feat: support more LoRA models (#520)

* Refactor: wtype per tensor

* Fix default args

* refactor: fix flux

* Refactor photmaker v2 support

* unet: refactor the refactoring

* Refactor: fix controlnet and tae

* refactor: upscaler

* Refactor: fix runtime type override

* upscaler: use fp16 again

* Refactor: Flexible sd3 arch

* Refactor: Flexible Flux arch

* format code

---------

Co-authored-by: leejet <leejet714@gmail.com>

.clang-format

@@ -3,11 +3,10 @@ UseTab: Never
 IndentWidth: 4
 TabWidth: 4
 AllowShortIfStatementsOnASingleLine: false
-IndentCaseLabels: false
 ColumnLimit: 0
 AccessModifierOffset: -4
 NamespaceIndentation: All
-FixNamespaceComments: false 
+FixNamespaceComments: false
 AlignAfterOpenBracket: true
 AlignConsecutiveAssignments: true
 IndentCaseLabels: true

.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-2019
+
+    env:
+      VULKAN_VERSION: 1.3.261.1
 
     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_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DSD_BUILD_SHARED_LIBS=ON"
+          - build: "avx2"
+            defines: "-DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
+          - build: "avx"
+            defines: "-DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS=ON"
+          - build: "avx512"
+            defines: "-DGGML_AVX512=ON -DSD_BUILD_SHARED_LIBS=ON"
+          - build: "cuda12"
+            defines: "-DSD_CUBLAS=ON -DSD_BUILD_SHARED_LIBS=ON"
+          # - 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@v0.2.11
+        with:
+          cuda: "12.2.0"
+          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' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
+          & "$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: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' && matrix.build == 'cuda12' ) || 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: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' && matrix.build == 'cuda12' ) || 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
 
@@ -164,7 +290,11 @@ jobs:
     steps:
       - 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 hash
         id: commit

.gitignore

@@ -10,5 +10,4 @@ test/
 *.gguf
 output*.png
 models*
-!taesd-model.gguf
 *.log

.gitmodules

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

CMakeLists.txt

@@ -25,31 +25,92 @@ endif()
 #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_FLASH_ATTN                 "sd: use flash attention for x4 less memory usage" OFF)
-option(BUILD_SHARED_LIBS             "sd: build shared libs" OFF)
+option(SD_HIPBLAS                    "sd: rocm backend" OFF)
+option(SD_METAL                      "sd: metal backend" OFF)
+option(SD_VULKAN                     "sd: vulkan backend" OFF)
+option(SD_SYCL                       "sd: sycl backend" OFF)
+option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
+option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
 
 if(SD_CUBLAS)
-	message("Use CUBLAS as backend stable-diffusion")
-    set(GGML_CUBLAS ON)
+    message("-- Use CUBLAS as backend stable-diffusion")
+    set(GGML_CUDA ON)
     add_definitions(-DSD_USE_CUBLAS)
 endif()
 
-if(SD_FLASH_ATTN)
-    message("Use Flash Attention for memory optimization")
-    add_definitions(-DSD_USE_FLASH_ATTENTION)
+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 ()
 
-set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
-# deps
-add_subdirectory(ggml)
-
-add_subdirectory(thirdparty)
+if (SD_HIPBLAS)
+    message("-- Use HIPBLAS as backend stable-diffusion")
+    set(GGML_HIPBLAS ON)
+    add_definitions(-DSD_USE_CUBLAS)
+    if(SD_FAST_SOFTMAX)
+        set(GGML_CUDA_FAST_SOFTMAX ON)
+    endif()
+endif ()
 
 set(SD_LIB stable-diffusion)
 
-add_library(${SD_LIB} stable-diffusion.h stable-diffusion.cpp model.h model.cpp util.h util.cpp)
+file(GLOB SD_LIB_SOURCES 
+    "*.h"
+    "*.cpp"
+    "*.hpp"
+)
+
+# we can get only one share lib
+if(SD_BUILD_SHARED_LIBS)
+    message("-- Build shared library")
+    message(${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 ${SD_LIB_SOURCES})
+endif()
+
+if(SD_SYCL)
+    message("-- Use SYCL as backend stable-diffusion")
+    set(GGML_SYCL ON)
+    add_definitions(-DSD_USE_SYCL)
+    # disable fast-math on host, see:
+    # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html
+    if (WIN32)
+        set(SYCL_COMPILE_OPTIONS /fp:precise)
+    else()
+        set(SYCL_COMPILE_OPTIONS -fp-model=precise)
+    endif()
+    message("-- Turn off fast-math for host in SYCL backend")
+    target_compile_options(${SD_LIB} PRIVATE ${SYCL_COMPILE_OPTIONS})
+endif()
+
+set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
+
+# see https://github.com/ggerganov/ggml/pull/682
+add_definitions(-DGGML_MAX_NAME=128)
+
+# deps
+# Only add ggml if it hasn't been added yet
+if (NOT TARGET ggml)
+    add_subdirectory(ggml)
+endif()
+
+add_subdirectory(thirdparty)
+
 target_link_libraries(${SD_LIB} PUBLIC ggml zip)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
 target_compile_features(${SD_LIB} PUBLIC cxx_std_11)

README.md

@@ -1,31 +1,39 @@
 <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++
+Inference of Stable Diffusion and Flux in pure C/C++
 
 ## 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 and SD2.x support
+- SD1.x, SD2.x, SDXL and [SD3/SD3.5](./docs/sd3.md) support
+    - !!!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).
+- [Flux-dev/Flux-schnell Support](./docs/flux.md)
+
+- [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo) and [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo) support
+- [PhotoMaker](https://github.com/TencentARC/PhotoMaker) support.
 - 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 backend for GPU acceleration.
+- Full CUDA, Metal, Vulkan 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)
+- Flash Attention for memory usage optimization
 - Original `txt2img` and `img2img` mode
 - 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
+- Control Net support with SD 1.5
 - Sampling method
     - `Euler A`
     - `Euler`
@@ -48,14 +56,14 @@ Inference of [Stable Diffusion](https://github.com/CompVis/stable-diffusion) in
 - [ ] 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 BPE Tokenizer
-- [ ] Implement [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN/tree/master) upscaler
-- [ ] k-quants support
+- [ ] Implement Inpainting support
 
 ## 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
 
 ```
@@ -78,11 +86,13 @@ 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
@@ -112,65 +122,152 @@ cmake .. -DSD_CUBLAS=ON
 cmake --build . --config Release
 ```
 
-### Using Flash Attention
+##### Using HipBLAS
+This provides BLAS acceleration using the ROCm cores of your AMD GPU. Make sure to have the ROCm toolkit installed.
 
-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.
+Windows User Refer to [docs/hipBLAS_on_Windows.md](docs%2FhipBLAS_on_Windows.md) for a comprehensive guide.
 
 ```
-cmake .. -DSD_FLASH_ATTN=ON
+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
 ```
 
+
+##### Using Metal
+
+Using Metal makes the computation run on the GPU. Currently, there are some issues with Metal when performing operations on very large matrices, making it highly inefficient at the moment. Performance improvements are expected in the near future.
+
+```
+cmake .. -DSD_METAL=ON
+cmake --build . --config Release
+```
+
+##### Using Vulkan
+
+Install Vulkan SDK from https://www.lunarg.com/vulkan-sdk/.
+
+```
+cmake .. -DSD_VULKAN=ON
+cmake --build . --config Release
+```
+
+##### 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
 
 ```
-usage: sd [arguments]
+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 [MODEL]                 run mode (txt2img or img2img or convert, default: txt2img)
+  -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
+  -m, --model [MODEL]                path to full model
+  --diffusion-model                  path to the standalone diffusion model
+  --clip_l                           path to the clip-l text encoder
+  --clip_g                           path to the clip-l text encoder
+  --t5xxl                            path to the the t5xxl text encoder
   --vae [VAE]                        path to vae
   --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
-  --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
+  --stacked-id-embd-dir [DIR]        path to PHOTOMAKER stacked id embeddings
+  --input-id-images-dir [DIR]        path to PHOTOMAKER input id images dir
+  --normalize-input                  normalize PHOTOMAKER input id images
+  --upscale-model [ESRGAN_PATH]      path to esrgan model. 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 (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
   --lora-model-dir [DIR]             lora model directory
   -i, --init-img [IMAGE]             path to the input image, required by img2img
+  --control-image [IMAGE]            path to image condition, control net
   -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)
   --strength STRENGTH                strength for noising/unnoising (default: 0.75)
+  --style-ratio STYLE-RATIO          strength for keeping input identity (default: 20%)
+  --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 {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, 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)
+  -b, --batch-count COUNT            number of images to generate
+  --schedule {discrete, karras, exponential, ays, gits} Denoiser sigma schedule (default: discrete)
+  --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-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.
+  --control-net-cpu                  keep controlnet in cpu (for low vram)
+  --canny                            apply canny preprocessor (edge detection)
+  --color                            Colors the logging tags according to level
   -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
+# ./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
+# ./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
 ```
 
 Using formats of different precisions will yield results of varying quality.
@@ -192,76 +289,31 @@ Using formats of different precisions will yield results of varying quality.
   <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: [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
+* 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)
 
-`../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
-```
-
-### 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)
 
 ## Contributors
 
@@ -269,11 +321,19 @@ 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)
 - [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)

clip.hpp

@@ -0,0 +1,952 @@
+#ifndef __CLIP_HPP__
+#define __CLIP_HPP__
+
+#include "ggml_extend.hpp"
+#include "model.h"
+
+/*================================================== CLIPTokenizer ===================================================*/
+
+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);
+}
+
+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, 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, 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, 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, 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;
+
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type token_wtype    = (tensor_types.find(prefix + "token_embedding.weight") != tensor_types.end()) ? tensor_types[prefix + "token_embedding.weight"] : GGML_TYPE_F32;
+        enum ggml_type position_wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "position_embedding.weight") != tensor_types.end()) ? tensor_types[prefix + "position_embedding.weight"] : 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)
+        : embed_dim(embed_dim),
+          vocab_size(vocab_size),
+          num_positions(num_positions) {
+    }
+
+    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, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type patch_wtype    = GGML_TYPE_F16;  // tensor_types.find(prefix + "patch_embedding.weight") != tensor_types.end() ? tensor_types[prefix + "patch_embedding.weight"] : GGML_TYPE_F16;
+        enum ggml_type class_wtype    = GGML_TYPE_F32;  // tensor_types.find(prefix + "class_embedding") != tensor_types.end() ? tensor_types[prefix + "class_embedding"] : GGML_TYPE_F32;
+        enum ggml_type position_wtype = GGML_TYPE_F32;  // tensor_types.find(prefix + "position_embedding.weight") != tensor_types.end() ? tensor_types[prefix + "position_embedding.weight"] : 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, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        if (version == OPEN_CLIP_VIT_BIGG_14) {
+            enum ggml_type wtype      = GGML_TYPE_F32;  // tensor_types.find(prefix + "text_projection") != tensor_types.end() ? tensor_types[prefix + "text_projection"] : 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
+    int32_t clip_skip         = -1;
+    bool with_final_ln        = true;
+
+    CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
+                  int clip_skip_value = -1,
+                  bool with_final_ln  = true)
+        : 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;
+        }
+        set_clip_skip(clip_skip_value);
+
+        blocks["embeddings"]       = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token));
+        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));
+    }
+
+    void set_clip_skip(int skip) {
+        if (skip <= 0) {
+            return;
+        }
+        clip_skip = skip;
+    }
+
+    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,
+                                struct ggml_tensor* input_ids,
+                                struct ggml_tensor* tkn_embeddings,
+                                size_t max_token_idx = 0,
+                                bool return_pooled   = false) {
+        // input_ids: [N, n_token]
+        auto embeddings       = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
+        auto encoder          = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
+        auto final_layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["final_layer_norm"]);
+
+        auto x = embeddings->forward(ctx, input_ids, tkn_embeddings);  // [N, n_token, hidden_size]
+        x      = encoder->forward(ctx, x, return_pooled ? -1 : clip_skip, true);
+        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("Missing text_projection matrix, assuming identity...");
+            }
+            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, struct ggml_tensor* pixel_values, bool return_pooled = true) {
+        // 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, x, -1, 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, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type wtype = tensor_types.find(prefix + "weight") != tensor_types.end() ? tensor_types[prefix + "weight"] : 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, struct ggml_tensor* pixel_values) {
+        // pixel_values: [N, num_channels, image_size, image_size]
+        // return: [N, projection_dim]
+        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, pixel_values);  // [N, hidden_size]
+        x      = visual_projection->forward(ctx, x);        // [N, projection_dim]
+
+        return x;  // [N, projection_dim]
+    }
+};
+
+struct CLIPTextModelRunner : public GGMLRunner {
+    CLIPTextModel model;
+
+    CLIPTextModelRunner(ggml_backend_t backend,
+                        std::map<std::string, enum ggml_type>& tensor_types,
+                        const std::string prefix,
+                        CLIPVersion version = OPENAI_CLIP_VIT_L_14,
+                        int clip_skip_value = 1,
+                        bool with_final_ln  = true)
+        : GGMLRunner(backend), model(version, clip_skip_value, with_final_ln) {
+        model.init(params_ctx, tensor_types, prefix);
+    }
+
+    std::string get_desc() {
+        return "clip";
+    }
+
+    void set_clip_skip(int clip_skip) {
+        model.set_clip_skip(clip_skip);
+    }
+
+    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,
+                                struct ggml_tensor* input_ids,
+                                struct ggml_tensor* embeddings,
+                                size_t max_token_idx = 0,
+                                bool return_pooled   = false) {
+        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, input_ids, embeddings, max_token_idx, return_pooled);
+    }
+
+    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) {
+        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, input_ids, embeddings, max_token_idx, return_pooled);
+
+        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,
+                 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);
+        };
+        GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
+    }
+};
+
+#endif  // __CLIP_HPP__

common.hpp

@@ -0,0 +1,523 @@
+#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);  // [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 GGMLBlock {
+protected:
+    int64_t dim_in;
+    int64_t dim_out;
+
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
+        enum ggml_type wtype      = (tensor_types.find(prefix + "proj.weight") != tensor_types.end()) ? tensor_types[prefix + "proj.weight"] : GGML_TYPE_F32;
+        enum ggml_type bias_wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "proj.bias") != tensor_types.end()) ? tensor_types[prefix + "proj.bias"] : 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 FeedForward : public GGMLBlock {
+public:
+    FeedForward(int64_t dim,
+                int64_t dim_out,
+                int64_t mult = 4) {
+        int64_t inner_dim = dim * mult;
+
+        blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GEGLU(dim, inner_dim));
+        // net_1 is nn.Dropout(), skip for inference
+        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out));
+    }
+
+    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<GEGLU>(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, 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, 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, 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, x, x);  // self-attention
+        x      = ggml_add(ctx, x, r);
+        r      = x;
+        x      = norm2->forward(ctx, x);
+        x      = attn2->forward(ctx, 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, 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, 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, std::map<std::string, enum ggml_type>& 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;  //(tensor_types.ypes.find(prefix + "mix_factor") != tensor_types.end()) ? tensor_types[prefix + "mix_factor"] : 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 +1,458 @@
+#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 (version == VERSION_SD2) {
+            context_dim       = 1024;
+            num_head_channels = 64;
+            num_heads         = -1;
+        } else if (version == VERSION_SDXL) {
+            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 (version == VERSION_SDXL || 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,
+                                                struct ggml_tensor* x,
+                                                struct ggml_tensor* context) {
+        auto block = std::dynamic_pointer_cast<SpatialTransformer>(blocks[name]);
+        return block->forward(ctx, x, context);
+    }
+
+    struct ggml_tensor* input_hint_block_forward(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,
+                                             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, 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, 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,
+               std::map<std::string, enum ggml_type>& tensor_types,
+               SDVersion version = VERSION_SD1)
+        : GGMLRunner(backend), control_net(version) {
+        control_net.init(params_ctx, tensor_types, "");
+    }
+
+    ~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, 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,
+                                        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) {
+        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, backend, ignore_tensors);
+
+        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 +1,181 @@
+#ifndef __DIFFUSION_MODEL_H__
+#define __DIFFUSION_MODEL_H__
+
+#include "flux.hpp"
+#include "mmdit.hpp"
+#include "unet.hpp"
+
+struct DiffusionModel {
+    virtual void compute(int n_threads,
+                         struct ggml_tensor* x,
+                         struct ggml_tensor* timesteps,
+                         struct ggml_tensor* context,
+                         struct ggml_tensor* c_concat,
+                         struct ggml_tensor* y,
+                         struct ggml_tensor* guidance,
+                         int num_video_frames                      = -1,
+                         std::vector<struct ggml_tensor*> controls = {},
+                         float control_strength                    = 0.f,
+                         struct ggml_tensor** output               = NULL,
+                         struct ggml_context* output_ctx           = NULL,
+                         std::vector<int> skip_layers              = std::vector<int>())             = 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,
+              std::map<std::string, enum ggml_type>& tensor_types,
+              SDVersion version = VERSION_SD1,
+              bool flash_attn   = false)
+        : unet(backend, tensor_types, "model.diffusion_model", version, flash_attn) {
+    }
+
+    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,
+                 struct ggml_tensor* x,
+                 struct ggml_tensor* timesteps,
+                 struct ggml_tensor* context,
+                 struct ggml_tensor* c_concat,
+                 struct ggml_tensor* y,
+                 struct ggml_tensor* guidance,
+                 int num_video_frames                      = -1,
+                 std::vector<struct ggml_tensor*> controls = {},
+                 float control_strength                    = 0.f,
+                 struct ggml_tensor** output               = NULL,
+                 struct ggml_context* output_ctx           = NULL,
+                 std::vector<int> skip_layers              = std::vector<int>()) {
+        (void)skip_layers;  // SLG doesn't work with UNet models
+        return unet.compute(n_threads, x, timesteps, context, c_concat, y, num_video_frames, controls, control_strength, output, output_ctx);
+    }
+};
+
+struct MMDiTModel : public DiffusionModel {
+    MMDiTRunner mmdit;
+
+    MMDiTModel(ggml_backend_t backend,
+               std::map<std::string, enum ggml_type>& tensor_types)
+        : mmdit(backend, tensor_types, "model.diffusion_model") {
+    }
+
+    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,
+                 struct ggml_tensor* x,
+                 struct ggml_tensor* timesteps,
+                 struct ggml_tensor* context,
+                 struct ggml_tensor* c_concat,
+                 struct ggml_tensor* y,
+                 struct ggml_tensor* guidance,
+                 int num_video_frames                      = -1,
+                 std::vector<struct ggml_tensor*> controls = {},
+                 float control_strength                    = 0.f,
+                 struct ggml_tensor** output               = NULL,
+                 struct ggml_context* output_ctx           = NULL,
+                 std::vector<int> skip_layers              = std::vector<int>()) {
+        return mmdit.compute(n_threads, x, timesteps, context, y, output, output_ctx, skip_layers);
+    }
+};
+
+struct FluxModel : public DiffusionModel {
+    Flux::FluxRunner flux;
+
+    FluxModel(ggml_backend_t backend,
+              std::map<std::string, enum ggml_type>& tensor_types,
+              bool flash_attn = false)
+        : flux(backend, tensor_types, "model.diffusion_model", flash_attn) {
+    }
+
+    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,
+                 struct ggml_tensor* x,
+                 struct ggml_tensor* timesteps,
+                 struct ggml_tensor* context,
+                 struct ggml_tensor* c_concat,
+                 struct ggml_tensor* y,
+                 struct ggml_tensor* guidance,
+                 int num_video_frames                      = -1,
+                 std::vector<struct ggml_tensor*> controls = {},
+                 float control_strength                    = 0.f,
+                 struct ggml_tensor** output               = NULL,
+                 struct ggml_context* output_ctx           = NULL,
+                 std::vector<int> skip_layers              = std::vector<int>()) {
+        return flux.compute(n_threads, x, timesteps, context, y, guidance, output, output_ctx, skip_layers);
+    }
+};
+
+#endif

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
+```
+
+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
+```
+
+| 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
+```
+
+![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/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 +1,13 @@
+## 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

docs/photo_maker.md

@@ -0,0 +1,54 @@
+## 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 `--stacked-id-embd-dir PATH` parameter.
+- Specify the input images path using the `--input-id-images-dir PATH` parameter.
+  - input images **must** have the same width and height for preprocessing (to be improved)
+
+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:
+
+- ```--style-ratio  (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 --stacked-id-embd-dir ../models/photomaker-v1.safetensors --input-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 --style-ratio 10 --vae-on-cpu -o output.png
+```
+
+## 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/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
+```
+
+![](../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
+```

esrgan.hpp

@@ -0,0 +1,197 @@
+#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;  // default RealESRGAN_x4plus_anime_6B
+    int num_block   = 6;  // default RealESRGAN_x4plus_anime_6B
+    int num_in_ch   = 3;
+    int num_out_ch  = 3;
+    int num_feat    = 64;  // default RealESRGAN_x4plus_anime_6B
+    int num_grow_ch = 32;  // default RealESRGAN_x4plus_anime_6B
+
+public:
+    RRDBNet() {
+        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}));
+        // upsample
+        blocks["conv_up1"]  = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        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}));
+    }
+
+    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*4, w*4]
+        auto conv_first = std::dynamic_pointer_cast<Conv2d>(blocks["conv_first"]);
+        auto conv_body  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_body"]);
+        auto conv_up1   = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up1"]);
+        auto conv_up2   = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up2"]);
+        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
+        feat     = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx, feat, 2)));
+        feat     = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx, feat, 2)));
+        auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat)));
+        return out;
+    }
+};
+
+struct ESRGAN : public GGMLRunner {
+    RRDBNet rrdb_net;
+    int scale     = 4;
+    int tile_size = 128;  // avoid cuda OOM for 4gb VRAM
+
+    ESRGAN(ggml_backend_t backend, std::map<std::string, enum ggml_type>& tensor_types)
+        : GGMLRunner(backend) {
+        rrdb_net.init(params_ctx, tensor_types, "");
+    }
+
+    std::string get_desc() {
+        return "esrgan";
+    }
+
+    bool load_from_file(const std::string& file_path) {
+        LOG_INFO("loading esrgan from '%s'", file_path.c_str());
+
+        alloc_params_buffer();
+        std::map<std::string, ggml_tensor*> esrgan_tensors;
+        rrdb_net.get_param_tensors(esrgan_tensors);
+
+        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;
+        }
+
+        bool success = model_loader.load_tensors(esrgan_tensors, backend);
+
+        if (!success) {
+            LOG_ERROR("load esrgan tensors from model loader failed");
+            return false;
+        }
+
+        LOG_INFO("esrgan model loaded");
+        return success;
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* x) {
+        struct ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        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/main.cpp

@@ -1,21 +1,26 @@
 #include <stdio.h>
-#include <ctime>
+#include <string.h>
+#include <time.h>
+#include <iostream>
 #include <random>
-#include "ggml/ggml.h"
+#include <string>
+#include <vector>
+
+// #include "preprocessing.hpp"
+#include "flux.hpp"
 #include "stable-diffusion.h"
-#include "util.h"
 
 #define STB_IMAGE_IMPLEMENTATION
+#define STB_IMAGE_STATIC
 #include "stb_image.h"
 
 #define STB_IMAGE_WRITE_IMPLEMENTATION
 #define STB_IMAGE_WRITE_STATIC
 #include "stb_image_write.h"
 
-#include <cstring>
-#include <iostream>
-#include <string>
-#include <vector>
+#define STB_IMAGE_RESIZE_IMPLEMENTATION
+#define STB_IMAGE_RESIZE_STATIC
+#include "stb_image_resize.h"
 
 const char* rng_type_to_str[] = {
     "std_default",
@@ -31,6 +36,8 @@ const char* sample_method_str[] = {
     "dpm++2s_a",
     "dpm++2m",
     "dpm++2mv2",
+    "ipndm",
+    "ipndm_v",
     "lcm",
 };
 
@@ -39,45 +46,85 @@ const char* schedule_str[] = {
     "default",
     "discrete",
     "karras",
+    "exponential",
+    "ays",
+    "gits",
 };
 
 const char* modes_str[] = {
     "txt2img",
     "img2img",
+    "img2vid",
+    "convert",
 };
 
 enum SDMode {
     TXT2IMG,
     IMG2IMG,
+    IMG2VID,
+    CONVERT,
     MODE_COUNT
 };
 
 struct SDParams {
     int n_threads = -1;
     SDMode mode   = TXT2IMG;
-
     std::string model_path;
+    std::string clip_l_path;
+    std::string clip_g_path;
+    std::string t5xxl_path;
+    std::string diffusion_model_path;
     std::string vae_path;
     std::string taesd_path;
-    ggml_type wtype = GGML_TYPE_COUNT;
+    std::string esrgan_path;
+    std::string controlnet_path;
+    std::string embeddings_path;
+    std::string stacked_id_embeddings_path;
+    std::string input_id_images_path;
+    sd_type_t wtype = SD_TYPE_COUNT;
     std::string lora_model_dir;
     std::string output_path = "output.png";
     std::string input_path;
+    std::string control_image_path;
 
     std::string prompt;
     std::string negative_prompt;
-    float cfg_scale = 7.0f;
-    int width       = 512;
-    int height      = 512;
-    int batch_count = 1;
+    float min_cfg     = 1.0f;
+    float cfg_scale   = 7.0f;
+    float guidance    = 3.5f;
+    float style_ratio = 20.f;
+    int clip_skip     = -1;  // <= 0 represents unspecified
+    int width         = 512;
+    int height        = 512;
+    int batch_count   = 1;
 
