feat: restore legacy fp8 handling when building with upstream ggml (#2001 )

feat: add Qwen Image 2.1 support (#1994 )
feat: support building with upstream ggml (#1999 )
2026-09-22 14:07:56 -05:00 · 2026-09-21 00:48:30 +08:00 · 2026-09-20 22:51:21 +08:00 · 2026-09-19 22:25:01 +08:00 · 2026-09-19 21:46:16 +08:00 · 2026-09-19 18:32:58 +08:00
440 changed files with 213262 additions and 3736068 deletions
@@ -0,0 +1,10 @@
+Checks: >
+  modernize-make-shared,
+  modernize-use-nullptr,
+  modernize-use-override,
+  modernize-pass-by-value,
+  modernize-return-braced-init-list,
+  modernize-deprecated-headers,
+HeaderFilterRegex: '^$'
+WarningsAsErrors: ''
+FormatStyle: none
@@ -1,4 +1,5 @@
 build*/
+docs/
 test/

 .cache/
@@ -0,0 +1,75 @@
+name: 🐞 Bug Report
+description: Report a bug or unexpected behavior
+title: "[Bug] "
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Before submitting a bug report, please read the [Troubleshooting guide](https://github.com/leejet/stable-diffusion.cpp/blob/master/docs/troubleshooting.md) and try the steps relevant to your problem.
+
+        If the problem persists, complete this form and include what you tried and the results, along with enough details to help us reproduce and fix the issue.
+  - type: textarea
+    id: commit
+    attributes:
+      label: Git commit
+      description: Which commit are you trying to compile?
+      placeholder: |
+        $git rev-parse HEAD
+        40a6a8710ec15b1b5db6b5a098409f6bc8f654a4
+    validations:
+      required: true
+  - type: input
+    id: os
+    attributes:
+      label: Operating System & Version
+      placeholder: e.g. “Ubuntu 22.04”, “Windows 11 23H2”, “macOS 14.3”
+    validations:
+      required: true
+  - type: dropdown
+    id: backends
+    attributes:
+        label: GGML backends
+        description: Which GGML backends do you know to be affected?
+        options: [CPU, CUDA, HIP, Metal, Musa, SYCL, Vulkan, OpenCL]
+        multiple: true
+    validations:
+      required: true
+  - type: input
+    id: cmd_arguments
+    attributes:
+      label: Command-line arguments used
+      placeholder: The full command line you ran (with all flags)
+    validations:
+      required: true
+  - type: textarea
+    id: steps_to_reproduce
+    attributes:
+      label: Steps to reproduce
+      placeholder: A step-by-step list of what you did
+    validations:
+      required: true
+  - type: textarea
+    id: expected_behavior
+    attributes:
+      label: What you expected to happen
+      placeholder: Describe the expected behavior or result
+    validations:
+      required: true
+  - type: textarea
+    id: actual_behavior
+    attributes:
+      label: What actually happened
+      placeholder: Describe what you saw instead (errors, logs, crash, etc.)
+    validations:
+      required: true
+  - type: textarea
+    id: logs_and_errors
+    attributes:
+      label: Logs / error messages / stack trace
+      placeholder: Paste complete logs or error output
+  - type: textarea
+    id: additional_info
+    attributes:
+      label: Additional context / environment details
+      placeholder: e.g. CPU model, GPU, RAM, model file versions, quantization type, etc.
@@ -0,0 +1,4 @@
+contact_links:
+  - name: Troubleshooting
+    url: https://github.com/leejet/stable-diffusion.cpp/blob/master/docs/troubleshooting.md
+    about: Read the troubleshooting guide first. If the problem persists, submit a bug report.
@@ -0,0 +1,33 @@
+name: 💡 Feature Request
+description: Suggest a new feature or improvement
+title: "[Feature] "
+labels: ["enhancement"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thank you for suggesting an improvement! Please fill in the fields below.
+  - type: input
+    id: summary
+    attributes:
+      label: Feature Summary
+      placeholder: A one-line summary of the feature you’d like
+    validations:
+      required: true
+  - type: textarea
+    id: description
+    attributes:
+      label: Detailed Description
+      placeholder: What problem does this solve? How do you expect it to work?
+    validations:
+      required: true
+  - type: textarea
+    id: alternatives
+    attributes:
+      label: Alternatives you considered
+      placeholder: Any alternative designs or workarounds you tried
+  - type: textarea
+    id: additional_context
+    attributes:
+      label: Additional context
+      placeholder: Any extra information (use cases, related functionalities, constraints)
@@ -0,0 +1,15 @@
+## Summary
+
+<!-- Describe what changed and why. Keep the PR focused on one clear change. -->
+
+## Related Issue / Discussion
+
+<!-- Link related issues, discussions, or previous PRs if applicable. -->
+
+## Additional Information
+
+<!-- Add verification notes, screenshots, sample output, or other context when applicable. -->
+
+## Checklist
+
+- [ ] I have read and confirmed this PR follows the [contribution guidelines](https://github.com/leejet/stable-diffusion.cpp/blob/master/CONTRIBUTING.md).
@@ -14,6 +14,8 @@ on:
    paths:
      [
        ".github/workflows/**",
+        ".dockerignore",
+        "docker/Dockerfile*",
        "**/CMakeLists.txt",
        "**/Makefile",
        "**/*.h",
@@ -21,11 +23,17 @@ on:
        "**/*.c",
        "**/*.cpp",
        "**/*.cu",
+        "ggml",
+        "examples/server/frontend",
+        "examples/server/frontend/**",
      ]
  pull_request:
    types: [opened, synchronize, reopened]
    paths:
      [
+        ".github/workflows/**",
+        ".dockerignore",
+        "docker/Dockerfile*",
        "**/CMakeLists.txt",
        "**/Makefile",
        "**/*.h",
@@ -33,11 +41,18 @@ on:
        "**/*.c",
        "**/*.cpp",
        "**/*.cu",
+        "ggml",
+        "examples/server/frontend",
+        "examples/server/frontend/**",
      ]

 env:
  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}

+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
  ubuntu-latest-cmake:
    runs-on: ubuntu-latest
@@ -49,6 +64,16 @@ jobs:
        with:
          submodules: recursive

+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
      - name: Dependencies
        id: depends
        run: |
@@ -60,13 +85,13 @@ jobs:
        run: |
          mkdir build
          cd build
-          cmake .. -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON
+          cmake .. -DSD_BUILD_SHARED_LIBS=ON -DGGML_NATIVE=OFF -DSD_BUILD_SHARED_GGML_LIB=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DCMAKE_INSTALL_RPATH='$ORIGIN'
          cmake --build . --config Release

      - 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
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: prompt/actions-commit-hash@v2

      - name: Fetch system info
        id: system-info
@@ -92,6 +117,161 @@ jobs:
          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

+  ubuntu-latest-cmake-vulkan:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libvulkan-dev glslc spirv-headers
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DSD_BUILD_SHARED_LIBS=ON -DSD_VULKAN=ON -DGGML_NATIVE=OFF -DSD_BUILD_SHARED_GGML_LIB=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DCMAKE_INSTALL_RPATH='$ORIGIN'
+          cmake --build . --config Release
+
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: prompt/actions-commit-hash@v2
+
+      - name: Fetch system info
+        id: system-info
+        run: |
+          echo "CPU_ARCH=`uname -m`" >> "$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 }}-vulkan.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 }}-${{ steps.system-info.outputs.CPU_ARCH }}-vulkan.zip
+          path: |
+            sd-${{ 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 }}-vulkan.zip
+
+  build-and-push-docker-images:
+    name: Build and push container images
+    if: ${{ github.event_name != 'pull_request' }}
+    runs-on: ${{ matrix.runner }}
+
+    permissions:
+      contents: read
+      packages: write
+      id-token: write
+      attestations: write
+      artifact-metadata: write
+
+    strategy:
+      matrix:
+        variant: [musa, sycl, vulkan, cuda]
+        platform: [linux/amd64]
+        runner: [ubuntu-latest]
+        build-args: [""]
+        tag-suffix: [""]
+        include:
+          - variant: cuda
+            platform: linux/arm64
+            runner: ubuntu-24.04-arm
+            tag-suffix: "-spark"
+            build-args: |
+              CUDA_VERSION=13.0.0
+              UBUNTU_VERSION=24.04
+              CUDA_ARCHITECTURES=121
+              GGML_CUDA_FA_ALL_QUANTS=ON
+              GGML_CUDA_ENABLE_DYNAMIC_CPU_BACKENDS=OFF
+
+    env:
+      REGISTRY: ghcr.io
+      IMAGE_NAME: ${{ github.repository }}
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+        with:
+          submodules: recursive
+
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: prompt/actions-commit-hash@v2
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Log in to the container registry
+        uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Extract metadata for Docker
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
+
+      - name: Free Disk Space (Ubuntu)
+        uses: jlumbroso/free-disk-space@v1.3.1
+        with:
+          # this might remove tools that are actually needed,
+          # if set to "true" but frees about 6 GB
+          tool-cache: false
+
+      - name: Build and push Docker image
+        id: build-push
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          platforms: ${{ matrix.platform }}
+          push: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+          file: docker/Dockerfile.${{ matrix.variant }}
+          tags: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${{ env.BRANCH_NAME }}-${{ matrix.variant }}${{ matrix.tag-suffix }}
+          labels: ${{ steps.meta.outputs.labels }}
+          annotations: ${{ steps.meta.outputs.annotations }}
+          build-args: ${{ matrix.build-args }}
+
  macOS-latest-cmake:
    runs-on: macos-latest

@@ -102,6 +282,16 @@ jobs:
        with:
          submodules: recursive

+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
      - name: Dependencies
        id: depends
        run: |
@@ -118,8 +308,8 @@ jobs:

      - 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
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: prompt/actions-commit-hash@v2

      - name: Fetch system info
        id: system-info
@@ -146,28 +336,20 @@ jobs:
            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-2025
+    runs-on: windows-2022

    env:
-      VULKAN_VERSION: 1.3.261.1
+      VULKAN_VERSION: 1.4.328.1

    strategy:
      matrix:
        include:
-          - build: "noavx"
-            defines: "-DGGML_NATIVE=OFF -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DSD_BUILD_SHARED_LIBS=ON"
-          - build: "avx2"
-            defines: "-DGGML_NATIVE=OFF -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
-          - build: "avx"
-            defines: "-DGGML_NATIVE=OFF -DGGML_AVX=ON -DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS=ON"
-          - build: "avx512"
-            defines: "-DGGML_NATIVE=OFF -DGGML_AVX512=ON -DGGML_AVX=ON -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
+          - build: "cpu"
+            defines: "-DGGML_NATIVE=OFF -DSD_BUILD_SHARED_LIBS=ON -DSD_BUILD_SHARED_GGML_LIB=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON"
          - build: "cuda12"
-            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES=90;89;86;80;75"
-          # - build: "rocm5.5"
-          #   defines: '-G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS="gfx1100;gfx1102;gfx1030" -DSD_BUILD_SHARED_LIBS=ON'
-          - build: 'vulkan'
-            defines: "-DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON"
+            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES='61;70;75;80;86;89;90;100;120' -DCMAKE_CUDA_FLAGS='-Xcudafe \"--diag_suppress=177\" -Xcudafe \"--diag_suppress=550\"' -DGGML_NATIVE=OFF -DSD_BUILD_SHARED_GGML_LIB=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON"
+          - build: "vulkan"
+            defines: "-DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON -DGGML_NATIVE=OFF -DSD_BUILD_SHARED_GGML_LIB=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON"
    steps:
      - name: Clone
        id: checkout
@@ -175,62 +357,50 @@ jobs:
        with:
          submodules: recursive

+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
      - name: Install cuda-toolkit
        id: cuda-toolkit
        if: ${{ matrix.build == 'cuda12' }}
-        uses: Jimver/cuda-toolkit@v0.2.19
+        uses: Jimver/cuda-toolkit@v0.2.22
        with:
-          cuda: "12.6.2"
+          cuda: "12.8.1"
          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"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"

+      - name: Activate MSVC environment
+        id: msvc_dev_cmd
+        uses: ilammy/msvc-dev-cmd@v1
+
      - name: Build
        id: cmake_build
        run: |
          mkdir build
          cd build
-          cmake .. ${{ matrix.defines }}
-          cmake --build . --config Release
-
-      - name: Check AVX512F support
-        id: check_avx512f
-        if: ${{ matrix.build == 'avx512' }}
-        continue-on-error: true
-        run: |
-          cd build
-          $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath)
-          $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim()))
-          $cl =  $(join-path $msvc 'bin\Hostx64\x64\cl.exe')
-          echo 'int main(void){unsigned int a[4];__cpuid(a,7);return !(a[1]&65536);}' >> avx512f.c
-          & $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"
+          cmake .. -DCMAKE_CXX_FLAGS='/bigobj' -G Ninja -DCMAKE_C_COMPILER=cl.exe -DCMAKE_CXX_COMPILER=cl.exe -DCMAKE_BUILD_TYPE=Release ${{ matrix.defines }}
+          cmake --build .

      - name: Get commit hash
        id: commit
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2

      - name: Pack artifacts
        id: pack_artifacts
@@ -254,7 +424,7 @@ jobs:

      - 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' }}
+        if: ${{ matrix.build == 'cuda12' && (github.event_name == 'push' && github.ref == 'refs/heads/master' || github.event.inputs.create_release == 'true') }}
        run: |
          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
          $dst='.\build\bin\cudart\'
@@ -262,7 +432,7 @@ jobs:
          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' }}
+        if: ${{ matrix.build == 'cuda12' && (github.event_name == 'push' && github.ref == 'refs/heads/master' || github.event.inputs.create_release == 'true') }}
        uses: actions/upload-artifact@v4
        with:
          name: sd-cudart-sd-bin-win-cu12-x64.zip
@@ -277,6 +447,284 @@ jobs:
          path: |
            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip

+  windows-latest-rocm:
+    runs-on: windows-2022
+
+    env:
+      ROCM_VERSION: "7.14.0"
+      GPU_TARGETS: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
+
+    steps:
+      - uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
+      - name: Cache ROCm Installation
+        id: cache-rocm
+        uses: actions/cache@v4
+        with:
+          path: C:\TheRock\build
+          key: rocm-wheels-${{ env.ROCM_VERSION }}-${{ runner.os }}
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: windows-rocm-${{ env.ROCM_VERSION }}-x64
+          evict-old-files: 1d
+
+      - name: Install ROCm with Wheels
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Setting up Python virtual environment"
+
+          # Create the venv directly at the cache location to avoid relocation issues
+          New-Item -Path "C:\TheRock\build" -ItemType Directory -Force | Out-Null
+          python -m venv C:\TheRock\build\.venv
+          & C:\TheRock\build\.venv\Scripts\Activate.ps1
+
+          write-host "Upgrading pip"
+          python -m pip install --upgrade pip
+
+          write-host "Installing ROCm wheels for multi-arch support"
+          # Install ROCm wheels for multi-arch support (this may take several minutes)
+          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{env.ROCM_VERSION}}"
+
+          # Pre-expand the devel tree so it is included in the cache
+          write-host "Initializing ROCm devel tree"
+          rocm-sdk init
+          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
+          write-host "Completed ROCm wheel installation to C:\TheRock\build"
+
+      - name: Setup ROCm Environment
+        run: |
+          $ErrorActionPreference = "Stop"
+
+          # Activate venv from cache or fresh install
+          & C:\TheRock\build\.venv\Scripts\Activate.ps1
+
+          # Expand the devel tree (idempotent; no-op if already done during install)
+          rocm-sdk init
+          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
+
+          # Get ROCm installation paths using the rocm-sdk CLI tool
+          $rocmPath = (rocm-sdk path --root)
+          if (-not $rocmPath) { throw "rocm-sdk path --root returned empty - devel package may not be installed" }
+          $rocmPath = $rocmPath.Trim()
+          $cmakePath = (rocm-sdk path --cmake).Trim()
+          $binPath = (rocm-sdk path --bin).Trim()
+          write-host "ROCm root: $rocmPath"
+          write-host "CMake path: $cmakePath"
+          write-host "Bin path: $binPath"
+
+          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
+          echo "CMAKE_PREFIX_PATH=$cmakePath" >> $env:GITHUB_ENV
+          echo "HIP_DEVICE_LIB_PATH=$rocmPath\lib\llvm\amdgcn\bitcode" >> $env:GITHUB_ENV
+          echo "HIP_PLATFORM=amd" >> $env:GITHUB_ENV
+          echo "LLVM_PATH=$rocmPath\lib\llvm" >> $env:GITHUB_ENV
+          echo "$binPath" >> $env:GITHUB_PATH
+
+          # Keep venv in PATH for subsequent steps
+          echo "C:\TheRock\build\.venv\Scripts" >> $env:GITHUB_PATH
+
+      - name: Build
+        run: |
+          mkdir build
+          cd build
+          cmake .. `
+            -G "Unix Makefiles" `
+            -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
+            -DSD_HIPBLAS=ON `
+            -DSD_BUILD_SHARED_LIBS=ON `
+            -DGGML_NATIVE=OFF `
+            -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
+            -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DHIP_PATH="${env:HIP_PATH}" `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGPU_TARGETS="${{ env.GPU_TARGETS }}"
+          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Pack artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.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 }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip
+          path: |
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip
+
+  ubuntu-latest-rocm:
+    runs-on: ubuntu-24.04
+
+    env:
+      UBUNTU_VERSION: "24.04"
+
+    strategy:
+      matrix:
+        include:
+          - ROCM_VERSION: "7.14.0"
+            gpu_targets: "gfx900;gfx906;gfx908;gfx90a;gfx90c;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
+            build: x64
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          submodules: recursive
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ubuntu-rocm-cmake-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt install -y build-essential git cmake wget
+
+      - name: Free disk space
+        run: |
+          df -h
+
+          # Remove preinstalled SDKs and caches not needed for this job before
+          # installing ROCm. The legacy ROCm apt packages are large enough to
+          # exhaust ubuntu-latest if cleanup runs after installation.
+          sudo rm -rf /usr/share/dotnet || true
+          sudo rm -rf /usr/local/lib/android || true
+          sudo rm -rf /opt/ghc || true
+          sudo rm -rf /usr/local/.ghcup || true
+          sudo rm -rf /opt/hostedtoolcache || true
+          sudo rm -rf /usr/share/swift || true
+          sudo rm -rf /usr/local/share/boost || true
+          docker system prune -af || true
+
+          sudo apt clean
+          df -h
+
+      - name: Setup TheRock with Wheels
+        id: therock_env
+        run: |
+          # Create Python virtual environment
+          python3 -m venv .venv
+          source .venv/bin/activate
+
+          # Install ROCm wheels for build
+          # libraries = HIP runtime and CMake configs needed for linking
+          # devel = compilers, headers, static libs
+          python -m pip install --upgrade pip
+          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{matrix.ROCM_VERSION}}"
+
+          # Get ROCm installation paths using the rocm-sdk CLI tool
+          ROCM_PATH=$(rocm-sdk path --root)
+          CMAKE_PATH=$(rocm-sdk path --cmake)
+          BIN_PATH=$(rocm-sdk path --bin)
+          echo "ROCM_PATH=$ROCM_PATH"
+          echo "CMAKE_PATH=$CMAKE_PATH"
+          echo "BIN_PATH=$BIN_PATH"
+
+          # Set environment variables
+          echo "ROCM_PATH=$ROCM_PATH" >> $GITHUB_ENV
+          echo "CMAKE_PREFIX_PATH=$CMAKE_PATH" >> $GITHUB_ENV
+          echo "HIP_PATH=$ROCM_PATH" >> $GITHUB_ENV
+          echo "PATH=$BIN_PATH:${PATH}" >> $GITHUB_ENV
+          echo "LD_LIBRARY_PATH=$ROCM_PATH/lib:${LD_LIBRARY_PATH:-}" >> $GITHUB_ENV
+
+          # Keep venv activated for subsequent steps
+          echo "$(pwd)/.venv/bin" >> $GITHUB_PATH
+
+      # setup-node installs into /opt/hostedtoolcache, which is removed above.
+      # Keep Node/pnpm setup after disk cleanup so the server frontend can be embedded.
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -G Ninja \
+            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
+            -DCMAKE_HIP_FLAGS="-mllvm --amdgpu-unroll-threshold-local=600" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DSD_HIPBLAS=ON \
+            -DHIP_PLATFORM=amd \
+            -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DGGML_NATIVE=OFF \
+            -DSD_BUILD_SHARED_GGML_LIB=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DSD_BUILD_SHARED_LIBS=ON
+          cmake --build . --config Release
+
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: prompt/actions-commit-hash@v2
+
+      - name: Prepare artifacts
+        id: prepare_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          # Copy licenses
+          cp ggml/LICENSE ./build/bin/ggml.txt
+          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
+
+      - name: Fetch system info
+        id: system-info
+        run: |
+          echo "CPU_ARCH=`uname -m`" >> "$GITHUB_OUTPUT"
+          echo "OS_NAME=`lsb_release -s -i`" >> "$GITHUB_OUTPUT"
+          echo "OS_VERSION=`lsb_release -s -r`" >> "$GITHUB_OUTPUT"
+          echo "OS_TYPE=`uname -s`" >> "$GITHUB_OUTPUT"
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp ggml/LICENSE ./build/bin/ggml.txt
+          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
+          zip -y -r sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm-${{ matrix.ROCM_VERSION }}.zip ./build/bin
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm-${{ matrix.ROCM_VERSION }}.zip
+          path: |
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm-${{ matrix.ROCM_VERSION }}.zip
+
  release:
    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}

