert(split_k_d_name);
+                            applied_lora_tensors.insert(split_v_d_name);
+                        }
+                    } else if (is_qkvm_split) {
+                        auto split_q_d_name = full_key + "attn.to_q" + lora_downs[type] + ".weight";
+                        if (lora_tensors.find(split_q_d_name) != lora_tensors.end()) {
+                            // print_ggml_tensor(it.second, true);  //[3072, 21504, 1, 1]
+                            // find qkv and mlp up parts in LoRA model
+                            auto split_k_d_name = full_key + "attn.to_k" + lora_downs[type] + ".weight";
+                            auto split_v_d_name = full_key + "attn.to_v" + lora_downs[type] + ".weight";
+
+                            auto split_q_u_name = full_key + "attn.to_q" + lora_ups[type] + ".weight";
+                            auto split_k_u_name = full_key + "attn.to_k" + lora_ups[type] + ".weight";
+                            auto split_v_u_name = full_key + "attn.to_v" + lora_ups[type] + ".weight";
+
+                            auto split_m_d_name = full_key + "proj_mlp" + lora_downs[type] + ".weight";
+                            auto split_m_u_name = full_key + "proj_mlp" + lora_ups[type] + ".weight";
+
+                            auto split_q_scale_name = full_key + "attn.to_q" + ".scale";
+                            auto split_k_scale_name = full_key + "attn.to_k" + ".scale";
+                            auto split_v_scale_name = full_key + "attn.to_v" + ".scale";
+                            auto split_m_scale_name = full_key + "proj_mlp" + ".scale";
+
+                            auto split_q_alpha_name = full_key + "attn.to_q" + ".alpha";
+                            auto split_k_alpha_name = full_key + "attn.to_k" + ".alpha";
+                            auto split_v_alpha_name = full_key + "attn.to_v" + ".alpha";
+                            auto split_m_alpha_name = full_key + "proj_mlp" + ".alpha";
+
+                            ggml_tensor* lora_q_down = NULL;
+                            ggml_tensor* lora_q_up   = NULL;
+                            ggml_tensor* lora_k_down = NULL;
+                            ggml_tensor* lora_k_up   = NULL;
+                            ggml_tensor* lora_v_down = NULL;
+                            ggml_tensor* lora_v_up   = NULL;
+
+                            ggml_tensor* lora_m_down = NULL;
+                            ggml_tensor* lora_m_up   = NULL;
+
+                            lora_q_up = to_f32(compute_ctx, lora_tensors[split_q_u_name]);
+
+                            if (lora_tensors.find(split_q_d_name) != lora_tensors.end()) {
+                                lora_q_down = to_f32(compute_ctx, lora_tensors[split_q_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_q_u_name) != lora_tensors.end()) {
+                                lora_q_up = to_f32(compute_ctx, lora_tensors[split_q_u_name]);
+                            }
+
+                            if (lora_tensors.find(split_k_d_name) != lora_tensors.end()) {
+                                lora_k_down = to_f32(compute_ctx, lora_tensors[split_k_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_k_u_name) != lora_tensors.end()) {
+                                lora_k_up = to_f32(compute_ctx, lora_tensors[split_k_u_name]);
+                            }
+
+                            if (lora_tensors.find(split_v_d_name) != lora_tensors.end()) {
+                                lora_v_down = to_f32(compute_ctx, lora_tensors[split_v_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_v_u_name) != lora_tensors.end()) {
+                                lora_v_up = to_f32(compute_ctx, lora_tensors[split_v_u_name]);
+                            }
+
+                            if (lora_tensors.find(split_m_d_name) != lora_tensors.end()) {
+                                lora_m_down = to_f32(compute_ctx, lora_tensors[split_m_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_m_u_name) != lora_tensors.end()) {
+                                lora_m_up = to_f32(compute_ctx, lora_tensors[split_m_u_name]);
+                            }
+
+                            float q_rank = lora_q_up->ne[0];
+                            float k_rank = lora_k_up->ne[0];
+                            float v_rank = lora_v_up->ne[0];
+                            float m_rank = lora_v_up->ne[0];
+
+                            float lora_q_scale = 1;
+                            float lora_k_scale = 1;
+                            float lora_v_scale = 1;
+                            float lora_m_scale = 1;
+
+                            if (lora_tensors.find(split_q_scale_name) != lora_tensors.end()) {
+                                lora_q_scale = ggml_backend_tensor_get_f32(lora_tensors[split_q_scale_name]);
+                                applied_lora_tensors.insert(split_q_scale_name);
+                            }
+                            if (lora_tensors.find(split_k_scale_name) != lora_tensors.end()) {
+                                lora_k_scale = ggml_backend_tensor_get_f32(lora_tensors[split_k_scale_name]);
+                                applied_lora_tensors.insert(split_k_scale_name);
+                            }
+                            if (lora_tensors.find(split_v_scale_name) != lora_tensors.end()) {
+                                lora_v_scale = ggml_backend_tensor_get_f32(lora_tensors[split_v_scale_name]);
+                                applied_lora_tensors.insert(split_v_scale_name);
+                            }
+                            if (lora_tensors.find(split_m_scale_name) != lora_tensors.end()) {
+                                lora_m_scale = ggml_backend_tensor_get_f32(lora_tensors[split_m_scale_name]);
+                                applied_lora_tensors.insert(split_m_scale_name);
+                            }
+
+                            if (lora_tensors.find(split_q_alpha_name) != lora_tensors.end()) {
+                                float lora_q_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_q_alpha_name]);
+                                applied_lora_tensors.insert(split_q_alpha_name);
+                                lora_q_scale = lora_q_alpha / q_rank;
+                            }
+                            if (lora_tensors.find(split_k_alpha_name) != lora_tensors.end()) {
+                                float lora_k_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_k_alpha_name]);
+                                applied_lora_tensors.insert(split_k_alpha_name);
+                                lora_k_scale = lora_k_alpha / k_rank;
+                            }
+                            if (lora_tensors.find(split_v_alpha_name) != lora_tensors.end()) {
+                                float lora_v_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_v_alpha_name]);
+                                applied_lora_tensors.insert(split_v_alpha_name);
+                                lora_v_scale = lora_v_alpha / v_rank;
+                            }
+                            if (lora_tensors.find(split_m_alpha_name) != lora_tensors.end()) {
+                                float lora_m_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_m_alpha_name]);
+                                applied_lora_tensors.insert(split_m_alpha_name);
+                                lora_m_scale = lora_m_alpha / m_rank;
+                            }
+
+                            ggml_scale_inplace(compute_ctx, lora_q_down, lora_q_scale);
+                            ggml_scale_inplace(compute_ctx, lora_k_down, lora_k_scale);
+                            ggml_scale_inplace(compute_ctx, lora_v_down, lora_v_scale);
+                            ggml_scale_inplace(compute_ctx, lora_m_down, lora_m_scale);
+
+                            // print_ggml_tensor(lora_q_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_k_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_v_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_m_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_q_up, true);  //[R, 3072, 1, 1]
+                            // print_ggml_tensor(lora_k_up, true);  //[R, 3072, 1, 1]
+                            // print_ggml_tensor(lora_v_up, true);  //[R, 3072, 1, 1]
+                            // print_ggml_tensor(lora_m_up, true);  //[R, 12288, 1, 1]
+
+                            // these need to be stitched together this way:
+                            //                                 |q_up,0   ,0   ,0   |
+                            //                                 |0   ,k_up,0   ,0   |
+                            //                                 |0   ,0   ,v_up,0   |
+                            //                                 |0   ,0   ,0   ,m_up|
+                            // (q_down,k_down,v_down,m_down) . (q   ,k   ,v   ,m)
+
+                            // up_concat will be [21504, R*4, 1, 1]
+                            // down_concat will be [R*4, 3072, 1, 1]
+
+                            ggml_tensor* lora_down_concat = ggml_concat(compute_ctx, ggml_concat(compute_ctx, lora_q_down, lora_k_down, 1), ggml_concat(compute_ctx, lora_v_down, lora_m_down, 1), 1);
+                            // print_ggml_tensor(lora_down_concat, true);  //[3072, R*4, 1, 1]
+
+                            // this also means that if rank is bigger than 672, it is less memory efficient to do it this way (should be fine)
+                            // print_ggml_tensor(lora_q_up, true);  //[3072, R, 1, 1]
+                            ggml_tensor* z     = ggml_dup_tensor(compute_ctx, lora_q_up);
+                            ggml_tensor* mlp_z = ggml_dup_tensor(compute_ctx, lora_m_up);
+                            ggml_scale(compute_ctx, z, 0);
+                            ggml_scale(compute_ctx, mlp_z, 0);
+                            ggml_tensor* zz = ggml_concat(compute_ctx, z, z, 1);
+
+                            ggml_tensor* q_up = ggml_concat(compute_ctx, ggml_concat(compute_ctx, lora_q_up, zz, 1), mlp_z, 1);
+                            ggml_tensor* k_up = ggml_concat(compute_ctx, ggml_concat(compute_ctx, z, lora_k_up, 1), ggml_concat(compute_ctx, z, mlp_z, 1), 1);
+                            ggml_tensor* v_up = ggml_concat(compute_ctx, ggml_concat(compute_ctx, zz, lora_v_up, 1), mlp_z, 1);
+                            ggml_tensor* m_up = ggml_concat(compute_ctx, ggml_concat(compute_ctx, zz, z, 1), lora_m_up, 1);
+                            // print_ggml_tensor(q_up, true);  //[R, 21504, 1, 1]
+                            // print_ggml_tensor(k_up, true);  //[R, 21504, 1, 1]
+                            // print_ggml_tensor(v_up, true);  //[R, 21504, 1, 1]
+                            // print_ggml_tensor(m_up, true);  //[R, 21504, 1, 1]
+
+                            ggml_tensor* lora_up_concat = ggml_concat(compute_ctx, ggml_concat(compute_ctx, q_up, k_up, 0), ggml_concat(compute_ctx, v_up, m_up, 0), 0);
+                            // print_ggml_tensor(lora_up_concat, true);  //[R*4, 21504, 1, 1]
+
+                            lora_down = ggml_cont(compute_ctx, lora_down_concat);
+                            lora_up   = ggml_cont(compute_ctx, lora_up_concat);
+
+                            applied_lora_tensors.insert(split_q_u_name);
+                            applied_lora_tensors.insert(split_k_u_name);
+                            applied_lora_tensors.insert(split_v_u_name);
+                            applied_lora_tensors.insert(split_m_u_name);
+
+                            applied_lora_tensors.insert(split_q_d_name);
+                            applied_lora_tensors.insert(split_k_d_name);
+                            applied_lora_tensors.insert(split_v_d_name);
+                            applied_lora_tensors.insert(split_m_d_name);
+                        }
+                    } else {
+                        lora_up_name   = full_key + lora_ups[type] + ".weight";
+                        lora_down_name = full_key + lora_downs[type] + ".weight";
+                        lora_mid_name  = full_key + ".lora_mid.weight";
+
+                        alpha_name = full_key + ".alpha";
+                        scale_name = full_key + ".scale";
+
+                        if (lora_tensors.find(lora_up_name) != lora_tensors.end()) {
+                            lora_up = to_f32(compute_ctx, lora_tensors[lora_up_name]);
+                            applied_lora_tensors.insert(lora_up_name);
+                        }
+
+                        if (lora_tensors.find(lora_down_name) != lora_tensors.end()) {
+                            lora_down = to_f32(compute_ctx, lora_tensors[lora_down_name]);
+                            applied_lora_tensors.insert(lora_down_name);
+                        }
+
+                        if (lora_tensors.find(lora_mid_name) != lora_tensors.end()) {
+                            lora_mid = to_f32(compute_ctx, lora_tensors[lora_mid_name]);
+                            applied_lora_tensors.insert(lora_mid_name);
+                        }
+                    }
+
+                    if (lora_up == NULL || lora_down == NULL) {
+                        continue;
+                    }
+                    // calc_scale
+                    // TODO: .dora_scale?
+                    int64_t rank = lora_down->ne[ggml_n_dims(lora_down) - 1];
+                    if (lora_tensors.find(scale_name) != lora_tensors.end()) {
+                        scale_value = ggml_backend_tensor_get_f32(lora_tensors[scale_name]);
+                        applied_lora_tensors.insert(scale_name);
+                    } else if (lora_tensors.find(alpha_name) != lora_tensors.end()) {
+                        float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
+                        scale_value = alpha / rank;
+                        // LOG_DEBUG("rank %s %ld %.2f %.2f", alpha_name.c_str(), rank, alpha, scale_value);
+                        applied_lora_tensors.insert(alpha_name);
+                    }
+
+                    updown = ggml_merge_lora(compute_ctx, lora_down, lora_up, lora_mid);
+                }
+                scale_value *= multiplier;
+                ggml_tensor* original_tensor = model_tensor;
+                if (!ggml_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
+                    model_tensor = ggml_dup_tensor(compute_ctx, model_tensor);
+                    set_backend_tensor_data(model_tensor, original_tensor->data);
+                }
+                updown = ggml_reshape(compute_ctx, updown, model_tensor);
+                GGML_ASSERT(ggml_nelements(updown) == ggml_nelements(model_tensor));
+                updown = ggml_scale_inplace(compute_ctx, updown, scale_value);
+                ggml_tensor* final_tensor;
+                if (model_tensor->type != GGML_TYPE_F32 && model_tensor->type != GGML_TYPE_F16) {
+                    final_tensor = to_f32(compute_ctx, model_tensor);
+                    final_tensor = ggml_add_inplace(compute_ctx, final_tensor, updown);
+                    final_tensor = ggml_cpy(compute_ctx, final_tensor, model_tensor);
+                } else {
+                    final_tensor = ggml_add_inplace(compute_ctx, model_tensor, updown);
+                }
+                ggml_build_forward_expand(gf, final_tensor);
+                if (!ggml_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
+                    original_tensor_to_final_tensor[original_tensor] = final_tensor;
+                }
+                break;
+            }
+        }
+        size_t total_lora_tensors_count   = 0;
+        size_t applied_lora_tensors_count = 0;
+
+        for (auto& kv : lora_tensors) {
+            total_lora_tensors_count++;
+            if (applied_lora_tensors.find(kv.first) == applied_lora_tensors.end()) {
+                LOG_WARN("unused lora tensor |%s|", kv.first.c_str());
+                print_ggml_tensor(kv.second, true);
+                // exit(0);
+            } else {
+                applied_lora_tensors_count++;
+            }
+        }
+        /* Don't worry if this message shows up twice in the logs per LoRA,
+         * this function is called once to calculate the required buffer size
+         * and then again to actually generate a graph to be used */
+        if (applied_lora_tensors_count != total_lora_tensors_count) {
+            LOG_WARN("Only (%lu / %lu) LoRA tensors will be applied",
+                     applied_lora_tensors_count, total_lora_tensors_count);
+        } else {
+            LOG_DEBUG("(%lu / %lu) LoRA tensors will be applied",
+                      applied_lora_tensors_count, total_lora_tensors_count);
+        }
+
+        return gf;
+    }
+
+    void apply(std::map<std::string, struct ggml_tensor*> model_tensors, SDVersion version, int n_threads) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_lora_graph(model_tensors, version);
+        };
+        GGMLRunner::compute(get_graph, n_threads, false);
+        for (auto item : original_tensor_to_final_tensor) {
+            ggml_tensor* original_tensor = item.first;
+            ggml_tensor* final_tensor    = item.second;
+
+            ggml_backend_tensor_copy(final_tensor, original_tensor);
+        }
+        original_tensor_to_final_tensor.clear();
+        GGMLRunner::free_compute_buffer();
+    }
+};
+
+#endif  // __LORA_HPP__