-    SampleMethod sample_method = EULER_A;
-    Schedule schedule          = DEFAULT;
-    int sample_steps           = 20;
-    float strength             = 0.75f;
-    RNGType rng_type           = CUDA_RNG;
-    int64_t seed               = 42;
-    bool verbose               = false;
+    int video_frames         = 6;
+    int motion_bucket_id     = 127;
+    int fps                  = 6;
+    float augmentation_level = 0.f;
+
+    sample_method_t sample_method = EULER_A;
+    schedule_t schedule           = DEFAULT;
+    int sample_steps              = 20;
+    float strength                = 0.75f;
+    float control_strength        = 0.9f;
+    rng_type_t rng_type           = CUDA_RNG;
+    int64_t seed                  = 42;
+    bool verbose                  = false;
+    bool vae_tiling               = false;
+    bool control_net_cpu          = false;
+    bool normalize_input          = false;
+    bool clip_on_cpu              = false;
+    bool vae_on_cpu               = false;
+    bool diffusion_flash_attn     = false;
+    bool canny_preprocess         = false;
+    bool color                    = false;
+    int upscale_repeats           = 1;
+
+    std::vector<int> skip_layers = {7, 8, 9};
+    float slg_scale              = 0.;
+    float skip_layer_start       = 0.01;
+    float skip_layer_end         = 0.2;
 };
 
 void print_params(SDParams params) {
@@ -85,14 +132,35 @@ void print_params(SDParams params) {
     printf("    n_threads:         %d\n", params.n_threads);
     printf("    mode:              %s\n", modes_str[params.mode]);
     printf("    model_path:        %s\n", params.model_path.c_str());
-    printf("    wtype:             %s\n", params.wtype < GGML_TYPE_COUNT ? ggml_type_name(params.wtype) : "unspecified");
+    printf("    wtype:             %s\n", params.wtype < SD_TYPE_COUNT ? sd_type_name(params.wtype) : "unspecified");
+    printf("    clip_l_path:       %s\n", params.clip_l_path.c_str());
+    printf("    clip_g_path:       %s\n", params.clip_g_path.c_str());
+    printf("    t5xxl_path:        %s\n", params.t5xxl_path.c_str());
+    printf("    diffusion_model_path:   %s\n", params.diffusion_model_path.c_str());
     printf("    vae_path:          %s\n", params.vae_path.c_str());
     printf("    taesd_path:        %s\n", params.taesd_path.c_str());
+    printf("    esrgan_path:       %s\n", params.esrgan_path.c_str());
+    printf("    controlnet_path:   %s\n", params.controlnet_path.c_str());
+    printf("    embeddings_path:   %s\n", params.embeddings_path.c_str());
+    printf("    stacked_id_embeddings_path:   %s\n", params.stacked_id_embeddings_path.c_str());
+    printf("    input_id_images_path:   %s\n", params.input_id_images_path.c_str());
+    printf("    style ratio:       %.2f\n", params.style_ratio);
+    printf("    normalize input image :  %s\n", params.normalize_input ? "true" : "false");
     printf("    output_path:       %s\n", params.output_path.c_str());
     printf("    init_img:          %s\n", params.input_path.c_str());
+    printf("    control_image:     %s\n", params.control_image_path.c_str());
+    printf("    clip on cpu:       %s\n", params.clip_on_cpu ? "true" : "false");
+    printf("    controlnet cpu:    %s\n", params.control_net_cpu ? "true" : "false");
+    printf("    vae decoder on cpu:%s\n", params.vae_on_cpu ? "true" : "false");
+    printf("    diffusion flash attention:%s\n", params.diffusion_flash_attn ? "true" : "false");
+    printf("    strength(control): %.2f\n", params.control_strength);
     printf("    prompt:            %s\n", params.prompt.c_str());
     printf("    negative_prompt:   %s\n", params.negative_prompt.c_str());
+    printf("    min_cfg:           %.2f\n", params.min_cfg);
     printf("    cfg_scale:         %.2f\n", params.cfg_scale);
+    printf("    slg_scale:         %.2f\n", params.slg_scale);
+    printf("    guidance:          %.2f\n", params.guidance);
+    printf("    clip_skip:         %d\n", params.clip_skip);
     printf("    width:             %d\n", params.width);
     printf("    height:            %d\n", params.height);
     printf("    sample_method:     %s\n", sample_method_str[params.sample_method]);
@@ -102,6 +170,8 @@ void print_params(SDParams params) {
     printf("    rng:               %s\n", rng_type_to_str[params.rng_type]);
     printf("    seed:              %ld\n", params.seed);
     printf("    batch_count:       %d\n", params.batch_count);
+    printf("    vae_tiling:        %s\n", params.vae_tiling ? "true" : "false");
+    printf("    upscale_repeats:   %d\n", params.upscale_repeats);
 }
 
 void print_usage(int argc, const char* argv[]) {
@@ -109,31 +179,62 @@ void print_usage(int argc, const char* argv[]) {
     printf("\n");
     printf("arguments:\n");
     printf("  -h, --help                         show this help message and exit\n");
-    printf("  -M, --mode [txt2img or img2img]    generation mode (default: txt2img)\n");
-    printf("  -t, --threads N                    number of threads to use during computation (default: -1).\n");
+    printf("  -M, --mode [MODEL]                 run mode (txt2img or img2img or convert, default: txt2img)\n");
+    printf("  -t, --threads N                    number of threads to use during computation (default: -1)\n");
     printf("                                     If threads <= 0, then threads will be set to the number of CPU physical cores\n");
-    printf("  -m, --model [MODEL]                path to model\n");
+    printf("  -m, --model [MODEL]                path to full model\n");
+    printf("  --diffusion-model                  path to the standalone diffusion model\n");
+    printf("  --clip_l                           path to the clip-l text encoder\n");
+    printf("  --clip_g                           path to the clip-g text encoder\n");
+    printf("  --t5xxl                            path to the the t5xxl text encoder\n");
     printf("  --vae [VAE]                        path to vae\n");
     printf("  --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)\n");
-    printf("  --type [TYPE]                      weight type (f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0)\n");
-    printf("                                     If not specified, the default is the type of the weight file.\n");
+    printf("  --control-net [CONTROL_PATH]       path to control net model\n");
+    printf("  --embd-dir [EMBEDDING_PATH]        path to embeddings\n");
+    printf("  --stacked-id-embd-dir [DIR]        path to PHOTOMAKER stacked id embeddings\n");
+    printf("  --input-id-images-dir [DIR]        path to PHOTOMAKER input id images dir\n");
+    printf("  --normalize-input                  normalize PHOTOMAKER input id images\n");
+    printf("  --upscale-model [ESRGAN_PATH]      path to esrgan model. Upscale images after generate, just RealESRGAN_x4plus_anime_6B supported by now\n");
+    printf("  --upscale-repeats                  Run the ESRGAN upscaler this many times (default 1)\n");
+    printf("  --type [TYPE]                      weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K)\n");
+    printf("                                     If not specified, the default is the type of the weight file\n");
     printf("  --lora-model-dir [DIR]             lora model directory\n");
     printf("  -i, --init-img [IMAGE]             path to the input image, required by img2img\n");
+    printf("  --control-image [IMAGE]            path to image condition, control net\n");
     printf("  -o, --output OUTPUT                path to write result image to (default: ./output.png)\n");
     printf("  -p, --prompt [PROMPT]              the prompt to render\n");
     printf("  -n, --negative-prompt PROMPT       the negative prompt (default: \"\")\n");
     printf("  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)\n");
+    printf("  --slg-scale SCALE                  skip layer guidance (SLG) scale, only for DiT models: (default: 0)\n");
+    printf("                                     0 means disabled, a value of 2.5 is nice for sd3.5 medium\n");
+    printf("  --skip_layers LAYERS               Layers to skip for SLG steps: (default: [7,8,9])\n");
+    printf("  --skip_layer_start START           SLG enabling point: (default: 0.01)\n");
+    printf("  --skip_layer_end END               SLG disabling point: (default: 0.2)\n");
+    printf("                                     SLG will be enabled at step int([STEPS]*[START]) and disabled at int([STEPS]*[END])\n");
     printf("  --strength STRENGTH                strength for noising/unnoising (default: 0.75)\n");
+    printf("  --style-ratio STYLE-RATIO          strength for keeping input identity (default: 20%%)\n");
+    printf("  --control-strength STRENGTH        strength to apply Control Net (default: 0.9)\n");
     printf("                                     1.0 corresponds to full destruction of information in init image\n");
     printf("  -H, --height H                     image height, in pixel space (default: 512)\n");
     printf("  -W, --width W                      image width, in pixel space (default: 512)\n");
-    printf("  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, lcm}\n");
+    printf("  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm}\n");
     printf("                                     sampling method (default: \"euler_a\")\n");
     printf("  --steps  STEPS                     number of sample steps (default: 20)\n");
     printf("  --rng {std_default, cuda}          RNG (default: cuda)\n");
     printf("  -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)\n");
-    printf("  -b, --batch-count COUNT            number of images to generate.\n");
-    printf("  --schedule {discrete, karras}      Denoiser sigma schedule (default: discrete)\n");
+    printf("  -b, --batch-count COUNT            number of images to generate\n");
+    printf("  --schedule {discrete, karras, exponential, ays, gits} Denoiser sigma schedule (default: discrete)\n");
+    printf("  --clip-skip N                      ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)\n");
+    printf("                                     <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x\n");
+    printf("  --vae-tiling                       process vae in tiles to reduce memory usage\n");
+    printf("  --vae-on-cpu                       keep vae in cpu (for low vram)\n");
+    printf("  --clip-on-cpu                      keep clip in cpu (for low vram)\n");
+    printf("  --diffusion-fa                     use flash attention in the diffusion model (for low vram)\n");
+    printf("                                     Might lower quality, since it implies converting k and v to f16.\n");
+    printf("                                     This might crash if it is not supported by the backend.\n");
+    printf("  --control-net-cpu                  keep controlnet in cpu (for low vram)\n");
+    printf("  --canny                            apply canny preprocessor (edge detection)\n");
+    printf("  --color                            Colors the logging tags according to level\n");
     printf("  -v, --verbose                      print extra info\n");
 }
 
@@ -162,7 +263,8 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 }
             }
             if (mode_found == -1) {
-                fprintf(stderr, "error: invalid mode %s, must be one of [txt2img, img2img]\n",
+                fprintf(stderr,
+                        "error: invalid mode %s, must be one of [txt2img, img2img, img2vid, convert]\n",
                         mode_selected);
                 exit(1);
             }
@@ -173,6 +275,30 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.model_path = argv[i];
+        } else if (arg == "--clip_l") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.clip_l_path = argv[i];
+        } else if (arg == "--clip_g") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.clip_g_path = argv[i];
+        } else if (arg == "--t5xxl") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.t5xxl_path = argv[i];
+        } else if (arg == "--diffusion-model") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.diffusion_model_path = argv[i];
         } else if (arg == "--vae") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -185,29 +311,65 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.taesd_path = argv[i];
+        } else if (arg == "--control-net") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.controlnet_path = argv[i];
+        } else if (arg == "--upscale-model") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.esrgan_path = argv[i];
+        } else if (arg == "--embd-dir") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.embeddings_path = argv[i];
+        } else if (arg == "--stacked-id-embd-dir") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.stacked_id_embeddings_path = argv[i];
+        } else if (arg == "--input-id-images-dir") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.input_id_images_path = argv[i];
         } else if (arg == "--type") {
             if (++i >= argc) {
                 invalid_arg = true;
                 break;
             }
-            std::string type = argv[i];
-            if (type == "f32") {
-                params.wtype = GGML_TYPE_F32;
-            } else if (type == "f16") {
-                params.wtype = GGML_TYPE_F16;
-            } else if (type == "q4_0") {
-                params.wtype = GGML_TYPE_Q4_0;
-            } else if (type == "q4_1") {
-                params.wtype = GGML_TYPE_Q4_1;
-            } else if (type == "q5_0") {
-                params.wtype = GGML_TYPE_Q5_0;
-            } else if (type == "q5_1") {
-                params.wtype = GGML_TYPE_Q5_1;
-            } else if (type == "q8_0") {
-                params.wtype = GGML_TYPE_Q8_0;
-            } else {
-                fprintf(stderr, "error: invalid weight format %s, must be one of [f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0]\n",
-                        type.c_str());
+            std::string type        = argv[i];
+            bool found              = false;
+            std::string valid_types = "";
+            for (size_t i = 0; i < SD_TYPE_COUNT; i++) {
+                auto trait = ggml_get_type_traits((ggml_type)i);
+                std::string name(trait->type_name);
+                if (name == "f32" || trait->to_float && trait->type_size) {
+                    if (i)
+                        valid_types += ", ";
+                    valid_types += name;
+                    if (type == name) {
+                        if (ggml_quantize_requires_imatrix((ggml_type)i)) {
+                            printf("\033[35;1m[WARNING]\033[0m: type %s requires imatrix to work properly. A dummy imatrix will be used, expect poor quality.\n", trait->type_name);
+                        }
+                        params.wtype = (enum sd_type_t)i;
+                        found        = true;
+                        break;
+                    }
+                }
+            }
+            if (!found) {
+                fprintf(stderr, "error: invalid weight format %s, must be one of [%s]\n",
+                        type.c_str(),
+                        valid_types.c_str());
                 exit(1);
             }
         } else if (arg == "--lora-model-dir") {
@@ -222,6 +384,12 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.input_path = argv[i];
+        } else if (arg == "--control-image") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.control_image_path = argv[i];
         } else if (arg == "-o" || arg == "--output") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -234,6 +402,16 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.prompt = argv[i];
+        } else if (arg == "--upscale-repeats") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.upscale_repeats = std::stoi(argv[i]);
+            if (params.upscale_repeats < 1) {
+                fprintf(stderr, "error: upscale multiplier must be at least 1\n");
+                exit(1);
+            }
         } else if (arg == "-n" || arg == "--negative-prompt") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -246,12 +424,30 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.cfg_scale = std::stof(argv[i]);
+        } else if (arg == "--guidance") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.guidance = std::stof(argv[i]);
         } else if (arg == "--strength") {
             if (++i >= argc) {
                 invalid_arg = true;
                 break;
             }
             params.strength = std::stof(argv[i]);
+        } else if (arg == "--style-ratio") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.style_ratio = std::stof(argv[i]);
+        } else if (arg == "--control-strength") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.control_strength = std::stof(argv[i]);
         } else if (arg == "-H" || arg == "--height") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -270,6 +466,26 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.sample_steps = std::stoi(argv[i]);
+        } else if (arg == "--clip-skip") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.clip_skip = std::stoi(argv[i]);
+        } else if (arg == "--vae-tiling") {
+            params.vae_tiling = true;
+        } else if (arg == "--control-net-cpu") {
+            params.control_net_cpu = true;
+        } else if (arg == "--normalize-input") {
+            params.normalize_input = true;
+        } else if (arg == "--clip-on-cpu") {
+            params.clip_on_cpu = true;  // will slow down get_learned_condiotion but necessary for low MEM GPUs
+        } else if (arg == "--vae-on-cpu") {
+            params.vae_on_cpu = true;  // will slow down latent decoding but necessary for low MEM GPUs
+        } else if (arg == "--diffusion-fa") {
+            params.diffusion_flash_attn = true;  // can reduce MEM significantly
+        } else if (arg == "--canny") {
+            params.canny_preprocess = true;
         } else if (arg == "-b" || arg == "--batch-count") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -306,7 +522,7 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 invalid_arg = true;
                 break;
             }
-            params.schedule = (Schedule)schedule_found;
+            params.schedule = (schedule_t)schedule_found;
         } else if (arg == "-s" || arg == "--seed") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -329,12 +545,69 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 invalid_arg = true;
                 break;
             }
-            params.sample_method = (SampleMethod)sample_method_found;
+            params.sample_method = (sample_method_t)sample_method_found;
         } else if (arg == "-h" || arg == "--help") {
             print_usage(argc, argv);
             exit(0);
         } else if (arg == "-v" || arg == "--verbose") {
             params.verbose = true;
+        } else if (arg == "--color") {
+            params.color = true;
+        } else if (arg == "--slg-scale") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.slg_scale = std::stof(argv[i]);
+        } else if (arg == "--skip-layers") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            if (argv[i][0] != '[') {
+                invalid_arg = true;
+                break;
+            }
+            std::string layers_str = argv[i];
+            while (layers_str.back() != ']') {
+                if (++i >= argc) {
+                    invalid_arg = true;
+                    break;
+                }
+                layers_str += " " + std::string(argv[i]);
+            }
+            layers_str = layers_str.substr(1, layers_str.size() - 2);
+
+            std::regex regex("[, ]+");
+            std::sregex_token_iterator iter(layers_str.begin(), layers_str.end(), regex, -1);
+            std::sregex_token_iterator end;
+            std::vector<std::string> tokens(iter, end);
+            std::vector<int> layers;
+            for (const auto& token : tokens) {
+                try {
+                    layers.push_back(std::stoi(token));
+                } catch (const std::invalid_argument& e) {
+                    invalid_arg = true;
+                    break;
+                }
+            }
+            params.skip_layers = layers;
+
+            if (invalid_arg) {
+                break;
+            }
+        } else if (arg == "--skip-layer-start") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.skip_layer_start = std::stof(argv[i]);
+        } else if (arg == "--skip-layer-end") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.skip_layer_end = std::stof(argv[i]);
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             print_usage(argc, argv);
@@ -350,19 +623,19 @@ void parse_args(int argc, const char** argv, SDParams& params) {
         params.n_threads = get_num_physical_cores();
     }
 
-    if (params.prompt.length() == 0) {
+    if (params.mode != CONVERT && params.mode != IMG2VID && params.prompt.length() == 0) {
         fprintf(stderr, "error: the following arguments are required: prompt\n");
         print_usage(argc, argv);
         exit(1);
     }
 
-    if (params.model_path.length() == 0) {
-        fprintf(stderr, "error: the following arguments are required: model_path\n");
+    if (params.model_path.length() == 0 && params.diffusion_model_path.length() == 0) {
+        fprintf(stderr, "error: the following arguments are required: model_path/diffusion_model\n");
         print_usage(argc, argv);
         exit(1);
     }
 
-    if (params.mode == IMG2IMG && params.input_path.length() == 0) {
+    if ((params.mode == IMG2IMG || params.mode == IMG2VID) && params.input_path.length() == 0) {
         fprintf(stderr, "error: when using the img2img mode, the following arguments are required: init-img\n");
         print_usage(argc, argv);
         exit(1);
@@ -398,6 +671,24 @@ void parse_args(int argc, const char** argv, SDParams& params) {
         srand((int)time(NULL));
         params.seed = rand();
     }
+
+    if (params.mode == CONVERT) {
+        if (params.output_path == "output.png") {
+            params.output_path = "output.gguf";
+        }
+    }
+}
+
+static std::string sd_basename(const std::string& path) {
+    size_t pos = path.find_last_of('/');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    pos = path.find_last_of('\\');
+    if (pos != std::string::npos) {
+        return path.substr(pos + 1);
+    }
+    return path;
 }
 
 std::string get_image_params(SDParams params, int64_t seed) {
@@ -407,9 +698,20 @@ std::string get_image_params(SDParams params, int64_t seed) {
     }
     parameter_string += "Steps: " + std::to_string(params.sample_steps) + ", ";
     parameter_string += "CFG scale: " + std::to_string(params.cfg_scale) + ", ";
+    if (params.slg_scale != 0 && params.skip_layers.size() != 0) {
+        parameter_string += "SLG scale: " + std::to_string(params.cfg_scale) + ", ";
+        parameter_string += "Skip layers: [";
+        for (const auto& layer : params.skip_layers) {
+            parameter_string += std::to_string(layer) + ", ";
+        }
+        parameter_string += "], ";
+        parameter_string += "Skip layer start: " + std::to_string(params.skip_layer_start) + ", ";
+        parameter_string += "Skip layer end: " + std::to_string(params.skip_layer_end) + ", ";
+    }
+    parameter_string += "Guidance: " + std::to_string(params.guidance) + ", ";
     parameter_string += "Seed: " + std::to_string(seed) + ", ";
     parameter_string += "Size: " + std::to_string(params.width) + "x" + std::to_string(params.height) + ", ";
-    parameter_string += "Model: " + basename(params.model_path) + ", ";
+    parameter_string += "Model: " + sd_basename(params.model_path) + ", ";
     parameter_string += "RNG: " + std::string(rng_type_to_str[params.rng_type]) + ", ";
     parameter_string += "Sampler: " + std::string(sample_method_str[params.sample_method]);
     if (params.schedule == KARRAS) {
@@ -420,86 +722,335 @@ std::string get_image_params(SDParams params, int64_t seed) {
     return parameter_string;
 }
 
+/* Enables Printing the log level tag in color using ANSI escape codes */
+void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
+    SDParams* params = (SDParams*)data;
+    int tag_color;
+    const char* level_str;
+    FILE* out_stream = (level == SD_LOG_ERROR) ? stderr : stdout;
+
+    if (!log || (!params->verbose && level <= SD_LOG_DEBUG)) {
+        return;
+    }
+
+    switch (level) {
+        case SD_LOG_DEBUG:
+            tag_color = 37;
+            level_str = "DEBUG";
+            break;
+        case SD_LOG_INFO:
+            tag_color = 34;
+            level_str = "INFO";
+            break;
+        case SD_LOG_WARN:
+            tag_color = 35;
+            level_str = "WARN";
+            break;
+        case SD_LOG_ERROR:
+            tag_color = 31;
+            level_str = "ERROR";
+            break;
+        default: /* Potential future-proofing */
+            tag_color = 33;
+            level_str = "?????";
+            break;
+    }
+
+    if (params->color == true) {
+        fprintf(out_stream, "\033[%d;1m[%-5s]\033[0m ", tag_color, level_str);
+    } else {
+        fprintf(out_stream, "[%-5s] ", level_str);
+    }
+    fputs(log, out_stream);
+    fflush(out_stream);
+}
+
 int main(int argc, const char* argv[]) {
     SDParams params;
+
     parse_args(argc, argv, params);
 
+    sd_set_log_callback(sd_log_cb, (void*)&params);
+
     if (params.verbose) {
         print_params(params);
-        printf("%s", sd_get_system_info().c_str());
-        set_sd_log_level(SDLogLevel::DEBUG);
+        printf("%s", sd_get_system_info());
     }
 
-    bool vae_decode_only        = true;
-    uint8_t* input_image_buffer = NULL;
-    if (params.mode == IMG2IMG) {
+    if (params.mode == CONVERT) {
+        bool success = convert(params.model_path.c_str(), params.vae_path.c_str(), params.output_path.c_str(), params.wtype);
+        if (!success) {
+            fprintf(stderr,
+                    "convert '%s'/'%s' to '%s' failed\n",
+                    params.model_path.c_str(),
+                    params.vae_path.c_str(),
+                    params.output_path.c_str());
+            return 1;
+        } else {
+            printf("convert '%s'/'%s' to '%s' success\n",
+                   params.model_path.c_str(),
+                   params.vae_path.c_str(),
+                   params.output_path.c_str());
+            return 0;
+        }
+    }
+
+    if (params.mode == IMG2VID) {
+        fprintf(stderr, "SVD support is broken, do not use it!!!\n");
+        return 1;
+    }
+
+    bool vae_decode_only          = true;
+    uint8_t* input_image_buffer   = NULL;
+    uint8_t* control_image_buffer = NULL;
+    if (params.mode == IMG2IMG || params.mode == IMG2VID) {
         vae_decode_only = false;
 
         int c              = 0;
-        input_image_buffer = stbi_load(params.input_path.c_str(), &params.width, &params.height, &c, 3);
+        int width          = 0;
+        int height         = 0;
+        input_image_buffer = stbi_load(params.input_path.c_str(), &width, &height, &c, 3);
         if (input_image_buffer == NULL) {
             fprintf(stderr, "load image from '%s' failed\n", params.input_path.c_str());
             return 1;
         }
-        if (c != 3) {
-            fprintf(stderr, "input image must be a 3 channels RGB image, but got %d channels\n", c);
+        if (c < 3) {
+            fprintf(stderr, "the number of channels for the input image must be >= 3, but got %d channels\n", c);
             free(input_image_buffer);
             return 1;
         }
-        if (params.width <= 0 || params.width % 64 != 0) {
-            fprintf(stderr, "error: the width of image must be a multiple of 64\n");
+        if (width <= 0) {
+            fprintf(stderr, "error: the width of image must be greater than 0\n");
             free(input_image_buffer);
             return 1;
         }
-        if (params.height <= 0 || params.height % 64 != 0) {
-            fprintf(stderr, "error: the height of image must be a multiple of 64\n");
+        if (height <= 0) {
+            fprintf(stderr, "error: the height of image must be greater than 0\n");
             free(input_image_buffer);
             return 1;
         }
+
+        // Resize input image ...
+        if (params.height != height || params.width != width) {
+            printf("resize input image from %dx%d to %dx%d\n", width, height, params.width, params.height);
+            int resized_height = params.height;
+            int resized_width  = params.width;
+
+            uint8_t* resized_image_buffer = (uint8_t*)malloc(resized_height * resized_width * 3);
+            if (resized_image_buffer == NULL) {
+                fprintf(stderr, "error: allocate memory for resize input image\n");
+                free(input_image_buffer);
+                return 1;
+            }
+            stbir_resize(input_image_buffer, width, height, 0,
+                         resized_image_buffer, resized_width, resized_height, 0, STBIR_TYPE_UINT8,
+                         3 /*RGB channel*/, STBIR_ALPHA_CHANNEL_NONE, 0,
+                         STBIR_EDGE_CLAMP, STBIR_EDGE_CLAMP,
+                         STBIR_FILTER_BOX, STBIR_FILTER_BOX,
+                         STBIR_COLORSPACE_SRGB, nullptr);
+
+            // Save resized result
+            free(input_image_buffer);
+            input_image_buffer = resized_image_buffer;
+        }
     }
 
-    StableDiffusion sd(params.n_threads, vae_decode_only, params.taesd_path, true, params.lora_model_dir, params.rng_type);
+    sd_ctx_t* sd_ctx = new_sd_ctx(params.model_path.c_str(),
+                                  params.clip_l_path.c_str(),
+                                  params.clip_g_path.c_str(),
+                                  params.t5xxl_path.c_str(),
+                                  params.diffusion_model_path.c_str(),
+                                  params.vae_path.c_str(),
+                                  params.taesd_path.c_str(),
+                                  params.controlnet_path.c_str(),
+                                  params.lora_model_dir.c_str(),
+                                  params.embeddings_path.c_str(),
+                                  params.stacked_id_embeddings_path.c_str(),
+                                  vae_decode_only,
+                                  params.vae_tiling,
+                                  true,
+                                  params.n_threads,
+                                  params.wtype,
+                                  params.rng_type,
+                                  params.schedule,
+                                  params.clip_on_cpu,
+                                  params.control_net_cpu,
+                                  params.vae_on_cpu,
+                                  params.diffusion_flash_attn);
 
-    if (!sd.load_from_file(params.model_path, params.vae_path, params.wtype, params.schedule)) {
+    if (sd_ctx == NULL) {
+        printf("new_sd_ctx_t failed\n");
         return 1;
     }
 