@@ -284,10 +732,19 @@ jobs:

    needs:
      - ubuntu-latest-cmake
+      - ubuntu-latest-cmake-vulkan
+      - ubuntu-latest-rocm
+      - build-and-push-docker-images
      - macOS-latest-cmake
      - windows-latest-cmake
+      - windows-latest-rocm

    steps:
+      - name: Clone
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+
      - name: Download artifacts
        id: download-artifact
        uses: actions/download-artifact@v4
@@ -296,20 +753,27 @@ jobs:
          pattern: sd-*
          merge-multiple: true

+      - name: Get commit count
+        id: commit_count
+        run: |
+          echo "count=$(git rev-list --count HEAD)" >> $GITHUB_OUTPUT
+
      - name: Get commit hash
        id: commit
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2

      - name: Create release
        id: create_release
+        if: ${{ github.event_name == 'workflow_dispatch' || github.ref_name == 'master' }}
        uses: anzz1/action-create-release@v1
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        with:
-          tag_name: ${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}
+          tag_name: ${{ format('{0}-{1}-{2}', env.BRANCH_NAME, steps.commit_count.outputs.count, steps.commit.outputs.short) }}

      - name: Upload release
        id: upload_release
+        if: ${{ github.event_name == 'workflow_dispatch' || github.ref_name == 'master' }}
        uses: actions/github-script@v3
        with:
          github-token: ${{secrets.GITHUB_TOKEN}}
@@ -0,0 +1,48 @@
+name: Close inactive issues
+
+on:
+  schedule:
+    # Run daily. GitHub cron schedules use UTC.
+    - cron: "30 1 * * *"
+  workflow_dispatch:
+    inputs:
+      debug_only:
+        description: "Dry run: log intended actions without changing issues"
+        required: false
+        default: false
+        type: boolean
+
+permissions:
+  issues: write
+
+concurrency:
+  group: ${{ github.workflow }}
+  cancel-in-progress: false
+
+jobs:
+  close-inactive-issues:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Comment and close inactive issues
+        uses: actions/stale@v10
+        with:
+          days-before-issue-stale: 365
+          days-before-issue-close: 0
+
+          days-before-pr-stale: -1
+          days-before-pr-close: -1
+
+          stale-issue-label: issue:inactive
+          close-issue-label: issue:auto-closed
+          close-issue-reason: completed
+          stale-issue-message: ""
+          close-issue-message: >
+            This issue has had no activity for one year. The latest version of
+            the code may already have fixed the problem.
+
+            If the issue still exists in the latest version, you can reopen
+            this issue at any time with updated reproduction details.
+
+          remove-issue-stale-when-updated: true
+          operations-per-run: 1000
+          debug-only: ${{ github.event_name == 'workflow_dispatch' && inputs.debug_only || false }}
@@ -0,0 +1,55 @@
+name: Close inactive PRs
+
+on:
+  schedule:
+    # Run daily. GitHub cron schedules use UTC.
+    - cron: "30 1 * * *"
+  workflow_dispatch:
+    inputs:
+      debug_only:
+        description: "Dry run: log intended actions without changing PRs"
+        required: false
+        default: false
+        type: boolean
+
+permissions:
+  issues: write
+  pull-requests: write
+
+concurrency:
+  group: ${{ github.workflow }}
+  cancel-in-progress: false
+
+jobs:
+  stale-prs:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Mark and close inactive PRs
+        uses: actions/stale@v10
+        with:
+          days-before-issue-stale: -1
+          days-before-issue-close: -1
+
+          days-before-pr-stale: 365
+          days-before-pr-close: 7
+
+          stale-pr-label: pr:inactive
+          close-pr-label: pr:auto-closed
+          exempt-pr-labels: pr:keep-open
+
+          stale-pr-message: >
+            This PR has been inactive for 365 days. If there is no new activity
+            within 7 days, it will be closed automatically. Comment, push new
+            commits, or remove the pr:inactive label to keep it open. Add
+            pr:keep-open to exempt it from future inactive PR cleanup.
+
+          close-pr-message: >
+            Closing this PR because it has had no activity for 7 days after
+            being marked inactive. If this is still useful or ready to move
+            forward, feel free to reopen it with fresh context or updated
+            details. Sorry for any inconvenience.
+
+          remove-pr-stale-when-updated: true
+          delete-branch: false
+          operations-per-run: 100
+          debug-only: ${{ github.event_name == 'workflow_dispatch' && inputs.debug_only || false }}
@@ -1,13 +1,20 @@
 build*/
+cmake-build-*/
 test/
+tests/
 .vscode/
+.idea/
 .cache/
 *.swp
-.vscode/
 *.bat
 *.bin
 *.exe
 *.gguf
 output*.png
 models*
-*.log
+*.log
+preview.png
+.claude/
+CLAUDE.local.md
+.agents/
+.codex/
@@ -1,3 +1,12 @@
 [submodule "ggml"]
    path = ggml
-	url = https://github.com/ggml-org/ggml.git
+	url = https://github.com/leejet/ggml.git
+[submodule "examples/server/frontend"]
+	path = examples/server/frontend
+	url = https://github.com/leejet/sdcpp-webui.git
+[submodule "thirdparty/libwebp"]
+	path = thirdparty/libwebp
+	url = https://github.com/webmproject/libwebp.git
+[submodule "thirdparty/libwebm"]
+	path = thirdparty/libwebm
+	url = https://github.com/webmproject/libwebm.git
@@ -0,0 +1,184 @@
+# Instructions for stable-diffusion.cpp
+
+This document is for AI coding agents working in this repository. It should
+describe agent-specific workflow, repository routing, editing boundaries, and
+project-specific pitfalls.
+
+For general contribution rules, including PR scope, commit conventions, code
+style, dependency updates, security hygiene, and AI-assisted contribution policy,
+see `CONTRIBUTING.md`.
+
+---
+
+## Agent Operating Rules
+
+Before analyzing or modifying the repository:
+
+1. Read this file.
+2. Use `rg` / `rg --files` or directory listing commands to confirm the current
+   tree before relying on a path.
+3. Start from `src/` and relevant `docs/` for runtime behavior.
+4. Read the relevant code before editing.
+5. Prefer the smallest change that fits the existing architecture.
+6. Report focused verification and mention any tests not run.
+
+Agents must not:
+
+* Run `git push`, create PRs, or submit issue/PR comments on the user's behalf.
+* Create commits unless the user explicitly requests that specific commit.
+* Modify `ggml/`, `thirdparty/`, or `examples/server/frontend/` unless
+  explicitly requested and necessary.
+* Read large local model files or tokenizer vocabulary files.
+* Rewrite unrelated code for style-only reasons.
+* Add secrets, model weights, generated binaries, local absolute paths, or
+  machine-specific output.
+
+When a change is large, architectural, or likely to affect public behavior,
+pause and present a short plan before editing.
+
+---
+
+## Repository Map and Editing Boundaries
+
+This is a routing map for agents, not a full architecture document. The layout
+can change, so verify paths before using them. Do not inspect excluded
+large-data directories while checking the tree.
+
+### Primary Project Code
+
+Core implementation lives under `src/`.
+
+Current source layout includes:
+
+* `src/core/` - shared tensor, ggml integration, backend, graph, RNG, and utility
+  code.
+* `src/model/` - model families and model components.
+* `src/model_io/` - model file loading, GGUF, safetensors, pickle, and related
+  serialization helpers.
+* `src/runtime/` - sampling, denoising, guidance, caching, preprocessing, and
+  runtime execution helpers.
+* `src/tokenizers/` - tokenizer implementations.
+* `src/conditioning/` - conditioning and prompt-related implementation.
+* `src/extensions/` - optional feature extensions.
+* top-level `src/*.cpp` and `src/*.h` files - public implementation entry
+  points, model loading, conversion, versioning, and shared managers.
+
+`src/tokenizers/vocab/` contains large tokenizer vocabulary data. Do not read or
+parse files in this directory; reference the path only when necessary.
+
+### Public API
+
+`include/` contains the C API exposed by the project. Currently the primary
+public header is `include/stable-diffusion.h`.
+
+Treat public headers as stable API. Avoid breaking compatibility unless the user
+explicitly requests it. If public behavior changes, update relevant examples or
+documentation.
+
+### Examples
+
+`examples/` contains programs demonstrating library usage.
+
+* `examples/cli/` - command line program for running models, testing features,
+  and debugging.
+* `examples/common/` - shared example support code.
+* `examples/server/` - server application built on top of the library.
+* `examples/server/frontend/` - git submodule containing independent frontend
+  code. Avoid modifying it unless explicitly requested.
+
+### Documentation and Tooling
+
+* `docs/` - documentation for supported models, build options, behavior, and
+  workflows.
+* `scripts/` - development, model processing, build automation, formatting, and
+  tooling scripts.
+* `cmake/` - CMake support modules.
+* `docker/` - Docker-related project files.
+* `assets/` - documentation assets; not runtime code.
+
+### External, Local, and Generated State
+
+* `ggml/` - git submodule for the ggml dependency.
+* `thirdparty/` - vendored third-party dependencies.
+* `models/` - local model storage. Ignore this directory and do not read model
+  files.
+* `test/` - local testing scripts. Use only when relevant to the task.
+* `build/`, `build_*`, and similar directories - generated build outputs.
+  Inspect them only when debugging a build result.
+
+---
+
+## Agent Workflow for Code Changes
+
+1. Identify the relevant modules under `src/`.
+2. Check whether the change touches the public API in `include/`.
+3. Consult relevant `docs/` and examples before changing user-facing behavior.
+4. Follow existing local patterns before adding new abstractions.
+5. Keep edits scoped to the requested behavior.
+6. Run the narrowest useful build, test, or inspection command available.
+
+Follow `CONTRIBUTING.md` for formatting, naming, PR expectations, dependency
+update policy, and security rules. For tokenizer additions, follow its embedded-data
+allowlist and default to an external `tokenizer.json`.
+
+---
+
+## Code Comments
+
+Keep comments rare and useful.
+
+Do not add comments that only describe what the code does. Add comments only
+when the code cannot fully express the logic, the logic is unusually complex, or
+there are historical reasons, invariants, constraints, compatibility concerns,
+or known pitfalls that future maintainers need to understand.
+
+Do not add task-specific comments that will be meaningless after review.
+
+Examples from the current codebase:
+
+```cpp
+// GOOD: explains a safety constraint that is not obvious from the assignment.
+// From src/model_io/pickle_io.cpp.
+// Non-tensor checkpoint metadata can use REDUCE for arbitrary
+// Python objects. Do not execute it; keep stack shape only.
+stack.push_back(make_none_value());
+
+// BAD: describes only what the next line does.
+// Set the token count to zero.
+token_count = 0;
+```
+
+---
+
+## Text File Encoding
+
+When reading or editing repository text files:
+
+* Prefer UTF-8 with LF for Markdown, frontend source, JSON, and other text-first
+  project files unless the file already clearly uses a different encoding.
+* Do not assume terminal output encoding matches file encoding on Windows.
+* A file that looks garbled in PowerShell output may still be valid UTF-8.
+* When inspecting UTF-8 files in PowerShell, prefer explicit UTF-8 reads such as:
+  * `[Console]::OutputEncoding = [System.Text.Encoding]::UTF8`
+  * `Get-Content -Encoding utf8 <path>`
+* Avoid rewriting a file purely because console output looked garbled; verify
+  the actual file encoding first.
+
+---
+
+## Tensor and Layout Notes
+
+Additional tensor/layout rules for this codebase:
+
+* `sd::Tensor` shape order is not PyTorch/NumPy-style. `shape()[0]` is the
+  lowest and most contiguous dimension, and higher indices are higher
+  dimensions.
+* Broadcasting for `sd::Tensor` must align dimensions from low to high dimension
+  indices. If one tensor has fewer dimensions, append implicit `1`s at the
+  higher-dimension end.
+* `ggml_n_dims` / `ggml_n_dims(tensor)` can drop trailing singleton high
+  dimensions. Do not assume a logical trailing dimension of `1` will still be
+  counted in ggml metadata.
+* Internal tensor-returning interfaces use an empty `sd::Tensor` to represent
+  null, absent, or failure states. Do not add `std::optional<sd::Tensor<...>>`
+  for internal APIs unless a distinct semantic state is truly required.
@@ -0,0 +1,11 @@
+@AGENTS.md
+
+## Claude Code
+
+Follow `AGENTS.md` as the shared repository instructions.
+
+Do not duplicate contribution, style, PR, dependency, or security policy here;
+use `CONTRIBUTING.md` as the canonical source for those rules.
+
+Keep Claude-specific project notes in this file only when they do not apply to
+other coding agents.
@@ -8,15 +8,72 @@ if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
 endif()

+if (MSVC)
+    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
+    add_compile_definitions(_SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING)
+    # /MP is MSVC-only: icx rejects it outright once offloading is enabled.
+    add_compile_options(
+        $<$<AND:$<COMPILE_LANGUAGE:C>,$<C_COMPILER_ID:MSVC>>:/MP>
+        $<$<COMPILE_LANGUAGE:C>:/utf-8>
+        $<$<AND:$<COMPILE_LANGUAGE:CXX>,$<CXX_COMPILER_ID:MSVC>>:/MP>
+        $<$<COMPILE_LANGUAGE:CXX>:/utf-8>
+    )
+endif()
+
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)

+if(APPLE)
+    function(sd_set_macos_rpaths target)
+        get_target_property(target_type ${target} TYPE)
+        if(target_type STREQUAL "EXECUTABLE")
+            set(runtime_paths "@executable_path" "@executable_path/../lib")
+        elseif(target_type STREQUAL "SHARED_LIBRARY" OR target_type STREQUAL "MODULE_LIBRARY")
+            set(runtime_paths "@loader_path" "@loader_path/../lib")
+            set_target_properties(${target} PROPERTIES
+                MACOSX_RPATH ON
+                INSTALL_NAME_DIR "@rpath"
+                BUILD_WITH_INSTALL_NAME_DIR ON
+            )
+        else()
+            return()
+        endif()
+
+        # Release artifacts zip the build output directly, so keep macOS rpaths relocatable.
+        set_target_properties(${target} PROPERTIES
+            BUILD_RPATH "${runtime_paths}"
+            INSTALL_RPATH "${runtime_paths}"
+            BUILD_WITH_INSTALL_RPATH ON
+        )
+    endfunction()
+endif()
+
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
    set(SD_STANDALONE ON)
 else()
    set(SD_STANDALONE OFF)
 endif()