ltxv.hpp

@@ -0,0 +1,74 @@
+#ifndef __LTXV_HPP__
+#define __LTXV_HPP__
+
+#include "common.hpp"
+#include "ggml_extend.hpp"
+
+namespace LTXV {
+
+    class CausalConv3d : public GGMLBlock {
+    protected:
+        int time_kernel_size;
+
+    public:
+        CausalConv3d(int64_t in_channels,
+                     int64_t out_channels,
+                     int kernel_size        = 3,
+                     std::tuple<int> stride = {1, 1, 1},
+                     int dilation           = 1,
+                     bool bias              = true) {
+            time_kernel_size = kernel_size / 2;
+            blocks["conv"]   = std::shared_ptr<GGMLBlock>(new Conv3d(in_channels,
+                                                                     out_channels,
+                                                                     {kernel_size, kernel_size, kernel_size},
+                                                                     stride,
+                                                                     {0, kernel_size / 2, kernel_size / 2},
+                                                                     {dilation, 1, 1},
+                                                                     bias));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct 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"]);
+            if (causal) {
+                auto h               = ggml_cont(ctx, ggml_permute(ctx, x, 0, 1, 3, 2));                                                  // [ID, N*IC, IH, IW]
+                auto first_frame     = ggml_view_3d(ctx, h, h->ne[0], h->ne[1], h->ne[2], h->nb[1], h->nb[2], 0);                         // [N*IC, IH, IW]
+                first_frame          = ggml_reshape_4d(ctx, first_frame, first_frame->ne[0], first_frame->ne[1], 1, first_frame->ne[2]);  // [N*IC, 1, IH, IW]
+                auto first_frame_pad = first_frame;
+                for (int i = 1; i < time_kernel_size - 1; i++) {
+                    first_frame_pad = ggml_concat(ctx, first_frame_pad, first_frame, 2);
+                }
+                x = ggml_concat(ctx, first_frame_pad, x, 2);
+            } else {
+                auto h         = ggml_cont(ctx, ggml_permute(ctx, x, 0, 1, 3, 2));  // [ID, N*IC, IH, IW]
+                int64_t offset = h->nb[2] * h->ne[2];
+
+                auto first_frame     = ggml_view_3d(ctx, h, h->ne[0], h->ne[1], h->ne[2], h->nb[1], h->nb[2], 0);                         // [N*IC, IH, IW]
+                first_frame          = ggml_reshape_4d(ctx, first_frame, first_frame->ne[0], first_frame->ne[1], 1, first_frame->ne[2]);  // [N*IC, 1, IH, IW]
+                auto first_frame_pad = first_frame;
+                for (int i = 1; i < (time_kernel_size - 1) / 2; i++) {
+                    first_frame_pad = ggml_concat(ctx, first_frame_pad, first_frame, 2);
+                }
+
+                auto last_frame     = ggml_view_3d(ctx, h, h->ne[0], h->ne[1], h->ne[2], h->nb[1], h->nb[2], offset * (h->ne[3] - 1));  // [N*IC, IH, IW]
+                last_frame          = ggml_reshape_4d(ctx, last_frame, last_frame->ne[0], last_frame->ne[1], 1, last_frame->ne[2]);     // [N*IC, 1, IH, IW]
+                auto last_frame_pad = last_frame;
+                for (int i = 1; i < (time_kernel_size - 1) / 2; i++) {
+                    last_frame_pad = ggml_concat(ctx, last_frame_pad, last_frame, 2);
+                }
+
+                x = ggml_concat(ctx, first_frame_pad, x, 2);
+                x = ggml_concat(ctx, x, last_frame_pad, 2);
+            }
+
+            x = conv->forward(ctx, x);
+            return x;
+        }
+    };
+
+};
+
+#endif

model.h

@@ -4,35 +4,143 @@
 #include <functional>
 #include <map>
 #include <memory>
+#include <set>
+#include <sstream>
 #include <string>
+#include <tuple>
 #include <vector>
 
-#include "ggml/ggml-backend.h"
-#include "ggml/ggml.h"
+#include "ggml-backend.h"
+#include "ggml.h"
+#include "gguf.h"
 #include "json.hpp"
 #include "zip.h"
 