-    std::vector<uint8_t*> results;
+    sd_image_t* control_image = NULL;
+    if (params.controlnet_path.size() > 0 && params.control_image_path.size() > 0) {
+        int c                = 0;
+        control_image_buffer = stbi_load(params.control_image_path.c_str(), &params.width, &params.height, &c, 3);
+        if (control_image_buffer == NULL) {
+            fprintf(stderr, "load image from '%s' failed\n", params.control_image_path.c_str());
+            return 1;
+        }
+        control_image = new sd_image_t{(uint32_t)params.width,
+                                       (uint32_t)params.height,
+                                       3,
+                                       control_image_buffer};
+        if (params.canny_preprocess) {  // apply preprocessor
+            control_image->data = preprocess_canny(control_image->data,
+                                                   control_image->width,
+                                                   control_image->height,
+                                                   0.08f,
+                                                   0.08f,
+                                                   0.8f,
+                                                   1.0f,
+                                                   false);
+        }
+    }
+
+    sd_image_t* results;
     if (params.mode == TXT2IMG) {
-        results = sd.txt2img(params.prompt,
-                             params.negative_prompt,
-                             params.cfg_scale,
-                             params.width,
-                             params.height,
-                             params.sample_method,
-                             params.sample_steps,
-                             params.seed,
-                             params.batch_count);
+        results = txt2img(sd_ctx,
+                          params.prompt.c_str(),
+                          params.negative_prompt.c_str(),
+                          params.clip_skip,
+                          params.cfg_scale,
+                          params.guidance,
+                          params.width,
+                          params.height,
+                          params.sample_method,
+                          params.sample_steps,
+                          params.seed,
+                          params.batch_count,
+                          control_image,
+                          params.control_strength,
+                          params.style_ratio,
+                          params.normalize_input,
+                          params.input_id_images_path.c_str(),
+                          params.skip_layers.data(),
+                          params.skip_layers.size(),
+                          params.slg_scale,
+                          params.skip_layer_start,
+                          params.skip_layer_end);
     } else {
-        results = sd.img2img(input_image_buffer,
-                             params.prompt,
-                             params.negative_prompt,
-                             params.cfg_scale,
-                             params.width,
-                             params.height,
-                             params.sample_method,
-                             params.sample_steps,
-                             params.strength,
-                             params.seed);
+        sd_image_t input_image = {(uint32_t)params.width,
+                                  (uint32_t)params.height,
+                                  3,
+                                  input_image_buffer};
+
+        if (params.mode == IMG2VID) {
+            results = img2vid(sd_ctx,
+                              input_image,
+                              params.width,
+                              params.height,
+                              params.video_frames,
+                              params.motion_bucket_id,
+                              params.fps,
+                              params.augmentation_level,
+                              params.min_cfg,
+                              params.cfg_scale,
+                              params.sample_method,
+                              params.sample_steps,
+                              params.strength,
+                              params.seed);
+            if (results == NULL) {
+                printf("generate failed\n");
+                free_sd_ctx(sd_ctx);
+                return 1;
+            }
+            size_t last            = params.output_path.find_last_of(".");
+            std::string dummy_name = last != std::string::npos ? params.output_path.substr(0, last) : params.output_path;
+            for (int i = 0; i < params.video_frames; i++) {
+                if (results[i].data == NULL) {
+                    continue;
+                }
+                std::string final_image_path = i > 0 ? dummy_name + "_" + std::to_string(i + 1) + ".png" : dummy_name + ".png";
+                stbi_write_png(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel,
+                               results[i].data, 0, get_image_params(params, params.seed + i).c_str());
+                printf("save result image to '%s'\n", final_image_path.c_str());
+                free(results[i].data);
+                results[i].data = NULL;
+            }
+            free(results);
+            free_sd_ctx(sd_ctx);
+            return 0;
+        } else {
+            results = img2img(sd_ctx,
+                              input_image,
+                              params.prompt.c_str(),
+                              params.negative_prompt.c_str(),
+                              params.clip_skip,
+                              params.cfg_scale,
+                              params.guidance,
+                              params.width,
+                              params.height,
+                              params.sample_method,
+                              params.sample_steps,
+                              params.strength,
+                              params.seed,
+                              params.batch_count,
+                              control_image,
+                              params.control_strength,
+                              params.style_ratio,
+                              params.normalize_input,
+                              params.input_id_images_path.c_str(),
+                              params.skip_layers.data(),
+                              params.skip_layers.size(),
+                              params.slg_scale,
+                              params.skip_layer_start,
+                              params.skip_layer_end);
+        }
     }
 
-    if (results.size() == 0 || results.size() != params.batch_count) {
-        LOG_ERROR("generate failed");
+    if (results == NULL) {
+        printf("generate failed\n");
+        free_sd_ctx(sd_ctx);
         return 1;
     }
 
+    int upscale_factor = 4;  // unused for RealESRGAN_x4plus_anime_6B.pth
+    if (params.esrgan_path.size() > 0 && params.upscale_repeats > 0) {
+        upscaler_ctx_t* upscaler_ctx = new_upscaler_ctx(params.esrgan_path.c_str(),
+                                                        params.n_threads);
+
+        if (upscaler_ctx == NULL) {
+            printf("new_upscaler_ctx failed\n");
+        } else {
+            for (int i = 0; i < params.batch_count; i++) {
+                if (results[i].data == NULL) {
+                    continue;
+                }
+                sd_image_t current_image = results[i];
+                for (int u = 0; u < params.upscale_repeats; ++u) {
+                    sd_image_t upscaled_image = upscale(upscaler_ctx, current_image, upscale_factor);
+                    if (upscaled_image.data == NULL) {
+                        printf("upscale failed\n");
+                        break;
+                    }
+                    free(current_image.data);
+                    current_image = upscaled_image;
+                }
+                results[i] = current_image;  // Set the final upscaled image as the result
+            }
+        }
+    }
+
     size_t last            = params.output_path.find_last_of(".");
     std::string dummy_name = last != std::string::npos ? params.output_path.substr(0, last) : params.output_path;
     for (int i = 0; i < params.batch_count; i++) {
+        if (results[i].data == NULL) {
+            continue;
+        }
         std::string final_image_path = i > 0 ? dummy_name + "_" + std::to_string(i + 1) + ".png" : dummy_name + ".png";
-        stbi_write_png(final_image_path.c_str(), params.width, params.height, 3, results[i], 0, get_image_params(params, params.seed + i).c_str());
-        LOG_INFO("save result image to '%s'", final_image_path.c_str());
+        stbi_write_png(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel,
+                       results[i].data, 0, get_image_params(params, params.seed + i).c_str());
+        printf("save result image to '%s'\n", final_image_path.c_str());
+        free(results[i].data);
+        results[i].data = NULL;
     }
+    free(results);
+    free_sd_ctx(sd_ctx);
+    free(control_image_buffer);
+    free(input_image_buffer);
 
     return 0;
 }

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 +1,2 @@
-clang-format -style=file -i *.cpp *.h
+clang-format -style=file -i *.cpp *.h *.hpp
 clang-format -style=file -i examples/cli/*.cpp

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 +1,680 @@
+#ifndef __LORA_HPP__
+#define __LORA_HPP__
+
+#include "ggml_extend.hpp"
+
+#define LORA_GRAPH_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
+        // 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::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) {
+        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 = false) {
+        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;
+        }
+
+        bool dry_run          = true;
+        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, "lora")) {
+                // LOG_INFO("skipping LoRA tesnor '%s'", name.c_str());
+                return true;
+            }
+            // LOG_INFO("%s", name.c_str());
+            for (int i = 0; i < LORA_TYPE_COUNT; i++) {
+                if (name.find(type_fingerprints[i]) != std::string::npos) {
+                    type = (lora_t)i;
+                    break;
+                }
+            }
+
+            if (dry_run) {
+                struct ggml_tensor* real = ggml_new_tensor(params_ctx,
+                                                           tensor_storage.type,
+                                                           tensor_storage.n_dims,
+                                                           tensor_storage.ne);
+                lora_tensors[name]       = real;
+            } else {
+                auto real   = lora_tensors[name];
+                *dst_tensor = real;
+            }
+
+            return true;
+        };
+
+        model_loader.load_tensors(on_new_tensor_cb, backend);
+        alloc_params_buffer();
+        // exit(0);
+        dry_run = false;
+        model_loader.load_tensors(on_new_tensor_cb, backend);
+
+        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);
+        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");
+            }
+
+            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, '.', '_');
+            ret.push_back(key);
+        }
+        return ret;
+    }
+
+    struct ggml_cgraph* build_lora_graph(std::map<std::string, struct ggml_tensor*> model_tensors, SDVersion version) {
+        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);
+
+        std::set<std::string> applied_lora_tensors;
+        for (auto it : model_tensors) {
+            std::string k_tensor       = it.first;
+            struct ggml_tensor* weight = model_tensors[it.first];
+
+            std::vector<std::string> keys = to_lora_keys(k_tensor, version);
+            if (keys.size() == 0)
+                continue;
+            ggml_tensor* lora_up   = NULL;
+            ggml_tensor* lora_down = NULL;
+            for (auto& key : keys) {
+                std::string alpha_name         = "";
+                std::string scale_name         = "";
+                std::string split_q_scale_name = "";
+                std::string lora_down_name     = "";
+                std::string lora_up_name       = "";
+
+                if (starts_with(key, "SPLIT|")) {
+                    key = key.substr(sizeof("SPLIT|") - 1);
+                    // TODO: Handle alphas
+                    std::string suffix  = "";
+                    auto split_q_d_name = lora_pre[type] + key + "q" + suffix + lora_downs[type] + ".weight";
+
+                    if (lora_tensors.find(split_q_d_name) == lora_tensors.end()) {
+                        suffix         = "_proj";
+                        split_q_d_name = lora_pre[type] + 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 = lora_pre[type] + key + "k" + suffix + lora_downs[type] + ".weight";
+                        auto split_v_d_name = lora_pre[type] + key + "v" + suffix + lora_downs[type] + ".weight";
+
+                        auto split_q_u_name = lora_pre[type] + key + "q" + suffix + lora_ups[type] + ".weight";
+                        auto split_k_u_name = lora_pre[type] + key + "k" + suffix + lora_ups[type] + ".weight";
+                        auto split_v_u_name = lora_pre[type] + key + "v" + suffix + lora_ups[type] + ".weight";
+
+                        auto split_q_scale_name = lora_pre[type] + key + "q" + suffix + ".scale";
+                        auto split_k_scale_name = lora_pre[type] + key + "k" + suffix + ".scale";
+                        auto split_v_scale_name = lora_pre[type] + key + "v" + suffix + ".scale";
+
+                        auto split_q_alpha_name = lora_pre[type] + key + "q" + suffix + ".alpha";
+                        auto split_k_alpha_name = lora_pre[type] + key + "k" + suffix + ".alpha";
+                        auto split_v_alpha_name = lora_pre[type] + 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);
+                    }
+                }
+                if (starts_with(key, "SPLIT_L|")) {
+                    key = key.substr(sizeof("SPLIT_L|") - 1);
+
+                    auto split_q_d_name = lora_pre[type] + 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 = lora_pre[type] + key + "attn.to_k" + lora_downs[type] + ".weight";
+                        auto split_v_d_name = lora_pre[type] + key + "attn.to_v" + lora_downs[type] + ".weight";
+
+                        auto split_q_u_name = lora_pre[type] + key + "attn.to_q" + lora_ups[type] + ".weight";
+                        auto split_k_u_name = lora_pre[type] + key + "attn.to_k" + lora_ups[type] + ".weight";
+                        auto split_v_u_name = lora_pre[type] + key + "attn.to_v" + lora_ups[type] + ".weight";
+
+                        auto split_m_d_name = lora_pre[type] + key + "proj_mlp" + lora_downs[type] + ".weight";
+                        auto split_m_u_name = lora_pre[type] + key + "proj_mlp" + lora_ups[type] + ".weight";
+
+                        auto split_q_scale_name = lora_pre[type] + key + "attn.to_q" + ".scale";
+                        auto split_k_scale_name = lora_pre[type] + key + "attn.to_k" + ".scale";
+                        auto split_v_scale_name = lora_pre[type] + key + "attn.to_v" + ".scale";
+                        auto split_m_scale_name = lora_pre[type] + key + "proj_mlp" + ".scale";
+
+                        auto split_q_alpha_name = lora_pre[type] + key + "attn.to_q" + ".alpha";
+                        auto split_k_alpha_name = lora_pre[type] + key + "attn.to_k" + ".alpha";
+                        auto split_v_alpha_name = lora_pre[type] + key + "attn.to_v" + ".alpha";
+                        auto split_m_alpha_name = lora_pre[type] + 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);
+                    }
+                }
+                if (lora_up == NULL || lora_down == NULL) {
+                    lora_up_name = lora_pre[type] + key + lora_ups[type] + ".weight";
+                    if (lora_tensors.find(lora_up_name) == lora_tensors.end()) {
+                        if (key == "model_diffusion_model_output_blocks_2_2_conv") {
+                            // fix for some sdxl lora, like lcm-lora-xl
+                            key          = "model_diffusion_model_output_blocks_2_1_conv";
+                            lora_up_name = lora_pre[type] + key + lora_ups[type] + ".weight";
+                        }
+                    }
+
+                    lora_down_name = lora_pre[type] + key + lora_downs[type] + ".weight";
+                    alpha_name     = lora_pre[type] + key + ".alpha";
+                    scale_name     = lora_pre[type] + key + ".scale";
+
+                    if (lora_tensors.find(lora_up_name) != lora_tensors.end()) {
+                        lora_up = lora_tensors[lora_up_name];
+                    }
+
+                    if (lora_tensors.find(lora_down_name) != lora_tensors.end()) {
+                        lora_down = lora_tensors[lora_down_name];
+                    }
+                    applied_lora_tensors.insert(lora_up_name);
+                    applied_lora_tensors.insert(lora_down_name);
+                    applied_lora_tensors.insert(alpha_name);
+                    applied_lora_tensors.insert(scale_name);
+                }
+
+                if (lora_up == NULL || lora_down == NULL) {
+                    continue;
+                }
+                // calc_scale
+                int64_t dim       = lora_down->ne[ggml_n_dims(lora_down) - 1];
+                float scale_value = 1.0f;
+                if (lora_tensors.find(scale_name) != lora_tensors.end()) {
+                    scale_value = ggml_backend_tensor_get_f32(lora_tensors[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 / dim;
+                }
+                scale_value *= multiplier;
+
+                // flat lora tensors to multiply it
+                int64_t lora_up_rows   = lora_up->ne[ggml_n_dims(lora_up) - 1];
+                lora_up                = ggml_reshape_2d(compute_ctx, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
+                int64_t lora_down_rows = lora_down->ne[ggml_n_dims(lora_down) - 1];
+                lora_down              = ggml_reshape_2d(compute_ctx, lora_down, ggml_nelements(lora_down) / lora_down_rows, lora_down_rows);
+
+                // ggml_mul_mat requires tensor b transposed
+                lora_down                  = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, lora_down));
+                struct ggml_tensor* updown = ggml_mul_mat(compute_ctx, lora_up, lora_down);
+                updown                     = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, updown));
+                updown                     = ggml_reshape(compute_ctx, updown, weight);
+                GGML_ASSERT(ggml_nelements(updown) == ggml_nelements(weight));
+                updown = ggml_scale_inplace(compute_ctx, updown, scale_value);
+                ggml_tensor* final_weight;
+                if (weight->type != GGML_TYPE_F32 && weight->type != GGML_TYPE_F16) {
+                    // final_weight = ggml_new_tensor(compute_ctx, GGML_TYPE_F32, ggml_n_dims(weight), weight->ne);
+                    // final_weight = ggml_cpy(compute_ctx, weight, final_weight);
+                    final_weight = to_f32(compute_ctx, weight);
+                    final_weight = ggml_add_inplace(compute_ctx, final_weight, updown);
+                    final_weight = ggml_cpy(compute_ctx, final_weight, weight);
+                } else {
+                    final_weight = ggml_add_inplace(compute_ctx, weight, updown);
+                }
+                // final_weight = ggml_add_inplace(compute_ctx, weight, updown);  // apply directly
+                ggml_build_forward_expand(gf, final_weight);
+                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 have been applied",
+                     applied_lora_tensors_count, total_lora_tensors_count);
+        } else {
+            LOG_DEBUG("(%lu / %lu) LoRA tensors applied successfully",
+                      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, true);
+    }
+};
+
+#endif  // __LORA_HPP__

model.h

@@ -4,27 +4,63 @@
 #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 "json.hpp"
 #include "zip.h"
 
+#define SD_MAX_DIMS 5
+
 enum SDVersion {
-    VERSION_1_x,
-    VERSION_2_x,
-    VERSION_XL,
+    VERSION_SD1,
+    VERSION_SD2,
+    VERSION_SDXL,
+    VERSION_SVD,
+    VERSION_SD3,
+    VERSION_FLUX,
     VERSION_COUNT,
 };
 
+static inline bool sd_version_is_flux(SDVersion version) {
+    if (version == VERSION_FLUX) {
+        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_dit(SDVersion version) {
+    if (sd_version_is_flux(version) || sd_version_is_sd3(version)) {
+        return true;
+    }
+    return false;
+}
+
+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;
+    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
@@ -40,7 +76,11 @@ 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 {
@@ -48,7 +88,7 @@ struct TensorStorage {
     }
 
     int64_t nbytes_to_read() const {
-        if (is_bf16) {
+        if (is_bf16 || is_f8_e4m3 || is_f8_e5m2) {
             return nbytes() / 2;
         } else {
             return nbytes();
@@ -68,6 +108,7 @@ struct TensorStorage {
     std::vector<TensorStorage> chunk(size_t n) {
         std::vector<TensorStorage> chunks;
         size_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++) {
             TensorStorage chunk_i = *this;
@@ -81,7 +122,7 @@ 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 +130,31 @@ 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";
+        }
+        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;
 
 class ModelLoader {
 protected:
@@ -104,7 +166,7 @@ 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 +174,27 @@ protected:
     bool init_from_diffusers_file(const std::string& file_path, const std::string& prefix = "");
 
 public:
+    std::map<std::string, enum ggml_type> tensor_storages_types;
+
     bool init_from_file(const std::string& file_path, const std::string& prefix = "");
     SDVersion get_sd_version();
     ggml_type get_sd_wtype();
-    bool load_vocab(on_new_token_cb_t on_new_token_cb);
-    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb);
-    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, ggml_backend_t backend);
+    bool load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
+                      ggml_backend_t backend,
+                      std::set<std::string> ignore_tensors = {});
+
+    bool save_to_gguf_file(const std::string& file_path, ggml_type type);
+    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_t5_tokenizer_json();
 };
-#endif  // __MODEL_H__
+
+#endif  // __MODEL_H__

pmid.hpp

@@ -0,0 +1,845 @@
+#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;
+    }
+};
+
+/*
+class QFormerPerceiver(nn.Module):
+    def __init__(self, id_embeddings_dim, cross_attention_dim, num_tokens, embedding_dim=1024, use_residual=True, ratio=4):
+        super().__init__()
+
+        self.num_tokens = num_tokens
+        self.cross_attention_dim = cross_attention_dim
+        self.use_residual = use_residual
+        print(cross_attention_dim*num_tokens)
+        self.token_proj = nn.Sequential(
+            nn.Linear(id_embeddings_dim, id_embeddings_dim*ratio),
+            nn.GELU(),
+            nn.Linear(id_embeddings_dim*ratio, cross_attention_dim*num_tokens),
+        )
+        self.token_norm = nn.LayerNorm(cross_attention_dim)
+        self.perceiver_resampler = FacePerceiverResampler(
+            dim=cross_attention_dim,
+            depth=4,
+            dim_head=128,
+            heads=cross_attention_dim // 128,
+            embedding_dim=embedding_dim,
+            output_dim=cross_attention_dim,
+            ff_mult=4,
+        )
+
+    def forward(self, x, last_hidden_state):
+        x = self.token_proj(x)
+        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
+        x = self.token_norm(x) # cls token
+        out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
+        if self.use_residual: # TODO: if use_residual is not true
+            out = x + 1.0 * out
+        return out
+*/
+
+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];
+        // print_ggml_tensor(x, true, "PerceiverAttention reshape x 0: ");
+        // printf("heads = %d \n", heads);
+        // x = ggml_view_4d(ctx, x, x->ne[0], x->ne[1], heads, x->ne[2]/heads,
+        //                          x->nb[1], x->nb[2], x->nb[3], 0);
+        x = ggml_reshape_4d(ctx, x, x->ne[0] / heads, heads, x->ne[1], x->ne[2]);
+        // x = ggml_view_4d(ctx, x, x->ne[0]/heads, heads, x->ne[1], x->ne[2],
+        //                          x->nb[1], x->nb[2], x->nb[3], 0);
+        // x = ggml_cont(ctx, x);
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));
+        // print_ggml_tensor(x, true, "PerceiverAttention reshape x 1: ");
+        // x  = ggml_reshape_4d(ctx, x, ne[0], heads, ne[1], ne[2]/heads);
+        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));
+    }
+
+    /*
+    def forward(self, x, last_hidden_state):
+        x = self.token_proj(x)
+        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
+        x = self.token_norm(x) # cls token
+        out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
+        if self.use_residual: # TODO: if use_residual is not true
+            out = x + 1.0 * out
+        return out
+    */
+
+    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;
+    }
+};
+
+/*
+class FacePerceiverResampler(torch.nn.Module):
+    def __init__(
+        self,
+        *,
+        dim=768,
+        depth=4,
+        dim_head=64,
+        heads=16,
+        embedding_dim=1280,
+        output_dim=768,
+        ff_mult=4,
+    ):
+        super().__init__()
+
+        self.proj_in = torch.nn.Linear(embedding_dim, dim)
+        self.proj_out = torch.nn.Linear(dim, output_dim)
+        self.norm_out = torch.nn.LayerNorm(output_dim)
+        self.layers = torch.nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                torch.nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, latents, x):
+        x = self.proj_in(x)
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+*/
+
+/*
+
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(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"]);
+
+        // print_ggml_tensor(id_embeds, true, "Fuseblock id_embeds: ");
+        // print_ggml_tensor(prompt_embeds, true, "Fuseblock prompt_embeds: ");
+
+        // auto prompt_embeds0 = ggml_cont(ctx, ggml_permute(ctx, prompt_embeds, 2, 0, 1, 3));
+        // auto id_embeds0     = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 2, 0, 1, 3));
+        // print_ggml_tensor(id_embeds0, true, "Fuseblock id_embeds0: ");
+        // print_ggml_tensor(prompt_embeds0, true, "Fuseblock prompt_embeds0: ");
+        // concat is along dim 2
+        // auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds0, id_embeds0, 2);
+        auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds, id_embeds, 0);
+        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 0: ");
+        // stacked_id_embeds      = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 1, 2, 0, 3));
+        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 1: ");
+        // 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 = ggml_nn_layer_norm(ctx, stacked_id_embeds, ln_w, ln_b);
+
+        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);
+
+        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 1: ");
+
+        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
+        // print_ggml_tensor(class_tokens_mask_pos, false);
+        ggml_set_name(class_tokens_mask_pos, "class_tokens_mask_pos");
+        ggml_set_name(prompt_embeds, "prompt_embeds");
+        // print_ggml_tensor(valid_id_embeds, true, "valid_id_embeds");
+        // print_ggml_tensor(class_tokens_mask_pos, true, "class_tokens_mask_pos");
+        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);
+
+        // stacked_id_embeds = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        // print_ggml_tensor(stacked_id_embeds, true, "AA stacked_id_embeds");
+        // print_ggml_tensor(left, true, "AA left");
+        // print_ggml_tensor(right, true, "AA right");
+        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);
+        }
+        // print_ggml_tensor(stacked_id_embeds, true, "BB stacked_id_embeds");
+        // stacked_id_embeds                         = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        // print_ggml_tensor(stacked_id_embeds, true, "CC stacked_id_embeds");
+        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");
+        // print_ggml_tensor(updated_prompt_embeds, true, "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,
+                                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, 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));
+        /*
+        cross_attention_dim = 2048
+        # projection
+        self.num_tokens = 2
+        self.cross_attention_dim = cross_attention_dim
+        self.qformer_perceiver = QFormerPerceiver(
+                                    id_embeddings_dim,
+                                    cross_attention_dim,
+                                    self.num_tokens,
+                                )*/
+        blocks["qformer_perceiver"] = std::shared_ptr<GGMLBlock>(new QFormerPerceiver(id_embeddings_dim,
+                                                                                      cross_attention_dim,
+                                                                                      num_tokens));
+    }
+
+    /*
+    def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds):
+        b, num_inputs, c, h, w = id_pixel_values.shape
+        id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w)
+
+        last_hidden_state = self.vision_model(id_pixel_values)[0]
+        id_embeds = id_embeds.view(b * num_inputs, -1)
+
+        id_embeds = self.qformer_perceiver(id_embeds, last_hidden_state)
+        id_embeds = id_embeds.view(b, num_inputs, self.num_tokens, -1)
+        updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask)
+    */
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* id_pixel_values,
+                                struct ggml_tensor* prompt_embeds,
+                                struct ggml_tensor* class_tokens_mask,
+                                struct ggml_tensor* class_tokens_mask_pos,
+                                struct ggml_tensor* id_embeds,
+                                struct ggml_tensor* left,
+                                struct ggml_tensor* right) {
+        // x: [N, channels, h, w]
+        auto vision_model      = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
+        auto fuse_module       = std::dynamic_pointer_cast<FuseModule>(blocks["fuse_module"]);
+        auto qformer_perceiver = std::dynamic_pointer_cast<QFormerPerceiver>(blocks["qformer_perceiver"]);
+
+        // struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
+        struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values, false);  // [N, hidden_size]
+        id_embeds                             = qformer_perceiver->forward(ctx, id_embeds, last_hidden_state);
+
+        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, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix, SDVersion version = VERSION_SDXL, PMVersion pm_v = PM_VERSION_1, float sty = 20.f)
+        : GGMLRunner(backend),
+          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,
+                                                       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,
+                                                        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,
+                      ModelLoader* ml,
+                      const std::string& file_path = "",
+                      const std::string& prefix    = "")
+        : file_path(file_path), GGMLRunner(backend), 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 = false) {
+        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;
+        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) {
+                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, backend);
+        alloc_params_buffer();
+
+        dry_run = false;
+        model_loader->load_tensors(on_new_tensor_cb, backend);
+
+        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 +1,227 @@
+#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                 = 20 * 1024 * 1024;  // 10
+    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);
+                }
+            }
+        }
+    }
+}
+
+uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
+    struct ggml_init_params params;
+    params.mem_size               = static_cast<size_t>(10 * 1024 * 1024);  // 10
+    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 NULL;
+    }
+
+    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, width, height, 3, 1);
+    struct ggml_tensor* image_gray = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, width, 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 < height; iy++) {
+        for (int ix = 0; ix < 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);
+        }
+    }
+    free(img);
+    uint8_t* output = sd_tensor_to_image(image);
+    ggml_free(work_ctx);
+    return output;
+}
+
+#endif  // __PREPROCESSING_HPP__

rng_philox.hpp

@@ -4,7 +4,7 @@
 #include <cmath>
 #include <vector>
 
-#include "rng.h"
+#include "rng.hpp"
 
 // RNG imitiating torch cuda randn on CPU.
 // Port from: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/5ef669de080814067961f28357256e8fe27544f4/modules/rng_philox.py

stable-diffusion.h

@@ -1,16 +1,39 @@
 #ifndef __STABLE_DIFFUSION_H__
 #define __STABLE_DIFFUSION_H__
 
-#include <memory>
-#include <string>
-#include <vector>
+#if defined(_WIN32) || defined(__CYGWIN__)
+#ifndef SD_BUILD_SHARED_LIB
+#define SD_API
+#else
+#ifdef SD_BUILD_DLL
+#define SD_API __declspec(dllexport)
+#else
+#define SD_API __declspec(dllimport)
+#endif
+#endif
+#else
+#if __GNUC__ >= 4
+#define SD_API __attribute__((visibility("default")))
+#else
+#define SD_API
+#endif
+#endif
 
-enum RNGType {
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+
+enum rng_type_t {
     STD_DEFAULT_RNG,
     CUDA_RNG
 };
 
-enum SampleMethod {
+enum sample_method_t {
     EULER_A,
     EULER,
     HEUN,
@@ -18,58 +41,198 @@ enum SampleMethod {
     DPMPP2S_A,
     DPMPP2M,
     DPMPP2Mv2,
+    IPNDM,
+    IPNDM_V,
     LCM,
     N_SAMPLE_METHODS
 };
 
-enum Schedule {
+enum schedule_t {
     DEFAULT,
     DISCRETE,
     KARRAS,
+    EXPONENTIAL,
+    AYS,
+    GITS,
     N_SCHEDULES
 };
 