+set(SD_SUBMODULE_WEBP FALSE)
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/thirdparty/libwebp/CMakeLists.txt")
+    set(SD_SUBMODULE_WEBP TRUE)
+endif()
+if(SD_SUBMODULE_WEBP)
+    set(SD_WEBP_DEFAULT ON)
+else()
+    set(SD_WEBP_DEFAULT ${SD_USE_SYSTEM_WEBP})
+endif()
+
+set(SD_SUBMODULE_WEBM FALSE)
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/thirdparty/libwebm/CMakeLists.txt")
+    set(SD_SUBMODULE_WEBM TRUE)
+endif()
+if(SD_SUBMODULE_WEBM)
+    set(SD_WEBM_DEFAULT ON)
+else()
+    set(SD_WEBM_DEFAULT ${SD_USE_SYSTEM_WEBM})
+endif()
+
 #
 # Option list
 #
@@ -24,6 +81,10 @@ endif()
 # general
 #option(SD_BUILD_TESTS                "sd: build tests"    ${SD_STANDALONE})
 option(SD_BUILD_EXAMPLES             "sd: build examples" ${SD_STANDALONE})
+option(SD_WEBP                       "sd: enable WebP image I/O support" ${SD_WEBP_DEFAULT})
+option(SD_USE_SYSTEM_WEBP            "sd: link against system libwebp" OFF)
+option(SD_WEBM                       "sd: enable WebM video output support" ${SD_WEBM_DEFAULT})
+option(SD_USE_SYSTEM_WEBM            "sd: link against system libwebm" OFF)
 option(SD_CUDA                       "sd: cuda backend" OFF)
 option(SD_HIPBLAS                    "sd: rocm backend" OFF)
 option(SD_METAL                      "sd: metal backend" OFF)
@@ -31,81 +92,222 @@ option(SD_VULKAN                     "sd: vulkan backend" OFF)
 option(SD_OPENCL                     "sd: opencl backend" OFF)
 option(SD_SYCL                       "sd: sycl backend" OFF)
 option(SD_MUSA                       "sd: musa backend" OFF)
-option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
 option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
+option(SD_BUILD_SHARED_GGML_LIB      "sd: build ggml as a separate shared lib" OFF)
 option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
+option(SD_USE_UPSTREAM_GGML          "sd: build with upstream GGML instead of the patched GGML extensions" OFF)
+set(SD_GGML_SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/ggml" CACHE PATH "sd: ggml source directory (also supplies private headers for system ggml)")
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)

+set(CMAKE_C_STANDARD 11)
+set(CMAKE_C_STANDARD_REQUIRED true)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED true)
+
 if(SD_CUDA)
    message("-- Use CUDA as backend stable-diffusion")
    set(GGML_CUDA ON)
-    add_definitions(-DSD_USE_CUDA)
 endif()

 if(SD_METAL)
    message("-- Use Metal as backend stable-diffusion")
    set(GGML_METAL ON)
-    add_definitions(-DSD_USE_METAL)
 endif()

 if (SD_VULKAN)
    message("-- Use Vulkan as backend stable-diffusion")
    set(GGML_VULKAN ON)
-    add_definitions(-DSD_USE_VULKAN)
 endif ()

 if (SD_OPENCL)
    message("-- Use OpenCL as backend stable-diffusion")
    set(GGML_OPENCL ON)
-    add_definitions(-DSD_USE_OPENCL)
 endif ()

 if (SD_HIPBLAS)
    message("-- Use HIPBLAS as backend stable-diffusion")
    set(GGML_HIP ON)
-    add_definitions(-DSD_USE_CUDA)
-    if(SD_FAST_SOFTMAX)
-        set(GGML_CUDA_FAST_SOFTMAX ON)
-    endif()
+    # ggml-hip's device-stub objects must be position-independent, or the
+    # default-PIE sd-cli link fails with `relocation R_X86_64_32 ... cannot be
+    # used when making a PIE object` on distros that default to PIE
+    # (Ubuntu 24.04, Fedora 40+, Debian 12+). The shared-library branch below
+    # already sets this; the static build (the HIP default) did not.
+    set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 endif ()

 if(SD_MUSA)
    message("-- Use MUSA as backend stable-diffusion")
    set(GGML_MUSA ON)
-    add_definitions(-DSD_USE_CUDA)
-    if(SD_FAST_SOFTMAX)
-        set(GGML_CUDA_FAST_SOFTMAX ON)
+endif()
+
+if(SD_WEBP)
+    if(NOT SD_SUBMODULE_WEBP AND NOT SD_USE_SYSTEM_WEBP)
+        message(FATAL_ERROR "WebP support enabled but no source found.
+          Either initialize the submodule:\n  git submodule update --init thirdparty/libwebp\n\n"
+          "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBP=ON")
+    endif()
+    if(SD_USE_SYSTEM_WEBP)
+        find_package(WebP)
+        if(WebP_FOUND)
+            add_library(webp ALIAS WebP::webp)
+            # libwebp CMake target naming is not consistent across versions/distros.
+            # Some export WebP::libwebpmux, others export WebP::webpmux.
+            if(TARGET WebP::libwebpmux)
+                add_library(libwebpmux ALIAS WebP::libwebpmux)
+            elseif(TARGET WebP::webpmux)
+                add_library(libwebpmux ALIAS WebP::webpmux)
+            else()
+                message(FATAL_ERROR
+                    "Could not find a compatible webpmux target in system WebP package. "
+                    "Expected WebP::libwebpmux or WebP::webpmux."
+                )
+            endif()
+        else()
+            find_package(PkgConfig REQUIRED)
+            pkg_check_modules(WebP REQUIRED IMPORTED_TARGET GLOBAL libwebp)
+            pkg_check_modules(WebPMux REQUIRED IMPORTED_TARGET GLOBAL libwebpmux)
+            link_libraries(PkgConfig::WebP)
+            link_libraries(PkgConfig::WebPMux)
+            add_library(libwebpmux ALIAS PkgConfig::WebPMux)
+        endif()
    endif()
 endif()

+if(SD_WEBM)
+    if(NOT SD_WEBP)
+        message(FATAL_ERROR "SD_WEBM requires SD_WEBP because WebM output reuses libwebp VP8 encoding.")
+    endif()
+    if(NOT SD_SUBMODULE_WEBM AND NOT SD_USE_SYSTEM_WEBM)
+        message(FATAL_ERROR "WebM support enabled but no source found.
+          Either initialize the submodule:\n  git submodule update --init thirdparty/libwebm\n\n"
+          "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBM=ON")
+    endif()
+    if(SD_USE_SYSTEM_WEBM)
+        find_package(PkgConfig)
+        if(PkgConfig_FOUND)
+            pkg_check_modules(WebM REQUIRED IMPORTED_TARGET GLOBAL libwebm)
+        endif()
+        if(PkgConfig_FOUND AND WebM_FOUND)
+            link_libraries(PkgConfig::WebM)
+        else()
+            find_path(WEBM_INCLUDE_DIR
+                NAMES mkvmuxer/mkvmuxer.h mkvparser/mkvparser.h common/webmids.h
+                PATH_SUFFIXES webm
+                REQUIRED)
+            find_library(WEBM_LIBRARY
+                NAMES webm libwebm
+                REQUIRED)
+
+            add_library(webm UNKNOWN IMPORTED)
+            set_target_properties(webm PROPERTIES
+                IMPORTED_LOCATION "${WEBM_LIBRARY}"
+                INTERFACE_INCLUDE_DIRECTORIES "${WEBM_INCLUDE_DIR}")
+        endif()
+    endif()
+endif()
+
+if (SD_RPC)
+    message("-- Use RPC as backend stable-diffusion")
+    set(GGML_RPC ON)
+    add_definitions(-DSD_USE_RPC)
+endif ()
+
 set(SD_LIB stable-diffusion)

-file(GLOB SD_LIB_SOURCES
-    "*.h"
-    "*.cpp"
-    "*.hpp"
+file(GLOB SD_LIB_SOURCES CONFIGURE_DEPENDS
+    "src/*.h"
+    "src/*.cpp"
+    "src/*.hpp"
+    "src/conditioning/*.h"
+    "src/conditioning/*.cpp"
+    "src/conditioning/*.hpp"
+    "src/core/*.h"
+    "src/core/*.cpp"
+    "src/core/*.hpp"
+    "src/extensions/*.h"
+    "src/extensions/*.cpp"
+    "src/extensions/*.hpp"
+    "src/model/*/*.h"
+    "src/model/*/*.cpp"
+    "src/model/*/*.hpp"
+    "src/pipeline/*.h"
+    "src/pipeline/*.cpp"
+    "src/runtime/*.h"
+    "src/runtime/*.cpp"
+    "src/runtime/*.hpp"
+    "src/model_io/*.h"
+    "src/model_io/*.cpp"
+    "src/tokenizers/*.h"
+    "src/tokenizers/*.cpp"
+    "src/tokenizers/vocab/*.h"
+    "src/tokenizers/vocab/*.cpp"
+)
+
+find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
+if(GIT_EXE)
+    execute_process(COMMAND ${GIT_EXE} describe --tags --abbrev=7 --dirty=+
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE SDCPP_BUILD_VERSION
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        ERROR_QUIET
+    )
+    execute_process(COMMAND ${GIT_EXE} rev-parse --short HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE SDCPP_BUILD_COMMIT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        ERROR_QUIET
+    )
+endif()
+
+if(NOT SDCPP_BUILD_VERSION)
+    set(SDCPP_BUILD_VERSION unknown)
+endif()
+message(STATUS "stable-diffusion.cpp version ${SDCPP_BUILD_VERSION}")
+
+if(NOT SDCPP_BUILD_COMMIT)
+    set(SDCPP_BUILD_COMMIT unknown)
+endif()
+message(STATUS "stable-diffusion.cpp commit ${SDCPP_BUILD_COMMIT}")
+
+set_property(
+  SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/src/version.cpp
+  APPEND PROPERTY COMPILE_DEFINITIONS
+  SDCPP_BUILD_COMMIT=${SDCPP_BUILD_COMMIT} SDCPP_BUILD_VERSION=${SDCPP_BUILD_VERSION}
 )

-# we can get only one share lib
 if(SD_BUILD_SHARED_LIBS)
    message("-- Build shared library")
    message(${SD_LIB_SOURCES})
-    set(BUILD_SHARED_LIBS OFF)
+    if(NOT SD_BUILD_SHARED_GGML_LIB)
+        set(BUILD_SHARED_LIBS OFF)
+    endif()
    add_library(${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)
+    if(NOT SD_BUILD_SHARED_GGML_LIB)
+        set(BUILD_SHARED_LIBS OFF)
+    endif()
    add_library(${SD_LIB} STATIC ${SD_LIB_SOURCES})
 endif()

+if(MSVC)
+    target_compile_options(${SD_LIB} PRIVATE $<$<COMPILE_LANGUAGE:CXX>:/bigobj>)
+    # ggml backends can throw C++ exceptions through their C API.
+    target_compile_options(${SD_LIB} PRIVATE $<$<AND:$<COMPILE_LANGUAGE:CXX>,$<CXX_COMPILER_ID:MSVC>>:/EHsc->)
+endif()
+
+if(APPLE)
+    sd_set_macos_rpaths(${SD_LIB})
+endif()
+
 if(SD_SYCL)
    message("-- Use SYCL as backend stable-diffusion")
    set(GGML_SYCL ON)
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
-    add_definitions(-DSD_USE_SYCL)
    # disable fast-math on host, see:
    # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html
    if (WIN32)
@@ -121,31 +323,87 @@ set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)

 if (NOT SD_USE_SYSTEM_GGML)
    # see https://github.com/ggerganov/ggml/pull/682
-    add_definitions(-DGGML_MAX_NAME=128)
+    add_definitions(-DGGML_MAX_NAME=160)
 endif()

 # deps
-# Only add ggml if it hasn't been added yet
-if (NOT TARGET ggml)
-    if (SD_USE_SYSTEM_GGML)
-        find_package(ggml REQUIRED)
-        if (NOT ggml_FOUND)
-            message(FATAL_ERROR "System-installed GGML library not found.")
-        endif()
-        add_library(ggml ALIAS ggml::ggml)
-    else()
-        add_subdirectory(ggml)
-    endif()
-endif()
+include(cmake/ggml.cmake)

 add_subdirectory(thirdparty)

-target_link_libraries(${SD_LIB} PUBLIC ggml zip)
+target_sources(${SD_LIB} PRIVATE $<TARGET_OBJECTS:zip>)
+target_link_libraries(${SD_LIB} PUBLIC ggml)
+find_package(Threads REQUIRED)
+target_link_libraries(${SD_LIB} PRIVATE Threads::Threads)
+target_link_libraries(${SD_LIB} PRIVATE onig sd-utf8proc)
+if (SD_CUDA)
+    find_package(CUDAToolkit REQUIRED)
+    # Keep the driver stub on downstream link lines when no driver is installed.
+    target_link_libraries(${SD_LIB} PUBLIC CUDA::cuda_driver)
+    set_property(SOURCE src/core/ggml_extend_backend.cpp APPEND PROPERTY COMPILE_DEFINITIONS SD_USE_CUDA)
+endif()
+target_include_directories(${SD_LIB} PUBLIC . src include)
+target_include_directories(${SD_LIB} PRIVATE src/core)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
-target_compile_features(${SD_LIB} PUBLIC cxx_std_11)
+target_compile_features(${SD_LIB} PUBLIC c_std_11 cxx_std_17)


 if (SD_BUILD_EXAMPLES)
    add_subdirectory(examples)
 endif()

+
+
+#
+# install
+#
+
+include(CMakePackageConfigHelpers)
+include(GNUInstallDirs)
+
+set(SD_INSTALL_VERSION "${SDCPP_BUILD_VERSION}")
+set(SD_INSTALL_COMMIT  "${SDCPP_BUILD_COMMIT}")
+set(SD_SHARED_LIB      ${SD_BUILD_SHARED_LIBS})
+
+set(SD_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
+set(SD_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
+set(SD_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
+
+set(SD_PUBLIC_HEADERS
+    ${CMAKE_CURRENT_SOURCE_DIR}/include/stable-diffusion.h)
+
+set_target_properties(${SD_LIB}
+    PROPERTIES
+        PUBLIC_HEADER "${SD_PUBLIC_HEADERS}")
+
+
+install(TARGETS ${SD_LIB}
+        ARCHIVE
+        LIBRARY
+        RUNTIME
+        PUBLIC_HEADER)
+
+
+configure_package_config_file(
+        ${CMAKE_CURRENT_SOURCE_DIR}/cmake/stable-diffusion-config.cmake.in
+        ${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion-config.cmake
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/stable-diffusion
+    PATH_VARS SD_INCLUDE_INSTALL_DIR
+              SD_LIB_INSTALL_DIR
+              SD_BIN_INSTALL_DIR )
+
+write_basic_package_version_file(
+        ${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion-version.cmake
+    VERSION ${SD_INSTALL_VERSION}
+    COMPATIBILITY SameMajorVersion)
+
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion-config.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion-version.cmake
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/stable-diffusion)
+
+configure_file(cmake/stable-diffusion.pc.in
+        "${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion.pc"
+        @ONLY)
+
+install(FILES "${CMAKE_CURRENT_BINARY_DIR}/stable-diffusion.pc"
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig)
@@ -0,0 +1,82 @@
+# Contributing
+
+This document collects general contribution conventions for this repository.
+
+## Before You Start
+
+Before opening a PR, please search existing PRs to avoid duplicating ongoing work.
+
+For large-scale refactors or changes with broad impact, please open an issue first to discuss the approach before submitting a PR.
+
+If you want to update a third-party dependency, please open an issue first instead of submitting a direct PR. See [Dependency Updates](#dependency-updates) for details.
+
+## Pull Requests
+
+Keep each PR focused on one clear change. Large or overly complex PRs are harder to review and may not be merged.
+
+Follow Conventional Commit-style subjects seen in history: `feat:`, `fix:`, `refactor:`, `ci:`, `docs:`, `chore:`. Keep subjects imperative and scoped.
+
+PRs should include:
+
+- What changed and why (short problem/solution summary).
+- Verification evidence when applicable (commands and key outputs).
+- Linked issue/PR context when applicable.
+- Screenshots or sample outputs for UI/visual behavior changes.
+
+## Code Style
+
+Format code according to the repository style before submitting changes.
+
+Formatting follows `.clang-format` (Chromium base, 4-space indent, no tabs). Run `scripts/format-code.sh` or `scripts/format-code.ps1` before opening a PR. Keep C++ standard at C++17-compatible patterns used in this repo.
+
+Naming conventions:
+
+- Use `PascalCase` for class/struct/type names.
+- In `PascalCase` names, preserve common abbreviations in uppercase, for example `SD`, `API`, `HTTP`, `JSON`, `RGB`, `VAE`, `TAE`, `LoRA`, and `WebP`.
+- Use `snake_case` for functions, methods, variables, and file names unless an existing API requires a different style.
+- Use a trailing underscore for private data member names, for example `hidden_size_` or `tokenizer_`.
+- Use `.h` for C and C++ header files. Do not introduce new `.hpp` headers.
+- Use macro-based header include guards instead of `#pragma once`.
+- Format header include guards as `__SD_{PATH}__`, where `{PATH}` is the header path in uppercase snake case without the file extension. For example, `src/sample.h` should use `__SD_SAMPLE_H__`.
+- Do not introduce anonymous namespaces in new or modified code; prefer `static` file-local functions/variables or an explicit named namespace when scoping is needed.
+- In `class`/`struct` definitions, place data members before member functions unless an existing type already clearly follows a different pattern.
+- Keep `test_*.cpp` / `test_*.py` naming for tests.
+
+Some older code in the project may not fully follow the current conventions. Please do not submit PRs that only rewrite existing code to match style rules.
+
+When adding or modifying model implementations, follow the model config and weight detection conventions in [docs/model_config.md](docs/model_config.md).
+
+## Tokenizer Data
+
+New model integrations must use an external `tokenizer.json` by default. Do not
+embed new vocabularies or merge tables solely for less widely used models;
+these tables increase the binary size for every user.
+
+The embedded-data allowlist is CLIP, T5/UMT5, Qwen 2/3, Mistral, and Gemma 3/4.
+Models may reuse an existing embedded tokenizer when its vocabulary and behavior
+match their text encoder. Gemma 2 and GPT-OSS require external JSON files.
+
+Adding to this allowlist requires maintainer approval, supported by the model's
+usage, reuse across models, and measured binary-size cost. Document the matching
+JSON and CLI option for models that require an external tokenizer, and fail
+initialization clearly when it is missing.
+
+## AI-Assisted Contributions
+
+AI tools may be used to assist development, but contributors are responsible for the quality and correctness of the submitted code.
+
+If any part of a contribution was generated with AI assistance, the contributor must perform a thorough human review before submitting the PR and understand every changed line.
+
+Do not list AI tools as co-authors. The human contributor is the sole responsible author of the submitted code.
+
+Please do not submit AI-generated code that you do not understand, and do not include meaningless experiments, temporary test code, or unrelated generated output in a PR.
+
+## Dependency Updates
+
+Do not submit PRs that update `ggml`. `ggml` updates are performed only after local validation by the maintainer.
+
+Other third-party dependencies are not updated unless necessary. If you want to update a dependency, please open an issue first instead of submitting a direct PR.
+
+## Security & Configuration
+
+Do not commit model weights, secrets, or local absolute paths. Keep large binaries out of git unless intentionally tracked release assets.
@@ -1,17 +0,0 @@
-ARG UBUNTU_VERSION=22.04
-
-FROM ubuntu:$UBUNTU_VERSION as build
-
-RUN apt-get update && apt-get install -y build-essential git cmake
-
-WORKDIR /sd.cpp
-
-COPY . .
-
-RUN mkdir build && cd build && cmake .. && cmake --build . --config Release
-
-FROM ubuntu:$UBUNTU_VERSION as runtime
-
-COPY --from=build /sd.cpp/build/bin/sd /sd
-
-ENTRYPOINT [ "/sd" ]
@@ -1,23 +0,0 @@
-ARG MUSA_VERSION=rc4.2.0
-ARG UBUNTU_VERSION=22.04
-
-FROM mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64 as build
-
-RUN apt-get update && apt-get install -y ccache cmake git
-
-WORKDIR /sd.cpp
-
-COPY . .
-
-RUN mkdir build && cd build && \
-    cmake .. -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
-        -DCMAKE_C_FLAGS="${CMAKE_C_FLAGS} -fopenmp -I/usr/lib/llvm-14/lib/clang/14.0.0/include -L/usr/lib/llvm-14/lib" \
-        -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -fopenmp -I/usr/lib/llvm-14/lib/clang/14.0.0/include -L/usr/lib/llvm-14/lib" \
-        -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release && \
-    cmake --build . --config Release
-
-FROM mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64 as runtime
-
-COPY --from=build /sd.cpp/build/bin/sd /sd
-
-ENTRYPOINT [ "/sd" ]
@@ -1,44 +1,105 @@
 <p align="center">
-  <img src="./assets/cat_with_sd_cpp_42.png" width="360x">
+  <img src="./assets/logo.png" width="360x">
 </p>