+#define SD_MAX_DIMS 5
+
 enum SDVersion {
-    VERSION_1_x,
-    VERSION_2_x,
-    VERSION_XL,
+    VERSION_SD1,
+    VERSION_SD1_INPAINT,
+    VERSION_SD1_PIX2PIX,
+    VERSION_SD2,
+    VERSION_SD2_INPAINT,
+    VERSION_SDXL,
+    VERSION_SDXL_INPAINT,
+    VERSION_SDXL_PIX2PIX,
+    VERSION_SVD,
+    VERSION_SD3,
+    VERSION_FLUX,
+    VERSION_FLUX_FILL,
+    VERSION_FLUX_CONTROLS,
+    VERSION_FLEX_2,
+    VERSION_WAN2,
+    VERSION_WAN2_2_I2V,
+    VERSION_WAN2_2_TI2V,
+    VERSION_QWEN_IMAGE,
     VERSION_COUNT,
 };
 
+static inline bool sd_version_is_sd1(SDVersion version) {
+    if (version == VERSION_SD1 || version == VERSION_SD1_INPAINT || version == VERSION_SD1_PIX2PIX) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_sd2(SDVersion version) {
+    if (version == VERSION_SD2 || version == VERSION_SD2_INPAINT) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_sdxl(SDVersion version) {
+    if (version == VERSION_SDXL || version == VERSION_SDXL_INPAINT || version == VERSION_SDXL_PIX2PIX) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_sd3(SDVersion version) {
+    if (version == VERSION_SD3) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_flux(SDVersion version) {
+    if (version == VERSION_FLUX || version == VERSION_FLUX_FILL || version == VERSION_FLUX_CONTROLS || version == VERSION_FLEX_2) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_wan(SDVersion version) {
+    if (version == VERSION_WAN2 || version == VERSION_WAN2_2_I2V || version == VERSION_WAN2_2_TI2V) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_qwen_image(SDVersion version) {
+    if (version == VERSION_QWEN_IMAGE) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_inpaint(SDVersion version) {
+    if (version == VERSION_SD1_INPAINT || version == VERSION_SD2_INPAINT || version == VERSION_SDXL_INPAINT || version == VERSION_FLUX_FILL || version == VERSION_FLEX_2) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_dit(SDVersion version) {
+    if (sd_version_is_flux(version) ||
+        sd_version_is_sd3(version) ||
+        sd_version_is_wan(version) ||
+        sd_version_is_qwen_image(version)) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_unet_edit(SDVersion version) {
+    return version == VERSION_SD1_PIX2PIX || version == VERSION_SDXL_PIX2PIX;
+}
+
+static inline bool sd_version_is_control(SDVersion version) {
+    return version == VERSION_FLUX_CONTROLS || version == VERSION_FLEX_2;
+}
+
+static bool sd_version_is_inpaint_or_unet_edit(SDVersion version) {
+    return sd_version_is_unet_edit(version) || sd_version_is_inpaint(version) || sd_version_is_control(version);
+}
+
+enum PMVersion {
+    PM_VERSION_1,
+    PM_VERSION_2,
+};
+
 struct TensorStorage {
     std::string name;
-    ggml_type type = GGML_TYPE_F32;
-    bool is_bf16   = false;
-    int64_t ne[4]  = {1, 1, 1, 1};
-    int n_dims     = 0;
+    ggml_type type          = GGML_TYPE_F32;
+    bool is_bf16            = false;
+    bool is_f8_e4m3         = false;
+    bool is_f8_e5m2         = false;
+    bool is_f64             = false;
+    bool is_i64             = false;
+    int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
+    int n_dims              = 0;
 
     size_t file_index = 0;
     int index_in_zip  = -1;  // >= means stored in a zip file
-    size_t offset     = 0;   // offset in file
+    uint64_t offset   = 0;   // offset in file
 
     TensorStorage() = default;
 
-    TensorStorage(const std::string& name, ggml_type type, int64_t* ne, int n_dims, size_t file_index, size_t offset = 0)
+    TensorStorage(const std::string& name, ggml_type type, const int64_t* ne, int n_dims, size_t file_index, size_t offset = 0)
         : name(name), type(type), n_dims(n_dims), file_index(file_index), offset(offset) {
         for (int i = 0; i < n_dims; i++) {
             this->ne[i] = ne[i];
@@ -40,7 +148,11 @@ struct TensorStorage {
     }
 
     int64_t nelements() const {
-        return ne[0] * ne[1] * ne[2] * ne[3];
+        int64_t n = 1;
+        for (int i = 0; i < SD_MAX_DIMS; i++) {
+            n *= ne[i];
+        }
+        return n;
     }
 
     int64_t nbytes() const {
@@ -48,8 +160,10 @@ struct TensorStorage {
     }
 
     int64_t nbytes_to_read() const {
-        if (is_bf16) {
+        if (is_bf16 || is_f8_e4m3 || is_f8_e5m2) {
             return nbytes() / 2;
+        } else if (is_f64 || is_i64) {
+            return nbytes() * 2;
         } else {
             return nbytes();
         }
@@ -67,9 +181,10 @@ struct TensorStorage {
 
     std::vector<TensorStorage> chunk(size_t n) {
         std::vector<TensorStorage> chunks;
-        size_t chunk_size = nbytes_to_read() / n;
+        uint64_t chunk_size = nbytes_to_read() / n;
+        // printf("%d/%d\n", chunk_size, nbytes_to_read());
         reverse_ne();
-        for (int i = 0; i < n; i++) {
+        for (size_t i = 0; i < n; i++) {
             TensorStorage chunk_i = *this;
             chunk_i.ne[0]         = ne[0] / n;
             chunk_i.offset        = offset + i * chunk_size;
@@ -81,7 +196,7 @@ struct TensorStorage {
     }
 
     void reverse_ne() {
-        int64_t new_ne[4] = {1, 1, 1, 1};
+        int64_t new_ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
         for (int i = 0; i < n_dims; i++) {
             new_ne[i] = ne[n_dims - 1 - i];
         }
@@ -89,10 +204,37 @@ struct TensorStorage {
             ne[i] = new_ne[i];
         }
     }
+
+    std::string to_string() const {
+        std::stringstream ss;
+        const char* type_name = ggml_type_name(type);
+        if (is_bf16) {
+            type_name = "bf16";
+        } else if (is_f8_e4m3) {
+            type_name = "f8_e4m3";
+        } else if (is_f8_e5m2) {
+            type_name = "f8_e5m2";
+        } else if (is_f64) {
+            type_name = "f64";
+        } else if (is_i64) {
+            type_name = "i64";
+        }
+        ss << name << " | " << type_name << " | ";
+        ss << n_dims << " [";
+        for (int i = 0; i < SD_MAX_DIMS; i++) {
+            ss << ne[i];
+            if (i != SD_MAX_DIMS - 1) {
+                ss << ", ";
+            }
+        }
+        ss << "]";
+        return ss.str();
+    }
 };
 
 typedef std::function<bool(const TensorStorage&, ggml_tensor**)> on_new_tensor_cb_t;
-typedef std::function<void(const std::string&, int32_t)> on_new_token_cb_t;
+
+typedef std::map<std::string, enum ggml_type> String2GGMLType;
 
 class ModelLoader {
 protected:
@@ -104,7 +246,7 @@ protected:
                         zip_t* zip,
                         std::string dir,
                         size_t file_index,
-                        const std::string& prefix);
+                        const std::string prefix);
 
     bool init_from_gguf_file(const std::string& file_path, const std::string& prefix = "");
     bool init_from_safetensors_file(const std::string& file_path, const std::string& prefix = "");
@@ -112,12 +254,38 @@ protected:
     bool init_from_diffusers_file(const std::string& file_path, const std::string& prefix = "");
 
 public:
+    String2GGMLType tensor_storages_types;
+
     bool init_from_file(const std::string& file_path, const std::string& prefix = "");
+    bool model_is_unet();
     SDVersion get_sd_version();
     ggml_type get_sd_wtype();
-    std::string load_merges();
-    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend_t backend);
-    int64_t cal_mem_size(ggml_backend_t backend);
+    ggml_type get_conditioner_wtype();
+    ggml_type get_diffusion_model_wtype();
+    ggml_type get_vae_wtype();
+    void set_wtype_override(ggml_type wtype, std::string prefix = "");
+    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads = 0);
+    bool load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
+                      std::set<std::string> ignore_tensors = {},
+                      int n_threads                        = 0);
+
+    std::vector<std::string> get_tensor_names() const {
+        std::vector<std::string> names;
+        for (const auto& ts : tensor_storages) {
+            names.push_back(ts.name);
+        }
+        return names;
+    }
+
+    bool save_to_gguf_file(const std::string& file_path, ggml_type type, const std::string& tensor_type_rules);
+    bool tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type);
+    int64_t get_params_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);
     ~ModelLoader() = default;
+
+    static std::string load_merges();
+    static std::string load_qwen2_merges();
+    static std::string load_t5_tokenizer_json();
+    static std::string load_umt5_tokenizer_json();
 };
-#endif  // __MODEL_H__
+
+#endif  // __MODEL_H__