-class StableDiffusionGGML;
-
-class StableDiffusion {
-private:
-    std::shared_ptr<StableDiffusionGGML> sd;
-
-public:
-    StableDiffusion(int n_threads                = -1,
-                    bool vae_decode_only         = false,
-                    std::string taesd_path       = "",
-                    bool free_params_immediately = false,
-                    std::string lora_model_dir   = "",
-                    RNGType rng_type             = STD_DEFAULT_RNG);
-    bool load_from_file(const std::string& model_path,
-                        const std::string& vae_path,
-                        ggml_type wtype,
-                        Schedule d = DEFAULT);
-    std::vector<uint8_t*> txt2img(
-        std::string prompt,
-        std::string negative_prompt,
-        float cfg_scale,
-        int width,
-        int height,
-        SampleMethod sample_method,
-        int sample_steps,
-        int64_t seed,
-        int batch_count);
-
-    std::vector<uint8_t*> img2img(
-        const uint8_t* init_img_data,
-        std::string prompt,
-        std::string negative_prompt,
-        float cfg_scale,
-        int width,
-        int height,
-        SampleMethod sample_method,
-        int sample_steps,
-        float strength,
-        int64_t seed);
+// same as enum ggml_type
+enum sd_type_t {
+    SD_TYPE_F32  = 0,
+    SD_TYPE_F16  = 1,
+    SD_TYPE_Q4_0 = 2,
+    SD_TYPE_Q4_1 = 3,
+    // SD_TYPE_Q4_2 = 4, support has been removed
+    // SD_TYPE_Q4_3 = 5, support has been removed
+    SD_TYPE_Q5_0     = 6,
+    SD_TYPE_Q5_1     = 7,
+    SD_TYPE_Q8_0     = 8,
+    SD_TYPE_Q8_1     = 9,
+    SD_TYPE_Q2_K     = 10,
+    SD_TYPE_Q3_K     = 11,
+    SD_TYPE_Q4_K     = 12,
+    SD_TYPE_Q5_K     = 13,
+    SD_TYPE_Q6_K     = 14,
+    SD_TYPE_Q8_K     = 15,
+    SD_TYPE_IQ2_XXS  = 16,
+    SD_TYPE_IQ2_XS   = 17,
+    SD_TYPE_IQ3_XXS  = 18,
+    SD_TYPE_IQ1_S    = 19,
+    SD_TYPE_IQ4_NL   = 20,
+    SD_TYPE_IQ3_S    = 21,
+    SD_TYPE_IQ2_S    = 22,
+    SD_TYPE_IQ4_XS   = 23,
+    SD_TYPE_I8       = 24,
+    SD_TYPE_I16      = 25,
+    SD_TYPE_I32      = 26,
+    SD_TYPE_I64      = 27,
+    SD_TYPE_F64      = 28,
+    SD_TYPE_IQ1_M    = 29,
+    SD_TYPE_BF16     = 30,
+    SD_TYPE_Q4_0_4_4 = 31,
+    SD_TYPE_Q4_0_4_8 = 32,
+    SD_TYPE_Q4_0_8_8 = 33,
+    SD_TYPE_TQ1_0    = 34,
+    SD_TYPE_TQ2_0    = 35,
+    SD_TYPE_COUNT,
 };
 
-std::string sd_get_system_info();
+SD_API const char* sd_type_name(enum sd_type_t type);
 
-#endif  // __STABLE_DIFFUSION_H__
+enum sd_log_level_t {
+    SD_LOG_DEBUG,
+    SD_LOG_INFO,
+    SD_LOG_WARN,
+    SD_LOG_ERROR
+};
+
+typedef void (*sd_log_cb_t)(enum sd_log_level_t level, const char* text, void* data);
+typedef void (*sd_progress_cb_t)(int step, int steps, float time, void* data);
+
+SD_API void sd_set_log_callback(sd_log_cb_t sd_log_cb, void* data);
+SD_API void sd_set_progress_callback(sd_progress_cb_t cb, void* data);
+SD_API int32_t get_num_physical_cores();
+SD_API const char* sd_get_system_info();
+
+typedef struct {
+    uint32_t width;
+    uint32_t height;
+    uint32_t channel;
+    uint8_t* data;
+} sd_image_t;
+
+typedef struct sd_ctx_t sd_ctx_t;
+
+SD_API sd_ctx_t* new_sd_ctx(const char* model_path,
+                            const char* clip_l_path,
+                            const char* clip_g_path,
+                            const char* t5xxl_path,
+                            const char* diffusion_model_path,
+                            const char* vae_path,
+                            const char* taesd_path,
+                            const char* control_net_path_c_str,
+                            const char* lora_model_dir,
+                            const char* embed_dir_c_str,
+                            const char* stacked_id_embed_dir_c_str,
+                            bool vae_decode_only,
+                            bool vae_tiling,
+                            bool free_params_immediately,
+                            int n_threads,
+                            enum sd_type_t wtype,
+                            enum rng_type_t rng_type,
+                            enum schedule_t s,
+                            bool keep_clip_on_cpu,
+                            bool keep_control_net_cpu,
+                            bool keep_vae_on_cpu,
+                            bool diffusion_flash_attn);
+
+SD_API void free_sd_ctx(sd_ctx_t* sd_ctx);
+
+SD_API sd_image_t* txt2img(sd_ctx_t* sd_ctx,
+                           const char* prompt,
+                           const char* negative_prompt,
+                           int clip_skip,
+                           float cfg_scale,
+                           float guidance,
+                           int width,
+                           int height,
+                           enum sample_method_t sample_method,
+                           int sample_steps,
+                           int64_t seed,
+                           int batch_count,
+                           const sd_image_t* control_cond,
+                           float control_strength,
+                           float style_strength,
+                           bool normalize_input,
+                           const char* input_id_images_path,
+                           int* skip_layers,
+                           size_t skip_layers_count,
+                           float slg_scale,
+                           float skip_layer_start,
+                           float skip_layer_end);
+
+SD_API sd_image_t* img2img(sd_ctx_t* sd_ctx,
+                           sd_image_t init_image,
+                           const char* prompt,
+                           const char* negative_prompt,
+                           int clip_skip,
+                           float cfg_scale,
+                           float guidance,
+                           int width,
+                           int height,
+                           enum sample_method_t sample_method,
+                           int sample_steps,
+                           float strength,
+                           int64_t seed,
+                           int batch_count,
+                           const sd_image_t* control_cond,
+                           float control_strength,
+                           float style_strength,
+                           bool normalize_input,
+                           const char* input_id_images_path,
+                           int* skip_layers,
+                           size_t skip_layers_count,
+                           float slg_scale,
+                           float skip_layer_start,
+                           float skip_layer_end);
+
+SD_API sd_image_t* img2vid(sd_ctx_t* sd_ctx,
+                           sd_image_t init_image,
+                           int width,
+                           int height,
+                           int video_frames,
+                           int motion_bucket_id,
+                           int fps,
+                           float augmentation_level,
+                           float min_cfg,
+                           float cfg_scale,
+                           enum sample_method_t sample_method,
+                           int sample_steps,
+                           float strength,
+                           int64_t seed);
+
+typedef struct upscaler_ctx_t upscaler_ctx_t;
+
+SD_API upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path,
+                                        int n_threads);
+SD_API void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx);
+
+SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx, sd_image_t input_image, uint32_t upscale_factor);
+
+SD_API bool convert(const char* input_path, const char* vae_path, const char* output_path, enum sd_type_t output_type);
+
+SD_API uint8_t* preprocess_canny(uint8_t* img,
+                                 int width,
+                                 int height,
+                                 float high_threshold,
+                                 float low_threshold,
+                                 float weak,
+                                 float strong,
+                                 bool inverse);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // __STABLE_DIFFUSION_H__

t5.hpp

@@ -0,0 +1,986 @@
+#ifndef __T5_HPP__
+#define __T5_HPP__
+
+#include <float.h>
+#include <limits>
+#include <map>
+#include <memory>
+#include <regex>
+#include <sstream>
+#include <string>
+#include <unordered_map>
+
+#include "darts.h"
+#include "ggml_extend.hpp"
+#include "json.hpp"
+#include "model.h"
+
+// Port from: https://github.com/google/sentencepiece/blob/master/src/unigram_model.h
+// and https://github.com/google/sentencepiece/blob/master/src/unigram_model.h.
+// Original License: https://github.com/google/sentencepiece/blob/master/LICENSE
+//
+// Since tokenization is not the bottleneck in SD, performance was not a major consideration
+// during the migration.
+class MetaspacePreTokenizer {
+private:
+    std::string replacement;
+    bool add_prefix_space;
+
+public:
+    MetaspacePreTokenizer(const std::string replacement = " ", bool add_prefix_space = true)
+        : replacement(replacement), add_prefix_space(add_prefix_space) {}
+
+    std::string tokenize(const std::string& input) const {
+        std::string tokens;
+        std::stringstream ss(input);
+
+        if (add_prefix_space) {
+            tokens += replacement;
+        }
+
+        std::string token;
+        bool firstToken = true;
+        while (std::getline(ss, token, ' ')) {
+            if (!firstToken)
+                tokens += replacement + token;
+            else
+                tokens += token;
+
+            firstToken = false;
+        }
+
+        return tokens;
+    }
+};
+
+using EncodeResult = std::vector<std::pair<std::string, int>>;
+class T5UniGramTokenizer {
+public:
+    enum Status {
+        OK,
+        NO_PIECES_LOADED,
+        NO_ENTRY_FOUND,
+        BUILD_DOUBLE_ARRAY_FAILED,
+        PIECE_ALREADY_DEFINED,
+        INVLIAD_JSON
+    };
+
+protected:
+    MetaspacePreTokenizer pre_tokenizer;
+
+    // all <piece, score> pairs
+    std::vector<std::pair<std::string, float>> piece_score_pairs;
+
+    float min_score_ = 0.0;
+    float max_score_ = 0.0;
+    std::unique_ptr<Darts::DoubleArray> trie_;
+
+    // Maximum size of the return value of Trie, which corresponds
+    // to the maximum size of shared common prefix in the sentence pieces.
+    int trie_results_size_;
+    // unknown id.
+    int unk_id_            = 2;
+    std::string eos_token_ = "</s>";
+    int eos_id_            = 1;
+    int pad_id_            = 0;
+    // status.
+    Status status_ = OK;
+
+    float kUnkPenalty = 10.0;
+
+    std::string replacement;
+    bool add_prefix_space = true;
+
+    void InitializePieces(const std::string& json_str) {
+        nlohmann::json data;
+
+        try {
+            data = nlohmann::json::parse(json_str);
+        } catch (const nlohmann::json::parse_error& e) {
+            status_ = INVLIAD_JSON;
+            return;
+        }
+        if (!data.contains("model")) {
+            status_ = INVLIAD_JSON;
+            return;
+        }
+        nlohmann::json model = data["model"];
+        if (!model.contains("vocab")) {
+            status_ = INVLIAD_JSON;
+            return;
+        }
+        if (model.contains("unk_id")) {
+            unk_id_ = model["unk_id"];
+        }
+
+        replacement      = data["pre_tokenizer"]["replacement"];
+        add_prefix_space = data["pre_tokenizer"]["add_prefix_space"];
+
+        pre_tokenizer = MetaspacePreTokenizer(replacement, add_prefix_space);
+
+        for (const auto& item : model["vocab"]) {
+            if (item.size() != 2 || !item[0].is_string() || !item[1].is_number_float()) {
+                status_ = INVLIAD_JSON;
+                return;
+            }
+            std::string piece = item[0];
+            float score       = item[1];
+            piece_score_pairs.emplace_back(piece, score);
+        }
+    }
+
+    // Builds a Trie index.
+    void BuildTrie(std::vector<std::pair<std::string, int>>* pieces) {
+        if (status_ != OK)
+            return;
+
+        if (pieces->empty()) {
+            status_ = NO_PIECES_LOADED;
+            return;
+        }
+
+        // sort by sentencepiece since DoubleArray::build()
+        // only accepts sorted strings.
+        sort(pieces->begin(), pieces->end());
+
+        // Makes key/value set for DoubleArrayTrie.
+        std::vector<const char*> key(pieces->size());
+        std::vector<int> value(pieces->size());
+        for (size_t i = 0; i < pieces->size(); ++i) {
+            key[i]   = (*pieces)[i].first.data();  // sorted piece.
+            value[i] = (*pieces)[i].second;        // vocab_id
+        }
+
+        trie_ = std::unique_ptr<Darts::DoubleArray>(new Darts::DoubleArray());
+        if (trie_->build(key.size(), const_cast<char**>(&key[0]), nullptr,
+                         &value[0]) != 0) {
+            status_ = BUILD_DOUBLE_ARRAY_FAILED;
+            return;
+        }
+
+        // Computes the maximum number of shared prefixes in the trie.
+        const int kMaxTrieResultsSize = 1024;
+        std::vector<Darts::DoubleArray::result_pair_type> results(
+            kMaxTrieResultsSize);
+        trie_results_size_ = 0;
+        for (const auto& p : *pieces) {
+            const int num_nodes = trie_->commonPrefixSearch(
+                p.first.data(), results.data(), results.size(), p.first.size());
+            trie_results_size_ = std::max(trie_results_size_, num_nodes);
+        }
+
+        if (trie_results_size_ == 0)
+            status_ = NO_ENTRY_FOUND;
+    }
+
+    // Non-virtual (inlined) implementation for faster execution.
+    inline float GetScoreInlined(int id) const {
+        return piece_score_pairs[id].second;
+    }
+
+    inline bool IsUnusedInlined(int id) const {
+        return false;  // TODO
+    }
+
+    inline bool IsUserDefinedInlined(int id) const {
+        return false;  // TODO
+    }
+
+    inline size_t OneCharLen(const char* src) const {
+        return "\1\1\1\1\1\1\1\1\1\1\1\1\2\2\3\4"[(*src & 0xFF) >> 4];
+    }
+
+    // The optimized Viterbi encode.
+    // Main differences from the original function:
+    // 1. Memorizes the best path at each postion so far,
+    // 2. No need to store the Lattice nodes,
+    // 3. Works in utf-8 directly,
+    // 4. Defines a new struct with fewer fields than Lattice,
+    // 5. Does not depend on `class Lattice` nor call `SetSentence()`,
+    // `PopulateNodes()`, or `Viterbi()`. It does everything in one function.
+    // For detailed explanations please see the comments inside the function body.
+    EncodeResult EncodeOptimized(const std::string& normalized) const {
+        // An optimized Viterbi algorithm for unigram language models. Benchmarking
+        // results show that it generates almost identical outputs and achieves 2.1x
+        // speedup on average for 102 languages compared to the original
+        // implementation. It's based on the following three ideas:
+        //
+        // 1. Because it uses the *unigram* model:
+        //     best_score(x1, x2, …, xt) = best_score(x1, x2, …, x{t-1}) + score(xt)
+        // Deciding the best path (and score) can be decoupled into two isolated
+        // terms: (a) the best path ended before the last token `best_score(x1, x2, …,
+        // x{t-1})`, and (b) the last token and its `score(xt)`. The two terms are
+        // not related to each other at all.
+        //
+        // Therefore, we can compute once and store the *best_path ending at
+        // each character position*. In this way, when we know best_path_ends_at[M],
+        // we can reuse it to compute all the best_path_ends_at_[...] where the last
+        // token starts at the same character position M.
+        //
+        // This improves the time complexity from O(n*k*k) to O(n*k) because it
+        // eliminates the extra loop of recomputing the best path ending at the same
+        // position, where n is the input length and k is the maximum number of tokens
+        // that can be recognized starting at each position.
+        //
+        // 2. Again, because it uses the *unigram* model, we don’t need to actually
+        // store the lattice nodes. We still recognize all the tokens and lattice
+        // nodes from the input, but along identifying them, we use and discard them
+        // on the fly. There is no need to actually store them for best path Viterbi
+        // decoding. The only thing we need to store is the best_path ending at
+        // each character position.
+        //
+        // This improvement reduces the things needed to store in memory from O(n*k)
+        // to O(n), where n is the input length and k is the maximum number of tokens
+        // that can be recognized starting at each position.
+        //
+        // It also avoids the need of dynamic-size lattice node pool, because the
+        // number of things to store is fixed as n.
+        //
+        // 3. SentencePiece is designed to work with unicode, taking utf-8 encoding
+        // inputs. In the original implementation, the lattice positions are based on
+        // unicode positions. A mapping from unicode position to the utf-8 position is
+        // maintained to recover the utf-8 string piece.
+        //
+        // We found that it is sufficient and beneficial to directly work with utf-8
+        // positions:
+        //
+        // Firstly, it saves the conversion and mapping between unicode positions and
+        // utf-8 positions.
+        //
+        // Secondly, it reduces the number of fields we need to maintain in the
+        // node/path structure. Specifically, there are 8 fields defined in
+        // `Lattice::Node` used by the original encoder, but here in the optimized
+        // encoder we only need to define 3 fields in `BestPathNode`.
+
+        if (status() != OK || normalized.empty()) {
+            return {};
+        }
+        // Represents the last node of the best path.
+        struct BestPathNode {
+            int id = -1;  // The vocab id. (maybe -1 for UNK)
+            float best_path_score =
+                0;  // The total score of the best path ending at this node.
+            int starts_at =
+                -1;  // The starting position (in utf-8) of this node. The entire best
+                     // path can be constructed by backtracking along this link.
+        };
+        const int size        = normalized.size();
+        const float unk_score = min_score() - kUnkPenalty;
+        // The ends are exclusive.
+        std::vector<BestPathNode> best_path_ends_at(size + 1);
+        // Generate lattice on-the-fly (not stored) and update best_path_ends_at.
+        int starts_at = 0;
+        while (starts_at < size) {
+            std::size_t node_pos = 0;
+            std::size_t key_pos  = starts_at;
+            const auto best_path_score_till_here =
+                best_path_ends_at[starts_at].best_path_score;
+            bool has_single_node = false;
+            const int mblen =
+                std::min<int>(OneCharLen(normalized.data() + starts_at),
+                              size - starts_at);
+            while (key_pos < size) {
+                const int ret =
+                    trie_->traverse(normalized.data(), node_pos, key_pos, key_pos + 1);
+                if (ret == -2)
+                    break;
+                if (ret >= 0) {
+                    if (IsUnusedInlined(ret))
+                        continue;
+                    // Update the best path node.
+                    auto& target_node = best_path_ends_at[key_pos];
+                    const auto length = (key_pos - starts_at);
+                    // User defined symbol receives extra bonus to always be selected.
+                    const auto score = IsUserDefinedInlined(ret)
+                                           ? (length * max_score_ - 0.1)
+                                           : GetScoreInlined(ret);
+                    const auto candidate_best_path_score =
+                        score + best_path_score_till_here;
+                    if (target_node.starts_at == -1 ||
+                        candidate_best_path_score > target_node.best_path_score) {
+                        target_node.best_path_score = candidate_best_path_score;
+                        target_node.starts_at       = starts_at;
+                        target_node.id              = ret;
+                    }
+                    if (!has_single_node && length == mblen) {
+                        has_single_node = true;
+                    }
+                }
+            }
+            if (!has_single_node) {
+                auto& target_node = best_path_ends_at[starts_at + mblen];
+                const auto candidate_best_path_score =
+                    unk_score + best_path_score_till_here;
+                if (target_node.starts_at == -1 ||
+                    candidate_best_path_score > target_node.best_path_score) {
+                    target_node.best_path_score = candidate_best_path_score;
+                    target_node.starts_at       = starts_at;
+                    target_node.id              = unk_id_;
+                }
+            }
+            // Move by one unicode character.
+            starts_at += mblen;
+        }
+        // Backtrack to identify the best path.
+        EncodeResult results;
+        int ends_at = size;
+        while (ends_at > 0) {
+            const auto& node = best_path_ends_at[ends_at];
+            results.emplace_back(
+                normalized.substr(node.starts_at, ends_at - node.starts_at), node.id);
+            ends_at = node.starts_at;
+        }
+        std::reverse(results.begin(), results.end());
+        return results;
+    }
+
+public:
+    explicit T5UniGramTokenizer(const std::string& json_str = "") {
+        if (json_str.size() != 0) {
+            InitializePieces(json_str);
+        } else {
+            InitializePieces(ModelLoader::load_t5_tokenizer_json());
+        }
+
+        min_score_ = FLT_MAX;
+        max_score_ = FLT_MIN;
+
+        std::vector<std::pair<std::string, int>> pieces;
+        for (int i = 0; i < piece_score_pairs.size(); i++) {
+            const auto& sp = piece_score_pairs[i];
+
+            min_score_ = std::min(min_score_, sp.second);
+            max_score_ = std::max(max_score_, sp.second);
+
+            pieces.emplace_back(sp.first, i);
+        }
+
+        BuildTrie(&pieces);
+    }
+    ~T5UniGramTokenizer(){};
+
+    std::string Normalize(const std::string& input) const {
+        // Ref: https://github.com/huggingface/tokenizers/blob/1ff56c0c70b045f0cd82da1af9ac08cd4c7a6f9f/bindings/python/py_src/tokenizers/implementations/sentencepiece_unigram.py#L29
+        // TODO: nmt-nfkc
+        std::string normalized = std::regex_replace(input, std::regex(" {2,}"), " ");
+        return normalized;
+    }
+
+    std::vector<int> Encode(const std::string& input, bool append_eos_if_not_present = true) const {
+        std::string normalized = Normalize(input);
+        normalized             = pre_tokenizer.tokenize(normalized);
+        EncodeResult result    = EncodeOptimized(normalized);
+        if (result.size() > 0 && append_eos_if_not_present) {
+            auto item = result[result.size() - 1];
+            if (item.first != eos_token_) {
+                result.emplace_back(eos_token_, eos_id_);
+            }
+        }
+        std::vector<int> tokens;
+        for (auto item : result) {
+            tokens.push_back(item.second);
+        }
+        return tokens;
+    }
+
+    void pad_tokens(std::vector<int>& tokens,
+                    std::vector<float>& weights,
+                    size_t max_length = 0,
+                    bool padding      = false) {
+        if (max_length > 0 && padding) {
+            size_t orig_token_num = tokens.size() - 1;
+            size_t n              = std::ceil(orig_token_num * 1.0 / (max_length - 1));
+            if (n == 0) {
+                n = 1;
+            }
+            size_t length = max_length * n;
+            LOG_DEBUG("token length: %llu", length);
+            std::vector<int> new_tokens;
+            std::vector<float> new_weights;
+            int token_idx = 0;
+            for (int i = 0; i < length; i++) {
+                if (token_idx >= orig_token_num) {
+                    break;
+                }
+                if (i % max_length == max_length - 1) {
+                    new_tokens.push_back(eos_id_);
+                    new_weights.push_back(1.0);
+                } else {
+                    new_tokens.push_back(tokens[token_idx]);
+                    new_weights.push_back(weights[token_idx]);
+                    token_idx++;
+                }
+            }
+
+            new_tokens.push_back(eos_id_);
+            new_weights.push_back(1.0);
+            tokens  = new_tokens;
+            weights = new_weights;
+
+            if (padding) {
+                int pad_token_id = pad_id_;
+                tokens.insert(tokens.end(), length - tokens.size(), pad_token_id);
+                weights.insert(weights.end(), length - weights.size(), 1.0);
+            }
+        }
+    }
+
+    // Returns the minimum score in sentence pieces.
+    // min_score() - 10 is used for the cost of unknown sentence.
+    float min_score() const { return min_score_; }
+
+    // Returns the maximum score in sentence pieces.
+    // max_score() is used for the cost of user defined symbols.
+    float max_score() const { return max_score_; }
+
+    Status status() const { return status_; }
+};
+
+class T5LayerNorm : public UnaryBlock {
+protected:
+    int64_t hidden_size;
+    float eps;
+
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
+        params["weight"]     = ggml_new_tensor_1d(ctx, wtype, hidden_size);
+    }
+
+public:
+    T5LayerNorm(int64_t hidden_size,
+                float eps = 1e-06f)
+        : hidden_size(hidden_size),
+          eps(eps) {}
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        struct ggml_tensor* w = params["weight"];
+        x                     = ggml_rms_norm(ctx, x, eps);
+        x                     = ggml_mul(ctx, x, w);
+        return x;
+    }
+};
+
+struct T5DenseActDense : public UnaryBlock {
+public:
+    T5DenseActDense(int64_t model_dim, int64_t ff_dim) {
+        blocks["wi"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false));
+        blocks["wo"] = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, n_token, model_dim]
+        auto wi = std::dynamic_pointer_cast<Linear>(blocks["wi"]);
+        auto wo = std::dynamic_pointer_cast<Linear>(blocks["wo"]);
+
+        x = wi->forward(ctx, x);
+        x = ggml_relu_inplace(ctx, x);
+        x = wo->forward(ctx, x);
+        return x;
+    }
+};
+
+struct T5DenseGatedActDense : public UnaryBlock {
+public:
+    T5DenseGatedActDense(int64_t model_dim, int64_t ff_dim) {
+        blocks["wi_0"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false));
+        blocks["wi_1"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false));
+        blocks["wo"]   = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, n_token, model_dim]
+        auto wi_0 = std::dynamic_pointer_cast<Linear>(blocks["wi_0"]);
+        auto wi_1 = std::dynamic_pointer_cast<Linear>(blocks["wi_1"]);
+        auto wo   = std::dynamic_pointer_cast<Linear>(blocks["wo"]);
+
+        auto hidden_gelu   = ggml_gelu_inplace(ctx, wi_0->forward(ctx, x));
+        auto hidden_linear = wi_1->forward(ctx, x);
+        x                  = ggml_mul_inplace(ctx, hidden_gelu, hidden_linear);
+        x                  = wo->forward(ctx, x);
+        return x;
+    }
+};
+
+struct T5LayerFF : public UnaryBlock {
+public:
+    T5LayerFF(int64_t model_dim, int64_t ff_dim) {
+        blocks["DenseReluDense"] = std::shared_ptr<GGMLBlock>(new T5DenseGatedActDense(model_dim, ff_dim));
+        blocks["layer_norm"]     = std::shared_ptr<GGMLBlock>(new T5LayerNorm(model_dim));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, n_token, model_dim]
+        auto DenseReluDense = std::dynamic_pointer_cast<T5DenseGatedActDense>(blocks["DenseReluDense"]);
+        auto layer_norm     = std::dynamic_pointer_cast<T5LayerNorm>(blocks["layer_norm"]);
+
+        auto forwarded_states = layer_norm->forward(ctx, x);
+        forwarded_states      = DenseReluDense->forward(ctx, forwarded_states);
+        x                     = ggml_add_inplace(ctx, forwarded_states, x);
+        return x;
+    }
+};
+
+class T5Attention : public GGMLBlock {
+protected:
+    int64_t model_dim;
+    int64_t inner_dim;
+    int64_t num_heads;
+    bool using_relative_attention_bias;
+    int64_t relative_attention_num_buckets  = 32;
+    int64_t relative_attention_max_distance = 128;
+
+public:
+    T5Attention(int64_t model_dim,
+                int64_t inner_dim,
+                int64_t num_heads,
+                bool using_relative_attention_bias = false)
+        : model_dim(model_dim),
+          inner_dim(inner_dim),
+          num_heads(num_heads),
+          using_relative_attention_bias(using_relative_attention_bias) {
+        blocks["q"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, inner_dim, false));
+        blocks["k"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, inner_dim, false));
+        blocks["v"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, inner_dim, false));
+        blocks["o"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, model_dim, false));
+        if (using_relative_attention_bias) {
+            blocks["relative_attention_bias"] = std::shared_ptr<GGMLBlock>(new Embedding(relative_attention_num_buckets, num_heads));
+        }
+    }
+
+    struct ggml_tensor* compute_bias(struct ggml_context* ctx,
+                                     struct ggml_tensor* relative_position_bucket) {
+        auto relative_attention_bias = std::dynamic_pointer_cast<Embedding>(blocks["relative_attention_bias"]);
+
+        auto values = relative_attention_bias->forward(ctx, relative_position_bucket);  // shape (query_length, key_length, num_heads)
+        values      = ggml_cont(ctx, ggml_permute(ctx, values, 2, 0, 1, 3));            // shape (1, num_heads, query_length, key_length)
+        return values;
+    }
+
+    // x: [N, n_token, model_dim]
+    std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
+                                                                struct ggml_tensor* x,
+                                                                struct ggml_tensor* past_bias                = NULL,
+                                                                struct ggml_tensor* mask                     = NULL,
+                                                                struct ggml_tensor* relative_position_bucket = NULL) {
+        auto q_proj   = std::dynamic_pointer_cast<Linear>(blocks["q"]);
+        auto k_proj   = std::dynamic_pointer_cast<Linear>(blocks["k"]);
+        auto v_proj   = std::dynamic_pointer_cast<Linear>(blocks["v"]);
+        auto out_proj = std::dynamic_pointer_cast<Linear>(blocks["o"]);
+
+        int64_t n_head = num_heads;
+        int64_t d_head = inner_dim / n_head;
+
+        auto q = q_proj->forward(ctx, x);
+        auto k = k_proj->forward(ctx, x);
+        auto v = v_proj->forward(ctx, x);
+
+        if (using_relative_attention_bias && relative_position_bucket != NULL) {
+            past_bias = compute_bias(ctx, relative_position_bucket);
+        }
+        if (past_bias != NULL) {
+            if (mask != NULL) {
+                mask = ggml_add(ctx, mask, past_bias);
+            } else {
+                mask = past_bias;
+            }
+        }
+
+        k = ggml_scale_inplace(ctx, k, sqrt(d_head));
+
+        x = ggml_nn_attention_ext(ctx, q, k, v, num_heads, mask);  // [N, n_token, d_head * n_head]
+
+        x = out_proj->forward(ctx, x);  // [N, n_token, model_dim]
+        return {x, past_bias};
+    }
+};
+
+struct T5LayerSelfAttention : public GGMLBlock {
+public:
+    T5LayerSelfAttention(int64_t model_dim,
+                         int64_t inner_dim,
+                         int64_t ff_dim,
+                         int64_t num_heads,
+                         bool using_relative_attention_bias) {
+        blocks["SelfAttention"] = std::shared_ptr<GGMLBlock>(new T5Attention(model_dim, inner_dim, num_heads, using_relative_attention_bias));
+        blocks["layer_norm"]    = std::shared_ptr<GGMLBlock>(new T5LayerNorm(model_dim));
+    }
+
+    std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
+                                                                struct ggml_tensor* x,
+                                                                struct ggml_tensor* past_bias                = NULL,
+                                                                struct ggml_tensor* mask                     = NULL,
+                                                                struct ggml_tensor* relative_position_bucket = NULL) {
+        // x: [N, n_token, model_dim]
+        auto SelfAttention = std::dynamic_pointer_cast<T5Attention>(blocks["SelfAttention"]);
+        auto layer_norm    = std::dynamic_pointer_cast<T5LayerNorm>(blocks["layer_norm"]);
+
+        auto normed_hidden_state = layer_norm->forward(ctx, x);
+        auto ret                 = SelfAttention->forward(ctx, normed_hidden_state, past_bias, mask, relative_position_bucket);
+        auto output              = ret.first;
+        past_bias                = ret.second;
+
+        x = ggml_add_inplace(ctx, output, x);
+        return {x, past_bias};
+    }
+};
+
+struct T5Block : public GGMLBlock {
+public:
+    T5Block(int64_t model_dim, int64_t inner_dim, int64_t ff_dim, int64_t num_heads, bool using_relative_attention_bias) {
+        blocks["layer.0"] = std::shared_ptr<GGMLBlock>(new T5LayerSelfAttention(model_dim, inner_dim, ff_dim, num_heads, using_relative_attention_bias));
+        blocks["layer.1"] = std::shared_ptr<GGMLBlock>(new T5LayerFF(model_dim, ff_dim));
+    }
+
+    std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
+                                                                struct ggml_tensor* x,
+                                                                struct ggml_tensor* past_bias                = NULL,
+                                                                struct ggml_tensor* mask                     = NULL,
+                                                                struct ggml_tensor* relative_position_bucket = NULL) {
+        // x: [N, n_token, model_dim]
+        auto layer_0 = std::dynamic_pointer_cast<T5LayerSelfAttention>(blocks["layer.0"]);
+        auto layer_1 = std::dynamic_pointer_cast<T5LayerFF>(blocks["layer.1"]);
+
+        auto ret  = layer_0->forward(ctx, x, past_bias, mask, relative_position_bucket);
+        x         = ret.first;
+        past_bias = ret.second;
+        x         = layer_1->forward(ctx, x);
+        return {x, past_bias};
+    }
+};
+
+struct T5Stack : public GGMLBlock {
+    int64_t num_layers;
+
+public:
+    T5Stack(int64_t num_layers,
+            int64_t model_dim,
+            int64_t inner_dim,
+            int64_t ff_dim,
+            int64_t num_heads)
+        : num_layers(num_layers) {
+        for (int i = 0; i < num_layers; i++) {
+            blocks["block." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new T5Block(model_dim, inner_dim, ff_dim, num_heads, i == 0));
+        }
+
+        blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new T5LayerNorm(model_dim));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* past_bias                = NULL,
+                                struct ggml_tensor* attention_mask           = NULL,
+                                struct ggml_tensor* relative_position_bucket = NULL) {
+        // x: [N, n_token, model_dim]
+        for (int i = 0; i < num_layers; i++) {
+            auto block = std::dynamic_pointer_cast<T5Block>(blocks["block." + std::to_string(i)]);
+
+            auto ret  = block->forward(ctx, x, past_bias, attention_mask, relative_position_bucket);
+            x         = ret.first;
+            past_bias = ret.second;
+        }
+
+        auto final_layer_norm = std::dynamic_pointer_cast<T5LayerNorm>(blocks["final_layer_norm"]);
+
+        x = final_layer_norm->forward(ctx, x);
+        return x;
+    }
+};
+
+struct T5 : public GGMLBlock {
+public:
+    T5(int64_t num_layers,
+       int64_t model_dim,
+       int64_t ff_dim,
+       int64_t num_heads,
+       int64_t vocab_size) {
+        blocks["encoder"] = std::shared_ptr<GGMLBlock>(new T5Stack(num_layers, model_dim, model_dim, ff_dim, num_heads));
+        blocks["shared"]  = std::shared_ptr<GGMLBlock>(new Embedding(vocab_size, model_dim));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* input_ids,
+                                struct ggml_tensor* past_bias                = NULL,
+                                struct ggml_tensor* attention_mask           = NULL,
+                                struct ggml_tensor* relative_position_bucket = NULL) {
+        // input_ids: [N, n_token]
+
+        auto shared  = std::dynamic_pointer_cast<Embedding>(blocks["shared"]);
+        auto encoder = std::dynamic_pointer_cast<T5Stack>(blocks["encoder"]);
+
+        auto x = shared->forward(ctx, input_ids);
+        x      = encoder->forward(ctx, x, past_bias, attention_mask, relative_position_bucket);
+        return x;
+    }
+};
+
+struct T5Runner : public GGMLRunner {
+    T5 model;
+    std::vector<int> relative_position_bucket_vec;
+
+    T5Runner(ggml_backend_t backend,
+             std::map<std::string, enum ggml_type>& tensor_types,
+             const std::string prefix,
+             int64_t num_layers = 24,
+             int64_t model_dim  = 4096,
+             int64_t ff_dim     = 10240,
+             int64_t num_heads  = 64,
+             int64_t vocab_size = 32128)
+        : GGMLRunner(backend), model(num_layers, model_dim, ff_dim, num_heads, vocab_size) {
+        model.init(params_ctx, tensor_types, prefix);
+    }
+
+    std::string get_desc() {
+        return "t5";
+    }
+
+    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,
+                                struct ggml_tensor* input_ids,
+                                struct ggml_tensor* relative_position_bucket) {
+        size_t N       = input_ids->ne[1];
+        size_t n_token = input_ids->ne[0];
+
+        auto hidden_states = model.forward(ctx, input_ids, NULL, NULL, relative_position_bucket);  // [N, n_token, model_dim]
+        return hidden_states;
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids) {
+        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+
+        input_ids = to_backend(input_ids);
+
+        relative_position_bucket_vec = compute_relative_position_bucket(input_ids->ne[0], input_ids->ne[0]);
+
+        // for (int i = 0; i < relative_position_bucket_vec.size(); i++) {
+        //     if (i % 77 == 0) {
+        //         printf("\n");
+        //     }
+        //     printf("%d ", relative_position_bucket_vec[i]);
+        // }
+
+        auto relative_position_bucket = ggml_new_tensor_2d(compute_ctx,
+                                                           GGML_TYPE_I32,
+                                                           input_ids->ne[0],
+                                                           input_ids->ne[0]);
+        set_backend_tensor_data(relative_position_bucket, relative_position_bucket_vec.data());
+
+        struct ggml_tensor* hidden_states = forward(compute_ctx, input_ids, relative_position_bucket);
+
+        ggml_build_forward_expand(gf, hidden_states);
+
+        return gf;
+    }
+
+    void compute(const int n_threads,
+                 struct ggml_tensor* input_ids,
+                 ggml_tensor** output,
+                 ggml_context* output_ctx = NULL) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(input_ids);
+        };
+        GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
+    }
+
+    static std::vector<int> _relative_position_bucket(const std::vector<int>& relative_position,
+                                                      bool bidirectional = true,
+                                                      int num_buckets    = 32,
+                                                      int max_distance   = 128) {
+        std::vector<int> relative_buckets(relative_position.size(), 0);
+        std::vector<int> abs_relative_position = relative_position;
+
+        if (bidirectional) {
+            num_buckets = num_buckets / 2;
+            for (size_t i = 0; i < relative_position.size(); ++i) {
+                if (relative_position[i] > 0) {
+                    relative_buckets[i] += num_buckets;
+                }
+                abs_relative_position[i] = std::abs(relative_position[i]);
+            }
+        } else {
+            for (size_t i = 0; i < relative_position.size(); ++i) {
+                abs_relative_position[i] = std::max(-relative_position[i], 0);
+            }
+        }
+
+        int max_exact = num_buckets / 2;
+        std::vector<int> relative_position_if_large(relative_position.size(), 0);
+
+        for (size_t i = 0; i < relative_position.size(); ++i) {
+            if (abs_relative_position[i] < max_exact) {
+                relative_buckets[i] += abs_relative_position[i];
+            } else {
+                float log_pos                 = std::log(static_cast<float>(abs_relative_position[i]) / max_exact);
+                float log_base                = std::log(static_cast<float>(max_distance) / max_exact);
+                relative_position_if_large[i] = max_exact + static_cast<int>((log_pos / log_base) * (num_buckets - max_exact));
+                relative_position_if_large[i] = std::min(relative_position_if_large[i], num_buckets - 1);
+                relative_buckets[i] += relative_position_if_large[i];
+            }
+        }
+
+        return relative_buckets;
+    }
+
+    std::vector<int> compute_relative_position_bucket(int query_length,
+                                                      int key_length) {
+        std::vector<int> context_position(query_length);
+        std::vector<int> memory_position(key_length);
+
+        for (int i = 0; i < query_length; ++i) {
+            context_position[i] = i;
+        }
+        for (int i = 0; i < key_length; ++i) {
+            memory_position[i] = i;
+        }
+
+        std::vector<std::vector<int>> relative_position(query_length, std::vector<int>(key_length, 0));
+        for (int i = 0; i < query_length; ++i) {
+            for (int j = 0; j < key_length; ++j) {
+                relative_position[i][j] = memory_position[j] - context_position[i];
+            }
+        }
+
+        std::vector<int> relative_position_bucket;
+        for (int i = 0; i < query_length; ++i) {
+            std::vector<int> result = _relative_position_bucket(relative_position[i], true);
+            relative_position_bucket.insert(relative_position_bucket.end(), result.begin(), result.end());
+        }
+
+        return relative_position_bucket;
+    }
+};
+
+struct T5Embedder {
+    T5UniGramTokenizer tokenizer;
+    T5Runner model;
+
+    static std::map<std::string, enum ggml_type> empty_tensor_types;
+
+    T5Embedder(ggml_backend_t backend,
+               std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types,
+               const std::string prefix                            = "",
+               int64_t num_layers                                  = 24,
+               int64_t model_dim                                   = 4096,
+               int64_t ff_dim                                      = 10240,
+               int64_t num_heads                                   = 64,
+               int64_t vocab_size                                  = 32128)
+        : model(backend, tensor_types, prefix, num_layers, model_dim, ff_dim, num_heads, vocab_size) {
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        model.get_param_tensors(tensors, prefix);
+    }
+
+    void alloc_params_buffer() {
+        model.alloc_params_buffer();
+    }
+
+    std::pair<std::vector<int>, std::vector<float>> tokenize(std::string text,
+                                                             size_t max_length = 0,
+                                                             bool padding      = false) {
+        auto parsed_attention = parse_prompt_attention(text);
+
+        {
+            std::stringstream ss;
+            ss << "[";
+            for (const auto& item : parsed_attention) {
+                ss << "['" << item.first << "', " << item.second << "], ";
+            }
+            ss << "]";
+            LOG_DEBUG("parse '%s' to %s", text.c_str(), ss.str().c_str());
+        }
+
+        std::vector<int> tokens;
+        std::vector<float> weights;
+        for (const auto& item : parsed_attention) {
+            const std::string& curr_text = item.first;
+            float curr_weight            = item.second;
+            std::vector<int> curr_tokens = tokenizer.Encode(curr_text, false);
+            tokens.insert(tokens.end(), curr_tokens.begin(), curr_tokens.end());
+            weights.insert(weights.end(), curr_tokens.size(), curr_weight);
+        }
+
+        int EOS_TOKEN_ID = 1;
+        tokens.push_back(EOS_TOKEN_ID);
+        weights.push_back(1.0);
+
+        tokenizer.pad_tokens(tokens, weights, max_length, padding);
+
+        // for (int i = 0; i < tokens.size(); i++) {
+        //     std::cout << tokens[i] << ":" << weights[i] << ", ";
+        // }
+        // std::cout << std::endl;
+
+        return {tokens, weights};
+    }
+
+    void test() {
+        struct ggml_init_params params;
+        params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
+        params.mem_buffer = NULL;
+        params.no_alloc   = false;
+
+        struct ggml_context* work_ctx = ggml_init(params);
+        GGML_ASSERT(work_ctx != NULL);
+
+        {
+            // cpu f16: pass
+            // cpu f32: pass
+            // cuda f16: nan
+            // cuda f32: pass
+            // cuda q8_0: nan
+            // TODO: fix cuda nan
+            std::string text("a lovely cat");
+            auto tokens_and_weights     = tokenize(text, 77, true);
+            std::vector<int>& tokens    = tokens_and_weights.first;
+            std::vector<float>& weights = tokens_and_weights.second;
+            for (auto token : tokens) {
+                printf("%d ", token);
+            }
+            printf("\n");
+            auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
+            struct ggml_tensor* out = NULL;
+
+            int t0 = ggml_time_ms();
+            model.compute(8, input_ids, &out, work_ctx);
+            int t1 = ggml_time_ms();
+
+            print_ggml_tensor(out);
+            LOG_DEBUG("t5 test done in %dms", t1 - t0);
+        }
+    }
+
+    static void load_from_file_and_test(const std::string& file_path) {
+        // ggml_backend_t backend    = ggml_backend_cuda_init(0);
+        ggml_backend_t backend         = ggml_backend_cpu_init();
+        ggml_type model_data_type      = GGML_TYPE_F32;
+        std::shared_ptr<T5Embedder> t5 = std::shared_ptr<T5Embedder>(new T5Embedder(backend));
+        {
+            LOG_INFO("loading from '%s'", file_path.c_str());
+
+            t5->alloc_params_buffer();
+            std::map<std::string, ggml_tensor*> tensors;
+            t5->get_param_tensors(tensors, "");
+
+            ModelLoader model_loader;
+            if (!model_loader.init_from_file(file_path)) {
+                LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
+                return;
+            }
+
+            bool success = model_loader.load_tensors(tensors, backend);
+
+            if (!success) {
+                LOG_ERROR("load tensors from model loader failed");
+                return;
+            }
+
+            LOG_INFO("t5 model loaded");
+        }
+        t5->test();
+    }
+};
+
+#endif  // __T5_HPP__