 # stable-diffusion.cpp

+<div align="center">
+<a href="https://trendshift.io/repositories/9714" target="_blank"><img src="https://trendshift.io/api/badge/repositories/9714" alt="leejet%2Fstable-diffusion.cpp | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+</div>
+
 Diffusion model(SD,Flux,Wan,...) inference in pure C/C++

 ***Note that this project is under active development. \
 API and command-line option may change frequently.***

+## 🔥Important News
+
+* **2026/09/20** 🚀 stable-diffusion.cpp adds **Day-0 support for Qwen-Image-2.1**
+* **2026/08/20** 🚀 stable-diffusion.cpp now supports **LTX-2.5**
+* **2026/08/04** 🚀 stable-diffusion.cpp adds **Day-1 support for MiniMax-H3**
+* **2026/06/25** 🚀 stable-diffusion.cpp now supports **Krea2**
+* **2026/06/04** 🚀 stable-diffusion.cpp now supports **Ideogram4**
+* **2026/05/31** 🚀 stable-diffusion.cpp now supports **PiD**
+* **2026/05/27** 🚀 stable-diffusion.cpp now supports **Lens**
+* **2026/05/17** 🚀 stable-diffusion.cpp now supports **LTX-2.3**
+* **2026/04/11** 🚀 stable-diffusion.cpp now uses a brand-new embedded web UI.  
+* **2026/01/18** 🚀 stable-diffusion.cpp now supports **FLUX.2-klein**  
+* **2025/12/01** 🚀 stable-diffusion.cpp now supports **Z-Image**  
+* **2025/11/30** 🚀 stable-diffusion.cpp now supports **FLUX.2-dev**  
+* **2025/10/13** 🚀 stable-diffusion.cpp now supports **Qwen-Image-Edit / Qwen-Image-Edit 2509**  
+* **2025/10/12** 🚀 stable-diffusion.cpp now supports **Qwen-Image**  
+* **2025/09/14** 🚀 stable-diffusion.cpp now supports **Wan2.1 Vace**  
+* **2025/09/06** 🚀 stable-diffusion.cpp now supports **Wan2.1 / Wan2.2**  
+
 ## 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)
+- Plain C/C++ implementation based on [ggml](https://github.com/ggml-org/ggml), working in the same way as [llama.cpp](https://github.com/ggml-org/llama.cpp)
 - Super lightweight and without external dependencies
 - Supported models
  - Image Models
-    - SD1.x, SD2.x, [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo)
-    - SDXL, [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo)
-      - !!!The VAE in SDXL encounters NaN issues under FP16, but unfortunately, the ggml_conv_2d only operates under FP16. Hence, a parameter is needed to specify the VAE that has fixed the FP16 NaN issue. You can find it here: [SDXL VAE FP16 Fix](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors).
+    - [SD1.x, SD2.x, SD-Turbo](./docs/sd.md)
+    - [SDXL, SDXL-Turbo](./docs/sd.md)
+    - [Some SD1.x and SDXL distilled models](./docs/distilled_sd.md)
    - [SD3/SD3.5](./docs/sd3.md)
-    - [Flux-dev/Flux-schnell](./docs/flux.md)
+    - [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
+    - [FLUX.2-dev/FLUX.2-klein](./docs/flux2.md)
+    - [Lens](./docs/lens.md)
    - [Chroma](./docs/chroma.md)
-  - Image Edit Models
+    - [Chroma1-Radiance](./docs/chroma_radiance.md)
+    - [Qwen Image](./docs/qwen_image.md)
+    - [Qwen Image 2.1](./docs/qwen_image_2.1.md)
+    - [PiD](./docs/pid.md)
+    - [LongCat Image](./docs/longcat_image.md)
+    - [Z-Image](./docs/z_image.md)
+    - [MiniT2I](./docs/minit2i.md)
+    - [SenseNova U1.5](./docs/sensenova_u1.md)
+    - [Ovis-Image](./docs/ovis_image.md)
+    - [Anima](./docs/anima.md)
+    - [ERNIE-Image](./docs/ernie_image.md)
+    - [Boogu Image](./docs/boogu_image.md)
+    - [Krea2](./docs/krea2.md)
+    - [Mage-Flow](./docs/mage_flow.md)
+    - [SeFi-Image](./docs/sefi_image.md)
+    - [HiDream-O1-Image](./docs/hidream_o1_image.md)
+    - [Ideogram4](./docs/ideogram4.md)
+  - [Image Edit Models](./docs/edit.md)
    - [FLUX.1-Kontext-dev](./docs/kontext.md)
+    - [Qwen Image Edit series](./docs/qwen_image_edit.md)
+    - [LongCat Image Edit](./docs/longcat_image.md)
+    - [Boogu Image Edit](./docs/boogu_image.md)
+    - [Mage-Flow-Edit](./docs/mage_flow.md#image-editing)
  - Video Models
    - [Wan2.1/Wan2.2](./docs/wan.md)
-  - [PhotoMaker](https://github.com/TencentARC/PhotoMaker) support.
+    - [MiniMax-H3](./docs/minimax_h3.md)
+    - [LTX-2.3/LTX-2.5](./docs/ltx2.md)
+    - [HunyuanVideo 1.5](./docs/hunyuan_video.md)
+    - [LingBot-Video](./docs/lingbot_video.md)
+  - [PhotoMaker](./docs/photo_maker.md) support.
+  - [IP-Adapter](./docs/ip_adapter.md) support (SD 1.5 and SDXL, including Plus)
  - Control Net support with SD 1.5
+  - [ADetailer](./docs/adetailer.md)
  - LoRA support, same as [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#lora)
  - Latent Consistency Models support (LCM/LCM-LoRA)
-  - Faster and memory efficient latent decoding with [TAESD](https://github.com/madebyollin/taesd)
-  - Upscale images generated with [ESRGAN](https://github.com/xinntao/Real-ESRGAN)
- 16-bit, 32-bit float support
- 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, Metal, Vulkan, OpenCL and SYCL backend for GPU acceleration.
- Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs models
-    - No need to convert to `.ggml` or `.gguf` anymore!
+  - Faster and memory efficient latent decoding with [TAESD](./docs/taesd.md)
+  - Upscale images generated with [ESRGAN](./docs/esrgan.md)
+- Supported backends
+  - CPU (AVX, AVX2 and AVX512 support for x86 architectures)
+  - CUDA
+  - Vulkan
+  - Metal
+  - OpenCL
+  - SYCL
+- Supported weight formats
+  - Pytorch checkpoint (`.ckpt` or `.pth` or `.pt`)
+  - Safetensors (`.safetensors`)
+  - GGUF (`.gguf`)
+- Convert mode supports converting model weights to `.gguf` or `.safetensors`
+- Supported platforms
+    - Linux
+    - Mac OS
+    - Windows
+    - Android (via Termux, [Local Diffusion](https://github.com/rmatif/Local-Diffusion))
 - Flash Attention for memory usage optimization
 - Negative prompt
 - [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) style tokenizer (not all the features, only token weighting for now)
@@ -51,378 +112,53 @@ API and command-line option may change frequently.***
    - `DPM++ 2M`
    - [`DPM++ 2M v2`](https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8457)
    - `DPM++ 2S a`
+    - `ER-SDE`
    - [`LCM`](https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/13952)
- Cross-platform reproducibility (`--rng cuda`, consistent with the `stable-diffusion-webui GPU RNG`)
+- Cross-platform reproducibility
+    - `--rng cuda`, default, consistent with the `stable-diffusion-webui GPU RNG`
+    - `--rng cpu`, consistent with the `comfyui RNG`
 - Embedds generation parameters into png output as webui-compatible text string
- Supported platforms
-    - Linux
-    - Mac OS
-    - Windows
-    - Android (via Termux, [Local Diffusion](https://github.com/rmatif/Local-Diffusion))

-### TODO
+## Quick Start

- [ ] More sampling methods
- [ ] Make inference faster
-    - The current implementation of ggml_conv_2d is slow and has high memory usage
- [ ] Continuing to reduce memory usage (quantizing the weights of ggml_conv_2d)
- [ ] Implement Inpainting support
+### Get the sd executable

-## Usage
+- Download pre-built binaries from the [releases page](https://github.com/leejet/stable-diffusion.cpp/releases)
+- Or build from source by following the [build guide](./docs/build.md)

-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.
+### Download model weights

-### Get the Code
+- download weights(.ckpt or .safetensors or .gguf). For example
+    - Stable Diffusion v1.5 from https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5 

-```
-git clone --recursive https://github.com/leejet/stable-diffusion.cpp
-cd stable-diffusion.cpp
-```
-
- If you have already cloned the repository, you can use the following command to update the repository to the latest code.
-
-```
-cd stable-diffusion.cpp
-git pull origin master
-git submodule init
-git submodule update
-```
-
-### Download weights
-
- download original weights(.ckpt or .safetensors). For example
-    - 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
+    ```sh
+    curl -L -O https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
    ```

-### Build
-
-#### Build from scratch
-
-```shell
-mkdir build
-cd build
-cmake ..
-cmake --build . --config Release
-```
-
-##### Using OpenBLAS
-
-```
-cmake .. -DGGML_OPENBLAS=ON
-cmake --build . --config Release
-```
-
-##### Using CUDA
-
-This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). Recommended to have at least 4 GB of VRAM.
-
-```
-cmake .. -DSD_CUDA=ON
-cmake --build . --config Release
-```
-
-##### Using HipBLAS
-This provides BLAS acceleration using the ROCm cores of your AMD GPU. Make sure to have the ROCm toolkit installed.
-
-Windows User Refer to [docs/hipBLAS_on_Windows.md](docs%2FhipBLAS_on_Windows.md) for a comprehensive guide.
-
-```
-cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=gfx1100 -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON
-cmake --build . --config Release
-```
-
-##### Using MUSA
-
-This provides BLAS acceleration using the MUSA cores of your Moore Threads GPU. Make sure to have the MUSA toolkit installed.
-
-```bash
-cmake .. -DCMAKE_C_COMPILER=/usr/local/musa/bin/clang -DCMAKE_CXX_COMPILER=/usr/local/musa/bin/clang++ -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release
-cmake --build . --config Release
-```
-
-##### 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 OpenCL (for Adreno GPU)
-
-Currently, it supports only Adreno GPUs and is primarily optimized for Q4_0 type
-
-To build for Windows ARM please refers to [Windows 11 Arm64
-](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/OPENCL.md#windows-11-arm64)
-
-Building for Android:
-
-  Android NDK:
-       Download and install the Android NDK from the [official Android developer site](https://developer.android.com/ndk/downloads).
-
-Setup OpenCL Dependencies for NDK:
-
-You need to provide OpenCL headers and the ICD loader library to your NDK sysroot.
-
-*   OpenCL Headers:
-    ```bash
-    # In a temporary working directory
-    git clone https://github.com/KhronosGroup/OpenCL-Headers
-    cd OpenCL-Headers
-    # Replace <YOUR_NDK_PATH> with your actual NDK installation path
-    # e.g., cp -r CL /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
-    sudo cp -r CL <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
-    cd ..
-    ```
-
-*   OpenCL ICD Loader:
-    ```bash
-    # In the same temporary working directory
-    git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
-    cd OpenCL-ICD-Loader
-    mkdir build_ndk && cd build_ndk
-
-    # Replace <YOUR_NDK_PATH> in the CMAKE_TOOLCHAIN_FILE and OPENCL_ICD_LOADER_HEADERS_DIR
-    cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \
-      -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
-      -DOPENCL_ICD_LOADER_HEADERS_DIR=<YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \
-      -DANDROID_ABI=arm64-v8a \
-      -DANDROID_PLATFORM=24 \
-      -DANDROID_STL=c++_shared
-
-    ninja
-    # Replace <YOUR_NDK_PATH>
-    # e.g., cp libOpenCL.so /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
-    sudo cp libOpenCL.so <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
-    cd ../..
-    ```
-
-Build `stable-diffusion.cpp` for Android with OpenCL:
-
-```bash
-mkdir build-android && cd build-android
-
-# Replace <YOUR_NDK_PATH> with your actual NDK installation path
-# e.g., -DCMAKE_TOOLCHAIN_FILE=/path/to/android-ndk-r26c/build/cmake/android.toolchain.cmake
-cmake .. -G Ninja \
-  -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
-  -DANDROID_ABI=arm64-v8a \
-  -DANDROID_PLATFORM=android-28 \
-  -DGGML_OPENMP=OFF \
-  -DSD_OPENCL=ON
-
-ninja
-```
-*(Note: Don't forget to include `LD_LIBRARY_PATH=/vendor/lib64` in your command line before running the binary)*
-
-##### Using SYCL
-
-Using SYCL makes the computation run on the Intel GPU. Please make sure you have installed the related driver and [Intel® oneAPI Base toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) before start. More details and steps can refer to [llama.cpp SYCL backend](https://github.com/ggerganov/llama.cpp/blob/master/docs/backend/SYCL.md#linux).
-
-```
-# Export relevant ENV variables
-source /opt/intel/oneapi/setvars.sh
-
-# Option 1: Use FP32 (recommended for better performance in most cases)
-cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-# Option 2: Use FP16
-cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
-
-cmake --build . --config Release
-```
-
-Example of text2img by using SYCL backend:
-
- download `stable-diffusion` model weight, refer to [download-weight](#download-weights).
-
- run `./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors --cfg-scale 5 --steps 30 --sampling-method euler  -H 1024 -W 1024 --seed 42 -p "fantasy medieval village world inside a glass sphere , high detail, fantasy, realistic, light effect, hyper detail, volumetric lighting, cinematic, macro, depth of field, blur, red light and clouds from the back, highly detailed epic cinematic concept art cg render made in maya, blender and photoshop, octane render, excellent composition, dynamic dramatic cinematic lighting, aesthetic, very inspirational, world inside a glass sphere by james gurney by artgerm with james jean, joe fenton and tristan eaton by ross tran, fine details, 4k resolution"`
-
-<p align="center">
-  <img src="./assets/sycl_sd3_output.png" width="360x">
-</p>
-
-
-
-##### Using Flash Attention
-
-Enabling flash attention for the diffusion model reduces memory usage by varying amounts of MB.
-eg.:
- - flux 768x768 ~600mb
- - SD2 768x768 ~1400mb
-
-For most backends, it slows things down, but for cuda it generally speeds it up too.
-At the moment, it is only supported for some models and some backends (like cpu, cuda/rocm, metal).
-
-Run by adding `--diffusion-fa` to the arguments and watch for:
-```
-[INFO ] stable-diffusion.cpp:312  - Using flash attention in the diffusion model
-```
-and the compute buffer shrink in the debug log:
-```
-[DEBUG] ggml_extend.hpp:1004 - flux compute buffer size: 650.00 MB(VRAM)
-```
-
-### Run
-
-```
-usage: ./bin/sd [arguments]
-
-arguments:
-  -h, --help                         show this help message and exit
-  -M, --mode [MODE]                  run mode, one of: [img_gen, vid_gen, convert], default: img_gen
-  -t, --threads N                    number of threads to use during computation (default: -1)
-                                     If threads <= 0, then threads will be set to the number of CPU physical cores
-  --offload-to-cpu                   place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
-  -m, --model [MODEL]                path to full model
-  --diffusion-model                  path to the standalone diffusion model
-  --high-noise-diffusion-model       path to the standalone high noise diffusion model
-  --clip_l                           path to the clip-l text encoder
-  --clip_g                           path to the clip-g text encoder
-  --clip_vision                      path to the clip-vision encoder
-  --t5xxl                            path to the t5xxl text encoder
-  --vae [VAE]                        path to vae
-  --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
-  --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 (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K)
-                                     If not specified, the default is the type of the weight file
-  --tensor-type-rules [EXPRESSION]   weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
-  --lora-model-dir [DIR]             lora model directory
-  -i, --init-img [IMAGE]             path to the init image, required by img2img
-  --mask [MASK]                      path to the mask image, required by img2img with mask
-  -i, --end-img [IMAGE]              path to the end image, required by flf2v
-  --control-image [IMAGE]            path to image condition, control net
-  -r, --ref-image [PATH]             reference image for Flux Kontext models (can be used multiple times)
-  -o, --output OUTPUT                path to write result image to (default: ./output.png)
-  -p, --prompt [PROMPT]              the prompt to render
-  -n, --negative-prompt PROMPT       the negative prompt (default: "")
-  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)
-  --img-cfg-scale SCALE              image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
-  --guidance SCALE                   distilled guidance scale for models with guidance input (default: 3.5)
-  --slg-scale SCALE                  skip layer guidance (SLG) scale, only for DiT models: (default: 0)
-                                     0 means disabled, a value of 2.5 is nice for sd3.5 medium
-  --eta SCALE                        eta in DDIM, only for DDIM and TCD: (default: 0)
-  --skip-layers LAYERS               Layers to skip for SLG steps: (default: [7,8,9])
-  --skip-layer-start START           SLG enabling point: (default: 0.01)
-  --skip-layer-end END               SLG disabling point: (default: 0.2)
-  --scheduler {discrete, karras, exponential, ays, gits} Denoiser sigma scheduler (default: discrete)
-  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
-                                     sampling method (default: "euler_a")
-  --steps  STEPS                     number of sample steps (default: 20)
-  --high-noise-cfg-scale SCALE       (high noise) unconditional guidance scale: (default: 7.0)
-  --high-noise-img-cfg-scale SCALE   (high noise) image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
-  --high-noise-guidance SCALE        (high noise) distilled guidance scale for models with guidance input (default: 3.5)
-  --high-noise-slg-scale SCALE       (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
-                                     0 means disabled, a value of 2.5 is nice for sd3.5 medium
-  --high-noise-eta SCALE             (high noise) eta in DDIM, only for DDIM and TCD: (default: 0)
-  --high-noise-skip-layers LAYERS    (high noise) Layers to skip for SLG steps: (default: [7,8,9])
-  --high-noise-skip-layer-start      (high noise) SLG enabling point: (default: 0.01)
-  --high-noise-skip-layer-end END    (high noise) SLG disabling point: (default: 0.2)
-  --high-noise-scheduler {discrete, karras, exponential, ays, gits} Denoiser sigma scheduler (default: discrete)
-  --high-noise-sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
-                                     (high noise) sampling method (default: "euler_a")
-  --high-noise-steps  STEPS          (high noise) number of sample steps (default: -1 = auto)
-                                     SLG will be enabled at step int([STEPS]*[START]) and disabled at int([STEPS]*[END])
-  --strength STRENGTH                strength for noising/unnoising (default: 0.75)
-  --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)
-  --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
-  --clip-skip N                      ignore last_dot_pos 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.
-  --diffusion-conv-direct            use Conv2d direct in the diffusion model
-                                     This might crash if it is not supported by the backend.
-  --vae-conv-direct                  use Conv2d direct in the vae model (should improve the performance)
-                                     This might crash if it is not supported by the backend.
-  --control-net-cpu                  keep controlnet in cpu (for low vram)
-  --canny                            apply canny preprocessor (edge detection)
-  --color                            colors the logging tags according to level
-  --chroma-disable-dit-mask          disable dit mask for chroma
-  --chroma-enable-t5-mask            enable t5 mask for chroma
-  --chroma-t5-mask-pad  PAD_SIZE     t5 mask pad size of chroma
-  --video-frames                     video frames (default: 1)
-  --fps                              fps (default: 24)
-  --moe-boundary BOUNDARY            timestep boundary for Wan2.2 MoE model. (default: 0.875)
-                                     only enabled if `--high-noise-steps` is set to -1
-  --flow-shift SHIFT                 shift value for Flow models like SD3.x or WAN (default: auto)
-  -v, --verbose                      print extra info
-```
-
-#### txt2img example
+### Generate an image with just one command