pmid.hpp

@@ -0,0 +1,645 @@
+#ifndef __PMI_HPP__
+#define __PMI_HPP__
+
+#include "ggml_extend.hpp"
+
+#include "clip.hpp"
+#include "lora.hpp"
+
+struct FuseBlock : public GGMLBlock {
+    // network hparams
+    int in_dim;
+    int out_dim;
+    int hidden_dim;
+    bool use_residue;
+
+public:
+    FuseBlock(int i_d, int o_d, int h_d, bool use_residue = true)
+        : in_dim(i_d), out_dim(o_d), hidden_dim(h_d), use_residue(use_residue) {
+        blocks["fc1"]       = std::shared_ptr<GGMLBlock>(new Linear(in_dim, hidden_dim, true));
+        blocks["fc2"]       = std::shared_ptr<GGMLBlock>(new Linear(hidden_dim, out_dim, true));
+        blocks["layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(in_dim));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct 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;
+        // x = ggml_nn_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, x);
+        x = fc2->forward(ctx, x);
+        // x = ggml_add(ctx, ggml_mul_mat(ctx, fc2_w, x),  fc2_b);
+        if (use_residue)
+            x = ggml_add(ctx, x, r);
+        return x;
+    }
+};
+
+struct PMFeedForward : public GGMLBlock {
+    // network hparams
+    int dim;
+
+public:
+    PMFeedForward(int d, int multi = 4)
+        : dim(d) {
+        int inner_dim = dim * multi;
+        blocks["0"]   = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["1"]   = std::shared_ptr<GGMLBlock>(new Mlp(dim, inner_dim, dim, false));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x) {
+        auto norm = std::dynamic_pointer_cast<LayerNorm>(blocks["0"]);
+        auto ff   = std::dynamic_pointer_cast<Mlp>(blocks["1"]);
+
+        x = norm->forward(ctx, x);
+        x = ff->forward(ctx, x);
+        return x;
+    }
+};
+
+struct PerceiverAttention : public GGMLBlock {
+    // network hparams
+    float scale;   // = dim_head**-0.5
+    int dim_head;  // = dim_head
+    int heads;     // = heads
+public:
+    PerceiverAttention(int dim, int dim_h = 64, int h = 8)
+        : scale(powf(dim_h, -0.5)), dim_head(dim_h), heads(h) {
+        int inner_dim    = dim_head * heads;
+        blocks["norm1"]  = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["norm2"]  = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["to_q"]   = std::shared_ptr<GGMLBlock>(new Linear(dim, inner_dim, false));
+        blocks["to_kv"]  = std::shared_ptr<GGMLBlock>(new Linear(dim, inner_dim * 2, false));
+        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) {
+        int64_t ne[4];
+        for (int i = 0; i < 4; ++i)
+            ne[i] = x->ne[i];
+        x = ggml_reshape_4d(ctx, x, x->ne[0] / heads, heads, x->ne[1], x->ne[2]);
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));
+        return x;
+    }
+
+    std::vector<struct ggml_tensor*> chunk_half(struct ggml_context* ctx,
+                                                struct 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(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* latents) {
+        // x (torch.Tensor): image features
+        //     shape (b, n1, D)
+        // latent (torch.Tensor): latent features
+        //     shape (b, n2, D)
+        int64_t ne[4];
+        for (int i = 0; i < 4; ++i)
+            ne[i] = latents->ne[i];
+
+        auto norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
+        auto norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm2"]);
+        x          = norm1->forward(ctx, x);
+        latents    = norm2->forward(ctx, latents);
+        auto to_q  = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
+        auto q     = to_q->forward(ctx, latents);
+
+        auto kv_input = ggml_concat(ctx, x, latents, 1);
+        auto to_kv    = std::dynamic_pointer_cast<Linear>(blocks["to_kv"]);
+        auto kv       = to_kv->forward(ctx, kv_input);
+        auto k        = ggml_view_4d(ctx, kv, kv->ne[0] / 2, kv->ne[1], kv->ne[2], kv->ne[3], kv->nb[1] / 2, kv->nb[2] / 2, kv->nb[3] / 2, 0);
+        auto v        = ggml_view_4d(ctx, kv, kv->ne[0] / 2, kv->ne[1], kv->ne[2], kv->ne[3], kv->nb[1] / 2, kv->nb[2] / 2, kv->nb[3] / 2, kv->nb[0] * (kv->ne[0] / 2));
+        k             = ggml_cont(ctx, k);
+        v             = ggml_cont(ctx, v);
+        q             = reshape_tensor(ctx, q, heads);
+        k             = reshape_tensor(ctx, k, heads);
+        v             = reshape_tensor(ctx, v, heads);
+        scale         = 1.f / sqrt(sqrt((float)dim_head));
+        k             = ggml_scale_inplace(ctx, k, scale);
+        q             = ggml_scale_inplace(ctx, q, scale);
+        // auto weight = ggml_mul_mat(ctx, q, k);
+        auto weight = ggml_mul_mat(ctx, k, q);  // NOTE order of mul is opposite to pytorch
+
+        // GGML's softmax() is equivalent to pytorch's softmax(x, dim=-1)
+        // in this case, dimension along which Softmax will be computed is the last dim
+        // in torch and the first dim in GGML, consistent with the convention that pytorch's
+        // last dimension (varying most rapidly) corresponds to GGML's first (varying most rapidly).
+        // weight = ggml_soft_max(ctx, weight);
+        weight = ggml_soft_max_inplace(ctx, weight);
+        v      = ggml_cont(ctx, ggml_transpose(ctx, v));
+        // auto out = ggml_mul_mat(ctx, weight, v);
+        auto out    = ggml_mul_mat(ctx, v, weight);  // NOTE order of mul is opposite to pytorch
+        out         = ggml_cont(ctx, ggml_permute(ctx, out, 0, 2, 1, 3));
+        out         = ggml_reshape_3d(ctx, out, ne[0], ne[1], ggml_nelements(out) / (ne[0] * ne[1]));
+        auto to_out = std::dynamic_pointer_cast<Linear>(blocks["to_out"]);
+        out         = to_out->forward(ctx, out);
+        return out;
+    }
+};
+
+struct FacePerceiverResampler : public GGMLBlock {
+    // network hparams
+    int depth;
+
+public:
+    FacePerceiverResampler(int dim           = 768,
+                           int d             = 4,
+                           int dim_head      = 64,
+                           int heads         = 16,
+                           int embedding_dim = 1280,
+                           int output_dim    = 768,
+                           int ff_mult       = 4)
+        : depth(d) {
+        blocks["proj_in"]  = std::shared_ptr<GGMLBlock>(new Linear(embedding_dim, dim, true));
+        blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(dim, output_dim, true));
+        blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new LayerNorm(output_dim));
+
+        for (int i = 0; i < depth; i++) {
+            std::string name = "layers." + std::to_string(i) + ".0";
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new PerceiverAttention(dim, dim_head, heads));
+            name             = "layers." + std::to_string(i) + ".1";
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new PMFeedForward(dim, ff_mult));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* latents,
+                                struct 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"]);
+        auto norm_out = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_out"]);
+
+        x = proj_in->forward(ctx, x);
+        for (int i = 0; i < depth; i++) {
+            std::string name = "layers." + std::to_string(i) + ".0";
+            auto attn        = std::dynamic_pointer_cast<PerceiverAttention>(blocks[name]);
+            name             = "layers." + std::to_string(i) + ".1";
+            auto ff          = std::dynamic_pointer_cast<PMFeedForward>(blocks[name]);
+            auto t           = attn->forward(ctx, x, latents);
+            latents          = ggml_add(ctx, t, latents);
+            t                = ff->forward(ctx, latents);
+            latents          = ggml_add(ctx, t, latents);
+        }
+        latents = proj_out->forward(ctx, latents);
+        latents = norm_out->forward(ctx, latents);
+        return latents;
+    }
+};
+
+struct QFormerPerceiver : public GGMLBlock {
+    // network hparams
+    int num_tokens;
+    int cross_attention_dim;
+    bool use_residul;
+
+public:
+    QFormerPerceiver(int id_embeddings_dim, int cross_attention_d, int num_t, int embedding_dim = 1024, bool use_r = true, int ratio = 4)
+        : cross_attention_dim(cross_attention_d), num_tokens(num_t), use_residul(use_r) {
+        blocks["token_proj"]          = std::shared_ptr<GGMLBlock>(new Mlp(id_embeddings_dim,
+                                                                           id_embeddings_dim * ratio,
+                                                                           cross_attention_dim * num_tokens,
+                                                                           true));
+        blocks["token_norm"]          = std::shared_ptr<GGMLBlock>(new LayerNorm(cross_attention_d));
+        blocks["perceiver_resampler"] = std::shared_ptr<GGMLBlock>(new FacePerceiverResampler(
+            cross_attention_dim,
+            4,
+            128,
+            cross_attention_dim / 128,
+            embedding_dim,
+            cross_attention_dim,
+            4));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct 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, 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);
+        if (use_residul)
+            out = ggml_add(ctx, x, out);
+        return out;
+    }
+};
+
+struct FuseModule : public GGMLBlock {
+    // network hparams
+    int embed_dim;
+
+public:
+    FuseModule(int imb_d)
+        : embed_dim(imb_d) {
+        blocks["mlp1"]       = std::shared_ptr<GGMLBlock>(new FuseBlock(imb_d * 2, imb_d, imb_d, false));
+        blocks["mlp2"]       = std::shared_ptr<GGMLBlock>(new FuseBlock(imb_d, imb_d, imb_d, true));
+        blocks["layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(embed_dim));
+    }
+
+    struct ggml_tensor* fuse_fn(struct ggml_context* ctx,
+                                struct ggml_tensor* prompt_embeds,
+                                struct 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"]);
+
+        auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds, id_embeds, 0);
+
+        stacked_id_embeds = mlp1->forward(ctx, stacked_id_embeds);
+        stacked_id_embeds = ggml_add(ctx, stacked_id_embeds, prompt_embeds);
+        stacked_id_embeds = mlp2->forward(ctx, stacked_id_embeds);
+        stacked_id_embeds = layer_norm->forward(ctx, stacked_id_embeds);
+
+        return stacked_id_embeds;
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* 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) {
+        // x: [N, channels, h, w]
+
+        struct 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, prompt_embeds, class_tokens_mask_pos);
+        ggml_set_name(image_token_embeds, "image_token_embeds");
+        valid_id_embeds                       = ggml_reshape_2d(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);
+
+        if (left && right) {
+            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 1);
+            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 1);
+        } else if (left) {
+            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 1);
+        } else if (right) {
+            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 1);
+        }
+
+        class_tokens_mask                         = ggml_cont(ctx, ggml_transpose(ctx, class_tokens_mask));