tae.hpp

@@ -0,0 +1,252 @@
+#ifndef __TAE_HPP__
+#define __TAE_HPP__
+
+#include "ggml_extend.hpp"
+
+#include "model.h"
+
+/*
+    ===================================    TinyAutoEncoder  ===================================
+    References:
+    https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/autoencoders/vae.py
+    https://github.com/madebyollin/taesd/blob/main/taesd.py
+
+*/
+
+class TAEBlock : public UnaryBlock {
+protected:
+    int n_in;
+    int n_out;
+
+public:
+    TAEBlock(int n_in, int n_out)
+        : n_in(n_in), n_out(n_out) {
+        blocks["conv.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_in, n_out, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_out, n_out, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.4"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_out, n_out, {3, 3}, {1, 1}, {1, 1}));
+        if (n_in != n_out) {
+            blocks["skip"] = std::shared_ptr<GGMLBlock>(new Conv2d(n_in, n_out, {1, 1}, {1, 1}, {1, 1}, {1, 1}, false));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, n_in, h, w]
+        // return: [n, n_out, h, w]
+
+        auto conv_0 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.0"]);
+        auto conv_2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.2"]);
+        auto conv_4 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.4"]);
+
+        auto h = conv_0->forward(ctx, x);
+        h      = ggml_relu_inplace(ctx, h);
+        h      = conv_2->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx, h);
+        h      = conv_4->forward(ctx, h);
+
+        if (n_in != n_out) {
+            auto skip = std::dynamic_pointer_cast<Conv2d>(blocks["skip"]);
+            LOG_DEBUG("skip");
+            x = skip->forward(ctx, x);
+        }
+
+        h = ggml_add(ctx, h, x);
+        h = ggml_relu_inplace(ctx, h);
+        return h;
+    }
+};
+
+class TinyEncoder : public UnaryBlock {
+    int in_channels = 3;
+    int channels    = 64;
+    int z_channels  = 4;
+    int num_blocks  = 3;
+
+public:
+    TinyEncoder() {
+        int index                       = 0;
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, channels, {3, 3}, {1, 1}, {1, 1}));
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, z_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [n, in_channels, h, w]
+        // return: [n, z_channels, h/8, w/8]
+
+        for (int i = 0; i < num_blocks * 3 + 6; i++) {
+            auto block = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(i)]);
+
+            x = block->forward(ctx, x);
+        }
+
+        return x;
+    }
+};
+
+class TinyDecoder : public UnaryBlock {
+    int z_channels   = 4;
+    int channels     = 64;
+    int out_channels = 3;
+    int num_blocks   = 3;
+
+public:
+    TinyDecoder(int index = 0) {
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, channels, {3, 3}, {1, 1}, {1, 1}));
+        index++;  // nn.ReLU()
+
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+        index++;  // nn.Upsample()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+        index++;  // nn.Upsample()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+
+        for (int i = 0; i < num_blocks; i++) {
+            blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        }
+        index++;  // nn.Upsample()
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels, {3, 3}, {1, 1}, {1, 1}, {1, 1}, false));
+
+        blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
+        blocks[std::to_string(index++)] = 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* z) {
+        // z: [n, z_channels, h, w]
+        // return: [n, out_channels, h*8, w*8]
+
+        auto h = ggml_scale(ctx, z, 1.0f / 3.0f);
+        h      = ggml_tanh_inplace(ctx, h);
+        h      = ggml_scale(ctx, h, 3.0f);
+
+        for (int i = 0; i < num_blocks * 3 + 10; i++) {
+            if (blocks.find(std::to_string(i)) == blocks.end()) {
+                if (i == 1) {
+                    h = ggml_relu_inplace(ctx, h);
+                } else {
+                    h = ggml_upscale(ctx, h, 2);
+                }
+                continue;
+            }
+            auto block = std::dynamic_pointer_cast<UnaryBlock>(blocks[std::to_string(i)]);
+
+            h = block->forward(ctx, h);
+        }
+
+        return h;
+    }
+};
+
+class TAESD : public GGMLBlock {
+protected:
+    bool decode_only;
+
+public:
+    TAESD(bool decode_only = true)
+        : decode_only(decode_only) {
+        blocks["decoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyDecoder());
+
+        if (!decode_only) {
+            blocks["encoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyEncoder());
+        }
+    }
+
+    struct ggml_tensor* decode(struct ggml_context* ctx, struct ggml_tensor* z) {
+        auto decoder = std::dynamic_pointer_cast<TinyDecoder>(blocks["decoder.layers"]);
+        return decoder->forward(ctx, z);
+    }
+
+    struct ggml_tensor* encode(struct ggml_context* ctx, struct ggml_tensor* x) {
+        auto encoder = std::dynamic_pointer_cast<TinyEncoder>(blocks["encoder.layers"]);
+        return encoder->forward(ctx, x);
+    }
+};
+
+struct TinyAutoEncoder : public GGMLRunner {
+    TAESD taesd;
+    bool decode_only = false;
+
+    TinyAutoEncoder(ggml_backend_t backend,
+                    std::map<std::string, enum ggml_type>& tensor_types,
+                    const std::string prefix,
+                    bool decoder_only = true)
+        : decode_only(decoder_only),
+          taesd(decode_only),
+          GGMLRunner(backend) {
+        taesd.init(params_ctx, tensor_types, prefix);
+    }
+
+    std::string get_desc() {
+        return "taesd";
+    }
+
+    bool load_from_file(const std::string& file_path) {
+        LOG_INFO("loading taesd from '%s', decode_only = %s", file_path.c_str(), decode_only ? "true" : "false");
+        alloc_params_buffer();
+        std::map<std::string, ggml_tensor*> taesd_tensors;
+        taesd.get_param_tensors(taesd_tensors);
+        std::set<std::string> ignore_tensors;
+        if (decode_only) {
+            ignore_tensors.insert("encoder.");
+        }
+
+        ModelLoader model_loader;
+        if (!model_loader.init_from_file(file_path)) {
+            LOG_ERROR("init taesd model loader from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        bool success = model_loader.load_tensors(taesd_tensors, backend, ignore_tensors);
+
+        if (!success) {
+            LOG_ERROR("load tae tensors from model loader failed");
+            return false;
+        }
+
+        LOG_INFO("taesd model loaded");
+        return success;
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
+        struct ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        z                       = to_backend(z);
+        struct ggml_tensor* out = decode_graph ? taesd.decode(compute_ctx, z) : taesd.encode(compute_ctx, z);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    void compute(const int n_threads,
+                 struct ggml_tensor* z,
+                 bool decode_graph,
+                 struct ggml_tensor** output,
+                 struct ggml_context* output_ctx = NULL) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(z, decode_graph);
+        };
+
+        GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+    }
+};
+
+#endif  // __TAE_HPP__

thirdparty/.clang-format

@@ -0,0 +1,2 @@
+DisableFormat: true
+SortIncludes: Never

thirdparty/LICENSE.darts_clone.txt

@@ -0,0 +1,10 @@
+Copyright (c) 2008-2011, Susumu Yata
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
+
+- Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. 
+- Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. 
+- Neither the name of the <ORGANIZATION> nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

thirdparty/README.md

@@ -1,2 +1,3 @@
 - json.hpp library from: https://github.com/nlohmann/json
-- ZIP Library from: https://github.com/kuba--/zip
+- ZIP Library from: https://github.com/kuba--/zip
+- darts.h from: https://github.com/google/sentencepiece/tree/master/third_party/darts_clone

thirdparty/zip.c

@@ -36,6 +36,7 @@
 #include <unistd.h>
 #endif
 
+#define USE_EXTERNAL_MZCRC
 #include "miniz.h"
 #include "zip.h"
 