 ```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
+./bin/sd-cli -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
 ```

-Using formats of different precisions will yield results of varying quality.
+***For detailed command-line arguments, check out [cli doc](./examples/cli/README.md).***

-| f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
-| ----  |----  |----  |----  |----  |----  |----  |
-| ![](./assets/f32.png) |![](./assets/f16.png) |![](./assets/q8_0.png) |![](./assets/q5_0.png) |![](./assets/q5_1.png) |![](./assets/q4_0.png) |![](./assets/q4_1.png) |
+## Performance

-#### img2img example
-
- `./output.png` is the image generated from the above txt2img pipeline
-
-
-```
-./bin/sd -m ../models/sd-v1-4.ckpt -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
-```
-
-<p align="center">
-  <img src="./assets/img2img_output.png" width="256x">
-</p>
+If you want to improve performance or reduce VRAM/RAM usage, please refer to [performance guide](./docs/performance.md).
+For runtime and parameter backend placement, see the [backend selection guide](./docs/backend.md).

 ## More Guides

+- [Troubleshooting](./docs/troubleshooting.md)
+- [Backend selection](./docs/backend.md)
+- [RPC](./docs/rpc.md)
 - [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)
+- [INT8 convrot safetensors](./docs/int8_convrot.md)
+- [Inference acceleration via caching](./docs/caching.md)

 ## Bindings

@@ -430,6 +166,7 @@ These projects wrap `stable-diffusion.cpp` for easier use in other languages/fra

 * Golang (non-cgo): [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
 * Golang (cgo): [Binozo/GoStableDiffusion](https://github.com/Binozo/GoStableDiffusion)
+* Golang (non-cgo): [l8bloom/gosd](https://github.com/l8bloom/gosd)
 * 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)
@@ -439,34 +176,18 @@ These projects wrap `stable-diffusion.cpp` for easier use in other languages/fra

 These projects use `stable-diffusion.cpp` as a backend for their image generation.

+- [GIMP Plugins](https://github.com/themanyone/gimp-plugins)
 - [Jellybox](https://jellybox.com)
 - [Stable Diffusion GUI](https://github.com/fszontagh/sd.cpp.gui.wx)
 - [Stable Diffusion CLI-GUI](https://github.com/piallai/stable-diffusion.cpp)
 - [Local Diffusion](https://github.com/rmatif/Local-Diffusion)
 - [sd.cpp-webui](https://github.com/daniandtheweb/sd.cpp-webui)
 - [LocalAI](https://github.com/mudler/LocalAI)
+- [Neural-Pixel](https://github.com/Luiz-Alcantara/Neural-Pixel)
+- [KoboldCpp](https://github.com/LostRuins/koboldcpp)

 ## Contributors

 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)
- [Wan2.1](https://github.com/Wan-Video/Wan2.1)
- [Wan2.2](https://github.com/Wan-Video/Wan2.2)
@@ -1,962 +0,0 @@
-#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, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
-        enum ggml_type token_wtype    = GGML_TYPE_F32;
-        enum ggml_type position_wtype = GGML_TYPE_F32;
-
-        params["token_embedding.weight"]    = ggml_new_tensor_2d(ctx, token_wtype, embed_dim, vocab_size);
-        params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
-    }
-
-public:
-    CLIPEmbeddings(int64_t embed_dim,
-                   int64_t vocab_size    = 49408,
-                   int64_t num_positions = 77)
-        : 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, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
-        enum ggml_type patch_wtype    = GGML_TYPE_F16;
-        enum ggml_type class_wtype    = GGML_TYPE_F32;
-        enum ggml_type position_wtype = GGML_TYPE_F32;
-
-        params["patch_embedding.weight"]    = ggml_new_tensor_4d(ctx, patch_wtype, patch_size, patch_size, num_channels, embed_dim);
-        params["class_embedding"]           = ggml_new_tensor_1d(ctx, class_wtype, embed_dim);
-        params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
-    }
-
-public:
-    CLIPVisionEmbeddings(int64_t embed_dim,
-                         int64_t num_channels = 3,
-                         int64_t patch_size   = 14,
-                         int64_t image_size   = 224)
-        : embed_dim(embed_dim),
-          num_channels(num_channels),
-          patch_size(patch_size),
-          image_size(image_size) {
-        num_patches   = (image_size / patch_size) * (image_size / patch_size);
-        num_positions = num_patches + 1;
-    }
-
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* pixel_values) {
-        // pixel_values: [N, num_channels, image_size, image_size]
-        // return: [N, num_positions, embed_dim]
-        GGML_ASSERT(pixel_values->ne[0] == image_size && pixel_values->ne[1] == image_size && pixel_values->ne[2] == num_channels);
-
-        auto patch_embed_weight    = params["patch_embedding.weight"];
-        auto class_embed_weight    = params["class_embedding"];
-        auto position_embed_weight = params["position_embedding.weight"];
-
-        // concat(patch_embedding, class_embedding) + position_embedding
-        struct ggml_tensor* patch_embedding;
-        int64_t N       = pixel_values->ne[3];
-        patch_embedding = ggml_nn_conv_2d(ctx, pixel_values, patch_embed_weight, NULL, patch_size, patch_size);  // [N, embed_dim, image_size // pacht_size, image_size // pacht_size]
-        patch_embedding = ggml_reshape_3d(ctx, patch_embedding, num_patches, embed_dim, N);                      // [N, embed_dim, num_patches]
-        patch_embedding = ggml_cont(ctx, ggml_permute(ctx, patch_embedding, 1, 0, 2, 3));                        // [N, num_patches, embed_dim]
-        patch_embedding = ggml_reshape_4d(ctx, patch_embedding, 1, embed_dim, num_patches, N);                   // [N, num_patches, embed_dim, 1]
-
-        struct ggml_tensor* class_embedding = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, embed_dim, N);
-        class_embedding                     = ggml_repeat(ctx, class_embed_weight, class_embedding);      // [N, embed_dim]
-        class_embedding                     = ggml_reshape_4d(ctx, class_embedding, 1, embed_dim, 1, N);  // [N, 1, embed_dim, 1]
-
-        struct ggml_tensor* x = ggml_concat(ctx, class_embedding, patch_embedding, 2);  // [N, num_positions, embed_dim, 1]
-        x                     = ggml_reshape_3d(ctx, x, embed_dim, num_positions, N);   // [N, num_positions, embed_dim]
-        x                     = ggml_add(ctx, x, position_embed_weight);
-        return x;  // [N, num_positions, embed_dim]
-    }
-};
-
-// OPENAI_CLIP_VIT_L_14: https://huggingface.co/openai/clip-vit-large-patch14/blob/main/config.json
-// OPEN_CLIP_VIT_H_14: https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K/blob/main/config.json
-// OPEN_CLIP_VIT_BIGG_14: https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k/blob/main/config.json (CLIPTextModelWithProjection)
-
-enum CLIPVersion {
-    OPENAI_CLIP_VIT_L_14,   // SD 1.x and SDXL
-    OPEN_CLIP_VIT_H_14,     // SD 2.x
-    OPEN_CLIP_VIT_BIGG_14,  // SDXL
-};
-
-class CLIPTextModel : public GGMLBlock {
-protected:
-    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
-        if (version == OPEN_CLIP_VIT_BIGG_14) {
-            enum ggml_type wtype      = GGML_TYPE_F32;
-            params["text_projection"] = ggml_new_tensor_2d(ctx, wtype, projection_dim, hidden_size);
-        }
-    }
-
-public:
-    CLIPVersion version = OPENAI_CLIP_VIT_L_14;
-    // network hparams
-    int32_t vocab_size        = 49408;
-    int32_t n_token           = 77;  // max_position_embeddings
-    int32_t hidden_size       = 768;
-    int32_t intermediate_size = 3072;
-    int32_t n_head            = 12;
-    int32_t n_layer           = 12;    // num_hidden_layers
-    int32_t projection_dim    = 1280;  // only for OPEN_CLIP_VIT_BIGG_14
-    int32_t clip_skip         = -1;
-    bool with_final_ln        = true;
-
-    CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
-                  bool with_final_ln  = true,
-                  int clip_skip_value = -1)
-        : 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) {
-            skip = -1;
-        }
-        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("identity projection");
-            }
-            return pooled;  // [hidden_size, 1, 1]
-        }
-
-        return x;  // [N, n_token, hidden_size]
-    }
-};
-
-class CLIPVisionModel : public GGMLBlock {
-public:
-    // network hparams
-    int32_t num_channels      = 3;
-    int32_t patch_size        = 14;
-    int32_t image_size        = 224;
-    int32_t num_positions     = 257;  // (image_size / patch_size)^2 + 1
-    int32_t hidden_size       = 1024;
-    int32_t intermediate_size = 4096;
-    int32_t n_head            = 16;
-    int32_t n_layer           = 24;
-
-public:
-    CLIPVisionModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14) {
-        if (version == OPEN_CLIP_VIT_H_14) {
-            hidden_size       = 1280;
-            intermediate_size = 5120;
-            n_head            = 16;
-            n_layer           = 32;
-        } else if (version == OPEN_CLIP_VIT_BIGG_14) {
-            hidden_size       = 1664;
-            intermediate_size = 8192;
-            n_head            = 16;
-            n_layer           = 48;
-        }
-
-        blocks["embeddings"]     = std::shared_ptr<GGMLBlock>(new CLIPVisionEmbeddings(hidden_size, num_channels, patch_size, image_size));
-        blocks["pre_layernorm"]  = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
-        blocks["encoder"]        = std::shared_ptr<GGMLBlock>(new CLIPEncoder(n_layer, hidden_size, n_head, intermediate_size));
-        blocks["post_layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
-    }
-
-    struct ggml_tensor* forward(struct ggml_context* ctx,
-                                struct ggml_tensor* pixel_values,
-                                bool return_pooled = true,
-                                int clip_skip      = -1) {
-        // pixel_values: [N, num_channels, image_size, image_size]
-        auto embeddings     = std::dynamic_pointer_cast<CLIPVisionEmbeddings>(blocks["embeddings"]);
-        auto pre_layernorm  = std::dynamic_pointer_cast<LayerNorm>(blocks["pre_layernorm"]);
-        auto encoder        = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
-        auto post_layernorm = std::dynamic_pointer_cast<LayerNorm>(blocks["post_layernorm"]);
-
-        auto x = embeddings->forward(ctx, pixel_values);  // [N, num_positions, embed_dim]
-        x      = pre_layernorm->forward(ctx, x);
-        x      = encoder->forward(ctx, x, clip_skip, false);
-        // print_ggml_tensor(x, true, "ClipVisionModel x: ");
-        auto last_hidden_state = x;
-        x                      = post_layernorm->forward(ctx, x);  // [N, n_token, hidden_size]
-
-        GGML_ASSERT(x->ne[3] == 1);
-        if (return_pooled) {
-            ggml_tensor* pooled = ggml_cont(ctx, ggml_view_2d(ctx, x, x->ne[0], x->ne[2], x->nb[2], 0));
-            return pooled;  // [N, hidden_size]
-        } else {
-            // return x;  // [N, n_token, hidden_size]
-            return last_hidden_state;  // [N, n_token, hidden_size]
-        }
-    }
-};
-
-class CLIPProjection : public UnaryBlock {
-protected:
-    int64_t in_features;
-    int64_t out_features;
-    bool transpose_weight;
-
-    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
-        enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
-        if (transpose_weight) {
-            params["weight"] = ggml_new_tensor_2d(ctx, wtype, out_features, in_features);
-        } else {
-            params["weight"] = ggml_new_tensor_2d(ctx, wtype, in_features, out_features);
-        }
-    }
-
-public:
-    CLIPProjection(int64_t in_features,
-                   int64_t out_features,
-                   bool transpose_weight = false)
-        : in_features(in_features),
-          out_features(out_features),
-          transpose_weight(transpose_weight) {}
-
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
-        struct ggml_tensor* w = params["weight"];
-        if (transpose_weight) {
-            w = ggml_cont(ctx, ggml_transpose(ctx, w));
-        }
-        return ggml_nn_linear(ctx, x, w, NULL);
-    }
-};
-
-class CLIPVisionModelProjection : public GGMLBlock {
-public:
-    int32_t hidden_size    = 1024;
-    int32_t projection_dim = 768;
-    int32_t image_size     = 224;
-
-public:
-    CLIPVisionModelProjection(CLIPVersion version   = OPENAI_CLIP_VIT_L_14,
-                              bool transpose_proj_w = false) {
-        if (version == OPEN_CLIP_VIT_H_14) {
-            hidden_size    = 1280;
-            projection_dim = 1024;
-        } else if (version == OPEN_CLIP_VIT_BIGG_14) {
-            hidden_size = 1664;
-        }
-
-        blocks["vision_model"]      = std::shared_ptr<GGMLBlock>(new CLIPVisionModel(version));
-        blocks["visual_projection"] = std::shared_ptr<GGMLBlock>(new CLIPProjection(hidden_size, projection_dim, transpose_proj_w));
-    }
-
-    struct ggml_tensor* forward(struct ggml_context* ctx,
-                                struct ggml_tensor* pixel_values,
-                                bool return_pooled = true,
-                                int clip_skip      = -1) {
-        // pixel_values: [N, num_channels, image_size, image_size]
-        // return: [N, projection_dim] if return_pooled else [N, n_token, hidden_size]
-        auto vision_model      = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
-        auto visual_projection = std::dynamic_pointer_cast<CLIPProjection>(blocks["visual_projection"]);
-
-        auto x = vision_model->forward(ctx, pixel_values, return_pooled, clip_skip);  // [N, hidden_size] or [N, n_token, hidden_size]
-
-        if (return_pooled) {
-            x = visual_projection->forward(ctx, x);  // [N, projection_dim]
-        }
-
-        return x;
-    }
-};
-
-struct CLIPTextModelRunner : public GGMLRunner {
-    CLIPTextModel model;
-
-    CLIPTextModelRunner(ggml_backend_t backend,
-                        bool offload_params_to_cpu,
-                        const String2GGMLType& tensor_types,
-                        const std::string prefix,
-                        CLIPVersion version = OPENAI_CLIP_VIT_L_14,
-                        bool with_final_ln  = true,
-                        int clip_skip_value = -1)
-        : GGMLRunner(backend, offload_params_to_cpu), model(version, with_final_ln, clip_skip_value) {
-        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__
@@ -0,0 +1,27 @@
+if(NOT TARGET ggml AND NOT TARGET ggml::ggml)
+    if(SD_USE_SYSTEM_GGML)
+        find_package(ggml REQUIRED)
+    else()
+        add_subdirectory("${SD_GGML_SOURCE_DIR}" "${CMAKE_CURRENT_BINARY_DIR}/ggml")
+    endif()
+endif()
+if(NOT TARGET ggml)
+    add_library(ggml ALIAS ggml::ggml)
+endif()
+
+get_target_property(sd_ggml_imported ggml IMPORTED)
+if(sd_ggml_imported)
+    set(sd_ggml_private_include "${SD_GGML_SOURCE_DIR}/src")
+else()
+    get_target_property(sd_ggml_private_include ggml SOURCE_DIR)
+endif()
+if(NOT EXISTS "${sd_ggml_private_include}/ggml-impl.h")
+    message(FATAL_ERROR "Set SD_GGML_SOURCE_DIR to the source tree matching the selected ggml library (ggml-impl.h is required).")
+endif()
+target_include_directories(${SD_LIB} PRIVATE "${sd_ggml_private_include}")
+set_property(TARGET ${SD_LIB} PROPERTY SD_GGML_PRIVATE_INCLUDE_DIR "${sd_ggml_private_include}")
+
+if(SD_USE_UPSTREAM_GGML)
+    target_compile_definitions(${SD_LIB} PUBLIC SD_USE_UPSTREAM_GGML)
+    message(WARNING "Using upstream GGML: INT8 tensorwise/convrot is disabled and FP8 weights are converted to F16 at load time. Some operators may be unsupported and performance may be lower than with patched GGML.")
+endif()
@@ -0,0 +1,45 @@
+set(SD_VERSION      "@SD_INSTALL_VERSION@")
+set(SD_BUILD_COMMIT "@SD_INSTALL_COMMIT@")
+set(SD_SHARED_LIB    @SD_SHARED_LIB@)
+
+@PACKAGE_INIT@
+
+set_and_check(SD_INCLUDE_DIR "@PACKAGE_SD_INCLUDE_INSTALL_DIR@")
+set_and_check(SD_LIB_DIR     "@PACKAGE_SD_LIB_INSTALL_DIR@")
+set(SD_BIN_DIR "@PACKAGE_SD_BIN_INSTALL_DIR@")
+
+include(CMakeFindDependencyMacro)
+find_dependency(ggml REQUIRED HINTS "${SD_LIB_DIR}/cmake")
+find_dependency(Threads REQUIRED)
+if(@SD_CUDA@)
+    find_dependency(CUDAToolkit REQUIRED)
+endif()
+
+if(NOT TARGET stable-diffusion)
+    find_library(stable-diffusion_LIBRARY stable-diffusion
+        REQUIRED
+        HINTS "${SD_LIB_DIR}"
+        NO_CMAKE_FIND_ROOT_PATH
+    )
+
+    add_library(stable-diffusion UNKNOWN IMPORTED)
+    set_target_properties(stable-diffusion
+        PROPERTIES
+            INTERFACE_INCLUDE_DIRECTORIES "${SD_INCLUDE_DIR}"
+            INTERFACE_LINK_LIBRARIES "ggml::ggml;Threads::Threads"
+            IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