+        class_tokens_mask                         = ggml_repeat(ctx, class_tokens_mask, prompt_embeds);
+        prompt_embeds                             = ggml_mul(ctx, prompt_embeds, class_tokens_mask);
+        struct ggml_tensor* updated_prompt_embeds = ggml_add(ctx, prompt_embeds, stacked_id_embeds);
+        ggml_set_name(updated_prompt_embeds, "updated_prompt_embeds");
+        return updated_prompt_embeds;
+    }
+};
+
+struct PhotoMakerIDEncoderBlock : public CLIPVisionModelProjection {
+    PhotoMakerIDEncoderBlock()
+        : CLIPVisionModelProjection(OPENAI_CLIP_VIT_L_14) {
+        blocks["visual_projection_2"] = std::shared_ptr<GGMLBlock>(new Linear(1024, 1280, false));
+        blocks["fuse_module"]         = std::shared_ptr<GGMLBlock>(new FuseModule(2048));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                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) {
+        // 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, backend, 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]
+
+        id_embeds   = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 2, 0, 1, 3));
+        id_embeds_2 = ggml_cont(ctx, ggml_permute(ctx, id_embeds_2, 2, 0, 1, 3));
+
+        id_embeds = ggml_concat(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_permute(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);
+        return updated_prompt_embeds;
+    }
+};
+
+struct PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock : public CLIPVisionModelProjection {
+    int cross_attention_dim;
+    int num_tokens;
+
+    PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock(int id_embeddings_dim = 512)
+        : CLIPVisionModelProjection(OPENAI_CLIP_VIT_L_14),
+          cross_attention_dim(2048),
+          num_tokens(2) {
+        blocks["visual_projection_2"] = std::shared_ptr<GGMLBlock>(new Linear(1024, 1280, false));
+        blocks["fuse_module"]         = std::shared_ptr<GGMLBlock>(new FuseModule(2048));
+        blocks["qformer_perceiver"]   = std::shared_ptr<GGMLBlock>(new QFormerPerceiver(id_embeddings_dim,
+                                                                                        cross_attention_dim,
+                                                                                        num_tokens));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                ggml_backend_t backend,
+                                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) {
+        // 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, backend, 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);
+        return updated_prompt_embeds;
+    }
+};
+
+struct PhotoMakerIDEncoder : public GGMLRunner {
+public:
+    SDVersion version    = VERSION_SDXL;
+    PMVersion pm_version = PM_VERSION_1;
+    PhotoMakerIDEncoderBlock id_encoder;
+    PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock id_encoder2;
+    float style_strength;
+
+    std::vector<float> ctm;
+    std::vector<ggml_fp16_t> ctmf16;
+    std::vector<int> ctmpos;
+
+    std::vector<ggml_fp16_t> zeros_left_16;
+    std::vector<float> zeros_left;
+    std::vector<ggml_fp16_t> zeros_right_16;
+    std::vector<float> zeros_right;
+
+public:
+    PhotoMakerIDEncoder(ggml_backend_t backend,
+                        bool offload_params_to_cpu,
+                        const String2GGMLType& tensor_types,
+                        const std::string prefix,
+                        SDVersion version = VERSION_SDXL,
+                        PMVersion pm_v    = PM_VERSION_1,
+                        float sty         = 20.f)
+        : GGMLRunner(backend, offload_params_to_cpu),
+          version(version),
+          pm_version(pm_v),
+          style_strength(sty) {
+        if (pm_version == PM_VERSION_1) {
+            id_encoder.init(params_ctx, tensor_types, prefix);
+        } else if (pm_version == PM_VERSION_2) {
+            id_encoder2.init(params_ctx, tensor_types, prefix);
+        }
+    }
+
+    std::string get_desc() {
+        return "pmid";
+    }
+
+    PMVersion get_version() const {
+        return pm_version;
+    }
+
+    void get_param_tensors(std::map<std::string, struct 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) {
+        ctm.clear();
+        ctmf16.clear();
+        ctmpos.clear();
+        zeros_left.clear();
+        zeros_left_16.clear();
+        zeros_right.clear();
+        zeros_right_16.clear();
+
+        ggml_context* ctx0 = compute_ctx;
+
+        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+
+        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(ctx0, 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  = NULL;
+        struct ggml_tensor* right = NULL;
+        for (int i = 0; i < class_tokens_mask.size(); i++) {
+            if (class_tokens_mask[i]) {
+                // printf(" 1,");
+                ctm.push_back(0.f);                        // here use 0.f instead of 1.f to make a scale mask
+                ctmf16.push_back(ggml_fp32_to_fp16(0.f));  // here use 0.f instead of 1.f to make a scale mask
+                ctmpos.push_back(i);
+            } else {
+                // printf(" 0,");
+                ctm.push_back(1.f);                        // here use 1.f instead of 0.f to make a scale mask
+                ctmf16.push_back(ggml_fp32_to_fp16(1.f));  // here use 0.f instead of 1.f to make a scale mask
+            }
+        }
+        // printf("\n");
+        if (ctmpos[0] > 0) {
+            // left = ggml_new_tensor_3d(ctx0, type, hidden_size, 1, ctmpos[0]);
+            left = ggml_new_tensor_3d(ctx0, type, hidden_size, ctmpos[0], 1);
+        }
+        if (ctmpos[ctmpos.size() - 1] < seq_length - 1) {
+            // right = ggml_new_tensor_3d(ctx0, type,
+            //                            hidden_size, 1, seq_length - ctmpos[ctmpos.size() - 1] - 1);
+            right = ggml_new_tensor_3d(ctx0, type,
+                                       hidden_size, seq_length - ctmpos[ctmpos.size() - 1] - 1, 1);
+        }
+        struct ggml_tensor* class_tokens_mask_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ctmpos.size());
+
+        {
+            if (type == GGML_TYPE_F16)
+                set_backend_tensor_data(class_tokens_mask_d, ctmf16.data());
+            else
+                set_backend_tensor_data(class_tokens_mask_d, ctm.data());
+            set_backend_tensor_data(class_tokens_mask_pos, ctmpos.data());
+            if (left) {
+                if (type == GGML_TYPE_F16) {
+                    for (int i = 0; i < ggml_nelements(left); ++i)
+                        zeros_left_16.push_back(ggml_fp32_to_fp16(0.f));
+                    set_backend_tensor_data(left, zeros_left_16.data());
+                } else {
+                    for (int i = 0; i < ggml_nelements(left); ++i)
+                        zeros_left.push_back(0.f);
+                    set_backend_tensor_data(left, zeros_left.data());
+                }
+            }
+            if (right) {
+                if (type == GGML_TYPE_F16) {
+                    for (int i = 0; i < ggml_nelements(right); ++i)
+                        zeros_right_16.push_back(ggml_fp32_to_fp16(0.f));
+                    set_backend_tensor_data(right, zeros_right_16.data());
+                } else {
+                    for (int i = 0; i < ggml_nelements(right); ++i)
+                        zeros_right.push_back(0.f);
+                    set_backend_tensor_data(right, zeros_right.data());
+                }
+            }
+        }
+        struct ggml_tensor* updated_prompt_embeds = NULL;
+        if (pm_version == PM_VERSION_1)
+            updated_prompt_embeds = id_encoder.forward(ctx0,
+                                                       runtime_backend,
+                                                       id_pixel_values_d,
+                                                       prompt_embeds_d,
+                                                       class_tokens_mask_d,
+                                                       class_tokens_mask_pos,
+                                                       left, right);
+        else if (pm_version == PM_VERSION_2)
+            updated_prompt_embeds = id_encoder2.forward(ctx0,
+                                                        runtime_backend,
+                                                        id_pixel_values_d,
+                                                        prompt_embeds_d,
+                                                        class_tokens_mask_d,
+                                                        class_tokens_mask_pos,
+                                                        id_embeds_d,
+                                                        left, right);
+
+        ggml_build_forward_expand(gf, updated_prompt_embeds);
+
+        return gf;
+    }
+
+    void 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);
+            return build_graph(id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds);
+        };
+
+        // GGMLRunner::compute(get_graph, n_threads, updated_prompt_embeds);
+        GGMLRunner::compute(get_graph, n_threads, true, updated_prompt_embeds, output_ctx);
+    }
+};
+
+struct PhotoMakerIDEmbed : public GGMLRunner {
+    std::map<std::string, struct ggml_tensor*> tensors;
+    std::string file_path;
+    ModelLoader* model_loader;
+    bool load_failed = false;
+    bool applied     = false;
+
+    PhotoMakerIDEmbed(ggml_backend_t backend,
+                      bool offload_params_to_cpu,
+                      ModelLoader* ml,
+                      const std::string& file_path = "",
+                      const std::string& prefix    = "")
+        : file_path(file_path), GGMLRunner(backend, offload_params_to_cpu), model_loader(ml) {
+        if (!model_loader->init_from_file(file_path, prefix)) {
+            load_failed = true;
+        }
+    }
+
+    std::string get_desc() {
+        return "id_embeds";
+    }
+
+    bool load_from_file(bool filter_tensor, int n_threads) {
+        LOG_INFO("loading PhotoMaker ID Embeds from '%s'", file_path.c_str());
+
+        if (load_failed) {
+            LOG_ERROR("init photomaker id embed from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        bool dry_run = true;
+        std::mutex tensor_mutex;
+        auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
+            const std::string& name = tensor_storage.name;
+
+            if (filter_tensor && !contains(name, "pmid.id_embeds")) {
+                // LOG_INFO("skipping LoRA tesnor '%s'", name.c_str());
+                return true;
+            }
+            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;
+            } else {
+                auto real   = tensors[name];
+                *dst_tensor = real;
+            }
+
+            return true;
+        };
+
+        model_loader->load_tensors(on_new_tensor_cb, n_threads);
+        alloc_params_buffer();
+
+        dry_run = false;
+        model_loader->load_tensors(on_new_tensor_cb, n_threads);
+
+        LOG_DEBUG("finished loading PhotoMaker ID Embeds ");
+        return true;
+    }
+
+    struct ggml_tensor* get() {
+        std::map<std::string, struct ggml_tensor*>::iterator pos;
+        pos = tensors.find("pmid.id_embeds");
+        if (pos != tensors.end())
+            return pos->second;
+        return NULL;
+    }
+};
+
+#endif  // __PMI_HPP__