@@ -1834,3 +1835,234 @@ int zip_extract(const char *zipname, const char *dir,
 
   return zip_archive_extract(&zip_archive, dir, on_extract, arg);
 }
+
+#if defined(__SSE4_2__) || defined(__AVX512F__)
+#include <immintrin.h>
+#endif
+
+// Phil Katz 32-Bit Cyclic Redundancy Check Uber Alles
+// Goes 73 GiB/s on an AMD Ryzen Threadripper PRO 7995WX
+// "Fast CRC Computation for Generic Polynomials Using PCLMULQDQ Instruction"
+//  V. Gopal, E. Ozturk, et al., 2009, http://intel.ly/2ySEwL0
+mz_ulong mz_crc32(mz_ulong init, const uint8_t *buf, size_t len) {
+  uint32_t crc = ~init;
+#if defined(__AVX512F__) && defined(__VPCLMULQDQ__) && defined(__PCLMUL__)
+  if (len >= 256) {
+    _Alignas(__m512) static const uint64_t k1k2[] = {
+        0x011542778a, 0x01322d1430, 0x011542778a, 0x01322d1430,
+        0x011542778a, 0x01322d1430, 0x011542778a, 0x01322d1430,
+    };
+    _Alignas(__m512) static const uint64_t k3k4[] = {
+        0x0154442bd4, 0x01c6e41596, 0x0154442bd4, 0x01c6e41596,
+        0x0154442bd4, 0x01c6e41596, 0x0154442bd4, 0x01c6e41596,
+    };
+    _Alignas(__m512) static const uint64_t k5k6[] = {
+        0x01751997d0,
+        0x00ccaa009e,
+    };
+    _Alignas(__m512) static const uint64_t k7k8[] = {
+        0x0163cd6124,
+        0x0000000000,
+    };
+    _Alignas(__m512) static const uint64_t poly[] = {
+        0x01db710641,
+        0x01f7011641,
+    };
+    __m512i x0, x1, x2, x3, x4, x5, x6, x7, x8, y5, y6, y7, y8;
+    __m128i a0, a1, a2, a3;
+    x1 = _mm512_loadu_si512((__m512i *)(buf + 0x00));
+    x2 = _mm512_loadu_si512((__m512i *)(buf + 0x40));
+    x3 = _mm512_loadu_si512((__m512i *)(buf + 0x80));
+    x4 = _mm512_loadu_si512((__m512i *)(buf + 0xC0));
+    x1 = _mm512_xor_si512(x1, _mm512_castsi128_si512(_mm_cvtsi32_si128(crc)));
+    x0 = _mm512_load_si512((__m512i *)k1k2);
+    buf += 256;
+    len -= 256;
+    while (len >= 256) {
+      x5 = _mm512_clmulepi64_epi128(x1, x0, 0x00);
+      x6 = _mm512_clmulepi64_epi128(x2, x0, 0x00);
+      x7 = _mm512_clmulepi64_epi128(x3, x0, 0x00);
+      x8 = _mm512_clmulepi64_epi128(x4, x0, 0x00);
+      x1 = _mm512_clmulepi64_epi128(x1, x0, 0x11);
+      x2 = _mm512_clmulepi64_epi128(x2, x0, 0x11);
+      x3 = _mm512_clmulepi64_epi128(x3, x0, 0x11);
+      x4 = _mm512_clmulepi64_epi128(x4, x0, 0x11);
+      y5 = _mm512_loadu_si512((__m512i *)(buf + 0x00));
+      y6 = _mm512_loadu_si512((__m512i *)(buf + 0x40));
+      y7 = _mm512_loadu_si512((__m512i *)(buf + 0x80));
+      y8 = _mm512_loadu_si512((__m512i *)(buf + 0xC0));
+      x1 = _mm512_xor_si512(x1, x5);
+      x2 = _mm512_xor_si512(x2, x6);
+      x3 = _mm512_xor_si512(x3, x7);
+      x4 = _mm512_xor_si512(x4, x8);
+      x1 = _mm512_xor_si512(x1, y5);
+      x2 = _mm512_xor_si512(x2, y6);
+      x3 = _mm512_xor_si512(x3, y7);
+      x4 = _mm512_xor_si512(x4, y8);
+      buf += 256;
+      len -= 256;
+    }
+    x0 = _mm512_load_si512((__m512i *)k3k4);
+    x5 = _mm512_clmulepi64_epi128(x1, x0, 0x00);
+    x1 = _mm512_clmulepi64_epi128(x1, x0, 0x11);
+    x1 = _mm512_xor_si512(x1, x2);
+    x1 = _mm512_xor_si512(x1, x5);
+    x5 = _mm512_clmulepi64_epi128(x1, x0, 0x00);
+    x1 = _mm512_clmulepi64_epi128(x1, x0, 0x11);
+    x1 = _mm512_xor_si512(x1, x3);
+    x1 = _mm512_xor_si512(x1, x5);
+    x5 = _mm512_clmulepi64_epi128(x1, x0, 0x00);
+    x1 = _mm512_clmulepi64_epi128(x1, x0, 0x11);
+    x1 = _mm512_xor_si512(x1, x4);
+    x1 = _mm512_xor_si512(x1, x5);
+    while (len >= 64) {
+      x2 = _mm512_loadu_si512((__m512i *)buf);
+      x5 = _mm512_clmulepi64_epi128(x1, x0, 0x00);
+      x1 = _mm512_clmulepi64_epi128(x1, x0, 0x11);
+      x1 = _mm512_xor_si512(x1, x2);
+      x1 = _mm512_xor_si512(x1, x5);
+      buf += 64;
+      len -= 64;
+    }
+    a0 = _mm_load_si128((__m128i *)k5k6);
+    a1 = _mm512_extracti32x4_epi32(x1, 0);
+    a2 = _mm512_extracti32x4_epi32(x1, 1);
+    a3 = _mm_clmulepi64_si128(a1, a0, 0x00);
+    a1 = _mm_clmulepi64_si128(a1, a0, 0x11);
+    a1 = _mm_xor_si128(a1, a3);
+    a1 = _mm_xor_si128(a1, a2);
+    a2 = _mm512_extracti32x4_epi32(x1, 2);
+    a3 = _mm_clmulepi64_si128(a1, a0, 0x00);
+    a1 = _mm_clmulepi64_si128(a1, a0, 0x11);
+    a1 = _mm_xor_si128(a1, a3);
+    a1 = _mm_xor_si128(a1, a2);
+    a2 = _mm512_extracti32x4_epi32(x1, 3);
+    a3 = _mm_clmulepi64_si128(a1, a0, 0x00);
+    a1 = _mm_clmulepi64_si128(a1, a0, 0x11);
+    a1 = _mm_xor_si128(a1, a3);
+    a1 = _mm_xor_si128(a1, a2);
+    a2 = _mm_clmulepi64_si128(a1, a0, 0x10);
+    a3 = _mm_setr_epi32(~0, 0, ~0, 0);
+    a1 = _mm_srli_si128(a1, 8);
+    a1 = _mm_xor_si128(a1, a2);
+    a0 = _mm_loadl_epi64((__m128i *)k7k8);
+    a2 = _mm_srli_si128(a1, 4);
+    a1 = _mm_and_si128(a1, a3);
+    a1 = _mm_clmulepi64_si128(a1, a0, 0x00);
+    a1 = _mm_xor_si128(a1, a2);
+    a0 = _mm_load_si128((__m128i *)poly);
+    a2 = _mm_and_si128(a1, a3);
+    a2 = _mm_clmulepi64_si128(a2, a0, 0x10);
+    a2 = _mm_and_si128(a2, a3);
+    a2 = _mm_clmulepi64_si128(a2, a0, 0x00);
+    a1 = _mm_xor_si128(a1, a2);
+    crc = _mm_extract_epi32(a1, 1);
+  }
+#endif
+#if defined(__SSE4_2__) && defined(__PCLMUL__)
+  if (len >= 64) {
+    _Alignas(__m128) static const uint64_t k1k2[] = {
+        0x0154442bd4,
+        0x01c6e41596,
+    };
+    _Alignas(__m128) static const uint64_t k3k4[] = {
+        0x01751997d0,
+        0x00ccaa009e,
+    };
+    _Alignas(__m128) static const uint64_t k5k0[] = {
+        0x0163cd6124,
+        0x0000000000,
+    };
+    _Alignas(__m128) static const uint64_t poly[] = {
+        0x01db710641,
+        0x01f7011641,
+    };
+    __m128i x0, x1, x2, x3, x4, x5, x6, x7, x8, y5, y6, y7, y8;
+    x1 = _mm_loadu_si128((__m128i *)(buf + 0x00));
+    x2 = _mm_loadu_si128((__m128i *)(buf + 0x10));
+    x3 = _mm_loadu_si128((__m128i *)(buf + 0x20));
+    x4 = _mm_loadu_si128((__m128i *)(buf + 0x30));
+    x1 = _mm_xor_si128(x1, _mm_cvtsi32_si128(crc));
+    x0 = _mm_load_si128((__m128i *)k1k2);
+    buf += 64;
+    len -= 64;
+    while (len >= 64) {
+      x5 = _mm_clmulepi64_si128(x1, x0, 0x00);
+      x6 = _mm_clmulepi64_si128(x2, x0, 0x00);
+      x7 = _mm_clmulepi64_si128(x3, x0, 0x00);
+      x8 = _mm_clmulepi64_si128(x4, x0, 0x00);
+      x1 = _mm_clmulepi64_si128(x1, x0, 0x11);
+      x2 = _mm_clmulepi64_si128(x2, x0, 0x11);
+      x3 = _mm_clmulepi64_si128(x3, x0, 0x11);
+      x4 = _mm_clmulepi64_si128(x4, x0, 0x11);
+      y5 = _mm_loadu_si128((__m128i *)(buf + 0x00));
+      y6 = _mm_loadu_si128((__m128i *)(buf + 0x10));
+      y7 = _mm_loadu_si128((__m128i *)(buf + 0x20));
+      y8 = _mm_loadu_si128((__m128i *)(buf + 0x30));
+      x1 = _mm_xor_si128(x1, x5);
+      x2 = _mm_xor_si128(x2, x6);
+      x3 = _mm_xor_si128(x3, x7);
+      x4 = _mm_xor_si128(x4, x8);
+      x1 = _mm_xor_si128(x1, y5);
+      x2 = _mm_xor_si128(x2, y6);
+      x3 = _mm_xor_si128(x3, y7);
+      x4 = _mm_xor_si128(x4, y8);
+      buf += 64;
+      len -= 64;
+    }
+    x0 = _mm_load_si128((__m128i *)k3k4);
+    x5 = _mm_clmulepi64_si128(x1, x0, 0x00);
+    x1 = _mm_clmulepi64_si128(x1, x0, 0x11);
+    x1 = _mm_xor_si128(x1, x2);
+    x1 = _mm_xor_si128(x1, x5);
+    x5 = _mm_clmulepi64_si128(x1, x0, 0x00);
+    x1 = _mm_clmulepi64_si128(x1, x0, 0x11);
+    x1 = _mm_xor_si128(x1, x3);
+    x1 = _mm_xor_si128(x1, x5);
+    x5 = _mm_clmulepi64_si128(x1, x0, 0x00);
+    x1 = _mm_clmulepi64_si128(x1, x0, 0x11);
+    x1 = _mm_xor_si128(x1, x4);
+    x1 = _mm_xor_si128(x1, x5);
+    while (len >= 16) {
+      x2 = _mm_loadu_si128((__m128i *)buf);
+      x5 = _mm_clmulepi64_si128(x1, x0, 0x00);
+      x1 = _mm_clmulepi64_si128(x1, x0, 0x11);
+      x1 = _mm_xor_si128(x1, x2);
+      x1 = _mm_xor_si128(x1, x5);
+      buf += 16;
+      len -= 16;
+    }
+    x2 = _mm_clmulepi64_si128(x1, x0, 0x10);
+    x3 = _mm_setr_epi32(~0, 0, ~0, 0);
+    x1 = _mm_srli_si128(x1, 8);
+    x1 = _mm_xor_si128(x1, x2);
+    x0 = _mm_loadl_epi64((__m128i *)k5k0);
+    x2 = _mm_srli_si128(x1, 4);
+    x1 = _mm_and_si128(x1, x3);
+    x1 = _mm_clmulepi64_si128(x1, x0, 0x00);
+    x1 = _mm_xor_si128(x1, x2);
+    x0 = _mm_load_si128((__m128i *)poly);
+    x2 = _mm_and_si128(x1, x3);
+    x2 = _mm_clmulepi64_si128(x2, x0, 0x10);
+    x2 = _mm_and_si128(x2, x3);
+    x2 = _mm_clmulepi64_si128(x2, x0, 0x00);
+    x1 = _mm_xor_si128(x1, x2);
+    crc = _mm_extract_epi32(x1, 1);
+  }
+#endif
+  static uint32_t tab[256];
+  if (!tab[255]) {
+    // generates table for byte-wise crc calculation on the polynomial
+    // x^32+x^26+x^23+x^22+x^16+x^12+x^11+x^10+x^8+x^7+x^5+x^4+x^2+x+1
+    uint32_t polynomial = 0xedb88320;  // bits are reversed
+    for (int d = 0; d < 256; ++d) {
+      uint32_t r = d;
+      for (int i = 0; i < 8; ++i)
+        r = r >> 1 ^ (r & 1 ? polynomial : 0);
+      tab[d] = r;
+    }
+  }
+  for (size_t i = 0; i < len; ++i)
+    crc = crc >> 8 ^ tab[(crc & 255) ^ buf[i]];
+  return ~crc & 0xffffffff;
+}

unet.hpp

@@ -0,0 +1,654 @@
+#ifndef __UNET_HPP__
+#define __UNET_HPP__
+
+#include "common.hpp"
+#include "ggml_extend.hpp"
+#include "model.h"
+
+/*==================================================== UnetModel =====================================================*/
+
+#define UNET_GRAPH_SIZE 10240
+
+class SpatialVideoTransformer : public SpatialTransformer {
+protected:
+    int64_t time_depth;
+    int64_t max_time_embed_period;
+
+public:
+    SpatialVideoTransformer(int64_t in_channels,
+                            int64_t n_head,
+                            int64_t d_head,
+                            int64_t depth,
+                            int64_t context_dim,
+                            int64_t time_depth            = 1,
+                            int64_t max_time_embed_period = 10000)
+        : SpatialTransformer(in_channels, n_head, d_head, depth, context_dim),
+          max_time_embed_period(max_time_embed_period) {
+        // We will convert unet transformer linear to conv2d 1x1 when loading the weights, so use_linear is always False
+        // use_spatial_context is always True
+        // merge_strategy is always learned_with_images
+        // merge_factor is loaded from weights
+        // time_context_dim is always None
+        // ff_in is always True
+        // disable_self_attn is always False
+        // disable_temporal_crossattention is always False
+
+        int64_t inner_dim = n_head * d_head;
+
+        GGML_ASSERT(depth == time_depth);
+        GGML_ASSERT(in_channels == inner_dim);
+
+        int64_t time_mix_d_head    = d_head;
+        int64_t n_time_mix_heads   = n_head;
+        int64_t time_mix_inner_dim = time_mix_d_head * n_time_mix_heads;  // equal to inner_dim
+        int64_t time_context_dim   = context_dim;
+
+        for (int i = 0; i < time_depth; i++) {
+            std::string name = "time_stack." + std::to_string(i);
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new BasicTransformerBlock(inner_dim,
+                                                                                    n_time_mix_heads,
+                                                                                    time_mix_d_head,
+                                                                                    time_context_dim,
+                                                                                    true));
+        }
+
+        int64_t time_embed_dim     = in_channels * 4;
+        blocks["time_pos_embed.0"] = std::shared_ptr<GGMLBlock>(new Linear(in_channels, time_embed_dim));
+        // time_pos_embed.1 is nn.SiLU()
+        blocks["time_pos_embed.2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, in_channels));
+
+        blocks["time_mixer"] = std::shared_ptr<GGMLBlock>(new AlphaBlender());
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* context,
+                                int timesteps) {
+        // x: [N, in_channels, h, w] aka [b*t, in_channels, h, w], t == timesteps
+        // context: [N, max_position(aka n_context), hidden_size(aka context_dim)] aka [b*t, n_context, context_dim], t == timesteps
+        // t_emb: [N, in_channels] aka [b*t, in_channels]
+        // timesteps is num_frames
+        // time_context is always None
+        // image_only_indicator is always tensor([0.])
+        // transformer_options is not used
+        // GGML_ASSERT(ggml_n_dims(context) == 3);
+
+        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 time_pos_embed_0 = std::dynamic_pointer_cast<Linear>(blocks["time_pos_embed.0"]);
+        auto time_pos_embed_2 = std::dynamic_pointer_cast<Linear>(blocks["time_pos_embed.2"]);
+        auto time_mixer       = std::dynamic_pointer_cast<AlphaBlender>(blocks["time_mixer"]);
+
+        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;
+
+        GGML_ASSERT(n == timesteps);  // We compute cond and uncond separately, so batch_size==1
+
+        auto time_context    = context;  // [b*t, n_context, context_dim]
+        auto spatial_context = context;
+        // time_context_first_timestep = time_context[::timesteps]
+        auto time_context_first_timestep = ggml_view_3d(ctx,
+                                                        time_context,
+                                                        time_context->ne[0],
+                                                        time_context->ne[1],
+                                                        1,
+                                                        time_context->nb[1],
+                                                        time_context->nb[2],
+                                                        0);  // [b, n_context, context_dim]
+        time_context                     = ggml_new_tensor_3d(ctx, GGML_TYPE_F32,
+                                                              time_context_first_timestep->ne[0],
+                                                              time_context_first_timestep->ne[1],
+                                                              time_context_first_timestep->ne[2] * h * w);
+        time_context                     = ggml_repeat(ctx, time_context_first_timestep, time_context);  // [b*h*w, n_context, context_dim]
+
+        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]
+
+        auto num_frames = ggml_arange(ctx, 0, timesteps, 1);
+        // since b is 1, no need to do repeat
+        auto t_emb = ggml_nn_timestep_embedding(ctx, num_frames, in_channels, max_time_embed_period);  // [N, in_channels]
+
+        auto emb = time_pos_embed_0->forward(ctx, t_emb);
+        emb      = ggml_silu_inplace(ctx, emb);
+        emb      = time_pos_embed_2->forward(ctx, emb);                   // [N, in_channels]
+        emb      = ggml_reshape_3d(ctx, emb, emb->ne[0], 1, emb->ne[1]);  // [N, 1, in_channels]
+
+        for (int i = 0; i < depth; i++) {
+            std::string transformer_name = "transformer_blocks." + std::to_string(i);
+            std::string time_stack_name  = "time_stack." + std::to_string(i);
+
+            auto block     = std::dynamic_pointer_cast<BasicTransformerBlock>(blocks[transformer_name]);
+            auto mix_block = std::dynamic_pointer_cast<BasicTransformerBlock>(blocks[time_stack_name]);
+
+            x = block->forward(ctx, x, spatial_context);  // [N, h * w, inner_dim]
+
+            // in_channels == inner_dim
+            auto x_mix = x;
+            x_mix      = ggml_add(ctx, x_mix, emb);  // [N, h * w, inner_dim]
+
+            int64_t N = x_mix->ne[2];
+            int64_t T = timesteps;
+            int64_t B = N / T;
+            int64_t S = x_mix->ne[1];
+            int64_t C = x_mix->ne[0];
+
+            x_mix = ggml_reshape_4d(ctx, x_mix, C, S, T, B);               // (b t) s c -> b t s c
+            x_mix = ggml_cont(ctx, ggml_permute(ctx, x_mix, 0, 2, 1, 3));  // b t s c -> b s t c
+            x_mix = ggml_reshape_3d(ctx, x_mix, C, T, S * B);              // b s t c -> (b s) t c
+
+            x_mix = mix_block->forward(ctx, x_mix, time_context);  // [B * h * w, T, inner_dim]
+
+            x_mix = ggml_reshape_4d(ctx, x_mix, C, T, S, B);               // (b s) t c -> b s t c
+            x_mix = ggml_cont(ctx, ggml_permute(ctx, x_mix, 0, 2, 1, 3));  // b s t c -> b t s c
+            x_mix = ggml_reshape_3d(ctx, x_mix, C, S, T * B);              // b t s c -> (b t) s c
+
+            x = time_mixer->forward(ctx, x, x_mix);  // [N, h * w, inner_dim]
+        }
+
+        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;
+    }
+};
+
+// ldm.modules.diffusionmodules.openaimodel.UNetModel
+class UnetModelBlock : public GGMLBlock {
+protected:
+    SDVersion version = VERSION_SD1;
+    // network hparams
+    int in_channels                        = 4;
+    int out_channels                       = 4;
+    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/SVD
+
+    UnetModelBlock(SDVersion version = VERSION_SD1, bool flash_attn = false)
+        : version(version) {
+        if (version == VERSION_SD2) {
+            context_dim       = 1024;
+            num_head_channels = 64;
+            num_heads         = -1;
+        } else if (version == VERSION_SDXL) {
+            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;
+        }
+        // dims is always 2
+        // use_temporal_attention is always True for SVD
+
+        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 (version == VERSION_SDXL || 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* {
+            if (version == VERSION_SVD) {
+                return new VideoResBlock(channels, emb_channels, out_channels);
+            } else {
+                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* {
+            if (version == VERSION_SVD) {
+                return new SpatialVideoTransformer(in_channels, n_head, d_head, depth, context_dim);
+            } else {
+                return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim, flash_attn);
+            }
+        };
+
+        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));
+                }
+                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));
+
+                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));
+
+        // output_blocks
+        int output_block_idx = 0;
+        for (int i = (int)len_mults - 1; i >= 0; i--) {
+            int mult = channel_mult[i];
+            for (int j = 0; j < num_res_blocks + 1; j++) {
+                int ich = input_block_chans.back();
+                input_block_chans.pop_back();
+
+                std::string name = "output_blocks." + std::to_string(output_block_idx) + ".0";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(get_resblock(ch + ich, time_embed_dim, mult * model_channels));
+
+                ch                = mult * model_channels;
+                int up_sample_idx = 1;
+                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 = "output_blocks." + std::to_string(output_block_idx) + ".1";
+                    blocks[name]     = std::shared_ptr<GGMLBlock>(get_attention_layer(ch, n_head, d_head, transformer_depth[i], context_dim));
+
+                    up_sample_idx++;
+                }
+
+                if (i > 0 && j == num_res_blocks) {
+                    std::string name = "output_blocks." + std::to_string(output_block_idx) + "." + std::to_string(up_sample_idx);
+                    blocks[name]     = std::shared_ptr<GGMLBlock>(new UpSampleBlock(ch, ch));
+
+                    ds /= 2;
+                }
+
+                output_block_idx += 1;
+            }
+        }
+
+        // out
+        blocks["out.0"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(ch));  // ch == model_channels
+        // out_1 is nn.SiLU()
+        blocks["out.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(model_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    struct ggml_tensor* resblock_forward(std::string name,
+                                         struct ggml_context* ctx,
+                                         struct ggml_tensor* x,
+                                         struct ggml_tensor* emb,
+                                         int num_video_frames) {
+        if (version == VERSION_SVD) {
+            auto block = std::dynamic_pointer_cast<VideoResBlock>(blocks[name]);
+
+            return block->forward(ctx, x, emb, num_video_frames);
+        } else {
+            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,
+                                                struct ggml_tensor* x,
+                                                struct ggml_tensor* context,
+                                                int timesteps) {
+        if (version == VERSION_SVD) {
+            auto block = std::dynamic_pointer_cast<SpatialVideoTransformer>(blocks[name]);
+
+            return block->forward(ctx, x, context, timesteps);
+        } else {
+            auto block = std::dynamic_pointer_cast<SpatialTransformer>(blocks[name]);
+
+            return block->forward(ctx, x, context);
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* timesteps,
+                                struct ggml_tensor* context,
+                                struct ggml_tensor* c_concat              = NULL,
+                                struct ggml_tensor* y                     = NULL,
+                                int num_video_frames                      = -1,
+                                std::vector<struct ggml_tensor*> controls = {},
+                                float control_strength                    = 0.f) {
+        // x: [N, in_channels, h, w] or [N, in_channels/2, h, w]
+        // timesteps: [N,]
+        // context: [N, max_position, hidden_size] or [1, max_position, hidden_size]. for example, [N, 77, 768]
+        // c_concat: [N, in_channels, h, w] or [1, in_channels, h, w]
+        // y: [N, adm_in_channels] or [1, adm_in_channels]
+        // return: [N, out_channels, h, w]
+        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 (c_concat != NULL) {
+            if (c_concat->ne[3] != x->ne[3]) {
+                c_concat = ggml_repeat(ctx, c_concat, x);
+            }
+            x = ggml_concat(ctx, x, c_concat, 2);
+        }
+
+        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 out_0 = std::dynamic_pointer_cast<GroupNorm32>(blocks["out.0"]);
+        auto out_2 = std::dynamic_pointer_cast<Conv2d>(blocks["out.2"]);
+
+        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]
+        }
+
+        // input_blocks
+        std::vector<struct ggml_tensor*> hs;
+
+        // input block 0
+        auto h = input_blocks_0_0->forward(ctx, x);
+
+        ggml_set_name(h, "bench-start");
+        hs.push_back(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, num_video_frames);  // [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, h, context, num_video_frames);  // [N, mult*model_channels, h, w]
+                }
+                hs.push_back(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))]
+                hs.push_back(h);
+            }
+        }
+        // [N, 4*model_channels, h/8, w/8]
+
+        // middle_block
+        h = resblock_forward("middle_block.0", ctx, h, emb, num_video_frames);             // [N, 4*model_channels, h/8, w/8]
+        h = attention_layer_forward("middle_block.1", ctx, h, context, num_video_frames);  // [N, 4*model_channels, h/8, w/8]
+        h = resblock_forward("middle_block.2", ctx, h, emb, num_video_frames);             // [N, 4*model_channels, h/8, w/8]
+
+        if (controls.size() > 0) {
+            auto cs = ggml_scale_inplace(ctx, controls[controls.size() - 1], control_strength);
+            h       = ggml_add(ctx, h, cs);  // middle control
+        }
+        int control_offset = controls.size() - 2;
+
+        // output_blocks
+        int output_block_idx = 0;
+        for (int i = (int)len_mults - 1; i >= 0; i--) {
+            for (int j = 0; j < num_res_blocks + 1; j++) {
+                auto h_skip = hs.back();
+                hs.pop_back();
+
+                if (controls.size() > 0) {
+                    auto cs = ggml_scale_inplace(ctx, controls[control_offset], control_strength);
+                    h_skip  = ggml_add(ctx, h_skip, cs);  // control net condition
+                    control_offset--;
+                }
+
+                h = ggml_concat(ctx, h, h_skip, 2);
+
+                std::string name = "output_blocks." + std::to_string(output_block_idx) + ".0";
+
+                h = resblock_forward(name, ctx, h, emb, num_video_frames);
+
+                int up_sample_idx = 1;
+                if (std::find(attention_resolutions.begin(), attention_resolutions.end(), ds) != attention_resolutions.end()) {
+                    std::string name = "output_blocks." + std::to_string(output_block_idx) + ".1";
+
+                    h = attention_layer_forward(name, ctx, h, context, num_video_frames);
+
+                    up_sample_idx++;
+                }
+
+                if (i > 0 && j == num_res_blocks) {
+                    std::string name = "output_blocks." + std::to_string(output_block_idx) + "." + std::to_string(up_sample_idx);
+                    auto block       = std::dynamic_pointer_cast<UpSampleBlock>(blocks[name]);
+
+                    h = block->forward(ctx, h);
+
+                    ds /= 2;
+                }
+
+                output_block_idx += 1;
+            }
+        }
+
+        // out
+        h = out_0->forward(ctx, h);
+        h = ggml_silu_inplace(ctx, h);
+        h = out_2->forward(ctx, h);
+        ggml_set_name(h, "bench-end");
+        return h;  // [N, out_channels, h, w]
+    }
+};
+
+struct UNetModelRunner : public GGMLRunner {
+    UnetModelBlock unet;
+
+    UNetModelRunner(ggml_backend_t backend,
+                    std::map<std::string, enum ggml_type>& tensor_types,
+                    const std::string prefix,
+                    SDVersion version = VERSION_SD1,
+                    bool flash_attn   = false)
+        : GGMLRunner(backend), unet(version, flash_attn) {
+        unet.init(params_ctx, tensor_types, prefix);
+    }
+
+    std::string get_desc() {
+        return "unet";
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        unet.get_param_tensors(tensors, prefix);
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+                                    struct ggml_tensor* timesteps,
+                                    struct ggml_tensor* context,
+                                    struct ggml_tensor* c_concat              = NULL,
+                                    struct ggml_tensor* y                     = NULL,
+                                    int num_video_frames                      = -1,
+                                    std::vector<struct ggml_tensor*> controls = {},
+                                    float control_strength                    = 0.f) {
+        struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, UNET_GRAPH_SIZE, false);
+
+        if (num_video_frames == -1) {
+            num_video_frames = x->ne[3];
+        }
+
+        x         = to_backend(x);
+        context   = to_backend(context);
+        y         = to_backend(y);
+        timesteps = to_backend(timesteps);
+
+        for (int i = 0; i < controls.size(); i++) {
+            controls[i] = to_backend(controls[i]);
+        }
+
+        struct ggml_tensor* out = unet.forward(compute_ctx,
+                                               x,
+                                               timesteps,
+                                               context,
+                                               c_concat,
+                                               y,
+                                               num_video_frames,
+                                               controls,
+                                               control_strength);
+
+        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,
+                 struct ggml_tensor* c_concat,
+                 struct ggml_tensor* y,
+                 int num_video_frames                      = -1,
+                 std::vector<struct ggml_tensor*> controls = {},
+                 float control_strength                    = 0.f,
+                 struct ggml_tensor** output               = 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]
+        // c_concat: [N, in_channels, h, w] or [1, in_channels, h, w]
+        // y: [N, adm_in_channels] or [1, adm_in_channels]
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(x, timesteps, context, c_concat, y, num_video_frames, controls, control_strength);
+        };
+
+        GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+    }
+
+    void test() {
+        struct ggml_init_params params;
+        params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
+        params.mem_buffer = NULL;
+        params.no_alloc   = false;
+
+        struct ggml_context* work_ctx = ggml_init(params);
+        GGML_ASSERT(work_ctx != NULL);
+
+        {
+            // CPU, num_video_frames = 1, x{num_video_frames, 8, 8, 8}: Pass
+            // CUDA, num_video_frames = 1, x{num_video_frames, 8, 8, 8}: Pass
+            // CPU, num_video_frames = 3, x{num_video_frames, 8, 8, 8}: Wrong result
+            // CUDA, num_video_frames = 3, x{num_video_frames, 8, 8, 8}: nan
+            int num_video_frames = 3;
+
+            auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 8, 8, 8, num_video_frames);
+            std::vector<float> timesteps_vec(num_video_frames, 999.f);
+            auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
+            ggml_set_f32(x, 0.5f);
+            // print_ggml_tensor(x);
+
+            auto context = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 1024, 1, num_video_frames);
+            ggml_set_f32(context, 0.5f);
+            // print_ggml_tensor(context);
+
+            auto y = ggml_new_tensor_2d(work_ctx, GGML_TYPE_F32, 768, num_video_frames);
+            ggml_set_f32(y, 0.5f);
+            // print_ggml_tensor(y);
+
+            struct ggml_tensor* out = NULL;
+
+            int t0 = ggml_time_ms();
+            compute(8, x, timesteps, context, NULL, y, num_video_frames, {}, 0.f, &out, work_ctx);
+            int t1 = ggml_time_ms();
+
+            print_ggml_tensor(out);
+            LOG_DEBUG("unet test done in %dms", t1 - t0);
+        }
+    }
+};
+
+#endif  // __UNET_HPP__