+            IMPORTED_LOCATION "${stable-diffusion_LIBRARY}"
+            INTERFACE_COMPILE_FEATURES "c_std_11;cxx_std_17"
+            POSITION_INDEPENDENT_CODE ON)
+
+    if(@SD_CUDA@)
+        set_property(TARGET stable-diffusion APPEND PROPERTY INTERFACE_LINK_LIBRARIES CUDA::cuda_driver)
+    endif()
+
+    if(SD_SHARED_LIB)
+        target_compile_definitions(stable-diffusion
+            INTERFACE SD_BUILD_SHARED_LIB)
+    endif()
+endif()
+
+check_required_components(stable-diffusion)
@@ -0,0 +1,11 @@
+prefix=@CMAKE_INSTALL_PREFIX@
+exec_prefix=${prefix}
+libdir=@CMAKE_INSTALL_FULL_LIBDIR@
+includedir=@CMAKE_INSTALL_FULL_INCLUDEDIR@
+
+Name: stable-diffusion
+Description: Diffusion model(SD,Flux,Wan,Qwen Image,Z-Image,...) inference in pure C/C++
+Version: @SDCPP_BUILD_VERSION@
+Libs: -L${libdir} -lstable-diffusion
+Libs.private: -lggml -lggml-base @CMAKE_THREAD_LIBS_INIT@
+Cflags: -I${includedir}
@@ -1,263 +0,0 @@
-#ifndef __DIFFUSION_MODEL_H__
-#define __DIFFUSION_MODEL_H__
-
-#include "flux.hpp"
-#include "mmdit.hpp"
-#include "unet.hpp"
-#include "wan.hpp"
-
-struct DiffusionModel {
-    virtual std::string get_desc()                                                      = 0;
-    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,
-                         std::vector<ggml_tensor*> ref_latents     = {},
-                         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,
-              bool offload_params_to_cpu,
-              const String2GGMLType& tensor_types = {},
-              SDVersion version                   = VERSION_SD1,
-              bool flash_attn                     = false)
-        : unet(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model", version, flash_attn) {
-    }
-
-    std::string get_desc() {
-        return unet.get_desc();
-    }
-
-    void alloc_params_buffer() {
-        unet.alloc_params_buffer();
-    }
-
-    void free_params_buffer() {
-        unet.free_params_buffer();
-    }
-
-    void free_compute_buffer() {
-        unet.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
-        unet.get_param_tensors(tensors, "model.diffusion_model");
-    }
-
-    size_t get_params_buffer_size() {
-        return unet.get_params_buffer_size();
-    }
-
-    int64_t get_adm_in_channels() {
-        return unet.unet.adm_in_channels;
-    }
-
-    void compute(int n_threads,
-                 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,
-                 std::vector<ggml_tensor*> ref_latents     = {},
-                 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,
-               bool offload_params_to_cpu,
-               const String2GGMLType& tensor_types = {})
-        : mmdit(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model") {
-    }
-
-    std::string get_desc() {
-        return mmdit.get_desc();
-    }
-
-    void alloc_params_buffer() {
-        mmdit.alloc_params_buffer();
-    }
-
-    void free_params_buffer() {
-        mmdit.free_params_buffer();
-    }
-
-    void free_compute_buffer() {
-        mmdit.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
-        mmdit.get_param_tensors(tensors, "model.diffusion_model");
-    }
-
-    size_t get_params_buffer_size() {
-        return mmdit.get_params_buffer_size();
-    }
-
-    int64_t get_adm_in_channels() {
-        return 768 + 1280;
-    }
-
-    void compute(int n_threads,
-                 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,
-                 std::vector<ggml_tensor*> ref_latents     = {},
-                 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,
-              bool offload_params_to_cpu,
-              const String2GGMLType& tensor_types = {},
-              SDVersion version                   = VERSION_FLUX,
-              bool flash_attn                     = false,
-              bool use_mask                       = false)
-        : flux(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model", version, flash_attn, use_mask) {
-    }
-
-    std::string get_desc() {
-        return flux.get_desc();
-    }
-
-    void alloc_params_buffer() {
-        flux.alloc_params_buffer();
-    }
-
-    void free_params_buffer() {
-        flux.free_params_buffer();
-    }
-
-    void free_compute_buffer() {
-        flux.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
-        flux.get_param_tensors(tensors, "model.diffusion_model");
-    }
-
-    size_t get_params_buffer_size() {
-        return flux.get_params_buffer_size();
-    }
-
-    int64_t get_adm_in_channels() {
-        return 768;
-    }
-
-    void compute(int n_threads,
-                 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,
-                 std::vector<ggml_tensor*> ref_latents     = {},
-                 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, c_concat, y, guidance, ref_latents, output, output_ctx, skip_layers);
-    }
-};
-
-struct WanModel : public DiffusionModel {
-    std::string prefix;
-    WAN::WanRunner wan;
-
-    WanModel(ggml_backend_t backend,
-             bool offload_params_to_cpu,
-             const String2GGMLType& tensor_types = {},
-             const std::string prefix            = "model.diffusion_model",
-             SDVersion version                   = VERSION_WAN2,
-             bool flash_attn                     = false)
-        : prefix(prefix), wan(backend, offload_params_to_cpu, tensor_types, prefix, version, flash_attn) {
-    }
-
-    std::string get_desc() {
-        return wan.get_desc();
-    }
-
-    void alloc_params_buffer() {
-        wan.alloc_params_buffer();
-    }
-
-    void free_params_buffer() {
-        wan.free_params_buffer();
-    }
-
-    void free_compute_buffer() {
-        wan.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
-        wan.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() {
-        return wan.get_params_buffer_size();
-    }
-
-    int64_t get_adm_in_channels() {
-        return 768;
-    }
-
-    void compute(int n_threads,
-                 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,
-                 std::vector<ggml_tensor*> ref_latents     = {},
-                 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 wan.compute(n_threads, x, timesteps, context, y, c_concat, NULL, output, output_ctx);
-    }
-};
-
-#endif
@@ -0,0 +1,43 @@
+ARG UBUNTU_VERSION=24.04
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN cmake . -B ./build \
+    -DSD_BUILD_SHARED_LIBS=ON \
+    -DGGML_NATIVE=OFF \
+    -DSD_BUILD_SHARED_GGML_LIB=ON \
+    -DGGML_BACKEND_DL=ON \
+    -DGGML_CPU_ALL_VARIANTS=ON \
+    -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+    -DCMAKE_INSTALL_RPATH='$ORIGIN'
+RUN cmake --build ./build --config Release --parallel
+
+FROM ubuntu:$UBUNTU_VERSION AS runtime
+
+RUN apt-get update && \
+    apt-get install --yes --no-install-recommends libgomp1 && \
+    apt-get clean
+
+COPY --from=build /sd.cpp/build/bin /sd.cpp/bin
+RUN printf '#!/bin/sh\nexec /sd.cpp/bin/sd-cli "$@"\n' > /sd-cli && \
+    printf '#!/bin/sh\nexec /sd.cpp/bin/sd-server "$@"\n' > /sd-server && \
+    chmod +x /sd-cli /sd-server
+
+ENTRYPOINT [ "/sd-cli" ]
@@ -0,0 +1,60 @@
+ARG CUDA_VERSION=12.6.3
+ARG UBUNTU_VERSION=24.04
+
+FROM nvidia/cuda:${CUDA_VERSION}-cudnn-devel-ubuntu${UBUNTU_VERSION} AS build
+
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git ccache cmake ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+ARG CUDACXX=/usr/local/cuda/bin/nvcc
+ARG CUDA_ARCHITECTURES=""
+ARG GGML_CUDA_FA_ALL_QUANTS=""
+ARG GGML_CUDA_ENABLE_DYNAMIC_CPU_BACKENDS=ON
+
+RUN set -- \
+        -DSD_CUDA=ON; \
+    if [ "${GGML_CUDA_ENABLE_DYNAMIC_CPU_BACKENDS}" = "ON" ]; then \
+        set -- "$@" \
+            -DSD_BUILD_SHARED_LIBS=ON \
+            -DGGML_NATIVE=OFF \
+            -DSD_BUILD_SHARED_GGML_LIB=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            "-DCMAKE_INSTALL_RPATH=\$ORIGIN"; \
+    fi; \
+    if [ -n "${CUDA_ARCHITECTURES}" ]; then \
+        set -- "$@" "-DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHITECTURES}"; \
+    fi; \
+    if [ -n "${GGML_CUDA_FA_ALL_QUANTS}" ]; then \
+        set -- "$@" "-DGGML_CUDA_FA_ALL_QUANTS=${GGML_CUDA_FA_ALL_QUANTS}"; \
+    fi; \
+    cmake . -B ./build "$@"
+RUN cmake --build ./build --config Release -j$(nproc)
+
+FROM nvidia/cuda:${CUDA_VERSION}-cudnn-runtime-ubuntu${UBUNTU_VERSION} AS runtime
+
+RUN apt-get update && \
+    apt-get install --yes --no-install-recommends libgomp1 && \
+    apt-get clean
+
+COPY --from=build /sd.cpp/build/bin /sd.cpp/bin
+RUN printf '#!/bin/sh\nexec /sd.cpp/bin/sd-cli "$@"\n' > /sd-cli && \
+    printf '#!/bin/sh\nexec /sd.cpp/bin/sd-server "$@"\n' > /sd-server && \
+    chmod +x /sd-cli /sd-server
+
+ENTRYPOINT [ "/sd-cli" ]
@@ -0,0 +1,45 @@
+ARG MUSA_VERSION=rc4.2.0
+ARG UBUNTU_VERSION=22.04
+
+FROM mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64 as build
+
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends ccache cmake git ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && \
+    cmake .. -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
+        -DCMAKE_C_FLAGS="${CMAKE_C_FLAGS} -fopenmp -I/usr/lib/llvm-14/lib/clang/14.0.0/include -L/usr/lib/llvm-14/lib" \
+        -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -fopenmp -I/usr/lib/llvm-14/lib/clang/14.0.0/include -L/usr/lib/llvm-14/lib" \
+        -DSD_MUSA=ON \
+        -DSD_BUILD_SHARED_LIBS=ON \
+        -DGGML_NATIVE=OFF \
+        -DSD_BUILD_SHARED_GGML_LIB=ON \
+        -DGGML_BACKEND_DL=ON \
+        -DGGML_CPU_ALL_VARIANTS=ON \
+        -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+        -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+        -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build . --config Release
+
+FROM mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64 as runtime
+
+COPY --from=build /sd.cpp/build/bin /sd.cpp/bin
+RUN printf '#!/bin/sh\nexec /sd.cpp/bin/sd-cli "$@"\n' > /sd-cli && \
+    printf '#!/bin/sh\nexec /sd.cpp/bin/sd-server "$@"\n' > /sd-server && \
+    chmod +x /sd-cli /sd-server
+
+ENTRYPOINT [ "/sd-cli" ]
@@ -0,0 +1,32 @@
+# ggml SYCL hardware detection uses BMG G31/WCL architecture enums added in oneAPI 2025.3.
+ARG SYCL_VERSION=2025.3.2-0
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS build
+
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends cmake ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && \
+    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DSD_SYCL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build . --config Release -j$(nproc)
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS runtime
+
+COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
+COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
+
+ENTRYPOINT [ "/sd-cli" ]
@@ -0,0 +1,44 @@
+ARG UBUNTU_VERSION=24.04
+
+FROM ubuntu:$UBUNTU_VERSION AS build
+
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake libvulkan-dev glslc spirv-headers ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN cmake . -B ./build \
+    -DSD_VULKAN=ON \
+    -DSD_BUILD_SHARED_LIBS=ON \
+    -DGGML_NATIVE=OFF \
+    -DSD_BUILD_SHARED_GGML_LIB=ON \
+    -DGGML_BACKEND_DL=ON \
+    -DGGML_CPU_ALL_VARIANTS=ON \
+    -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+    -DCMAKE_INSTALL_RPATH='$ORIGIN'
+RUN cmake --build ./build --config Release -j$(nproc)
+
+FROM ubuntu:$UBUNTU_VERSION AS runtime
+
+RUN apt-get update && \
+    apt-get install --yes --no-install-recommends libgomp1 libvulkan1 mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 && \
+    apt-get clean
+
+COPY --from=build /sd.cpp/build/bin /sd.cpp/bin
+RUN printf '#!/bin/sh\nexec /sd.cpp/bin/sd-cli "$@"\n' > /sd-cli && \
+    printf '#!/bin/sh\nexec /sd.cpp/bin/sd-server "$@"\n' > /sd-server && \
+    chmod +x /sd-cli /sd-server
+
+ENTRYPOINT [ "/sd-cli" ]
@@ -0,0 +1,110 @@
+# ADetailer
+
+`sd-cli` can run a YOLOv8 object detector on an existing or newly generated
+image and perform a cropped inpaint pass for every detected object. The first
+implementation supports YOLOv8 detection checkpoints. YOLOv8 segmentation and
+MediaPipe models are not supported yet.
+
+## Convert a detector
+
+Ultralytics checkpoints must be converted before use. The converter fuses
+BatchNorm into convolution layers and writes a safetensors file with the weight
+names expected by the native GGML implementation.
+
+```bash
+python scripts/convert_yolov8_to_safetensors.py face_yolov8n.pt face_yolov8n.safetensors
+```
+
+The converter requires Python packages `ultralytics`, `torch`, and
+`safetensors`.
+Only YOLOv8 detection checkpoints are accepted.
+PyTorch checkpoints use pickle internally, so only convert `.pt` files from a
+trusted source.
+
+## Repair an existing image
+
+Use the dedicated `adetailer` mode to detect and repair objects in an existing
+image:
+
+```bash
+./bin/sd-cli \
+  -M adetailer \
+  -m model.safetensors \
+  -i input.png \
+  -o repaired.png \
+  -p "detailed portrait photo" \
+  --negative-prompt "deformed face" \
+  --steps 24 \
+  --cfg-scale 6 \
+  --strength 0.4 \
+  --sampling-method dpm++2m \
+  --scheduler karras \
+  --ad-model face_yolov8n.safetensors \
+  --extra-ad-args "confidence=0.3,inpaint_padding=32,mask_blur=4"
+```
+
+This mode reuses the normal image-generation options for the detail pass:
+
+- `--init-img`, `--output`, `--prompt`, and `--negative-prompt`
+- `--steps`, `--cfg-scale`, `--sampling-method`, and `--scheduler`
+- `--strength`, `--seed`, LoRA settings, VAE tiling, and backend assignments
+- `--width` and `--height`, which also resize the input when specified
+
+`--ad-prompt` and `--ad-negative-prompt` optionally override the normal prompts.
+Values provided in `--extra-ad-args`, such as `steps`, `cfg_scale`,
+`denoising_strength`, or `inpaint_width`, take precedence over inherited values.
+
+## Repair generated images
+
+ADetailer can also run automatically after normal image generation:
+
+```bash
+./bin/sd-cli \
+  -m model.safetensors \
+  -p "portrait photo" \
+  --ad-model face_yolov8n.safetensors \
+  --ad-prompt "[PROMPT], detailed face" \
+  --ad-negative-prompt "" \
+  --extra-ad-args "confidence=0.3,denoising_strength=0.4,inpaint_width=512,inpaint_height=512"
+```
+
+An empty ADetailer prompt inherits the main prompt. `[PROMPT]` inserts the main
+prompt, `[SEP]` assigns different prompts to consecutive masks, and `[SKIP]`
+skips the corresponding mask.
+
+All settings other than the detector path and prompts are passed through
+`--extra-ad-args` as a comma-separated `key=value` list:
+
+| Key | Default | Description |
+| --- | ---: | --- |
+| `input_size` | `640` | Square YOLO input size; must be a multiple of 32 |
+| `confidence` | `0.3` | Detection confidence threshold |
+| `nms` | `0.45` | NMS IoU threshold |
+| `max_detections` | `100` | Maximum detections retained after NMS |
+| `mask_k_largest` | `0` | Keep only the largest K masks; zero keeps all |
+| `mask_min_ratio` | `0` | Minimum bbox area relative to the image |
+| `mask_max_ratio` | `1` | Maximum bbox area relative to the image |
+| `dilate_erode` | `4` | Positive values dilate; negative values erode |
+| `x_offset`, `y_offset` | `0` | Mask offset in pixels; positive Y moves upward |
+| `mask_mode` | `none` | `none`, `merge`, or `merge_invert` |
+| `merge_masks`, `invert_mask` | `false` | Boolean alternatives to `mask_mode` |
+| `mask_blur` | `4` | Final composite feather radius |
+| `inpaint_padding` | `32` | Padding around the detected region |
+| `inpaint_width`, `inpaint_height` | mode-specific | `512x512` after generation; input/output size in `adetailer` mode |
+| `denoising_strength` | mode-specific | `0.4` after generation; inherits `--strength` in `adetailer` mode |
+| `steps` | `0` | Detail steps; zero inherits the main generation |
+| `cfg_scale` | `-1` | Detail CFG; a negative value inherits the main generation |
+| `sample_method` | inherited | Detail sampler name |
+| `scheduler` | inherited | Detail scheduler name |
+| `sort_by` | `none` | `none`, `left_to_right`, `center_to_edge`, or `area` |
+
+Multiple masks are processed serially. Each completed inpaint becomes the input
+for the next mask, and the seed is incremented by the mask index. Use
+`mask_mode=merge` to process all detections in one inpaint pass.
+
+The detector uses the `detector` backend module. For example, keep detection on
+the CPU while diffusion runs on CUDA:
+
+```bash
+--backend "diffusion=cuda0,detector=cpu"
+```
@@ -0,0 +1,21 @@
+# How to Use
+
+## Download weights
+
+- Download Anima
+    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/Bedovyy/Anima-GGUF/tree/main
+    - gguf Anima2: https://huggingface.co/JusteLeo/Anima2-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae
+- Download Qwen3-0.6B-Base
+    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/mradermacher/Qwen3-0.6B-Base-GGUF/tree/main
+
+## Examples
+
+```sh
+.\bin\Release\sd-cli.exe --diffusion-model  ..\models\diffusion_models\anima-preview.safetensors --vae ..\models\vae\qwen_image_vae.safetensors  --llm ..\models\text_encoders\qwen_3_06b_base.safetensors  -p "a lovely cat holding a sign says 'anima.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="anima image example" src="../assets/anima/example.png" />
@@ -0,0 +1,171 @@
+# AnimateDiff (SD 1.5)
+
+AnimateDiff adds motion to a frozen Stable Diffusion 1.5 checkpoint by
+injecting a temporal-attention module at 20 UNet slots. The base SD 1.5
+model, VAE, and text encoder are unchanged; only the motion module produces
+the temporal residual that turns a batch of independent frames into a
+coherent animation. Reference: Guo et al., "AnimateDiff: Animate Your
+Personalized Text-to-Image Diffusion Models without Specific Tuning"
+(https://arxiv.org/abs/2307.04725).
+
+## Download weights
+
+- Motion module (v3, recommended)
+    - fp16 safetensors: https://huggingface.co/conrevo/AnimateDiff-A1111/resolve/main/motion_module/mm_sd15_v3.safetensors
+    - original checkpoint: https://huggingface.co/guoyww/animatediff/resolve/main/v3_sd15_mm.ckpt
+- SD 1.5 base model
+    - any SD 1.5 checkpoint works. `realisticVisionV60B1` and `toonyou_beta3`
+      are the ones used in guoyww's reference configs.
+- Domain Adapter LoRA (optional, v3 only, sharpens the base UNet's output
+  toward the motion module's trained distribution)
+    - ckpt: https://huggingface.co/guoyww/animatediff/resolve/main/v3_sd15_adapter.ckpt
+    - place under your `--lora-model-dir` and reference in the prompt as
+      `<lora:v3_sd15_adapter:1.0>`.
+
+The motion module is `~836 MB` and loads alongside the SD 1.5 UNet via
+`--motion-module`.
+
+## Motion module versions
+
+Per [animatediff.net/models](https://animatediff.net/models):
+
+| Module              | Base | Native res | Character |
+|---------------------|------|------------|-----------|
+| `mm_sd_v14.ckpt`    | 1.5  | 256x256    | earliest, more jittery |
+| `mm_sd_v15.ckpt`    | 1.5  | 256x256    | improved stability over v1.4 |
+| `mm_sd_v15_v2.ckpt` | 1.5  | 384x384    | significantly better motion dynamics |
+| `v3_sd15_mm.ckpt`   | 1.5  | 512x512    | smoothest, highest quality; pairs with a Domain Adapter LoRA |
+| `mm_sdxl_v10_beta`  | SDXL | 512x512    | experimental, not yet supported here |
+
+Match your `-H -W` to the module's native resolution for best results. v3 is
+trained at 512x512 - going smaller (e.g. 384x384) still works but the motion
+character is closer to v2.
+
+## Examples
+
+Generate an 8-frame animation at 512x512, seed 42, 20 steps. The sampler /
+scheduler / CFG values below match what mm_sd15_v3 was trained with; using
+SD 1.5 defaults (euler_a, low CFG) produces noise-like output.
+
+```
+.\bin\Release\sd-cli.exe -M vid_gen \
+    --model ..\models\checkpoints\realisticVisionV60B1.safetensors \
+    --motion-module ..\models\animatediff\mm_sd15_v3.safetensors \
+    --offload-to-cpu --diffusion-fa \
+    -p "a red apple on a wooden table" \
+    --cfg-scale 8.0 --sampling-method euler --scheduler discrete \
+    -H 512 -W 512 --video-frames 8 --fps 8 --steps 20 -s 42 \
+    -o out.avi
+```
+
+Generate at the motion module's native 16-frame context (recommended for
+best temporal quality). Needs more VRAM at 512x512, so drop to 384x384 or
+use layer streaming:
+
+```
+.\bin\Release\sd-cli.exe -M vid_gen \
+    --model ..\models\checkpoints\realisticVisionV60B1.safetensors \
+    --motion-module ..\models\animatediff\mm_sd15_v3.safetensors \
+    --offload-to-cpu --diffusion-fa \
+    -p "photo of coastline, rocks, storm weather, wind, waves, lightning" \
+    --cfg-scale 8.0 --sampling-method euler --scheduler discrete \
+    -H 384 -W 384 --video-frames 16 --fps 8 --steps 20 -s 42 \
+    -o out.avi
+```
+
+Low-VRAM streaming (verified with a 2 GiB cap on RTX 3060):
+
+```
+.\bin\Release\sd-cli.exe -M vid_gen \
+    --model ..\models\checkpoints\realisticVisionV60B1.safetensors \
+    --motion-module ..\models\animatediff\mm_sd15_v3.safetensors \
+    --max-vram 2.0 --diffusion-fa \
+    -p "photo of coastline, rocks, storm weather, wind, waves, lightning" \
+    --cfg-scale 8.0 --sampling-method euler --scheduler discrete \
+    -H 384 -W 384 --video-frames 8 --fps 8 --steps 20 -s 42 \
+    -o out.avi
+```
+
+## Reference-quality reproduction
+
+Using guoyww's official reference configs on this impl (RealisticVision v6.0
+base + `mm_sd15_v3` or `mm_sd_v15_v2` + native resolution + 16 frames + euler
+ 25 steps + CFG 8 + linear beta schedule) reproduces the reference
+AnimateDiff output style.
+
+### v3 (512x512, `mm_sd15_v3`)
+
+| Prompt                                | Sample |
+|---------------------------------------|--------|
+| B&W man on stormy coastline           | <img src="../assets/animatediff/v3_man.gif" width="256"/>    |
+| Close-up rabbit macro shot            | <img src="../assets/animatediff/v3_rabbit.gif" width="256"/> |
+| Coastline, storm, waves, lightning    | <img src="../assets/animatediff/v3_coast.gif" width="256"/>  |
+| Old house, storm, forest, night       | <img src="../assets/animatediff/v3_house.gif" width="256"/>  |
+
+### v2 (384x384, `mm_sd_v15_v2.ckpt`)
+
+| Prompt                                | Sample |
+|---------------------------------------|--------|
+| B&W man on stormy coastline           | <img src="../assets/animatediff/v2_man.gif" width="256"/>    |
+| Close-up rabbit macro shot            | <img src="../assets/animatediff/v2_rabbit.gif" width="256"/> |
+| Coastline, storm, waves, lightning    | <img src="../assets/animatediff/v2_coast.gif" width="256"/>  |
+| Old house, storm, forest, night       | <img src="../assets/animatediff/v2_house.gif" width="256"/>  |
+
+Motion is strong for scenes with motion cues in the prompt (storm/waves/wind)
+and subtle for static subjects (close-up macro), matching reference behavior.
+v2 has an additional motion module at the UNet middle block that v3 dropped;
+this impl auto-detects the topology from the checkpoint.
+
+### v3 + Domain Adapter LoRA
+
+Attaching the `v3_sd15_adapter` LoRA sharpens the base UNet output toward
+the training distribution the motion module was fine-tuned against. Same
+prompt, seed, config as above:
+
+<img src="../assets/animatediff/v3_rabbit_domain_lora.gif" width="256"/>
+
+Individual fur strands, glowing inner-ear, and richer forest detail become
+visible compared to the no-LoRA rendering.
+
+```
+sd-cli -M vid_gen --model realisticVisionV60B1.safetensors \
+       --motion-module mm_sd15_v3.safetensors \
+       --lora-model-dir ./loras \
+       -p "close up photo of a rabbit ...<lora:v3_sd15_adapter:1.0>" ...
+```
+
+## img2video
+
+Pass a pre-rendered image via `-i / --init-img` to animate FROM it. All N output frames start from the encoded init latent, then per-frame noise is added at `--strength`. Character identity, composition, and quality are anchored by the init image; the motion module adds subtle motion on top.
+
+Left: init image rendered with `-M img_gen`. Right: 8-frame vid_gen output.
+
+<img src="../assets/animatediff/img2video_demo.gif" width="512"/>
+
+```
+sd-cli -M img_gen ... -o init.png                  # any high-quality still
+sd-cli -M vid_gen --motion-module mm_sd15_v3.safetensors \
+    -i init.png --strength 0.75 \
+    --cfg-scale 7.0 --sampling-method euler --scheduler karras \
+    -H 512 -W 512 --video-frames 8 --steps 25 -s 42 \
+    -p "..." -o out.avi
+```
+
+`--strength` controls how far the motion module is allowed to deviate from the init image (higher = more motion, lower = more static).
+
+## Notes
+
+- The motion module was trained at `video_length=16`. Running with
+  `--video-frames 16` gives the best coherence; F=8 works but shows a shorter
+  motion arc. Frame counts up to 32 are supported by the positional encoding
+  but exceed the trained regime and produce more static output.
+- At `--video-frames 1` the motion module is skipped entirely and the output
+  is bit-identical to `-M img_gen`. This avoids the single-token
+  temporal-attention degeneracy that would otherwise emit an untrained-magnitude
+  residual on a single-frame sample.
+- The base UNet is frozen, so character identity and style follow the SD 1.5
+  checkpoint you pass to `--model`. LoRAs and prompt weighting attach to the
+  base model in the usual way.
+- No mid_block motion module in v3. `mm_sdxl_v10_beta` (SDXL variant) is not
+  supported yet.
+- Output is written as MJPEG AVI. Use `--fps` to set playback speed.
@@ -0,0 +1,315 @@
+# Backend selection
+
+`stable-diffusion.cpp` has two backend assignments:
+
+- `--backend` selects the runtime backend used to execute model graphs.
+- `--params-backend` selects where model parameters are kept.
+
+If `--params-backend` is not set, auto-fit chooses parameter placement. With
+`--auto-fit off`, parameters use the same backend as their module runtime backend.
+
+## Syntax
+
+A backend assignment can be a single backend name:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cpu
+```
+
+This applies to every module that does not have a more specific assignment.
+
+Assignments can also target individual modules:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend te=cpu,vae=cuda0,diffusion=vulkan0
+```
+
+The same syntax is used for parameter placement:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend te=cpu,vae=cpu
+```
+
+`--params-backend` also accepts the special value `disk`:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend disk
+```
+
+`--max-vram` can target resolved backend/device names:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend diffusion=cuda0,vae=vulkan0 --max-vram cuda0=6,vulkan0=2
+```
+
+The value is a shared per-device budget for managed weights and registered
+runner compute/cache buffers. Live free memory can lower the effective limit
+for each graph run. Driver contexts and allocations made outside the managed
+model runners are not part of this accounting, so it is not a hard physical
+VRAM cap.
+
+Module names are case-insensitive. Hyphens and underscores in module names are ignored, so `clip_vision`, `clip-vision`, and `clipvision` are equivalent.
+
+`all=`, `default=`, and `*=` can be used to set the default backend inside a mixed assignment:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend all=cuda0,te=cpu
+```
+
+## Multiple devices per module (layer split)
+
+A `--backend` module assignment can list several devices separated by `&`:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend "diffusion=cuda0&cuda1"
+```
+
+The module's transformer blocks are then distributed across the listed devices
+in contiguous ranges sized proportionally to each device's free memory (minus a
+compute-buffer headroom of about 2 GiB per device), and the
+module's graphs are executed with a `ggml_backend_sched` that runs each block
+on the device holding its weights, copying the residual stream at the range
+boundaries. The first device in the list is the module's main device: it also
+holds the non-block tensors (embeddings, final norms, small sub-runners such as
+CLIP models or projectors) and the graph inputs/outputs.
+
+Layer split is supported for the `diffusion` and `te` modules. For `te` it
+applies to the dominant text encoder (`t5xxl` or the LLM); other modules accept
+only a single device. If the module has no recognizable transformer blocks, the
+assignment falls back to the first listed device.
+
+`--params-backend` accepts no device lists. If the module has no explicit
+params assignment, each block range's parameters are loaded directly to (and,
+with `--params-backend diffusion=disk`, released directly from) its own device;
+an explicit assignment such as `te=cpu` keeps the parameters on that backend
+and stages each range to its device on demand.
+
+Layer split uses the fixed graph-cut plan to assign blocks across devices, but
+single-device segmented execution and next-segment prefetch are disabled for
+the split module. `--max-vram` can still provide the per-device limits used by
+layer split and auto-fit.
+
+Use `--list-devices` to see the device names available on the system.
+
+### Row split (`--split-mode row`)
+
+`--split-mode` selects how a multi-device module distributes its weights:
+`layer` (the default, described above) or `row`. It accepts a single mode or
+per-module assignments:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend "diffusion=cuda0&cuda1" --split-mode row
+sd-cli -m model.safetensors -p "a cat" --backend "diffusion=cuda0&cuda1,te=cuda0&cuda1" --split-mode diffusion=row,te=layer
+```
+
+In row mode the module keeps executing on its main (first listed) device, but
+its transformer-block matmul weights are allocated in the backend's row-split
+buffer type, which slices each weight's rows across the listed devices in
+proportion to free memory and runs those matmuls on all devices in parallel.
+Compared to a layer split this uses all GPUs within every layer (instead of
+sequentially device by device) at the cost of a cross-device reduction per
+matmul - usually the faster option when the devices have fast interconnect.
+
+Row split requires a compatible split-buffer export from the linked GGML
+backend. If it is unavailable (or the listed devices belong to different backend
+registries), the module falls back to a layer split.
+Embeddings, normalization weights, biases and other non-block tensors stay in
+regular buffers on the main device.
+
+Row-split execution can use graph segments, but split weights are loaded
+synchronously instead of using the normal single-device prefetch path. Because
+GGML does not expose exact shard allocation sizes, the managed budget currently
+counts a split buffer's full size on each participating device. This is a
+conservative bound and can reject otherwise feasible layouts.
+
+Direct ("immediately") LoRA application cannot patch row-split tensors; with
+`--split-mode row` the automatic LoRA mode selects runtime application, and an
+explicit `--lora-apply-mode immediately` skips the split tensors with a
+warning.
+
+## Automatic placement (`--auto-fit on|off`)
+
+`--auto-fit` requires `on` or `off` and defaults to `on` when omitted.
+Explicit `--params-backend` assignments disable auto-fit,
+regardless of argument order, even with `--auto-fit on`.
+
+Auto-fit preserves explicit `--backend` assignments, including per-module
+assignments and device lists. For modules without a runtime assignment, it chooses
+the GPU with the largest available memory budget (the first device on a tie).
+It then derives parameter placements from the model metadata, each module's
+compute devices, and the remaining memory budgets. The chosen backend
+specifications are printed.
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --auto-fit on
+sd-cli -m model.safetensors -p "a cat" --auto-fit on --max-vram cuda0=8,cuda1=14
+sd-cli -m model.safetensors -p "a cat" --backend cuda0
+sd-cli -m model.safetensors -p "a cat" --backend diffusion=cuda0,te=cpu,vae=cuda1
+sd-cli -m model.safetensors -p "a cat" --auto-fit off
+```
+
+Budgets reuse `--max-vram`: a positive per-device value caps what auto-fit
+plans with on that device, a negative value means "free memory minus that many
+GiB", and with no budget set each device's free memory minus a 512 MiB margin
+is used. These resolved GPU budgets, including the safety margin, also drive
+the runner's graph-cut capacity checks.
+
+Runtime capacity checks also leave 512 MiB of currently free device memory for
+backend scratch buffers and pipelines, including with explicit backend assignments.
+They cap stale free-memory reports by the device's total memory minus tracked
+resident allocations and reject reports that exceed the device's total memory.
+
+Components are considered in `diffusion`, `te`, `vae` order so that repeatedly
+used diffusion weights have priority. Each component's weights use the first
+storage location with enough remaining budget:
+
+1. The component's compute GPU, leaving estimated space for computation and weight staging.
+2. CPU RAM, reserving the larger of 2 GiB or 10% of available RAM for other work.
+3. Another GPU, choosing the one with the largest remaining budget that fits.
+4. Disk, reloading weights on demand.
+
+GPU cache space follows the same component priority. Before a lower-priority
+component can become permanently resident, the planner leaves room for the full
+weights and estimated compute space of higher-priority offloaded components.
+If offloaded diffusion already needs the entire main GPU budget, TE and VAE also