preprocessing.hpp

@@ -0,0 +1,226 @@
+#ifndef __PREPROCESSING_HPP__
+#define __PREPROCESSING_HPP__
+
+#include "ggml_extend.hpp"
+#define M_PI_ 3.14159265358979323846
+
+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               = NULL;
+    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   = 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 = -ks_mid + y;
+        for (int x = 0; x < kernel->ne[1]; x++) {
+            float gy = -ks_mid + x;
+            float k_ = expf(-((gx * gx + gy * gy) / (2.0f * powf(sigma, 2.0f)))) * normal;
+            ggml_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_tensor_get_f32(rgb_img, ix, iy);
+            float g    = ggml_tensor_get_f32(rgb_img, ix, iy, 1);
+            float b    = ggml_tensor_get_f32(rgb_img, ix, iy, 2);
+            float gray = 0.2989f * r + 0.5870f * g + 0.1140f * b;
+            ggml_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 = 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 = 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 = 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_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_tensor_get_f32(G, ix, iy + 1);
+                r = ggml_tensor_get_f32(G, ix, iy - 1);
+            }
+            // angle 45
+            else if (22.5f >= angle && angle < 67.5f) {
+                q = ggml_tensor_get_f32(G, ix + 1, iy - 1);
+                r = ggml_tensor_get_f32(G, ix - 1, iy + 1);
+            }
+            // angle 90
+            else if (67.5f >= angle && angle < 112.5) {
+                q = ggml_tensor_get_f32(G, ix + 1, iy);
+                r = ggml_tensor_get_f32(G, ix - 1, iy);
+            }
+            // angle 135
+            else if (112.5 >= angle && angle < 157.5f) {
+                q = ggml_tensor_get_f32(G, ix - 1, iy - 1);
+                r = ggml_tensor_get_f32(G, ix + 1, iy + 1);
+            }
+
+            float cur = ggml_tensor_get_f32(G, ix, iy);
+            if ((cur >= q) && (cur >= r)) {
+                ggml_tensor_set_f32(result, cur, ix, iy);
+            } else {
+                ggml_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 = 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_tensor_set_f32(img, ggml_tensor_get_f32(img, ix, iy), ix, iy);
+            } else {
+                ggml_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_tensor_get_f32(img, ix, iy);
+            if (imd_v == weak) {
+                if (ggml_tensor_get_f32(img, ix + 1, iy - 1) == strong || ggml_tensor_get_f32(img, ix + 1, iy) == strong ||
+                    ggml_tensor_get_f32(img, ix, iy - 1) == strong || ggml_tensor_get_f32(img, ix, iy + 1) == strong ||
+                    ggml_tensor_get_f32(img, ix - 1, iy - 1) == strong || ggml_tensor_get_f32(img, ix - 1, iy) == strong) {
+                    ggml_tensor_set_f32(img, strong, ix, iy);
+                } else {
+                    ggml_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             = NULL;
+    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_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 (int iy = 0; iy < img.height; iy++) {
+        for (int ix = 0; ix < img.width; ix++) {
+            float gray = ggml_tensor_get_f32(image_gray, ix, iy);
+            gray       = inverse ? 1.0f - gray : gray;
+            ggml_tensor_set_f32(image, gray, ix, iy);
+            ggml_tensor_set_f32(image, gray, ix, iy, 1);
+            ggml_tensor_set_f32(image, gray, ix, iy, 2);
+        }
+    }
+    sd_tensor_to_image(image, img.data);
+    ggml_free(work_ctx);
+    return true;
+}
+
+#endif  // __PREPROCESSING_HPP__