upscaler.cpp

@@ -0,0 +1,134 @@
+#include "esrgan.hpp"
+#include "ggml_extend.hpp"
+#include "model.h"
+#include "stable-diffusion.h"
+
+struct UpscalerGGML {
+    ggml_backend_t backend    = NULL;  // general backend
+    ggml_type model_data_type = GGML_TYPE_F16;
+    std::shared_ptr<ESRGAN> esrgan_upscaler;
+    std::string esrgan_path;
+    int n_threads;
+
+    UpscalerGGML(int n_threads)
+        : n_threads(n_threads) {
+    }
+
+    bool load_from_file(const std::string& esrgan_path) {
+#ifdef SD_USE_CUBLAS
+        LOG_DEBUG("Using CUDA backend");
+        backend = ggml_backend_cuda_init(0);
+#endif
+#ifdef SD_USE_METAL
+        LOG_DEBUG("Using Metal backend");
+        ggml_backend_metal_log_set_callback(ggml_log_callback_default, nullptr);
+        backend = ggml_backend_metal_init();
+#endif
+#ifdef SD_USE_VULKAN
+        LOG_DEBUG("Using Vulkan backend");
+        backend = ggml_backend_vk_init(0);
+#endif
+#ifdef SD_USE_SYCL
+        LOG_DEBUG("Using SYCL backend");
+        backend = ggml_backend_sycl_init(0);
+#endif
+        ModelLoader model_loader;
+        if (!model_loader.init_from_file(esrgan_path)) {
+            LOG_ERROR("init model loader from file failed: '%s'", esrgan_path.c_str());
+        }
+        model_loader.set_wtype_override(model_data_type);
+        if (!backend) {
+            LOG_DEBUG("Using CPU backend");
+            backend = ggml_backend_cpu_init();
+        }
+        LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
+        esrgan_upscaler = std::make_shared<ESRGAN>(backend, model_loader.tensor_storages_types);
+        if (!esrgan_upscaler->load_from_file(esrgan_path)) {
+            return false;
+        }
+        return true;
+    }
+
+    sd_image_t upscale(sd_image_t input_image, uint32_t upscale_factor) {
+        // upscale_factor, unused for RealESRGAN_x4plus_anime_6B.pth
+        sd_image_t upscaled_image = {0, 0, 0, NULL};
+        int output_width          = (int)input_image.width * esrgan_upscaler->scale;
+        int output_height         = (int)input_image.height * esrgan_upscaler->scale;
+        LOG_INFO("upscaling from (%i x %i) to (%i x %i)",
+                 input_image.width, input_image.height, output_width, output_height);
+
+        struct ggml_init_params params;
+        params.mem_size = output_width * output_height * 3 * sizeof(float) * 2;
+        params.mem_size += 2 * ggml_tensor_overhead();
+        params.mem_buffer = NULL;
+        params.no_alloc   = false;
+
+        // draft context
+        struct ggml_context* upscale_ctx = ggml_init(params);
+        if (!upscale_ctx) {
+            LOG_ERROR("ggml_init() failed");
+            return upscaled_image;
+        }
+        LOG_DEBUG("upscale work buffer size: %.2f MB", params.mem_size / 1024.f / 1024.f);
+        ggml_tensor* input_image_tensor = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, input_image.width, input_image.height, 3, 1);
+        sd_image_to_tensor(input_image.data, input_image_tensor);
+
+        ggml_tensor* upscaled = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, output_width, output_height, 3, 1);
+        auto on_tiling        = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
+            esrgan_upscaler->compute(n_threads, in, &out);
+        };
+        int64_t t0 = ggml_time_ms();
+        sd_tiling(input_image_tensor, upscaled, esrgan_upscaler->scale, esrgan_upscaler->tile_size, 0.25f, on_tiling);
+        esrgan_upscaler->free_compute_buffer();
+        ggml_tensor_clamp(upscaled, 0.f, 1.f);
+        uint8_t* upscaled_data = sd_tensor_to_image(upscaled);
+        ggml_free(upscale_ctx);
+        int64_t t3 = ggml_time_ms();
+        LOG_INFO("input_image_tensor upscaled, taking %.2fs", (t3 - t0) / 1000.0f);
+        upscaled_image = {
+            (uint32_t)output_width,
+            (uint32_t)output_height,
+            3,
+            upscaled_data,
+        };
+        return upscaled_image;
+    }
+};
+
+struct upscaler_ctx_t {
+    UpscalerGGML* upscaler = NULL;
+};
+
+upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path_c_str,
+                                 int n_threads) {
+    upscaler_ctx_t* upscaler_ctx = (upscaler_ctx_t*)malloc(sizeof(upscaler_ctx_t));
+    if (upscaler_ctx == NULL) {
+        return NULL;
+    }
+    std::string esrgan_path(esrgan_path_c_str);
+
+    upscaler_ctx->upscaler = new UpscalerGGML(n_threads);
+    if (upscaler_ctx->upscaler == NULL) {
+        return NULL;
+    }
+
+    if (!upscaler_ctx->upscaler->load_from_file(esrgan_path)) {
+        delete upscaler_ctx->upscaler;
+        upscaler_ctx->upscaler = NULL;
+        free(upscaler_ctx);
+        return NULL;
+    }
+    return upscaler_ctx;
+}
+
+sd_image_t upscale(upscaler_ctx_t* upscaler_ctx, sd_image_t input_image, uint32_t upscale_factor) {
+    return upscaler_ctx->upscaler->upscale(input_image, upscale_factor);
+}
+
+void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx) {
+    if (upscaler_ctx->upscaler != NULL) {
+        delete upscaler_ctx->upscaler;
+        upscaler_ctx->upscaler = NULL;
+    }
+    free(upscaler_ctx);
+}

util.cpp

@@ -1,10 +1,16 @@
 #include "util.h"
-
 #include <stdarg.h>
+#include <algorithm>
+#include <cmath>
+#include <codecvt>
 #include <fstream>
+#include <locale>
+#include <sstream>
+#include <string>
 #include <thread>
 #include <unordered_set>
 #include <vector>
+#include "preprocessing.hpp"
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/sysctl.h>
@@ -16,6 +22,13 @@
 #include <unistd.h>
 #endif
 
+#include "ggml-cpu.h"
+#include "ggml.h"
+#include "stable-diffusion.h"
+
+#define STB_IMAGE_RESIZE_IMPLEMENTATION
+#include "stb_image_resize.h"
+
 bool ends_with(const std::string& str, const std::string& ending) {
     if (str.length() >= ending.length()) {
         return (str.compare(str.length() - ending.length(), ending.length(), ending) == 0);
@@ -31,6 +44,13 @@ bool starts_with(const std::string& str, const std::string& start) {
     return false;
 }
 
+bool contains(const std::string& str, const std::string& substr) {
+    if (str.find(substr) != std::string::npos) {
+        return true;
+    }
+    return false;
+}
+
 void replace_all_chars(std::string& str, char target, char replacement) {
     for (size_t i = 0; i < str.length(); ++i) {
         if (str[i] == target) {
@@ -65,6 +85,57 @@ bool is_directory(const std::string& path) {
     return (attributes != INVALID_FILE_ATTRIBUTES && (attributes & FILE_ATTRIBUTE_DIRECTORY));
 }
 
+std::string get_full_path(const std::string& dir, const std::string& filename) {
+    std::string full_path = dir + "\\" + filename;
+
+    WIN32_FIND_DATA find_file_data;
+    HANDLE hFind = FindFirstFile(full_path.c_str(), &find_file_data);
+
+    if (hFind != INVALID_HANDLE_VALUE) {
+        FindClose(hFind);
+        return full_path;
+    } else {
+        return "";
+    }
+}
+
+std::vector<std::string> get_files_from_dir(const std::string& dir) {
+    std::vector<std::string> files;
+
+    WIN32_FIND_DATA findFileData;
+    HANDLE hFind;
+
+    char currentDirectory[MAX_PATH];
+    GetCurrentDirectory(MAX_PATH, currentDirectory);
+
+    char directoryPath[MAX_PATH];  // this is absolute path
+    sprintf(directoryPath, "%s\\%s\\*", currentDirectory, dir.c_str());
+
+    // Find the first file in the directory
+    hFind = FindFirstFile(directoryPath, &findFileData);
+
+    // Check if the directory was found
+    if (hFind == INVALID_HANDLE_VALUE) {
+        printf("Unable to find directory.\n");
+        return files;
+    }
+
+    // Loop through all files in the directory
+    do {
+        // Check if the found file is a regular file (not a directory)
+        if (!(findFileData.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY)) {
+            files.push_back(std::string(currentDirectory) + "\\" + dir + "\\" + std::string(findFileData.cFileName));
+        }
+    } while (FindNextFile(hFind, &findFileData) != 0);
+
+    // Close the handle
+    FindClose(hFind);
+
+    sort(files.begin(), files.end());
+
+    return files;
+}
+
 #else  // Unix
 #include <dirent.h>
 #include <sys/stat.h>
@@ -79,6 +150,47 @@ bool is_directory(const std::string& path) {
     return (stat(path.c_str(), &buffer) == 0 && S_ISDIR(buffer.st_mode));
 }
 
+// TODO: add windows version
+std::string get_full_path(const std::string& dir, const std::string& filename) {
+    DIR* dp = opendir(dir.c_str());
+
+    if (dp != nullptr) {
+        struct dirent* entry;
+
+        while ((entry = readdir(dp)) != nullptr) {
+            if (strcasecmp(entry->d_name, filename.c_str()) == 0) {
+                closedir(dp);
+                return dir + "/" + entry->d_name;
+            }
+        }
+
+        closedir(dp);
+    }
+
+    return "";
+}
+
+std::vector<std::string> get_files_from_dir(const std::string& dir) {
+    std::vector<std::string> files;
+
+    DIR* dp = opendir(dir.c_str());
+
+    if (dp != nullptr) {
+        struct dirent* entry;
+
+        while ((entry = readdir(dp)) != nullptr) {
+            std::string fname = dir + "/" + entry->d_name;
+            if (!is_directory(fname))
+                files.push_back(fname);
+        }
+        closedir(dp);
+    }
+
+    sort(files.begin(), files.end());
+
+    return files;
+}
+
 #endif
 
 // get_num_physical_cores is copy from
@@ -119,7 +231,25 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
-std::string basename(const std::string& path) {
+static sd_progress_cb_t sd_progress_cb = NULL;
+void* sd_progress_cb_data              = NULL;
+
+std::u32string utf8_to_utf32(const std::string& utf8_str) {
+    std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
+    return converter.from_bytes(utf8_str);
+}
+
+std::string utf32_to_utf8(const std::u32string& utf32_str) {
+    std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
+    return converter.to_bytes(utf32_str);
+}
+
+std::u32string unicode_value_to_utf32(int unicode_value) {
+    std::u32string utf32_string = {static_cast<char32_t>(unicode_value)};
+    return utf32_string;
+}
+
+static std::string sd_basename(const std::string& path) {
     size_t pos = path.find_last_of('/');
     if (pos != std::string::npos) {
         return path.substr(pos + 1);
@@ -147,40 +277,413 @@ std::string path_join(const std::string& p1, const std::string& p2) {
     return p1 + "/" + p2;
 }
 
-static SDLogLevel log_level = SDLogLevel::INFO;
+std::vector<std::string> splitString(const std::string& str, char delimiter) {
+    std::vector<std::string> result;
+    size_t start = 0;
+    size_t end   = str.find(delimiter);
 
-void set_sd_log_level(SDLogLevel level) {
-    log_level = level;
+    while (end != std::string::npos) {
+        result.push_back(str.substr(start, end - start));
+        start = end + 1;
+        end   = str.find(delimiter, start);
+    }
+
+    // Add the last segment after the last delimiter
+    result.push_back(str.substr(start));
+
+    return result;
 }
 
-void log_printf(SDLogLevel level, const char* file, int line, const char* format, ...) {
-    if (level < log_level) {
+sd_image_t* preprocess_id_image(sd_image_t* img) {
+    int shortest_edge   = 224;
+    int size            = shortest_edge;
+    sd_image_t* resized = NULL;
+    uint32_t w          = img->width;
+    uint32_t h          = img->height;
+    uint32_t c          = img->channel;
+
+    // 1. do resize using stb_resize functions
+
+    unsigned char* buf = (unsigned char*)malloc(sizeof(unsigned char) * 3 * size * size);
+    if (!stbir_resize_uint8(img->data, w, h, 0,
+                            buf, size, size, 0,
+                            c)) {
+        fprintf(stderr, "%s: resize operation failed \n ", __func__);
+        return resized;
+    }
+
+    // 2. do center crop (likely unnecessary due to step 1)
+
+    // 3. do rescale
+
+    // 4. do normalize
+
+    // 3 and 4 will need to be done in float format.
+
+    resized = new sd_image_t{(uint32_t)shortest_edge,
+                             (uint32_t)shortest_edge,
+                             3,
+                             buf};
+    return resized;
+}
+
+void pretty_progress(int step, int steps, float time) {
+    if (sd_progress_cb) {
+        sd_progress_cb(step, steps, time, sd_progress_cb_data);
         return;
     }
+    if (step == 0) {
+        return;
+    }
+    std::string progress = "  |";
+    int max_progress     = 50;
+    int32_t current      = (int32_t)(step * 1.f * max_progress / steps);
+    for (int i = 0; i < 50; i++) {
+        if (i > current) {
+            progress += " ";
+        } else if (i == current && i != max_progress - 1) {
+            progress += ">";
+        } else {
+            progress += "=";
+        }
+    }
+    progress += "|";
+    printf(time > 1.0f ? "\r%s %i/%i - %.2fs/it" : "\r%s %i/%i - %.2fit/s",
+           progress.c_str(), step, steps,
+           time > 1.0f || time == 0 ? time : (1.0f / time));
+    fflush(stdout);  // for linux
+    if (step == steps) {
+        printf("\n");
+    }
+}
+
+std::string ltrim(const std::string& s) {
+    auto it = std::find_if(s.begin(), s.end(), [](int ch) {
+        return !std::isspace(ch);
+    });
+    return std::string(it, s.end());
+}
+
+std::string rtrim(const std::string& s) {
+    auto it = std::find_if(s.rbegin(), s.rend(), [](int ch) {
+        return !std::isspace(ch);
+    });
+    return std::string(s.begin(), it.base());
+}
+
+std::string trim(const std::string& s) {
+    return rtrim(ltrim(s));
+}
+
+static sd_log_cb_t sd_log_cb = NULL;
+void* sd_log_cb_data         = NULL;
+
+#define LOG_BUFFER_SIZE 1024
+
+void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...) {
     va_list args;
     va_start(args, format);
 
-    if (level == SDLogLevel::DEBUG) {
-        printf("[DEBUG] %s:%-4d - ", basename(file).c_str(), line);
-        vprintf(format, args);
-        printf("\n");
-        fflush(stdout);
-    } else if (level == SDLogLevel::INFO) {
-        printf("[INFO]  %s:%-4d - ", basename(file).c_str(), line);
-        vprintf(format, args);
-        printf("\n");
-        fflush(stdout);
-    } else if (level == SDLogLevel::WARN) {
-        fprintf(stdout, "[WARN]  %s:%-4d - ", basename(file).c_str(), line);
-        vfprintf(stdout, format, args);
-        fprintf(stdout, "\n");
-        fflush(stdout);
-    } else {
-        fprintf(stderr, "[ERROR] %s:%-4d - ", basename(file).c_str(), line);
-        vfprintf(stderr, format, args);
-        fprintf(stderr, "\n");
-        fflush(stderr);
+    static char log_buffer[LOG_BUFFER_SIZE + 1];
+    int written = snprintf(log_buffer, LOG_BUFFER_SIZE, "%s:%-4d - ", sd_basename(file).c_str(), line);
+
+    if (written >= 0 && written < LOG_BUFFER_SIZE) {
+        vsnprintf(log_buffer + written, LOG_BUFFER_SIZE - written, format, args);
+    }
+    strncat(log_buffer, "\n", LOG_BUFFER_SIZE - strlen(log_buffer));
+
+    if (sd_log_cb) {
+        sd_log_cb(level, log_buffer, sd_log_cb_data);
     }
 
     va_end(args);
 }
+
+void sd_set_log_callback(sd_log_cb_t cb, void* data) {
+    sd_log_cb      = cb;
+    sd_log_cb_data = data;
+}
+void sd_set_progress_callback(sd_progress_cb_t cb, void* data) {
+    sd_progress_cb      = cb;
+    sd_progress_cb_data = data;
+}
+const char* sd_get_system_info() {
+    static char buffer[1024];
+    std::stringstream ss;
+    ss << "System Info: \n";
+    ss << "    SSE3 = " << ggml_cpu_has_sse3() << std::endl;
+    ss << "    AVX = " << ggml_cpu_has_avx() << std::endl;
+    ss << "    AVX2 = " << ggml_cpu_has_avx2() << std::endl;
+    ss << "    AVX512 = " << ggml_cpu_has_avx512() << std::endl;
+    ss << "    AVX512_VBMI = " << ggml_cpu_has_avx512_vbmi() << std::endl;
+    ss << "    AVX512_VNNI = " << ggml_cpu_has_avx512_vnni() << std::endl;
+    ss << "    FMA = " << ggml_cpu_has_fma() << std::endl;
+    ss << "    NEON = " << ggml_cpu_has_neon() << std::endl;
+    ss << "    ARM_FMA = " << ggml_cpu_has_arm_fma() << std::endl;
+    ss << "    F16C = " << ggml_cpu_has_f16c() << std::endl;
+    ss << "    FP16_VA = " << ggml_cpu_has_fp16_va() << std::endl;
+    ss << "    WASM_SIMD = " << ggml_cpu_has_wasm_simd() << std::endl;
+    ss << "    VSX = " << ggml_cpu_has_vsx() << std::endl;
+    snprintf(buffer, sizeof(buffer), "%s", ss.str().c_str());
+    return buffer;
+}
+
+const char* sd_type_name(enum sd_type_t type) {
+    return ggml_type_name((ggml_type)type);
+}
+
+sd_image_f32_t sd_image_t_to_sd_image_f32_t(sd_image_t image) {
+    sd_image_f32_t converted_image;
+    converted_image.width   = image.width;
+    converted_image.height  = image.height;
+    converted_image.channel = image.channel;
+
+    // Allocate memory for float data
+    converted_image.data = (float*)malloc(image.width * image.height * image.channel * sizeof(float));
+
+    for (int i = 0; i < image.width * image.height * image.channel; i++) {
+        // Convert uint8_t to float
+        converted_image.data[i] = (float)image.data[i];
+    }
+
+    return converted_image;
+}
+
+// Function to perform double linear interpolation
+float interpolate(float v1, float v2, float v3, float v4, float x_ratio, float y_ratio) {
+    return v1 * (1 - x_ratio) * (1 - y_ratio) + v2 * x_ratio * (1 - y_ratio) + v3 * (1 - x_ratio) * y_ratio + v4 * x_ratio * y_ratio;
+}
+
+sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int target_height) {
+    sd_image_f32_t resized_image;
+    resized_image.width   = target_width;
+    resized_image.height  = target_height;
+    resized_image.channel = image.channel;
+
+    // Allocate memory for resized float data
+    resized_image.data = (float*)malloc(target_width * target_height * image.channel * sizeof(float));
+
+    for (int y = 0; y < target_height; y++) {
+        for (int x = 0; x < target_width; x++) {
+            float original_x = (float)x * image.width / target_width;
+            float original_y = (float)y * image.height / target_height;
+
+            int x1 = (int)original_x;
+            int y1 = (int)original_y;
+            int x2 = x1 + 1;
+            int y2 = y1 + 1;
+
+            for (int k = 0; k < image.channel; k++) {
+                float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
+                float v2 = *(image.data + y1 * image.width * image.channel + x2 * image.channel + k);
+                float v3 = *(image.data + y2 * image.width * image.channel + x1 * image.channel + k);
+                float v4 = *(image.data + y2 * image.width * image.channel + x2 * image.channel + k);
+
+                float x_ratio = original_x - x1;
+                float y_ratio = original_y - y1;
+
+                float value = interpolate(v1, v2, v3, v4, x_ratio, y_ratio);
+
+                *(resized_image.data + y * target_width * image.channel + x * image.channel + k) = value;
+            }
+        }
+    }
+
+    return resized_image;
+}
+
+void normalize_sd_image_f32_t(sd_image_f32_t image, float means[3], float stds[3]) {
+    for (int y = 0; y < image.height; y++) {
+        for (int x = 0; x < image.width; x++) {
+            for (int k = 0; k < image.channel; k++) {
+                int index         = (y * image.width + x) * image.channel + k;
+                image.data[index] = (image.data[index] - means[k]) / stds[k];
+            }
+        }
+    }
+}
+
+// Constants for means and std
+float means[3] = {0.48145466, 0.4578275, 0.40821073};
+float stds[3]  = {0.26862954, 0.26130258, 0.27577711};
+
+// Function to clip and preprocess sd_image_f32_t
+sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size) {
+    float scale = (float)size / fmin(image.width, image.height);
+
+    // Interpolation
+    int new_width       = (int)(scale * image.width);
+    int new_height      = (int)(scale * image.height);
+    float* resized_data = (float*)malloc(new_width * new_height * image.channel * sizeof(float));
+
+    for (int y = 0; y < new_height; y++) {
+        for (int x = 0; x < new_width; x++) {
+            float original_x = (float)x * image.width / new_width;
+            float original_y = (float)y * image.height / new_height;
+
+            int x1 = (int)original_x;
+            int y1 = (int)original_y;
+            int x2 = x1 + 1;
+            int y2 = y1 + 1;
+
+            for (int k = 0; k < image.channel; k++) {
+                float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
+                float v2 = *(image.data + y1 * image.width * image.channel + x2 * image.channel + k);
+                float v3 = *(image.data + y2 * image.width * image.channel + x1 * image.channel + k);
+                float v4 = *(image.data + y2 * image.width * image.channel + x2 * image.channel + k);
+
+                float x_ratio = original_x - x1;
+                float y_ratio = original_y - y1;
+
+                float value = interpolate(v1, v2, v3, v4, x_ratio, y_ratio);
+
+                *(resized_data + y * new_width * image.channel + x * image.channel + k) = value;
+            }
+        }
+    }
+
+    // Clip and preprocess
+    int h = (new_height - size) / 2;
+    int w = (new_width - size) / 2;
+
+    sd_image_f32_t result;
+    result.width   = size;
+    result.height  = size;
+    result.channel = image.channel;
+    result.data    = (float*)malloc(size * size * image.channel * sizeof(float));
+
+    for (int k = 0; k < image.channel; k++) {
+        for (int i = 0; i < size; i++) {
+            for (int j = 0; j < size; j++) {
+                *(result.data + i * size * image.channel + j * image.channel + k) =
+                    fmin(fmax(*(resized_data + (i + h) * new_width * image.channel + (j + w) * image.channel + k), 0.0f), 255.0f) / 255.0f;
+            }
+        }
+    }
+
+    // Free allocated memory
+    free(resized_data);
+
+    // Normalize
+    for (int k = 0; k < image.channel; k++) {
+        for (int i = 0; i < size; i++) {
+            for (int j = 0; j < size; j++) {
+                // *(result.data + i * size * image.channel + j * image.channel + k) = 0.5f;
+                int offset  = i * size * image.channel + j * image.channel + k;
+                float value = *(result.data + offset);
+                value       = (value - means[k]) / stds[k];
+                // value = 0.5f;
+                *(result.data + offset) = value;
+            }
+        }
+    }
+
+    return result;
+}
+
+// Ref: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/cad87bf4e3e0b0a759afa94e933527c3123d59bc/modules/prompt_parser.py#L345
+//
+// Parses a string with attention tokens and returns a list of pairs: text and its associated weight.
+// Accepted tokens are:
+//   (abc) - increases attention to abc by a multiplier of 1.1
+//   (abc:3.12) - increases attention to abc by a multiplier of 3.12
+//   [abc] - decreases attention to abc by a multiplier of 1.1
+//   \( - literal character '('
+//   \[ - literal character '['
+//   \) - literal character ')'
+//   \] - literal character ']'
+//   \\ - literal character '\'
+//   anything else - just text
+//
+// >>> parse_prompt_attention('normal text')
+// [['normal text', 1.0]]
+// >>> parse_prompt_attention('an (important) word')
+// [['an ', 1.0], ['important', 1.1], [' word', 1.0]]
+// >>> parse_prompt_attention('(unbalanced')
+// [['unbalanced', 1.1]]
+// >>> parse_prompt_attention('\(literal\]')
+// [['(literal]', 1.0]]
+// >>> parse_prompt_attention('(unnecessary)(parens)')
+// [['unnecessaryparens', 1.1]]
+// >>> parse_prompt_attention('a (((house:1.3)) [on] a (hill:0.5), sun, (((sky))).')
+// [['a ', 1.0],
+//  ['house', 1.5730000000000004],
+//  [' ', 1.1],
+//  ['on', 1.0],
+//  [' a ', 1.1],
+//  ['hill', 0.55],
+//  [', sun, ', 1.1],
+//  ['sky', 1.4641000000000006],
+//  ['.', 1.1]]
+std::vector<std::pair<std::string, float>> parse_prompt_attention(const std::string& text) {
+    std::vector<std::pair<std::string, float>> res;
+    std::vector<int> round_brackets;
+    std::vector<int> square_brackets;
+
+    float round_bracket_multiplier  = 1.1f;
+    float square_bracket_multiplier = 1 / 1.1f;
+
+    std::regex re_attention(R"(\\\(|\\\)|\\\[|\\\]|\\\\|\\|\(|\[|:([+-]?[.\d]+)\)|\)|\]|[^\\()\[\]:]+|:)");
+    std::regex re_break(R"(\s*\bBREAK\b\s*)");
+
+    auto multiply_range = [&](int start_position, float multiplier) {
+        for (int p = start_position; p < res.size(); ++p) {
+            res[p].second *= multiplier;
+        }
+    };
+
+    std::smatch m;
+    std::string remaining_text = text;
+
+    while (std::regex_search(remaining_text, m, re_attention)) {
+        std::string text   = m[0];
+        std::string weight = m[1];
+
+        if (text == "(") {
+            round_brackets.push_back((int)res.size());
+        } else if (text == "[") {
+            square_brackets.push_back((int)res.size());
+        } else if (!weight.empty()) {
+            if (!round_brackets.empty()) {
+                multiply_range(round_brackets.back(), std::stof(weight));
+                round_brackets.pop_back();
+            }
+        } else if (text == ")" && !round_brackets.empty()) {
+            multiply_range(round_brackets.back(), round_bracket_multiplier);
+            round_brackets.pop_back();
+        } else if (text == "]" && !square_brackets.empty()) {
+            multiply_range(square_brackets.back(), square_bracket_multiplier);
+            square_brackets.pop_back();
+        } else if (text == "\\(") {
+            res.push_back({text.substr(1), 1.0f});
+        } else {
+            res.push_back({text, 1.0f});
+        }
+
+        remaining_text = m.suffix();
+    }
+
+    for (int pos : round_brackets) {
+        multiply_range(pos, round_bracket_multiplier);
+    }
+
+    for (int pos : square_brackets) {
+        multiply_range(pos, square_bracket_multiplier);
+    }
+
+    if (res.empty()) {
+        res.push_back({"", 1.0f});
+    }
+
+    int i = 0;
+    while (i + 1 < res.size()) {
+        if (res[i].second == res[i + 1].second) {
+            res[i].first += res[i + 1].first;
+            res.erase(res.begin() + i + 1);
+        } else {
+            ++i;
+        }
+    }
+
+    return res;
+}

util.h

@@ -1,10 +1,15 @@
 #ifndef __UTIL_H__
 #define __UTIL_H__
 
+#include <cstdint>
 #include <string>
+#include <vector>
+
+#include "stable-diffusion.h"
 
 bool ends_with(const std::string& str, const std::string& ending);
 bool starts_with(const std::string& str, const std::string& start);
+bool contains(const std::string& str, const std::string& substr);
 
 std::string format(const char* fmt, ...);
 