+use offloaded parameters. Their GPU copies can then be released after their
+phases, leaving more room to reuse diffusion weights across sampling steps.
+CPU parameter residency allows GPU weight caching; it does not force every
+weight to be copied again at every step.
+
+RAM and GPU budgets are shared across components. Each component uses a single
+parameter backend; several other GPUs' capacities are not combined to store
+one component. If available RAM cannot be queried, RAM residency is skipped.
+Weights stored on another GPU are copied to the component's compute devices for
+execution. CPU modules use RAM or disk. Compute reserves and cache priority are
+accounted for separately on each device, so a CPU module does not reserve GPU
+space. Storage on another module's GPU also leaves room for that module's work.
+
+Auto-fit does not select multi-GPU layer/row computation itself. Explicit device
+lists and `--split-mode` still control that computation. Before the runners have
+built their split plans, auto-fit conservatively counts the full component size
+on each listed GPU when checking residency and cache space. This can offload
+parameters even when a split layout would fit; use `--auto-fit off` to keep the
+default split-device parameter placement.
+
+For example, a diffusion model whose full weights exceed the main GPU's budget
+can use `--backend diffusion=cuda0 --params-backend diffusion=cpu` when RAM is
+sufficient. Automatic graph segmentation can then load the required weights
+for each segment and reclaim idle GPU copies. `--disable-segmented-compute`
+still disables segmentation.
+
+Initial compute reserves are estimates (2 GiB for diffusion and text encoders,
+1 GiB for VAE); higher-priority placements also leave staging space for the
+largest weight tensor of each lower-priority offloaded component. Actual segment
+weights, compute buffers and caches must
+still fit the runner's capacity checks. Offloading weights does not guarantee
+that every resolution or frame count will fit, and auto-fit does not change a
+component to CPU computation solely because its full weights exceed VRAM.
+If a VAE decode fails, decoding retries with spatial tiling even when `--auto-fit`
+is off; supported video decoders try temporal tiling first and can then add
+spatial tiling. Spatial retries use half-size tiles along each latent dimension.
+
+## Modules
+
+| Module | Purpose | Accepted names |
+| --- | --- | --- |
+| `diffusion` | UNet, DiT, MMDiT, Flux, Wan, Qwen Image, and other diffusion models | `diffusion`, `model`, `unet`, `dit` |
+| `te` | Text encoders and conditioners | `te`, `clip`, `text`, `textencoder`, `textencoders`, `conditioner`, `cond`, `llm`, `t5`, `t5xxl` |
+| `clip_vision` | CLIP vision encoder | `clip_vision`, `clipvision`, `clip-vision`, `vision` |
+| `vae` | VAE and TAE | `vae`, `firststage`, `autoencoder`, `tae` |
+| `controlnet` | ControlNet | `controlnet`, `control` |
+| `photomaker` | PhotoMaker ID encoder and PhotoMaker LoRA | `photomaker`, `photomakerid`, `pmid`, `photo` |
+| `upscaler` | ESRGAN upscaler | `upscaler`, `esrgan`, `hires` |
+| `detector` | ADetailer YOLOv8 detector | `detector`, `adetailer`, `yolo` |
+
+`te` is the preferred module name for text encoders. `clip` is kept as an accepted alias because many existing commands and model names use CLIP terminology.
+
+## Backend names
+
+Backend names are resolved against the GGML backend device list. Matching is case-insensitive and accepts exact names or unique prefixes, so common values include names such as:
+
+- `cpu`
+- `cuda0`
+- `vulkan0`
+- `metal`
+
+The special values `auto`, `default`, and an empty backend name select the default backend. The default preference is GPU, then integrated GPU, then CPU.
+
+The special value `gpu` selects the first GPU backend, falling back to the first integrated GPU backend.
+
+The special value `disk` is accepted only by `--params-backend`. `--backend disk` is invalid because `disk` is a parameter residency mode, not a runtime compute backend.
+
+## Runtime backend vs. parameter backend
+
+The runtime backend controls where graph execution runs. The parameter backend controls where model weights are allocated or whether they are reloaded from disk on demand.
+
+For example:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend cpu
+```
+
+This runs all modules on `cuda0`, but stores parameters in CPU RAM. During execution, parameters are moved to the runtime backend as needed.
+
+For example:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend disk
+```
+
+This runs all modules on `cuda0`, reloads parameters from the model file as needed, and releases those parameter buffers after use.
+
+Outside `--auto-fit`, `disk` is never selected implicitly. If `--params-backend` is not set, parameters use the runtime backend.
+
+Per-module assignments can be mixed:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend diffusion=cuda0,te=cpu,vae=cpu --params-backend diffusion=cuda0,te=cpu,vae=cpu
+```
+
+This keeps text encoding and VAE execution on CPU while the diffusion model runs on GPU.
+
+## Backend sharing and lifetime
+
+Backends are managed by `SDBackendManager`.
+
+Within one manager, backend instances are cached by resolved backend device name. If multiple modules request the same backend, they share the same `ggml_backend_t`.
+
+For example:
+
+```shell
+--backend te=cpu,vae=cpu
+```
+
+uses one shared CPU backend for both `te` and `vae` runtime execution.
+
+Runtime and parameter assignments also share the same backend cache. If `--backend diffusion=cuda0` and `--params-backend diffusion=cuda0` resolve to the same device, both use the same backend instance.
+
+`--params-backend disk` does not create a separate backend instance. Parameters are loaded lazily using the module runtime backend.
+
+`SDBackendManager` owns the backend instances and frees them when the context or upscaler is destroyed. Model runners receive non-owning runtime and parameter backend pointers and do not free them.
+
+## Compatibility flags
+
+The example CLI/server still accepts these older CPU placement flags as compatibility aliases:
+
+- `--clip-on-cpu`
+- `--vae-on-cpu`
+- `--control-net-cpu`
+- `--offload-to-cpu`
+
+`--clip-on-cpu`, `--vae-on-cpu`, and `--control-net-cpu` are deprecated. The example argument layer prepends `te=cpu`, `vae=cpu`, and `controlnet=cpu` to `--backend` before creating the context.
+
+`--offload-to-cpu` prepends a CPU default to the parameter assignment in the caller before creating the context:
+
+```shell
+--params-backend '*=cpu'
+```
+
+Because this default is inserted first, later explicit `--params-backend` entries can still override it, for example `--offload-to-cpu --params-backend te=disk` keeps non-TE parameters on CPU and reloads TE parameters from disk.
+
+Library callers should set `backend` and `params_backend` directly. `sd_ctx_params_init()`
+enables `auto_fit` by default; a nonempty `params_backend` assignment disables it.
+The `backend` assignment constrains auto-fit's compute placement.
+The old CPU/offload fields are no longer part of the C API. Explicit `--backend` and
+`--params-backend` assignments are preferred for new commands.
@@ -0,0 +1,31 @@
+# How to Use
+
+Boogu Image uses a Boogu diffusion transformer, the FLUX VAE, and Qwen3-VL as the LLM text and vision encoder.
+
+## Download weights
+
+- Download Boogu Image
+    - safetensors: https://huggingface.co/Comfy-Org/Boogu-Image/tree/main/diffusion_models
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download Qwen3-VL 8B
+    - gguf: https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-GGUF/tree/main
+        - For image editing with GGUF text encoders, also download the matching mmproj file and pass it with `--llm_vision`.
+
+## Examples
+
+### Boogu Image Base
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\boogu_image_base_bf16.safetensors --llm ..\models\text_encoders\Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae ..\models\vae\ae.sft -p "a lovely cat" --diffusion-fa -v --offload-to-cpu
+```
+
+<img width="256" alt="Boogu Image Base example" src="../assets/boogu/example.png" />
+
+### Boogu Image Edit
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\boogu_image_edit_bf16.safetensors --llm ..\models\text_encoders\Qwen3VL-8B-Instruct-Q4_K_M.gguf --llm_vision ..\models\text_encoders\mmproj-Qwen3VL-8B-Instruct-F16.gguf --vae ..\models\vae\ae.sft --diffusion-fa -v --offload-to-cpu -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'boogu.cpp'"
+```
+
+<img width="256" alt="Boogu Image Edit example" src="../assets/boogu/edit_example.png" />
@@ -0,0 +1,232 @@
+# Build from scratch
+
+## Get the Code
+
+```
+git clone --recursive https://github.com/leejet/stable-diffusion.cpp
+cd stable-diffusion.cpp
+```
+
+- If you have already cloned the repository, you can use the following command to update the repository to the latest code.
+
+```
+cd stable-diffusion.cpp
+git pull origin master
+git submodule init
+git submodule update
+```
+
+## Selecting a GGML source tree
+
+By default, sd.cpp builds the patched GGML submodule in `ggml/`. To build with
+an upstream GGML checkout instead, enable `SD_USE_UPSTREAM_GGML` and set
+`SD_GGML_SOURCE_DIR`:
+
+```shell
+cmake -S . -B build-upstream -DSD_USE_UPSTREAM_GGML=ON -DSD_GGML_SOURCE_DIR=../ggml-upstream
+cmake --build build-upstream --config Release
+```
+
+The selected source tree supplies both the library and its private headers.
+Backend options such as `-DSD_CUDA=ON` apply to the selected tree as usual.
+
+`SD_USE_UPSTREAM_GGML` defaults to `OFF`, which enables the patched GGML
+extensions. Set it to `ON` when using upstream GGML; it selects the compatibility
+mode and does not download or replace the GGML source tree. Upstream mode keeps
+the original FP8 safetensors handling: FP8 tensors are converted to F16 at load
+time (one byte per element in the file, two in RAM and VRAM). INT8
+tensorwise/convrot is disabled and its model files are rejected with an explicit
+error. FP8 GGUF files, FP8 weight type requests and tensor type rules are also
+rejected; no automatic conversion is performed.
+
+Upstream GGML may lack some operators and performance optimizations provided by
+the patched version. A warning is emitted during CMake configuration and when
+creating an inference context. Ordinary floating-point and shared GGML
+quantization types remain available, subject to backend operator support.
+
+`SD_USE_SYSTEM_GGML=ON` instead links an installed GGML CMake package, located
+with `ggml_DIR` or `CMAKE_PREFIX_PATH`. In that mode, `SD_GGML_SOURCE_DIR` must
+point to the matching source tree for private headers. The installed library
+must use the same ABI settings as sd.cpp, including `GGML_MAX_NAME`.
+Set `SD_USE_UPSTREAM_GGML=ON` as well if the installed package is upstream GGML.
+
+## WebP and WebM Support in Examples
+
+The example applications (`examples/cli` and `examples/server`) use `libwebp` to support WebP image I/O, and `examples/cli` can also use `libwebm` for `.webm` video output. Both are enabled by default. WebM output currently reuses `libwebp` to encode each frame as VP8 before muxing with `libwebm`.
+
+If you do not want WebP/WebM support, you can disable them at configure time:
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_WEBP=OFF -DSD_WEBM=OFF
+cmake --build . --config Release
+```
+
+If the submodules are not available, you can also link against system packages instead:
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_USE_SYSTEM_WEBP=ON -DSD_USE_SYSTEM_WEBM=ON
+cmake --build . --config Release
+```
+
+## Build (CPU only)
+
+If you don't have a GPU or CUDA installed, you can build a CPU-only version.
+
+```shell
+mkdir build && cd build
+cmake ..
+cmake --build . --config Release
+```
+
+## Build with OpenBLAS
+
+```shell
+mkdir build && cd build
+cmake .. -DGGML_OPENBLAS=ON
+cmake --build . --config Release
+```
+
+## Build with CUDA
+
+This provides GPU acceleration using NVIDIA GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). Recommended to have at least 4 GB of VRAM.
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_CUDA=ON
+cmake --build . --config Release
+```
+
+## Build with HipBLAS
+
+This provides GPU acceleration using AMD GPU. Make sure to have the ROCm toolkit installed.
+To build for another GPU architecture than installed in your system, set `$GFX_NAME` manually to the desired architecture (replace first command). This is also necessary if your GPU is not officially supported by ROCm, for example you have to set `$GFX_NAME` manually to `gfx1030` for consumer RDNA2 cards.
+
+Windows User Refer to [docs/hipBLAS_on_Windows.md](docs%2FhipBLAS_on_Windows.md) for a comprehensive guide.
+
+```shell
+mkdir build && cd build
+if command -v rocminfo; then export GFX_NAME=$(rocminfo | awk '/ *Name: +gfx[1-9]/ {print $2; exit}'); else echo "rocminfo missing!"; fi
+if [ -z "${GFX_NAME}" ]; then echo "Error: Couldn't detect GPU!"; else echo "Building for GPU: ${GFX_NAME}"; fi
+cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS=$GFX_NAME -DAMDGPU_TARGETS=$GFX_NAME -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+cmake --build . --config Release
+```
+
+## Build with MUSA
+
+This provides GPU acceleration using Moore Threads GPU. Make sure to have the MUSA toolkit installed.
+
+```shell
+mkdir build && cd build
+cmake .. -DCMAKE_C_COMPILER=/usr/local/musa/bin/clang -DCMAKE_CXX_COMPILER=/usr/local/musa/bin/clang++ -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release
+cmake --build . --config Release
+```
+
+## Build with 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.
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_METAL=ON
+cmake --build . --config Release
+```
+
+## Build with Vulkan
+
+Install Vulkan SDK from https://www.lunarg.com/vulkan-sdk/.
+On Ubuntu, install the Vulkan development packages and SPIR-V headers:
+
+```shell
+sudo apt-get install build-essential libvulkan-dev glslc spirv-headers
+```
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_VULKAN=ON
+cmake --build . --config Release
+```
+
+## Build with OpenCL (for Adreno GPU)
+
+Currently, it supports only Adreno GPUs and is primarily optimized for Q4_0 type
+
+To build for Windows ARM please refers to [Windows 11 Arm64](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/OPENCL.md#windows-11-arm64)
+
+Building for Android:
+
+  Android NDK:
+       Download and install the Android NDK from the [official Android developer site](https://developer.android.com/ndk/downloads).
+
+Setup OpenCL Dependencies for NDK:
+
+You need to provide OpenCL headers and the ICD loader library to your NDK sysroot.
+
+*   OpenCL Headers:
+    ```bash
+    # In a temporary working directory
+    git clone https://github.com/KhronosGroup/OpenCL-Headers
+    cd OpenCL-Headers
+    # Replace <YOUR_NDK_PATH> with your actual NDK installation path
+    # e.g., cp -r CL /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+    sudo cp -r CL <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+    cd ..
+    ```
+
+*   OpenCL ICD Loader:
+    ```shell
+    # In the same temporary working directory
+    git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
+    cd OpenCL-ICD-Loader
+    mkdir build_ndk && cd build_ndk
+
+    # Replace <YOUR_NDK_PATH> in the CMAKE_TOOLCHAIN_FILE and OPENCL_ICD_LOADER_HEADERS_DIR
+    cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \
+      -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
+      -DOPENCL_ICD_LOADER_HEADERS_DIR=<YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \
+      -DANDROID_ABI=arm64-v8a \
+      -DANDROID_PLATFORM=24 \
+      -DANDROID_STL=c++_shared
+
+    ninja
+    # Replace <YOUR_NDK_PATH>
+    # e.g., cp libOpenCL.so /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+    sudo cp libOpenCL.so <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+    cd ../..
+    ```
+
+Build `stable-diffusion.cpp` for Android with OpenCL:
+
+```shell
+mkdir build-android && cd build-android
+
+# Replace <YOUR_NDK_PATH> with your actual NDK installation path
+# e.g., -DCMAKE_TOOLCHAIN_FILE=/path/to/android-ndk-r26c/build/cmake/android.toolchain.cmake
+cmake .. -G Ninja \
+  -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
+  -DANDROID_ABI=arm64-v8a \
+  -DANDROID_PLATFORM=android-28 \
+  -DGGML_OPENMP=OFF \
+  -DSD_OPENCL=ON
+
+ninja
+```
+*(Note: Don't forget to include `LD_LIBRARY_PATH=/vendor/lib64` in your command line before running the binary)*
+
+## Build with 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/ggml-org/llama.cpp/blob/master/docs/backend/SYCL.md#linux).
+
+```shell
+# 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
+```
--- a/Show More
+++ b/Show More