qwen_image.hpp

@@ -0,0 +1,694 @@
+#ifndef __QWEN_IMAGE_HPP__
+#define __QWEN_IMAGE_HPP__
+
+#include "common.hpp"
+#include "flux.hpp"
+#include "ggml_extend.hpp"
+
+namespace Qwen {
+    constexpr int QWEN_IMAGE_GRAPH_SIZE = 20480;
+
+    struct TimestepEmbedding : public GGMLBlock {
+    public:
+        TimestepEmbedding(int64_t in_channels,
+                          int64_t time_embed_dim,
+                          int64_t out_dim       = 0,
+                          int64_t cond_proj_dim = 0,
+                          bool sample_proj_bias = true) {
+            blocks["linear_1"] = std::shared_ptr<GGMLBlock>(new Linear(in_channels, time_embed_dim, sample_proj_bias));
+            if (cond_proj_dim > 0) {
+                blocks["cond_proj"] = std::shared_ptr<GGMLBlock>(new Linear(cond_proj_dim, in_channels, false));
+            }
+            if (out_dim <= 0) {
+                out_dim = time_embed_dim;
+            }
+            blocks["linear_2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, out_dim, sample_proj_bias));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct ggml_tensor* sample,
+                                    struct ggml_tensor* condition = nullptr) {
+            if (condition != nullptr) {
+                auto cond_proj = std::dynamic_pointer_cast<Linear>(blocks["cond_proj"]);
+                sample         = ggml_add(ctx, sample, cond_proj->forward(ctx, condition));
+            }
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["linear_1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["linear_2"]);
+
+            sample = linear_1->forward(ctx, sample);
+            sample = ggml_silu_inplace(ctx, sample);
+            sample = linear_2->forward(ctx, sample);
+            return sample;
+        }
+    };
+
+    struct QwenTimestepProjEmbeddings : public GGMLBlock {
+    public:
+        QwenTimestepProjEmbeddings(int64_t embedding_dim) {
+            blocks["timestep_embedder"] = std::shared_ptr<GGMLBlock>(new TimestepEmbedding(256, embedding_dim));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct ggml_tensor* timesteps) {
+            // timesteps: [N,]
+            // return: [N, embedding_dim]
+            auto timestep_embedder = std::dynamic_pointer_cast<TimestepEmbedding>(blocks["timestep_embedder"]);
+
+            auto timesteps_proj = ggml_nn_timestep_embedding(ctx, timesteps, 256, 10000, 1.f);
+            auto timesteps_emb  = timestep_embedder->forward(ctx, timesteps_proj);
+            return timesteps_emb;
+        }
+    };
+
+    struct QwenImageAttention : public GGMLBlock {
+    protected:
+        int64_t dim_head;
+        bool flash_attn;
+
+    public:
+        QwenImageAttention(int64_t query_dim,
+                           int64_t dim_head,
+                           int64_t num_heads,
+                           int64_t out_dim         = 0,
+                           int64_t out_context_dim = 0,
+                           bool bias               = true,
+                           bool out_bias           = true,
+                           float eps               = 1e-6,
+                           bool flash_attn         = false)
+            : dim_head(dim_head), flash_attn(flash_attn) {
+            int64_t inner_dim = out_dim > 0 ? out_dim : dim_head * num_heads;
+            out_dim           = out_dim > 0 ? out_dim : query_dim;
+            out_context_dim   = out_context_dim > 0 ? out_context_dim : query_dim;
+
+            blocks["to_q"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["to_k"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["to_v"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+
+            blocks["norm_q"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+            blocks["norm_k"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+
+            blocks["add_q_proj"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["add_k_proj"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["add_v_proj"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+
+            blocks["norm_added_q"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+            blocks["norm_added_k"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+
+            float scale = 1.f / 32.f;
+            // The purpose of the scale here is to prevent NaN issues in certain situations.
+            // For example when using CUDA but the weights are k-quants (not all prompts).
+            blocks["to_out.0"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_dim, out_bias, false, false, scale));
+            // to_out.1 is nn.Dropout
+
+            blocks["to_add_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_context_dim, out_bias, false, false, scale));
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(struct ggml_context* ctx,
+                                                      ggml_backend_t backend,
+                                                      struct ggml_tensor* img,
+                                                      struct ggml_tensor* txt,
+                                                      struct ggml_tensor* pe,
+                                                      struct 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]
+            // return: ([N, n_img_token, hidden_size], [N, n_txt_token, hidden_size])
+
+            auto norm_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_q"]);
+            auto norm_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_k"]);
+
+            auto to_q     = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
+            auto to_k     = std::dynamic_pointer_cast<Linear>(blocks["to_k"]);
+            auto to_v     = std::dynamic_pointer_cast<Linear>(blocks["to_v"]);
+            auto to_out_0 = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+
+            auto norm_added_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_q"]);
+            auto norm_added_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_k"]);
+
+            auto add_q_proj = std::dynamic_pointer_cast<Linear>(blocks["add_q_proj"]);
+            auto add_k_proj = std::dynamic_pointer_cast<Linear>(blocks["add_k_proj"]);
+            auto add_v_proj = std::dynamic_pointer_cast<Linear>(blocks["add_v_proj"]);
+            auto to_add_out = std::dynamic_pointer_cast<Linear>(blocks["to_add_out"]);
+
+            int64_t N           = img->ne[2];
+            int64_t n_img_token = img->ne[1];
+            int64_t n_txt_token = txt->ne[1];
+
+            auto img_q        = to_q->forward(ctx, img);
+            int64_t num_heads = img_q->ne[0] / dim_head;
+            img_q             = ggml_reshape_4d(ctx, img_q, dim_head, num_heads, n_img_token, N);  // [N, n_img_token, n_head, d_head]
+            auto img_k        = to_k->forward(ctx, img);
+            img_k             = ggml_reshape_4d(ctx, img_k, dim_head, num_heads, n_img_token, N);  // [N, n_img_token, n_head, d_head]
+            auto img_v        = to_v->forward(ctx, img);
+            img_v             = ggml_reshape_4d(ctx, img_v, dim_head, num_heads, n_img_token, N);  // [N, n_img_token, n_head, d_head]
+
+            img_q = norm_q->forward(ctx, img_q);
+            img_k = norm_k->forward(ctx, img_k);
+
+            auto txt_q = add_q_proj->forward(ctx, txt);
+            txt_q      = ggml_reshape_4d(ctx, txt_q, dim_head, num_heads, n_txt_token, N);  // [N, n_txt_token, n_head, d_head]
+            auto txt_k = add_k_proj->forward(ctx, txt);
+            txt_k      = ggml_reshape_4d(ctx, txt_k, dim_head, num_heads, n_txt_token, N);  // [N, n_txt_token, n_head, d_head]
+            auto txt_v = add_v_proj->forward(ctx, txt);
+            txt_v      = ggml_reshape_4d(ctx, txt_v, dim_head, num_heads, n_txt_token, N);  // [N, n_txt_token, n_head, d_head]
+
+            txt_q = norm_added_q->forward(ctx, txt_q);
+            txt_k = norm_added_k->forward(ctx, txt_k);
+
+            auto q = ggml_concat(ctx, txt_q, img_q, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
+            auto k = ggml_concat(ctx, txt_k, img_k, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
+            auto v = ggml_concat(ctx, txt_v, img_v, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
+
+            auto attn         = Rope::attention(ctx, backend, q, k, v, pe, mask, flash_attn, (1.0f / 128.f));  // [N, n_txt_token + n_img_token, n_head*d_head]
+            attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));                           // [n_txt_token + n_img_token, N, hidden_size]
+            auto txt_attn_out = ggml_view_3d(ctx,
+                                             attn,
+                                             attn->ne[0],
+                                             attn->ne[1],
+                                             txt->ne[1],
+                                             attn->nb[1],
+                                             attn->nb[2],
+                                             0);                                              // [n_txt_token, N, hidden_size]
+            txt_attn_out      = ggml_cont(ctx, ggml_permute(ctx, txt_attn_out, 0, 2, 1, 3));  // [N, n_txt_token, hidden_size]
+            auto img_attn_out = ggml_view_3d(ctx,
+                                             attn,
+                                             attn->ne[0],
+                                             attn->ne[1],
+                                             img->ne[1],
+                                             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_permute(ctx, img_attn_out, 0, 2, 1, 3));  // [N, n_img_token, hidden_size]
+
+            img_attn_out = to_out_0->forward(ctx, img_attn_out);
+            txt_attn_out = to_add_out->forward(ctx, txt_attn_out);
+
+            return {img_attn_out, txt_attn_out};
+        }
+    };
+
+    class QwenImageTransformerBlock : public GGMLBlock {
+    public:
+        QwenImageTransformerBlock(int64_t dim,
+                                  int64_t num_attention_heads,
+                                  int64_t attention_head_dim,
+                                  float eps       = 1e-6,
+                                  bool flash_attn = false) {
+            // img_mod.0 is nn.SiLU()
+            blocks["img_mod.1"] = std::shared_ptr<GGMLBlock>(new Linear(dim, 6 * dim, true));
+
+            blocks["img_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
+            blocks["img_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
+            blocks["img_mlp"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim, 4, FeedForward::Activation::GELU, true));
+
+            // txt_mod.0 is nn.SiLU()
+            blocks["txt_mod.1"] = std::shared_ptr<GGMLBlock>(new Linear(dim, 6 * dim, true));
+
+            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["attn"] = std::shared_ptr<GGMLBlock>(new QwenImageAttention(dim,
+                                                                               attention_head_dim,
+                                                                               num_attention_heads,
+                                                                               0,     // out_dim
+                                                                               0,     // out_context-dim
+                                                                               true,  // bias
+                                                                               true,  // out_bias
+                                                                               eps,
+                                                                               flash_attn));
+        }
+
+        virtual std::pair<ggml_tensor*, ggml_tensor*> forward(struct ggml_context* ctx,
+                                                              ggml_backend_t backend,
+                                                              struct ggml_tensor* img,
+                                                              struct ggml_tensor* txt,
+                                                              struct ggml_tensor* t_emb,
+                                                              struct ggml_tensor* pe) {
+            // 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]
+            // return: ([N, n_img_token, hidden_size], [N, n_txt_token, hidden_size])
+
+            auto img_mod_1 = std::dynamic_pointer_cast<Linear>(blocks["img_mod.1"]);
+            auto img_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm1"]);
+            auto img_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm2"]);
+            auto img_mlp   = std::dynamic_pointer_cast<FeedForward>(blocks["img_mlp"]);
+
+            auto txt_mod_1 = std::dynamic_pointer_cast<Linear>(blocks["txt_mod.1"]);
+            auto txt_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["txt_norm1"]);
+            auto txt_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["txt_norm2"]);
+            auto txt_mlp   = std::dynamic_pointer_cast<FeedForward>(blocks["txt_mlp"]);
+
+            auto attn = std::dynamic_pointer_cast<QwenImageAttention>(blocks["attn"]);
+
+            auto img_mod_params    = ggml_silu(ctx, t_emb);
+            img_mod_params         = img_mod_1->forward(ctx, img_mod_params);
+            auto img_mod_param_vec = ggml_chunk(ctx, img_mod_params, 6, 0);
+
+            auto txt_mod_params    = ggml_silu(ctx, t_emb);
+            txt_mod_params         = txt_mod_1->forward(ctx, txt_mod_params);
+            auto txt_mod_param_vec = ggml_chunk(ctx, txt_mod_params, 6, 0);
+
+            auto img_normed    = img_norm1->forward(ctx, img);
+            auto img_modulated = Flux::modulate(ctx, img_normed, img_mod_param_vec[0], img_mod_param_vec[1]);
+            auto img_gate1     = img_mod_param_vec[2];
+
+            auto txt_normed    = txt_norm1->forward(ctx, txt);
+            auto txt_modulated = Flux::modulate(ctx, txt_normed, txt_mod_param_vec[0], txt_mod_param_vec[1]);
+            auto txt_gate1     = txt_mod_param_vec[2];
+
+            auto [img_attn_output, txt_attn_output] = attn->forward(ctx, backend, img_modulated, txt_modulated, pe);
+
+            img = ggml_add(ctx, img, ggml_mul(ctx, img_attn_output, img_gate1));
+            txt = ggml_add(ctx, txt, ggml_mul(ctx, txt_attn_output, txt_gate1));
+
+            auto img_normed2    = img_norm2->forward(ctx, img);
+            auto img_modulated2 = Flux::modulate(ctx, img_normed2, img_mod_param_vec[3], img_mod_param_vec[4]);
+            auto img_gate2      = img_mod_param_vec[5];
+
+            auto txt_normed2    = txt_norm2->forward(ctx, txt);
+            auto txt_modulated2 = Flux::modulate(ctx, txt_normed2, txt_mod_param_vec[3], txt_mod_param_vec[4]);
+            auto txt_gate2      = txt_mod_param_vec[5];
+
+            auto img_mlp_out = img_mlp->forward(ctx, img_modulated2);
+            auto txt_mlp_out = txt_mlp->forward(ctx, txt_modulated2);
+
+            img = ggml_add(ctx, img, ggml_mul(ctx, img_mlp_out, img_gate2));
+            txt = ggml_add(ctx, txt, ggml_mul(ctx, txt_mlp_out, txt_gate2));