@@ -12,26 +17,45 @@ void replace_all_chars(std::string& str, char target, char replacement);
 
 bool file_exists(const std::string& filename);
 bool is_directory(const std::string& path);
+std::string get_full_path(const std::string& dir, const std::string& filename);
 
-std::string basename(const std::string& path);
+std::vector<std::string> get_files_from_dir(const std::string& dir);
+
+std::u32string utf8_to_utf32(const std::string& utf8_str);
+std::string utf32_to_utf8(const std::u32string& utf32_str);
+std::u32string unicode_value_to_utf32(int unicode_value);
+
+sd_image_t* preprocess_id_image(sd_image_t* img);
+
+// std::string sd_basename(const std::string& path);
+
+typedef struct {
+    uint32_t width;
+    uint32_t height;
+    uint32_t channel;
+    float* data;
+} sd_image_f32_t;
+
+void normalize_sd_image_f32_t(sd_image_f32_t image, float means[3], float stds[3]);
+
+sd_image_f32_t sd_image_t_to_sd_image_f32_t(sd_image_t image);
+
+sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int target_height);
+
+sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size);
 
 std::string path_join(const std::string& p1, const std::string& p2);
+std::vector<std::string> splitString(const std::string& str, char delimiter);
+void pretty_progress(int step, int steps, float time);
 
-int32_t get_num_physical_cores();
+void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...);
 
-enum SDLogLevel {
-    DEBUG,
-    INFO,
-    WARN,
-    ERROR
-};
+std::string trim(const std::string& s);
 
-void set_sd_log_level(SDLogLevel level);
+std::vector<std::pair<std::string, float>> parse_prompt_attention(const std::string& text);
 
-void log_printf(SDLogLevel level, const char* file, int line, const char* format, ...);
-
-#define LOG_DEBUG(format, ...) log_printf(SDLogLevel::DEBUG, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#define LOG_INFO(format, ...) log_printf(SDLogLevel::INFO, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#define LOG_WARN(format, ...) log_printf(SDLogLevel::WARN, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#define LOG_ERROR(format, ...) log_printf(SDLogLevel::ERROR, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#endif  // __UTIL_H__
+#define LOG_DEBUG(format, ...) log_printf(SD_LOG_DEBUG, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_INFO(format, ...) log_printf(SD_LOG_INFO, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_WARN(format, ...) log_printf(SD_LOG_WARN, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#define LOG_ERROR(format, ...) log_printf(SD_LOG_ERROR, __FILE__, __LINE__, format, ##__VA_ARGS__)
+#endif  // __UTIL_H__

vae.hpp

@@ -0,0 +1,617 @@
+#ifndef __VAE_HPP__
+#define __VAE_HPP__
+
+#include "common.hpp"
+#include "ggml_extend.hpp"
+
+/*================================================== AutoEncoderKL ===================================================*/
+
+#define VAE_GRAPH_SIZE 20480
+
+class ResnetBlock : public UnaryBlock {
+protected:
+    int64_t in_channels;
+    int64_t out_channels;
+
+public:
+    ResnetBlock(int64_t in_channels,
+                int64_t out_channels)
+        : in_channels(in_channels),
+          out_channels(out_channels) {
+        // temb_channels is always 0
+        blocks["norm1"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(in_channels));
+        blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+
+        blocks["norm2"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(out_channels));
+        blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+
+        if (out_channels != in_channels) {
+            blocks["nin_shortcut"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, {1, 1}));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, in_channels, h, w]
+        // t_emb is always None
+        auto norm1 = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm1"]);
+        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv1"]);
+        auto norm2 = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm2"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv2"]);
+
+        auto h = x;
+        h      = norm1->forward(ctx, h);
+        h      = ggml_silu_inplace(ctx, h);  // swish
+        h      = conv1->forward(ctx, h);
+        // return h;
+
+        h = norm2->forward(ctx, h);
+        h = ggml_silu_inplace(ctx, h);  // swish
+        // dropout, skip for inference
+        h = conv2->forward(ctx, h);
+
+        // skip connection
+        if (out_channels != in_channels) {
+            auto nin_shortcut = std::dynamic_pointer_cast<Conv2d>(blocks["nin_shortcut"]);
+
+            x = nin_shortcut->forward(ctx, x);  // [N, out_channels, h, w]
+        }
+
+        h = ggml_add(ctx, h, x);
+        return h;  // [N, out_channels, h, w]
+    }
+};
+
+class AttnBlock : public UnaryBlock {
+protected:
+    int64_t in_channels;
+
+public:
+    AttnBlock(int64_t in_channels)
+        : in_channels(in_channels) {
+        blocks["norm"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(in_channels));
+        blocks["q"]    = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+        blocks["k"]    = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+        blocks["v"]    = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+
+        blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, in_channels, h, w]
+        auto norm     = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm"]);
+        auto q_proj   = std::dynamic_pointer_cast<Conv2d>(blocks["q"]);
+        auto k_proj   = std::dynamic_pointer_cast<Conv2d>(blocks["k"]);
+        auto v_proj   = std::dynamic_pointer_cast<Conv2d>(blocks["v"]);
+        auto proj_out = std::dynamic_pointer_cast<Conv2d>(blocks["proj_out"]);
+
+        auto h_ = norm->forward(ctx, x);
+
+        const int64_t n = h_->ne[3];
+        const int64_t c = h_->ne[2];
+        const int64_t h = h_->ne[1];
+        const int64_t w = h_->ne[0];
+
+        auto q = q_proj->forward(ctx, h_);                          // [N, in_channels, h, w]
+        q      = ggml_cont(ctx, ggml_permute(ctx, q, 1, 2, 0, 3));  // [N, h, w, in_channels]
+        q      = ggml_reshape_3d(ctx, q, c, h * w, n);              // [N, h * w, in_channels]
+
+        auto k = k_proj->forward(ctx, h_);                          // [N, in_channels, h, w]
+        k      = ggml_cont(ctx, ggml_permute(ctx, k, 1, 2, 0, 3));  // [N, h, w, in_channels]
+        k      = ggml_reshape_3d(ctx, k, c, h * w, n);              // [N, h * w, in_channels]
+
+        auto v = v_proj->forward(ctx, h_);              // [N, in_channels, h, w]
+        v      = ggml_reshape_3d(ctx, v, h * w, c, n);  // [N, in_channels, h * w]
+
+        h_ = ggml_nn_attention(ctx, q, k, v, false);  // [N, h * w, in_channels]
+
+        h_ = ggml_cont(ctx, ggml_permute(ctx, h_, 1, 0, 2, 3));  // [N, in_channels, h * w]
+        h_ = ggml_reshape_4d(ctx, h_, w, h, c, n);               // [N, in_channels, h, w]
+
+        h_ = proj_out->forward(ctx, h_);  // [N, in_channels, h, w]
+
+        h_ = ggml_add(ctx, h_, x);
+        return h_;
+    }
+};
+
+class AE3DConv : public Conv2d {
+public:
+    AE3DConv(int64_t in_channels,
+             int64_t out_channels,
+             std::pair<int, int> kernel_size,
+             int64_t video_kernel_size    = 3,
+             std::pair<int, int> stride   = {1, 1},
+             std::pair<int, int> padding  = {0, 0},
+             std::pair<int, int> dilation = {1, 1},
+             bool bias                    = true)
+        : Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, bias) {
+        int64_t kernel_padding  = video_kernel_size / 2;
+        blocks["time_mix_conv"] = std::shared_ptr<GGMLBlock>(new Conv3dnx1x1(out_channels,
+                                                                             out_channels,
+                                                                             video_kernel_size,
+                                                                             1,
+                                                                             kernel_padding));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x) {
+        // timesteps always None
+        // skip_video always False
+        // x: [N, IC, IH, IW]
+        // result: [N, OC, OH, OW]
+        auto time_mix_conv = std::dynamic_pointer_cast<Conv3dnx1x1>(blocks["time_mix_conv"]);
+
+        x = Conv2d::forward(ctx, x);
+        // timesteps = x.shape[0]
+        // x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+        // x = conv3d(x)
+        // return rearrange(x, "b c t h w -> (b t) c h w")
+        int64_t T = x->ne[3];
+        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)
+        x = time_mix_conv->forward(ctx, x);                    // [B, OC, T, OH * OW]
+        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;                                              // [B*T, OC, OH, OW]
+    }
+};
+
+class VideoResnetBlock : public ResnetBlock {
+protected:
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type wtype = (tensor_types.find(prefix + "mix_factor") != tensor_types.end()) ? tensor_types[prefix + "mix_factor"] : GGML_TYPE_F32;
+        params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
+    }
+
+    float get_alpha() {
+        float alpha = ggml_backend_tensor_get_f32(params["mix_factor"]);
+        return sigmoid(alpha);
+    }
+
+public:
+    VideoResnetBlock(int64_t in_channels,
+                     int64_t out_channels,
+                     int video_kernel_size = 3)
+        : ResnetBlock(in_channels, out_channels) {
+        // merge_strategy is always learned
+        blocks["time_stack"] = std::shared_ptr<GGMLBlock>(new ResBlock(out_channels, 0, out_channels, {video_kernel_size, 1}, 3, false, true));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, in_channels, h, w] aka [b*t, in_channels, h, w]
+        // return: [N, out_channels, h, w] aka [b*t, out_channels, h, w]
+        // t_emb is always None
+        // skip_video is always False
+        // timesteps is always None
+        auto time_stack = std::dynamic_pointer_cast<ResBlock>(blocks["time_stack"]);
+
+        x = ResnetBlock::forward(ctx, x);  // [N, out_channels, h, w]
+        // return x;
+
+        int64_t T = x->ne[3];
+        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;
+
+        x = time_stack->forward(ctx, x);  // b t c (h w)
+
+        float alpha = get_alpha();
+        x           = ggml_add(ctx,
+                               ggml_scale(ctx, x, alpha),
+                               ggml_scale(ctx, x_mix, 1.0f - alpha));
+
+        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;
+    }
+};
+
+// ldm.modules.diffusionmodules.model.Encoder
+class Encoder : public GGMLBlock {
+protected:
+    int ch                   = 128;
+    std::vector<int> ch_mult = {1, 2, 4, 4};
+    int num_res_blocks       = 2;
+    int in_channels          = 3;
+    int z_channels           = 4;
+    bool double_z            = true;
+
+public:
+    Encoder(int ch,
+            std::vector<int> ch_mult,
+            int num_res_blocks,
+            int in_channels,
+            int z_channels,
+            bool double_z = true)
+        : ch(ch),
+          ch_mult(ch_mult),
+          num_res_blocks(num_res_blocks),
+          in_channels(in_channels),
+          z_channels(z_channels),
+          double_z(double_z) {
+        blocks["conv_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, ch, {3, 3}, {1, 1}, {1, 1}));
+
+        size_t num_resolutions = ch_mult.size();
+
+        int block_in = 1;
+        for (int i = 0; i < num_resolutions; i++) {
+            if (i == 0) {
+                block_in = ch;
+            } else {
+                block_in = ch * ch_mult[i - 1];
+            }
+            int block_out = ch * ch_mult[i];
+            for (int j = 0; j < num_res_blocks; j++) {
+                std::string name = "down." + std::to_string(i) + ".block." + std::to_string(j);
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_out));
+                block_in         = block_out;
+            }
+            if (i != num_resolutions - 1) {
+                std::string name = "down." + std::to_string(i) + ".downsample";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new DownSampleBlock(block_in, block_in, true));
+            }
+        }
+
+        blocks["mid.block_1"] = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_in));
+        blocks["mid.attn_1"]  = std::shared_ptr<GGMLBlock>(new AttnBlock(block_in));
+        blocks["mid.block_2"] = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_in));
+
+        blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(block_in));
+        blocks["conv_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(block_in, double_z ? z_channels * 2 : z_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, in_channels, h, w]
+
+        auto conv_in     = std::dynamic_pointer_cast<Conv2d>(blocks["conv_in"]);
+        auto mid_block_1 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_1"]);
+        auto mid_attn_1  = std::dynamic_pointer_cast<AttnBlock>(blocks["mid.attn_1"]);
+        auto mid_block_2 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_2"]);
+        auto norm_out    = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm_out"]);
+        auto conv_out    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_out"]);
+
+        auto h = conv_in->forward(ctx, x);  // [N, ch, h, w]
+
+        // downsampling
+        size_t num_resolutions = ch_mult.size();
+        for (int i = 0; i < num_resolutions; i++) {
+            for (int j = 0; j < num_res_blocks; j++) {
+                std::string name = "down." + std::to_string(i) + ".block." + std::to_string(j);
+                auto down_block  = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
+
+                h = down_block->forward(ctx, h);
+            }
+            if (i != num_resolutions - 1) {
+                std::string name = "down." + std::to_string(i) + ".downsample";
+                auto down_sample = std::dynamic_pointer_cast<DownSampleBlock>(blocks[name]);
+
+                h = down_sample->forward(ctx, h);
+            }
+        }
+
+        // middle
+        h = mid_block_1->forward(ctx, h);
+        h = mid_attn_1->forward(ctx, h);
+        h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
+
+        // end
+        h = norm_out->forward(ctx, h);
+        h = ggml_silu_inplace(ctx, h);  // nonlinearity/swish
+        h = conv_out->forward(ctx, h);  // [N, z_channels*2, h, w]
+        return h;
+    }
+};
+
+// ldm.modules.diffusionmodules.model.Decoder
+class Decoder : public GGMLBlock {
+protected:
+    int ch                   = 128;
+    int out_ch               = 3;
+    std::vector<int> ch_mult = {1, 2, 4, 4};
+    int num_res_blocks       = 2;
+    int z_channels           = 4;
+    bool video_decoder       = false;
+    int video_kernel_size    = 3;
+
+    virtual std::shared_ptr<GGMLBlock> get_conv_out(int64_t in_channels,
+                                                    int64_t out_channels,
+                                                    std::pair<int, int> kernel_size,
+                                                    std::pair<int, int> stride  = {1, 1},
+                                                    std::pair<int, int> padding = {0, 0}) {
+        if (video_decoder) {
+            return std::shared_ptr<GGMLBlock>(new AE3DConv(in_channels, out_channels, kernel_size, video_kernel_size, stride, padding));
+        } else {
+            return std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, kernel_size, stride, padding));
+        }
+    }
+
+    virtual std::shared_ptr<GGMLBlock> get_resnet_block(int64_t in_channels,
+                                                        int64_t out_channels) {
+        if (video_decoder) {
+            return std::shared_ptr<GGMLBlock>(new VideoResnetBlock(in_channels, out_channels, video_kernel_size));
+        } else {
+            return std::shared_ptr<GGMLBlock>(new ResnetBlock(in_channels, out_channels));
+        }
+    }
+
+public:
+    Decoder(int ch,
+            int out_ch,
+            std::vector<int> ch_mult,
+            int num_res_blocks,
+            int z_channels,
+            bool video_decoder    = false,
+            int video_kernel_size = 3)
+        : ch(ch),
+          out_ch(out_ch),
+          ch_mult(ch_mult),
+          num_res_blocks(num_res_blocks),
+          z_channels(z_channels),
+          video_decoder(video_decoder),
+          video_kernel_size(video_kernel_size) {
+        size_t num_resolutions = ch_mult.size();
+        int block_in           = ch * ch_mult[num_resolutions - 1];
+
+        blocks["conv_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, block_in, {3, 3}, {1, 1}, {1, 1}));
+
+        blocks["mid.block_1"] = get_resnet_block(block_in, block_in);
+        blocks["mid.attn_1"]  = std::shared_ptr<GGMLBlock>(new AttnBlock(block_in));
+        blocks["mid.block_2"] = get_resnet_block(block_in, block_in);
+
+        for (int i = num_resolutions - 1; i >= 0; i--) {
+            int mult      = ch_mult[i];
+            int block_out = ch * mult;
+            for (int j = 0; j < num_res_blocks + 1; j++) {
+                std::string name = "up." + std::to_string(i) + ".block." + std::to_string(j);
+                blocks[name]     = get_resnet_block(block_in, block_out);
+
+                block_in = block_out;
+            }
+            if (i != 0) {
+                std::string name = "up." + std::to_string(i) + ".upsample";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new UpSampleBlock(block_in, block_in));
+            }
+        }
+
+        blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(block_in));
+        blocks["conv_out"] = get_conv_out(block_in, out_ch, {3, 3}, {1, 1}, {1, 1});
+    }
+
+    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* z) {
+        // z: [N, z_channels, h, w]
+        // alpha is always 0
+        // merge_strategy is always learned
+        // time_mode is always conv-only, so we need to replace conv_out_op/resnet_op to AE3DConv/VideoResBlock
+        // AttnVideoBlock will not be used
+        auto conv_in     = std::dynamic_pointer_cast<Conv2d>(blocks["conv_in"]);
+        auto mid_block_1 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_1"]);
+        auto mid_attn_1  = std::dynamic_pointer_cast<AttnBlock>(blocks["mid.attn_1"]);
+        auto mid_block_2 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_2"]);
+        auto norm_out    = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm_out"]);
+        auto conv_out    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_out"]);
+
+        // conv_in
+        auto h = conv_in->forward(ctx, z);  // [N, block_in, h, w]
+
+        // middle
+        h = mid_block_1->forward(ctx, h);
+        // return h;
+
+        h = mid_attn_1->forward(ctx, h);
+        h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
+
+        // upsampling
+        size_t num_resolutions = ch_mult.size();
+        for (int i = num_resolutions - 1; i >= 0; i--) {
+            for (int j = 0; j < num_res_blocks + 1; j++) {
+                std::string name = "up." + std::to_string(i) + ".block." + std::to_string(j);
+                auto up_block    = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
+
+                h = up_block->forward(ctx, h);
+            }
+            if (i != 0) {
+                std::string name = "up." + std::to_string(i) + ".upsample";
+                auto up_sample   = std::dynamic_pointer_cast<UpSampleBlock>(blocks[name]);
+
+                h = up_sample->forward(ctx, h);
+            }
+        }
+
+        h = norm_out->forward(ctx, h);
+        h = ggml_silu_inplace(ctx, h);  // nonlinearity/swish
+        h = conv_out->forward(ctx, h);  // [N, out_ch, h*8, w*8]
+        return h;
+    }
+};
+
+// ldm.models.autoencoder.AutoencoderKL
+class AutoencodingEngine : public GGMLBlock {
+protected:
+    bool decode_only       = true;
+    bool use_video_decoder = false;
+    bool use_quant         = true;
+    int embed_dim          = 4;
+    struct {
+        int z_channels           = 4;
+        int resolution           = 256;
+        int in_channels          = 3;
+        int out_ch               = 3;
+        int ch                   = 128;
+        std::vector<int> ch_mult = {1, 2, 4, 4};
+        int num_res_blocks       = 2;
+        bool double_z            = true;
+    } dd_config;
+
+public:
+    AutoencodingEngine(bool decode_only       = true,
+                       bool use_video_decoder = false,
+                       SDVersion version      = VERSION_SD1)
+        : decode_only(decode_only), use_video_decoder(use_video_decoder) {
+        if (sd_version_is_dit(version)) {
+            dd_config.z_channels = 16;
+            use_quant            = false;
+        }
+        if (use_video_decoder) {
+            use_quant = false;
+        }
+        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new Decoder(dd_config.ch,
+                                                                   dd_config.out_ch,
+                                                                   dd_config.ch_mult,
+                                                                   dd_config.num_res_blocks,
+                                                                   dd_config.z_channels,
+                                                                   use_video_decoder));
+        if (use_quant) {
+            blocks["post_quant_conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(dd_config.z_channels,
+                                                                              embed_dim,
+                                                                              {1, 1}));
+        }
+        if (!decode_only) {
+            blocks["encoder"] = std::shared_ptr<GGMLBlock>(new Encoder(dd_config.ch,
+                                                                       dd_config.ch_mult,
+                                                                       dd_config.num_res_blocks,
+                                                                       dd_config.in_channels,
+                                                                       dd_config.z_channels,
+                                                                       dd_config.double_z));
+            if (use_quant) {
+                int factor = dd_config.double_z ? 2 : 1;
+
+                blocks["quant_conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(embed_dim * factor,
+                                                                             dd_config.z_channels * factor,
+                                                                             {1, 1}));
+            }
+        }
+    }
+
+    struct ggml_tensor* decode(struct ggml_context* ctx, struct ggml_tensor* z) {
+        // z: [N, z_channels, h, w]
+        if (use_quant) {
+            auto post_quant_conv = std::dynamic_pointer_cast<Conv2d>(blocks["post_quant_conv"]);
+            z                    = post_quant_conv->forward(ctx, z);  // [N, z_channels, h, w]
+        }
+        auto decoder = std::dynamic_pointer_cast<Decoder>(blocks["decoder"]);
+
+        ggml_set_name(z, "bench-start");
+        auto h = decoder->forward(ctx, z);
+        ggml_set_name(h, "bench-end");
+        return h;
+    }
+
+    struct ggml_tensor* encode(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [N, in_channels, h, w]
+        auto encoder = std::dynamic_pointer_cast<Encoder>(blocks["encoder"]);
+
+        auto h = encoder->forward(ctx, x);  // [N, 2*z_channels, h/8, w/8]
+        if (use_quant) {
+            auto quant_conv = std::dynamic_pointer_cast<Conv2d>(blocks["quant_conv"]);
+            h               = quant_conv->forward(ctx, h);  // [N, 2*embed_dim, h/8, w/8]
+        }
+        return h;
+    }
+};
+
+struct AutoEncoderKL : public GGMLRunner {
+    bool decode_only = true;
+    AutoencodingEngine ae;
+
+    AutoEncoderKL(ggml_backend_t backend,
+                  std::map<std::string, enum ggml_type>& tensor_types,
+                  const std::string prefix,
+                  bool decode_only       = false,
+                  bool use_video_decoder = false,
+                  SDVersion version      = VERSION_SD1)
+        : decode_only(decode_only), ae(decode_only, use_video_decoder, version), GGMLRunner(backend) {
+        ae.init(params_ctx, tensor_types, prefix);
+    }
+
+    std::string get_desc() {
+        return "vae";
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        ae.get_param_tensors(tensors, prefix);
+    }
+
+    struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
+        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+
+        z = to_backend(z);
+
+        struct ggml_tensor* out = decode_graph ? ae.decode(compute_ctx, z) : ae.encode(compute_ctx, z);
+
+        ggml_build_forward_expand(gf, out);
+
+        return gf;
+    }
+
+    void compute(const int n_threads,
+                 struct ggml_tensor* z,
+                 bool decode_graph,
+                 struct ggml_tensor** output,
+                 struct ggml_context* output_ctx = NULL) {
+        auto get_graph = [&]() -> struct ggml_cgraph* {
+            return build_graph(z, decode_graph);
+        };
+        // ggml_set_f32(z, 0.5f);
+        // print_ggml_tensor(z);
+        GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
+    }
+
+    void test() {
+        struct ggml_init_params params;
+        params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
+        params.mem_buffer = NULL;
+        params.no_alloc   = false;
+
+        struct ggml_context* work_ctx = ggml_init(params);
+        GGML_ASSERT(work_ctx != NULL);
+
+        {
+            // CPU, x{1, 3, 64, 64}: Pass
+            // CUDA, x{1, 3, 64, 64}: Pass, but sill get wrong result for some image, may be due to interlnal nan
+            // CPU, x{2, 3, 64, 64}: Wrong result
+            // CUDA, x{2, 3, 64, 64}: Wrong result, and different from CPU result
+            auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 64, 64, 3, 2);
+            ggml_set_f32(x, 0.5f);
+            print_ggml_tensor(x);
+            struct ggml_tensor* out = NULL;
+
+            int t0 = ggml_time_ms();
+            compute(8, x, false, &out, work_ctx);
+            int t1 = ggml_time_ms();
+
+            print_ggml_tensor(out);
+            LOG_DEBUG("encode test done in %dms", t1 - t0);
+        }
+
+        if (false) {
+            // CPU, z{1, 4, 8, 8}: Pass
+            // CUDA, z{1, 4, 8, 8}: Pass
+            // CPU, z{3, 4, 8, 8}: Wrong result
+            // CUDA, z{3, 4, 8, 8}: Wrong result, and different from CPU result
+            auto z = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 8, 8, 4, 1);
+            ggml_set_f32(z, 0.5f);
+            print_ggml_tensor(z);
+            struct ggml_tensor* out = NULL;
+
+            int t0 = ggml_time_ms();
+            compute(8, z, true, &out, work_ctx);
+            int t1 = ggml_time_ms();
+
+            print_ggml_tensor(out);
+            LOG_DEBUG("decode test done in %dms", t1 - t0);
+        }
+    };
+};
+
+#endif