+
+            return {img, txt};
+        }
+    };
+
+    struct AdaLayerNormContinuous : public GGMLBlock {
+    public:
+        AdaLayerNormContinuous(int64_t embedding_dim,
+                               int64_t conditioning_embedding_dim,
+                               bool elementwise_affine = true,
+                               float eps               = 1e-5f,
+                               bool bias               = true) {
+            blocks["norm"]   = std::shared_ptr<GGMLBlock>(new LayerNorm(conditioning_embedding_dim, eps, elementwise_affine, bias));
+            blocks["linear"] = std::shared_ptr<GGMLBlock>(new Linear(conditioning_embedding_dim, embedding_dim * 2, bias));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* c) {
+            // x: [N, n_token, hidden_size]
+            // c: [N, hidden_size]
+            // return: [N, n_token, patch_size * patch_size * out_channels]
+
+            auto norm   = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+
+            auto emb   = linear->forward(ctx, ggml_silu(ctx, c));
+            auto mods  = ggml_chunk(ctx, emb, 2, 0);
+            auto scale = mods[0];
+            auto shift = mods[1];
+
+            x = norm->forward(ctx, x);
+            x = Flux::modulate(ctx, x, shift, scale);
+
+            return x;
+        }
+    };
+
+    struct QwenImageParams {
+        int64_t patch_size          = 2;
+        int64_t in_channels         = 64;
+        int64_t out_channels        = 16;
+        int64_t num_layers          = 60;
+        int64_t attention_head_dim  = 128;
+        int64_t num_attention_heads = 24;
+        int64_t joint_attention_dim = 3584;
+        float theta                 = 10000;
+        std::vector<int> axes_dim   = {16, 56, 56};
+        int64_t axes_dim_sum        = 128;
+        bool flash_attn             = false;
+    };
+
+    class QwenImageModel : public GGMLBlock {
+    protected:
+        QwenImageParams params;
+
+    public:
+        QwenImageModel() {}
+        QwenImageModel(QwenImageParams params)
+            : params(params) {
+            int64_t inner_dim         = params.num_attention_heads * params.attention_head_dim;
+            blocks["time_text_embed"] = std::shared_ptr<GGMLBlock>(new QwenTimestepProjEmbeddings(inner_dim));
+            blocks["txt_norm"]        = std::shared_ptr<GGMLBlock>(new RMSNorm(params.joint_attention_dim, 1e-6f));
+            blocks["img_in"]          = std::shared_ptr<GGMLBlock>(new Linear(params.in_channels, inner_dim));
+            blocks["txt_in"]          = std::shared_ptr<GGMLBlock>(new Linear(params.joint_attention_dim, inner_dim));
+
+            // blocks
+            for (int i = 0; i < params.num_layers; i++) {
+                auto block                                        = std::shared_ptr<GGMLBlock>(new QwenImageTransformerBlock(inner_dim,
+                                                                                                                             params.num_attention_heads,
+                                                                                                                             params.attention_head_dim,
+                                                                                                                             1e-6f,
+                                                                                                                             params.flash_attn));
+                blocks["transformer_blocks." + std::to_string(i)] = block;
+            }
+
+            blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new AdaLayerNormContinuous(inner_dim, inner_dim, false, 1e-6f));
+            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(struct ggml_context* 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_pad(ctx, x, pad_w, pad_h, 0, 0);  // [N, C, H + pad_h, W + pad_w]
+            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(struct ggml_context* ctx,
+                                        struct ggml_tensor* x) {
+            x = pad_to_patch_size(ctx, x);
+            x = patchify(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(struct ggml_context* ctx,
+                                         ggml_backend_t backend,
+                                         struct ggml_tensor* x,
+                                         struct ggml_tensor* timestep,
+                                         struct ggml_tensor* context,
+                                         struct ggml_tensor* pe) {
+            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"]);
+            auto txt_in          = std::dynamic_pointer_cast<Linear>(blocks["txt_in"]);
+            auto norm_out        = std::dynamic_pointer_cast<AdaLayerNormContinuous>(blocks["norm_out"]);
+            auto proj_out        = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
+
+            auto t_emb = time_text_embed->forward(ctx, timestep);
+            auto img   = img_in->forward(ctx, x);
+            auto txt   = txt_norm->forward(ctx, context);
+            txt        = txt_in->forward(ctx, txt);
+
+            for (int i = 0; i < params.num_layers; i++) {
+                auto block = std::dynamic_pointer_cast<QwenImageTransformerBlock>(blocks["transformer_blocks." + std::to_string(i)]);
+
+                auto result = block->forward(ctx, backend, img, txt, t_emb, pe);
+                img         = result.first;
+                txt         = result.second;
+            }
+
+            img = norm_out->forward(ctx, img, t_emb);
+            img = proj_out->forward(ctx, img);
+
+            return img;
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    ggml_backend_t backend,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* timestep,
+                                    struct ggml_tensor* context,
+                                    struct ggml_tensor* pe,
+                                    std::vector<ggml_tensor*> ref_latents = {}) {
+            // Forward pass of DiT.
+            // x: [N, C, H, W]
+            // timestep: [N,]
+            // context: [N, L, D]
+            // pe: [L, d_head/2, 2, 2]
+            // return: [N, C, H, W]
+
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+            int64_t C = x->ne[2];
+            int64_t N = x->ne[3];
+
+            auto img            = process_img(ctx, x);
+            uint64_t img_tokens = img->ne[1];
+
+            if (ref_latents.size() > 0) {
+                for (ggml_tensor* ref : ref_latents) {
+                    ref = process_img(ctx, ref);
+                    img = ggml_concat(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, backend, img, timestep, context, pe);  // [N, h_len*w_len, ph*pw*C]
+
+            if (out->ne[1] > img_tokens) {
+                out = ggml_cont(ctx, ggml_permute(ctx, out, 0, 2, 1, 3));  // [num_tokens, N, C * patch_size * patch_size]
+                out = ggml_view_3d(ctx, out, out->ne[0], out->ne[1], img_tokens, out->nb[1], out->nb[2], 0);
+                out = ggml_cont(ctx, ggml_permute(ctx, out, 0, 2, 1, 3));  // [N, h*w, C * patch_size * patch_size]
+            }
+
+            out = unpatchify(ctx, out, h_len, w_len);  // [N, C, H + pad_h, W + pad_w]
+
+            // slice
+            out = ggml_slice(ctx, out, 1, 0, H);  // [N, C, H, W + pad_w]
+            out = ggml_slice(ctx, out, 0, 0, W);  // [N, C, H, W]
+
+            return out;
+        }
+    };
+
+    struct QwenImageRunner : public GGMLRunner {
+    public:
+        QwenImageParams qwen_image_params;
+        QwenImageModel qwen_image;
+        std::vector<float> pe_vec;
+        SDVersion version;
+
+        QwenImageRunner(ggml_backend_t backend,
+                        bool offload_params_to_cpu,
+                        const String2GGMLType& tensor_types = {},
+                        const std::string prefix            = "",
+                        SDVersion version                   = VERSION_QWEN_IMAGE,
+                        bool flash_attn                     = false)
+            : GGMLRunner(backend, offload_params_to_cpu) {
+            qwen_image_params.flash_attn = flash_attn;
+            qwen_image_params.num_layers = 0;
+            for (auto pair : tensor_types) {
+                std::string tensor_name = pair.first;
+                if (tensor_name.find(prefix) == std::string::npos)
+                    continue;
+                size_t pos = tensor_name.find("transformer_blocks.");
+                if (pos != std::string::npos) {
+                    tensor_name = tensor_name.substr(pos);  // remove prefix
+                    auto items  = split_string(tensor_name, '.');
+                    if (items.size() > 1) {
+                        int block_index = atoi(items[1].c_str());
+                        if (block_index + 1 > qwen_image_params.num_layers) {
+                            qwen_image_params.num_layers = block_index + 1;
+                        }
+                    }
+                    continue;
+                }
+            }
+            LOG_ERROR("qwen_image_params.num_layers: %ld", qwen_image_params.num_layers);
+            qwen_image                   = QwenImageModel(qwen_image_params);
+            qwen_image.init(params_ctx, tensor_types, prefix);
+        }
+
+        std::string get_desc() {
+            return "qwen_image";
+        }
+
+        void get_param_tensors(std::map<std::string, struct 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_ASSERT(x->ne[3] == 1);
+            struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, QWEN_IMAGE_GRAPH_SIZE, false);
+
+            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]);
+            }
+
+            pe_vec      = Rope::gen_qwen_image_pe(x->ne[1],
+                                                  x->ne[0],
+                                                  qwen_image_params.patch_size,
+                                                  x->ne[3],
+                                                  context->ne[1],
+                                                  ref_latents,
+                                                  increase_ref_index,
+                                                  qwen_image_params.theta,
+                                                  qwen_image_params.axes_dim);
+            int pos_len = pe_vec.size() / qwen_image_params.axes_dim_sum / 2;
+            // LOG_DEBUG("pos_len %d", pos_len);
+            auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, qwen_image_params.axes_dim_sum / 2, pos_len);
+            // pe->data = pe_vec.data();
+            // print_ggml_tensor(pe, true, "pe");
+            // pe->data = NULL;
+            set_backend_tensor_data(pe, pe_vec.data());
+
+            struct ggml_tensor* out = qwen_image.forward(compute_ctx,
+                                                         runtime_backend,
+                                                         x,
+                                                         timesteps,
+                                                         context,
+                                                         pe,
+                                                         ref_latents);
+
+            ggml_build_forward_expand(gf, out);
+
+            return gf;
+        }
+
+        void 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           = NULL,
+                     struct ggml_context* output_ctx       = NULL) {
+            // x: [N, in_channels, h, w]
+            // timesteps: [N, ]
+            // context: [N, max_position, hidden_size]
+            auto get_graph = [&]() -> struct ggml_cgraph* {
+                return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
+            };
+
+            GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+        }
+
+        void test() {
+            struct ggml_init_params params;
+            params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
+            params.mem_buffer = NULL;
+            params.no_alloc   = false;
+
+            struct ggml_context* work_ctx = ggml_init(params);
+            GGML_ASSERT(work_ctx != NULL);
+
+            {
+                // auto x = ggml_new_tensor_4d(work_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);
+
+                std::vector<float> timesteps_vec(1, 1000.f);
+                auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
+
+                // auto context = ggml_new_tensor_3d(work_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);
+
+                struct ggml_tensor* out = NULL;
+
+                int t0 = ggml_time_ms();
+                compute(8, x, timesteps, context, {}, false, &out, work_ctx);
+                int t1 = ggml_time_ms();
+
+                print_ggml_tensor(out);
+                LOG_DEBUG("qwen_image test done in %dms", t1 - t0);
+            }
+        }
+
+        static void load_from_file_and_test(const std::string& file_path) {
+            // cuda q8: pass
+            // cuda q8 fa: nan
+            // ggml_backend_t backend    = ggml_backend_cuda_init(0);
+            ggml_backend_t backend    = ggml_backend_cpu_init();
+            ggml_type model_data_type = GGML_TYPE_Q8_0;
+
+            ModelLoader model_loader;
+            if (!model_loader.init_from_file(file_path, "model.diffusion_model.")) {
+                LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
+                return;
+            }
+
+            auto tensor_types = model_loader.tensor_storages_types;
+            for (auto& item : tensor_types) {
+                // LOG_DEBUG("%s %u", item.first.c_str(), item.second);
+                if (ends_with(item.first, "weight")) {
+                    item.second = model_data_type;
+                }
+            }
+
+            std::shared_ptr<QwenImageRunner> qwen_image = std::shared_ptr<QwenImageRunner>(new QwenImageRunner(backend,
+                                                                                                               false,
+                                                                                                               tensor_types,
+                                                                                                               "model.diffusion_model",
+                                                                                                               VERSION_QWEN_IMAGE,
+                                                                                                               true));
+
+            qwen_image->alloc_params_buffer();
+            std::map<std::string, ggml_tensor*> tensors;
+            qwen_image->get_param_tensors(tensors, "model.diffusion_model");
+
+            bool success = model_loader.load_tensors(tensors);
+
+            if (!success) {
+                LOG_ERROR("load tensors from model loader failed");
+                return;
+            }
+
+            LOG_INFO("qwen_image model loaded");
+            qwen_image->test();
+        }
+    };
+
+}  // namespace name
+
+#endif  // __QWEN_IMAGE_HPP__